feat(gpu): implement sub array search on GPU

The subtraction is done via addition of the negation,
the negation is done via unchecked_neg, this will make the
first block have a carry.
Then we called add_assign_with_carry_parallelized which did
a carry propagation on the rhs which here is the negated value,
meaning the subtraction would do 2 carry propagation.

To fix that we directly call the lower function.

.github/actionlint.yaml

@@ -3,6 +3,8 @@ self-hosted-runner:
   labels:
     - m1mac
     - 4090-desktop
+    - large_windows_16_latest
+    - large_ubuntu_16
 # Configuration variables in array of strings defined in your repository or
 # organization. `null` means disabling configuration variables check.
 # Empty array means no configuration variable is allowed.

.github/workflows/aws_tfhe_backward_compat_tests.yml

@@ -0,0 +1,120 @@
+# Run backward compatibility tests
+name: Backward compatibility Tests on CPU
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+
+jobs:
+  setup-instance:
+    name: Setup instance (backward-compat-tests)
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: cpu-small
+
+  backward-compat-tests:
+    name: Backward compatibility tests
+    needs: [ setup-instance ]
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: true
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Install git-lfs
+        run: |
+          sudo apt update && sudo apt -y install git-lfs
+
+      - name: Use specific data branch
+        if: ${{ contains(github.event.pull_request.labels.*.name, 'data_PR') }}
+        env:
+          PR_BRANCH: ${{ github.head_ref || github.ref_name }}
+        run: |
+          echo "BACKWARD_COMPAT_DATA_BRANCH=${PR_BRANCH}" >> "${GITHUB_ENV}"
+
+      - name: Get backward compat branch
+        id: backward_compat_branch
+        run: |
+          BRANCH="$(make backward_compat_branch)"
+          echo "branch=${BRANCH}" >> "${GITHUB_OUTPUT}"
+
+      - name: Clone test data
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+          repository: zama-ai/tfhe-backward-compat-data
+          path: tfhe/tfhe-backward-compat-data
+          lfs: 'true'
+          ref: ${{ steps.backward_compat_branch.outputs.branch }}
+
+      - name: Run backward compatibility tests
+        run: |
+          make test_backward_compatibility_ci
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Backward compatibility tests finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (backward-compat-tests)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, backward-compat-tests ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (backward-compat-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_fast_tests.yml

@@ -1,4 +1,4 @@
-# Run a small subset of shortint and integer tests to ensure quick feedback.
+# Run a small subset of tests to ensure quick feedback.
 name: Fast AWS Tests on CPU
 
 env:
@@ -11,6 +11,7 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
@@ -18,15 +19,112 @@ on:
   pull_request:
 
 jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      csprng_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.csprng_any_changed }}
+      zk_pok_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.zk_pok_any_changed }}
+      core_crypto_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.core_crypto_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      boolean_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.boolean_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      shortint_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.shortint_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      integer_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.integer_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      wasm_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.wasm_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      high_level_api_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.high_level_api_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      user_docs_test: ${{ env.IS_PULL_REQUEST == 'false' ||
+        steps.changed-files.outputs.user_docs_any_changed ||
+        steps.changed-files.outputs.dependencies_any_changed }}
+      any_file_changed: ${{ env.IS_PULL_REQUEST == 'false' || steps.aggregated-changes.outputs.any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            dependencies:
+              - tfhe/Cargo.toml
+              - concrete-csprng/**
+              - tfhe-zk-pok/**
+            csprng:
+              - concrete-csprng/**
+            zk_pok:
+              - tfhe-zk-pok/**
+            core_crypto:
+              - tfhe/src/core_crypto/**
+            boolean:
+              - tfhe/src/core_crypto/**
+              - tfhe/src/boolean/**
+            shortint:
+              - tfhe/src/core_crypto/**
+              - tfhe/src/shortint/**
+            integer:
+              - tfhe/src/core_crypto/**
+              - tfhe/src/shortint/**
+              - tfhe/src/integer/**
+            wasm:
+              - tfhe/src/**
+              - tfhe/js_on_wasm_tests/**
+              - tfhe/web_wasm_parallel_tests/**
+              - '!tfhe/src/c_api/**'
+              - '!tfhe/src/boolean/**'
+            high_level_api:
+              - tfhe/src/**
+              - '!tfhe/src/c_api/**'
+              - '!tfhe/src/boolean/**'
+              - '!tfhe/src/c_api/**'
+              - '!tfhe/src/js_on_wasm_api/**'
+            user_docs:
+              - tfhe/src/**
+              - '!tfhe/src/c_api/**'
+              - 'tfhe/docs/**.md'
+              - README.md
+
+      - name: Aggregate file changes
+        id: aggregated-changes
+        if: ( steps.changed-files.outputs.dependencies_any_changed == 'true' ||
+          steps.changed-files.outputs.csprng_any_changed == 'true' ||
+          steps.changed-files.outputs.zk_pok_any_changed == 'true' ||
+          steps.changed-files.outputs.core_crypto_any_changed == 'true' ||
+          steps.changed-files.outputs.boolean_any_changed == 'true' ||
+          steps.changed-files.outputs.shortint_any_changed == 'true' ||
+          steps.changed-files.outputs.integer_any_changed == 'true' ||
+          steps.changed-files.outputs.wasm_any_changed == 'true' ||
+          steps.changed-files.outputs.high_level_api_any_changed == 'true' ||
+          steps.changed-files.outputs.user_docs_any_changed == 'true')
+        run: |
+          echo "any_changed=true" >> "$GITHUB_OUTPUT"
+
   setup-instance:
     name: Setup instance (fast-tests)
+    if: github.event_name != 'pull_request' ||
+      needs.should-run.outputs.any_file_changed == 'true'
+    needs: should-run
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -37,14 +135,16 @@ jobs:
 
   fast-tests:
     name: Fast CPU tests
-    needs: setup-instance
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    needs: [ should-run, setup-instance ]
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -53,55 +153,58 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
       - name: Run concrete-csprng tests
+        if: needs.should-run.outputs.csprng_test == 'true'
         run: |
           make test_concrete_csprng
 
       - name: Run tfhe-zk-pok tests
+        if: needs.should-run.outputs.zk_pok_test == 'true'
         run: |
           make test_zk_pok
 
       - name: Run core tests
+        if: needs.should-run.outputs.core_crypto_test == 'true'
         run: |
           AVX512_SUPPORT=ON make test_core_crypto
 
       - name: Run boolean tests
+        if: needs.should-run.outputs.boolean_test == 'true'
         run: |
           make test_boolean
 
       - name: Run user docs tests
+        if: needs.should-run.outputs.user_docs_test == 'true'
         run: |
           make test_user_doc
 
       - name: Run js on wasm API tests
+        if: needs.should-run.outputs.wasm_test == 'true'
         run: |
           make test_nodejs_wasm_api_in_docker
 
       - name: Gen Keys if required
+        if: needs.should-run.outputs.shortint_test == 'true' ||
+          needs.should-run.outputs.integer_test == 'true'
         run: |
           make gen_key_cache
 
       - name: Run shortint tests
+        if: needs.should-run.outputs.shortint_test == 'true'
         run: |
           BIG_TESTS_INSTANCE=TRUE FAST_TESTS=TRUE make test_shortint_ci
 
       - name: Run integer tests
+        if: needs.should-run.outputs.integer_test == 'true'
         run: |
           BIG_TESTS_INSTANCE=TRUE FAST_TESTS=TRUE make test_integer_ci
 
-      - name: Run shortint multi-bit tests
-        run: |
-          BIG_TESTS_INSTANCE=TRUE FAST_TESTS=TRUE make test_shortint_multi_bit_ci
-
-      - name: Run integer multi-bit tests
-        run: |
-          BIG_TESTS_INSTANCE=TRUE FAST_TESTS=TRUE make test_integer_multi_bit_ci
-
       - name: Run high-level API tests
+        if: needs.should-run.outputs.high_level_api_test == 'true'
         run: |
           make test_high_level_api
 
@@ -125,7 +228,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/aws_tfhe_integer_tests.yml

@@ -10,24 +10,37 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  # We clear the cache to reduce memory pressure because of the numerous processes of cargo
+  # nextest
+  TFHE_RS_CLEAR_IN_MEMORY_KEY_CACHE: "1"
+  NO_BIG_PARAMS: FALSE
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
     types: [ labeled ]
+  push:
+    branches:
+      - main
+  schedule:
+    # Nightly tests @ 3AM after each work day
+    - cron: "0 3 * * MON-FRI"
 
 jobs:
   setup-instance:
     name: Setup instance (unsigned-integer-tests)
-    if: ${{ github.event_name == 'workflow_dispatch' || contains(github.event.label.name, 'approved') }}
+    if: (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') ||
+        (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+        (github.event_name == 'pull_request' && contains(github.event.label.name, 'approved')) ||
+        github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -40,12 +53,12 @@ jobs:
     name: Unsigned integer tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
-      cancel-in-progress: true
+      group: ${{ github.workflow }}_${{ github.ref }}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -54,10 +67,15 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
+      - name: Should skip big parameters set
+        if: github.event_name == 'pull_request'
+        run: |
+          echo "NO_BIG_PARAMS=TRUE" >> "${GITHUB_ENV}"
+
       - name: Gen Keys if required
         run: |
           make GEN_KEY_CACHE_MULTI_BIT_ONLY=TRUE gen_key_cache
@@ -72,7 +90,7 @@ jobs:
 
       - name: Run unsigned integer tests
         run: |
-          AVX512_SUPPORT=ON BIG_TESTS_INSTANCE=TRUE make test_unsigned_integer_ci
+          AVX512_SUPPORT=ON NO_BIG_PARAMS=${{ env.NO_BIG_PARAMS }} BIG_TESTS_INSTANCE=TRUE make test_unsigned_integer_ci
 
       - name: Slack Notification
         if: ${{ always() }}
@@ -90,7 +108,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/aws_tfhe_signed_integer_tests.yml

@@ -10,24 +10,37 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  # We clear the cache to reduce memory pressure because of the numerous processes of cargo
+  # nextest
+  TFHE_RS_CLEAR_IN_MEMORY_KEY_CACHE: "1"
+  NO_BIG_PARAMS: FALSE
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
     types: [ labeled ]
+  push:
+    branches:
+      - main
+  schedule:
+    # Nightly tests @ 3AM after each work day
+    - cron: "0 3 * * MON-FRI"
 
 jobs:
   setup-instance:
     name: Setup instance (signed-integer-tests)
-    if: ${{ github.event_name == 'workflow_dispatch' || contains(github.event.label.name, 'approved') }}
+    if: (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') ||
+        (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+        (github.event_name == 'pull_request' && contains(github.event.label.name, 'approved')) ||
+        github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -40,12 +53,12 @@ jobs:
     name: Signed integer tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
-      cancel-in-progress: true
+      group: ${{ github.workflow }}_${{ github.ref }}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -54,10 +67,15 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
+      - name: Should skip big parameters set
+        if: github.event_name == 'pull_request'
+        run: |
+          echo "NO_BIG_PARAMS=TRUE" >> "${GITHUB_ENV}"
+
       - name: Gen Keys if required
         run: |
           make GEN_KEY_CACHE_MULTI_BIT_ONLY=TRUE gen_key_cache
@@ -76,7 +94,7 @@ jobs:
 
       - name: Run signed integer tests
         run: |
-          AVX512_SUPPORT=ON BIG_TESTS_INSTANCE=TRUE make test_signed_integer_ci
+          AVX512_SUPPORT=ON NO_BIG_PARAMS=${{ env.NO_BIG_PARAMS }} BIG_TESTS_INSTANCE=TRUE make test_signed_integer_ci
 
       - name: Slack Notification
         if: ${{ always() }}
@@ -94,7 +112,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/aws_tfhe_tests.yml

@@ -24,6 +24,8 @@ on:
 jobs:
   should-run:
     runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     permissions:
       pull-requests: write
     outputs:
@@ -55,13 +57,13 @@ jobs:
       any_file_changed: ${{ env.IS_PULL_REQUEST == 'false' || steps.aggregated-changes.outputs.any_changed }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@0ad4b8fadaa221de15dcec353f45205ec38ea70b
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@03334d095e2739fa9ac4034ec16f66d5d01e9eba
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
         with:
           since_last_remote_commit: true
           files_yaml: |
@@ -84,6 +86,8 @@ jobs:
             high_level_api:
               - tfhe/src/**
               - '!tfhe/src/c_api/**'
+              - '!tfhe/src/boolean/**'
+              - '!tfhe/src/js_on_wasm_api/**'
             c_api:
               - tfhe/src/**
             examples:
@@ -119,7 +123,7 @@ jobs:
   setup-instance:
     name: Setup instance (cpu-tests)
     if: github.event_name != 'pull_request' ||
-      (github.event_name == 'pull_request' && needs.should-run.outputs.any_file_changed == 'true')
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.any_file_changed == 'true')
     needs: should-run
     runs-on: ubuntu-latest
     outputs:
@@ -127,7 +131,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -147,7 +151,7 @@ jobs:
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -156,7 +160,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -233,7 +237,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/aws_tfhe_wasm_tests.yml

@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -45,7 +45,7 @@ jobs:
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -54,7 +54,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -72,7 +72,7 @@ jobs:
 
       - name: Run parallel wasm tests
         run: |
-          make ci_test_web_js_api_parallel
+          make test_web_js_api_parallel_ci
 
       - name: Slack Notification
         if: ${{ always() }}
@@ -90,7 +90,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/boolean_benchmark.yml

@@ -3,30 +3,9 @@ name: Boolean benchmarks
 
 on:
   workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      # This input is not used in this workflow but still mandatory since a calling workflow could
-      # use it. If a triggering command include a user_inputs field, then the triggered workflow
-      # must include this very input, otherwise the workflow won't be called.
-      # See start_full_benchmarks.yml as example.
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
 
 env:
   CARGO_TERM_COLOR: always
@@ -34,36 +13,60 @@ env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
 
 jobs:
-  run-boolean-benchmarks:
-    name: Execute boolean benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
+  setup-instance:
+    name: Setup instance (boolean-benchmarks)
+    runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
 
+  boolean-benchmarks:
+    name: Execute boolean benchmarks in EC2
+    needs: setup-instance
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    continue-on-error: true
+    steps:
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
       - name: Set up home
         # "Install rust" step require root user to have a HOME directory which is not set.
         run: |
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
@@ -73,14 +76,12 @@ jobs:
 
       - name: Parse results
         run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
           python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
           --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
+          --hardware "hpc7a.96xlarge" \
+          --project-version "${{ env.COMMIT_HASH }}" \
           --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
+          --commit-date "${{ env.COMMIT_DATE }}" \
           --bench-date "${{ env.BENCH_DATE }}" \
           --walk-subdirs \
           --name-suffix avx512 \
@@ -97,13 +98,13 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_boolean
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -129,8 +130,28 @@ jobs:
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "Boolean benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+  teardown-instance:
+    name: Teardown instance (boolean-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, boolean-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (boolean-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/cargo_build.yml

@@ -19,11 +19,11 @@ jobs:
 
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest-large, windows-latest]
+        os: [large_ubuntu_16, macos-latest-large, large_windows_16_latest]
       fail-fast: false
 
     steps:
-      - uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+      - uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
 
       - name: Install and run newline linter checks
         if: matrix.os == 'ubuntu-latest'

.github/workflows/check_commit.yml

@@ -10,7 +10,7 @@ jobs:
       - name: Check first line
         uses: gsactions/commit-message-checker@16fa2d5de096ae0d35626443bcd24f1e756cafee
         with:
-          pattern: '^((feat|fix|chore|refactor|style|test|docs|doc)(\(\w+\))?\:) .+$'
+          pattern: '^((feat|fix|chore|refactor|style|test|docs|doc)(\([\w\-_]+\))?\!?\:) .+$'
           flags: "gs"
           error: 'Your first line has to contain a commit type and scope like "feat(my_feature): msg".'
           excludeDescription: "true" # optional: this excludes the description body of a pull request

.github/workflows/ci_lint.yml

@@ -13,7 +13,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
 
       - name: Get actionlint
         run: |

.github/workflows/code_coverage.yml

@@ -6,70 +6,58 @@ env:
   RUSTFLAGS: "-C target-cpu=native"
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
-    # All the inputs are provided by Slab
-    inputs:
-      instance_id:
-        description: "AWS instance ID"
-        type: string
-      instance_image_id:
-        description: "AWS instance AMI ID"
-        type: string
-      instance_type:
-        description: "AWS instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: 'Slab request ID'
-        type: string
-      fork_repo:
-        description: 'Name of forked repo as user/repo'
-        type: string
-      fork_git_sha:
-        description: 'Git SHA to checkout from fork'
-        type: string
+  # Code coverage workflow is only run via workflow_dispatch event since execution duration is not stabilized yet.
 
 jobs:
-  code-coverage:
-    concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}_${{ inputs.instance_image_id }}_${{ inputs.instance_type }}
-      cancel-in-progress: true
-    runs-on: ${{ inputs.runner_name }}
-    timeout-minutes: 11520 # 8 days
+  setup-instance:
+    name: Setup instance (code-coverage)
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
-      # Step used for log purpose.
-      - name: Instance configuration used
-        run: |
-          echo "ID: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-          echo "Fork repo: ${{ inputs.fork_repo }}"
-          echo "Fork git sha: ${{ inputs.fork_git_sha }}"
-
-      - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
-          repository: ${{ inputs.fork_repo }}
-          ref: ${{ inputs.fork_git_sha }}
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: cpu-small
+
+  code-coverage:
+    name: Code coverage tests
+    needs: setup-instance
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.event_name }}_${{ github.ref }}
+      cancel-in-progress: true
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    timeout-minutes: 5760 # 4 days
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
 
       - name: Set up home
         run: |
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@03334d095e2739fa9ac4034ec16f66d5d01e9eba
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
         with:
           files_yaml: |
             tfhe:
@@ -99,7 +87,7 @@ jobs:
           make test_shortint_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@125fc84a9a348dbcf27191600683ec096ec9021c
+        uses: codecov/codecov-action@e28ff129e5465c2c0dcc6f003fc735cb6ae0c673
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
@@ -113,7 +101,7 @@ jobs:
           make test_integer_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@125fc84a9a348dbcf27191600683ec096ec9021c
+        uses: codecov/codecov-action@e28ff129e5465c2c0dcc6f003fc735cb6ae0c673
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
@@ -127,8 +115,28 @@ jobs:
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "Code coverage finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+  teardown-instance:
+    name: Teardown instance (code-coverage)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, code-coverage ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (code-coverage) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/core_crypto_benchmark.yml

@@ -3,30 +3,6 @@ name: Core crypto benchmarks
 
 on:
   workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      # This input is not used in this workflow but still mandatory since a calling workflow could
-      # use it. If a triggering command include a user_inputs field, then the triggered workflow
-      # must include this very input, otherwise the workflow won't be called.
-      # See start_full_benchmarks.yml as example.
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
 
 env:
   CARGO_TERM_COLOR: always
@@ -34,67 +10,89 @@ env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
 
 jobs:
-  run-core-crypto-benchmarks:
-    name: Execute core crypto benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
+  setup-instance:
+    name: Setup instance (core-crypto-benchmarks)
+    runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
 
+  core-crypto-benchmarks:
+    name: Execute core crypto benchmarks in EC2
+    needs: setup-instance
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    steps:
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
       - name: Set up home
         # "Install rust" step require root user to have a HOME directory which is not set.
         run: |
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
       - name: Run benchmarks with AVX512
         run: |
           make bench_pbs
+          make bench_pbs128
           make bench_ks
 
       - name: Parse results
         run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
           python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
           --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
+          --hardware "hpc7a.96xlarge" \
+          --project-version "${{ env.COMMIT_HASH }}" \
           --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
+          --commit-date "${{ env.COMMIT_DATE }}" \
           --bench-date "${{ env.BENCH_DATE }}" \
           --name-suffix avx512 \
           --walk-subdirs \
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -120,8 +118,28 @@ jobs:
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "PBS benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+  teardown-instance:
+    name: Teardown instance (core-crypto-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, core-crypto-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (core-crypto-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/core_crypto_gpu_benchmark.yml

@@ -20,13 +20,14 @@ jobs:
   setup-instance:
     name: Setup instance (cuda-core-crypto-benchmarks)
     runs-on: ubuntu-latest
-    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -49,22 +50,13 @@ jobs:
             gcc: 11
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-      CMAKE_VERSION: 3.29.1
+      CMAKE_VERSION: 3.29.6
     steps:
       # Mandatory on hyperstack since a bootable volume is not re-usable yet.
       - name: Install dependencies
         run: |
           sudo apt update
-          sudo apt install ca-certificates curl
-          sudo install -m 0755 -d /etc/apt/keyrings
-          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
-          sudo chmod a+r /etc/apt/keyrings/docker.asc
-          echo \
-          "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
-          $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
-          sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
-          sudo apt update
-          sudo apt install -y checkinstall zlib1g-dev libssl-dev docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
           wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
           tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
           cd cmake-${{ env.CMAKE_VERSION }}
@@ -73,7 +65,7 @@ jobs:
           sudo make install
 
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -91,7 +83,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
@@ -136,13 +128,13 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -165,7 +157,7 @@ jobs:
   slack-notify:
     name: Slack Notification
     needs: [ setup-instance, cuda-core-crypto-benchmarks ]
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    runs-on: ubuntu-latest
     if: ${{ !success() && !cancelled() }}
     continue-on-error: true
     steps:
@@ -183,7 +175,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/csprng_randomness_tests.yml

@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -45,7 +45,7 @@ jobs:
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -54,7 +54,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -78,7 +78,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/data_pr_close.yml

@@ -0,0 +1,123 @@
+name: Close or Merge corresponding PR on the data repo
+
+# When a PR with the data_PR tag is closed or merged, this will close the corresponding PR in the data repo.
+
+env:
+  TARGET_REPO_API_URL: ${{ github.api_url }}/repos/zama-ai/tfhe-backward-compat-data
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  PR_BRANCH: ${{ github.head_ref || github.ref_name }}
+  CLOSE_TYPE: ${{ github.event.pull_request.merged && 'merge' || 'close' }}
+
+# only trigger on pull request closed events
+on:
+  pull_request:
+    types: [ closed ]
+
+# The same pattern is used for jobs that use the github api:
+# - save the result of the API call in the env var "GH_API_RES". Since the var is multiline
+# we use this trick: https://docs.github.com/en/actions/using-workflows/workflow-commands-for-github-actions#example-of-a-multiline-string
+# - "set +e" will make sure we reach the last "echo EOF" even in case of error
+# - "set -o" pipefail makes one line piped command return the error of the first failure
+# - 'RES="$?"' and 'exit $RES' are used to return the error code if a command failed. Without it, with "set +e"
+# the script will always return 0 because of the "echo EOF".
+
+
+jobs:
+  auto_close_job:
+    if: ${{ contains(github.event.pull_request.labels.*.name, 'data_PR') }}
+    runs-on: ubuntu-latest
+    steps:
+    - name: Find corresponding Pull Request in the data repo
+      run: |
+        {
+          set +e
+          set -o pipefail
+          echo 'TARGET_REPO_PR<<EOF'
+          curl --fail-with-body --no-progress-meter -L -X GET \
+          -H "Accept: application/vnd.github+json" \
+          -H "X-GitHub-Api-Version: 2022-11-28"  \
+          ${{ env.TARGET_REPO_API_URL }}/pulls\?head=${{ github.repository_owner }}:${{ env.PR_BRANCH }} | jq -e '.[0]' | sed 's/null/{ "message": "corresponding PR not found" }/'
+          RES="$?"
+          echo EOF
+        } >> "${GITHUB_ENV}"
+        exit $RES
+
+    - name: Comment on the PR to indicate the reason of the close
+      run: |
+        {
+          set +e
+          set -o pipefail
+          echo 'GH_API_RES<<EOF'
+          curl --fail-with-body --no-progress-meter -L -X POST \
+          -H "Accept: application/vnd.github+json" \
+          -H "Authorization: Bearer ${{ secrets.FHE_ACTIONS_TOKEN }}" \
+          -H "X-GitHub-Api-Version: 2022-11-28" \
+          ${{ fromJson(env.TARGET_REPO_PR).comments_url }} \
+          -d '{ "body": "PR ${{ env.CLOSE_TYPE }}d because the corresponding PR in main repo was ${{ env.CLOSE_TYPE }}d: ${{ github.repository }}#${{ github.event.number  }}" }'
+          RES="$?"
+          echo EOF
+        } >> "${GITHUB_ENV}"
+        exit $RES
+
+    - name: Merge the Pull Request in the data repo
+      if: ${{ github.event.pull_request.merged }}
+      run: |
+        {
+          set +e
+          set -o pipefail
+          echo 'GH_API_RES<<EOF'
+          curl --fail-with-body --no-progress-meter -L -X PUT \
+          -H "Accept: application/vnd.github+json" \
+          -H "Authorization: Bearer ${{ secrets.FHE_ACTIONS_TOKEN }}" \
+          -H "X-GitHub-Api-Version: 2022-11-28" \
+          ${{ fromJson(env.TARGET_REPO_PR).url }}/merge \
+          -d '{ "merge_method": "rebase" }'
+          RES="$?"
+          echo EOF
+        } >> "${GITHUB_ENV}"
+        exit $RES
+
+    - name: Close the Pull Request in the data repo
+      if: ${{ !github.event.pull_request.merged }}
+      run: |
+        {
+          set +e
+          set -o pipefail
+          echo 'GH_API_RES<<EOF'
+          curl --fail-with-body --no-progress-meter -L -X PATCH \
+          -H "Accept: application/vnd.github+json" \
+          -H "Authorization: Bearer ${{ secrets.FHE_ACTIONS_TOKEN }}" \
+          -H "X-GitHub-Api-Version: 2022-11-28" \
+          ${{ fromJson(env.TARGET_REPO_PR).url }} \
+          -d '{ "state": "closed" }'
+          RES="$?"
+          echo EOF
+        } >> "${GITHUB_ENV}"
+        exit $RES
+
+    - name: Delete the associated branch in the data repo
+      run: |
+        {
+          set +e
+          set -o pipefail
+          echo 'GH_API_RES<<EOF'
+          curl --fail-with-body --no-progress-meter -L -X DELETE \
+          -H "Accept: application/vnd.github+json" \
+          -H "Authorization: Bearer ${{ secrets.FHE_ACTIONS_TOKEN }}" \
+          -H "X-GitHub-Api-Version: 2022-11-28" \
+          ${{ env.TARGET_REPO_API_URL }}/git/refs/heads/${{ env.PR_BRANCH }}
+          RES="$?"
+          echo EOF
+        } >> "${GITHUB_ENV}"
+        exit $RES
+
+    - name: Slack Notification
+      if: ${{ always() && job.status == 'failure' }}
+      continue-on-error: true
+      uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+      env:
+        SLACK_COLOR: ${{ job.status }}
+        SLACK_MESSAGE: "Failed to auto-${{ env.CLOSE_TYPE }} PR on data repo: ${{ fromJson(env.GH_API_RES || env.TARGET_REPO_PR).message }}"

.github/workflows/gpu_4090_benchmark.yml

@@ -1,5 +1,5 @@
-# Run all benchmarks on an RTX 4090 machine and return parsed results to Slab CI bot.
-name: TFHE Cuda Backend - 4090 full benchmarks
+# Run benchmarks on an RTX 4090 machine and return parsed results to Slab CI bot.
+name: TFHE Cuda Backend - 4090 benchmarks
 
 env:
   CARGO_TERM_COLOR: always
@@ -11,6 +11,7 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_BENCH: TRUE
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
@@ -23,23 +24,22 @@ on:
 
 jobs:
   cuda-integer-benchmarks:
-    name: Cuda integer benchmarks for all operations flavor  (RTX 4090)
-    if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'schedule' || contains(github.event.label.name, '4090_bench') }}
+    name: Cuda integer benchmarks (RTX 4090)
+    if: ${{ github.event_name == 'workflow_dispatch' ||
+      github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs' ||
+      contains(github.event.label.name, '4090_bench') }}
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}_cuda_integer_bench
-      cancel-in-progress: true
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ["self-hosted", "4090-desktop"]
     timeout-minutes: 1440 # 24 hours
     strategy:
       fail-fast: false
       max-parallel: 1
-      matrix:
-        command: [integer, integer_multi_bit]
-        op_flavor: [default, unchecked]
 
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -50,14 +50,15 @@ jobs:
             echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
             echo "COMMIT_HASH=$(git describe --tags --dirty)";
           } >> "${GITHUB_ENV}"
+          echo "FAST_BENCH=TRUE" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -65,7 +66,7 @@ jobs:
 
       - name: Run integer benchmarks
         run: |
-          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_${{ matrix.command }}_gpu
+          make BENCH_OP_FLAVOR=default bench_integer_multi_bit_gpu
 
       - name: Parse results
         run: |
@@ -81,9 +82,9 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
-          name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
+          name: ${{ github.sha }}_integer_multi_bit_gpu_default
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Send data to Slab
@@ -114,13 +115,13 @@ jobs:
     needs: cuda-integer-benchmarks
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}_cuda_core_crypto_bench
-      cancel-in-progress: true
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ["self-hosted", "4090-desktop"]
     timeout-minutes: 1440 # 24 hours
 
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -133,18 +134,18 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
           token: ${{ secrets.FHE_ACTIONS_TOKEN }}
 
-      - name: Run integer benchmarks
+      - name: Run core crypto benchmarks
         run: |
           make bench_pbs_gpu
           make bench_ks_gpu
@@ -163,7 +164,7 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/gpu_4090_tests.yml

@@ -17,11 +17,16 @@ on:
   workflow_dispatch:
   pull_request:
     types: [ labeled ]
+  schedule:
+    # Nightly tests @ 1AM after each work day
+    - cron: "0 1 * * MON-FRI"
 
 jobs:
   cuda-tests-linux:
     name: CUDA tests (RTX 4090)
-    if: ${{ github.event_name == 'workflow_dispatch' || contains(github.event.label.name, '4090_test') }}
+    if: github.event_name == 'workflow_dispatch' ||
+      contains(github.event.label.name, '4090_test') ||
+      (github.event_name == 'schedule' &&  github.repository == 'zama-ai/tfhe-rs')
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}
       cancel-in-progress: true
@@ -29,12 +34,12 @@ jobs:
 
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 

.github/workflows/gpu_fast_h100_tests.yml

@@ -0,0 +1,199 @@
+# Compile and test tfhe-cuda-backend on an H100 VM on hyperstack
+name: TFHE Cuda Backend - Fast tests on H100
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+      types: [ labeled ]
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - Makefile
+              - '.github/workflows/gpu_fast_h100_tests.yml'
+              - scripts/**
+              - ci/**
+
+  setup-instance:
+    name: Setup instance (cuda-h100-tests)
+    needs: should-run
+    if: github.event_name != 'pull_request' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: single-h100
+
+  cuda-tests-linux:
+    name: CUDA H100 tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11 
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run core crypto and internal CUDA backend tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_core_crypto_gpu
+          BIG_TESTS_INSTANCE=TRUE make test_cuda_backend
+
+      - name: Run user docs tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_user_doc_gpu
+
+      - name: Test C API
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_c_api_gpu
+
+      - name: Run High Level API Tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_high_level_api_gpu
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
+          SLACK_MESSAGE: "Fast H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-h100-tests)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_fast_tests.yml

@@ -1,5 +1,5 @@
 # Compile and test tfhe-cuda-backend on an AWS instance
-name: TFHE Cuda Backend - Full tests
+name: TFHE Cuda Backend - Fast tests
 
 env:
   CARGO_TERM_COLOR: always
@@ -11,6 +11,7 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
@@ -18,94 +19,64 @@ on:
   pull_request:
 
 jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - '.github/workflows/gpu_fast_tests.yml'
+              - Makefile
+              - scripts/**
+              - ci/**
+
   setup-instance:
     name: Setup instance (cuda-tests)
+    needs: should-run
+    if: github.event_name != 'pull_request' ||
+      needs.should-run.outputs.gpu_test == 'true'
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
           slab-url: ${{ secrets.SLAB_BASE_URL }}
           job-secret: ${{ secrets.JOB_SECRET }}
-          backend: aws
+          backend: hyperstack
           profile: gpu-test
 
-  cuda-pcc:
-    name: CUDA post-commit checks
-    needs: setup-instance
-    concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
-      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
-    strategy:
-      fail-fast: false
-      # explicit include-based build matrix, of known valid options
-      matrix:
-        include:
-          - os: ubuntu-22.04
-            cuda: "12.2"
-            gcc: 9
-    env:
-      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-
-    steps:
-      - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          persist-credentials: 'false'
-
-      - name: Set up home
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: stable
-
-      - name: Export CUDA variables
-        if: ${{ !cancelled() }}
-        run: |
-          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
-          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
-          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
-          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
-
-      # Specify the correct host compilers
-      - name: Export gcc and g++ variables
-        if: ${{ !cancelled() }}
-        run: |
-          {
-            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
-            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
-            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
-            echo "HOME=/home/ubuntu";
-          } >> "${GITHUB_ENV}"
-
-      - name: Run fmt checks
-        run: |
-          make check_fmt_gpu
-
-      - name: Run clippy checks
-        run: |
-          make pcc_gpu
-
-      - name: Slack Notification
-        if: ${{ always() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "CUDA AWS post-commit checks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-
   cuda-tests-linux:
     name: CUDA tests
-    needs: [ setup-instance, cuda-pcc ]
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
@@ -117,13 +88,25 @@ jobs:
         include:
           - os: ubuntu-22.04
             cuda: "12.2"
-            gcc: 9
+            gcc: 11 
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-
+      CMAKE_VERSION: 3.29.6
     steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
@@ -132,7 +115,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -155,9 +138,14 @@ jobs:
             echo "HOME=/home/ubuntu";
           } >> "${GITHUB_ENV}"
 
-      - name: Run core crypto, integer and internal CUDA backend tests
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run core crypto and internal CUDA backend tests
         run: |
-          make test_gpu
+          make test_core_crypto_gpu
+          make test_cuda_backend
 
       - name: Run user docs tests
         run: |
@@ -171,23 +159,28 @@ jobs:
         run: |
           make test_high_level_api_gpu
 
-      - name: Slack Notification
-        if: ${{ always() }}
-        continue-on-error: true
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "CUDA AWS tests finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
+          SLACK_MESSAGE: "Base GPU tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.ACTION_RUN_URL }})"
 
   teardown-instance:
     name: Teardown instance (cuda-tests)
     if: ${{ always() && needs.setup-instance.result != 'skipped' }}
-    needs: [ setup-instance, cuda-pcc, cuda-tests-linux ]
+    needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_full_multi_gpu_tests.yml

@@ -0,0 +1,199 @@
+# Compile and test tfhe-cuda-backend on an AWS instance
+name: TFHE Cuda Backend - Full tests multi-GPU
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+    types: [ labeled ]
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - Makefile
+              - '.github/workflows/**_multi_gpu_tests.yml'
+              - scripts/**
+              - ci/**
+
+  setup-instance:
+    name: Setup instance (cuda-tests-multi-gpu)
+    needs: should-run
+    if: github.event_name != 'pull_request' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: multi-gpu-test
+
+  cuda-tests-linux:
+    name: CUDA multi-GPU tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11 
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      # No need to test core_crypto and classic PBS in integer since it's already tested on single GPU.
+      - name: Run multi-bit CUDA integer tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_integer_multi_bit_gpu_ci
+
+      - name: Run user docs tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_user_doc_gpu
+
+      - name: Test C API
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_c_api_gpu
+
+      - name: Run High Level API Tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_high_level_api_gpu
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
+          SLACK_MESSAGE: "Multi-GPU tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-tests-multi-gpu)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-tests-multi-gpu) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_pcc.yml

@@ -0,0 +1,126 @@
+# Perfom tfhe-cuda-backend post-commit checks on an AWS instance
+name: TFHE Cuda Backend - Post-commit Checks
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+on:
+  pull_request:
+
+jobs:
+  setup-instance:
+    name: Setup instance (cuda-pcc)
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: gpu-build
+
+  cuda-pcc:
+    name: CUDA post-commit checks
+    needs: setup-instance
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: true
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 9
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Run fmt checks
+        run: |
+          make check_fmt_gpu
+
+      - name: Run clippy checks
+        run: |
+          make pcc_gpu
+
+      - name: Slack Notification
+        if: ${{ always() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "CUDA AWS post-commit checks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-pcc)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-pcc ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-pcc) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_signed_integer_h100_tests.yml

@@ -1,5 +1,5 @@
-# Compile and test tfhe-cuda-backend on an H100 VM on hyperstack
-name: TFHE Cuda Backend - Full tests on H100
+# Signed integer GPU tests on an H100 VM on hyperstack
+name: TFHE Cuda Backend - Signed integer tests on H100
 
 env:
   CARGO_TERM_COLOR: always
@@ -11,22 +11,61 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
 
 on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
+      types: [ labeled ]
 
 jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - Makefile
+              - '.github/workflows/gpu_signed_integer_h100_tests.yml'
+              - scripts/**
+              - ci/**
+
   setup-instance:
     name: Setup instance (cuda-h100-tests)
+    needs: should-run
+    if: github.event_name != 'pull_request' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
     runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -36,8 +75,10 @@ jobs:
           profile: single-h100
 
   cuda-tests-linux:
-    name: CUDA H100 tests
-    needs: [ setup-instance ]
+    name: CUDA H100 signed integer tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
       group: ${{ github.workflow }}_${{ github.ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
@@ -52,22 +93,13 @@ jobs:
             gcc: 11 
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-      CMAKE_VERSION: 3.29.1
+      CMAKE_VERSION: 3.29.6
     steps:
       # Mandatory on hyperstack since a bootable volume is not re-usable yet.
       - name: Install dependencies
         run: |
           sudo apt update
-          sudo apt install ca-certificates curl
-          sudo install -m 0755 -d /etc/apt/keyrings
-          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
-          sudo chmod a+r /etc/apt/keyrings/docker.asc
-          echo \
-          "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
-           $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
-          sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
-          sudo apt update
-          sudo apt install -y checkinstall zlib1g-dev libssl-dev docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
           wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
           tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
           cd cmake-${{ env.CMAKE_VERSION }}
@@ -76,14 +108,14 @@ jobs:
           sudo make install
 
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
 
       - name: Set up home
         run: |
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -106,27 +138,23 @@ jobs:
             echo "HOME=/home/ubuntu";
           } >> "${GITHUB_ENV}"
 
-      - name: Run core crypto, integer and internal CUDA backend tests
-        run: |
-          make test_gpu
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
 
-      - name: Run user docs tests
+      - name: Run signed integer tests
         run: |
-          make test_user_doc_gpu
+          BIG_TESTS_INSTANCE=TRUE make test_signed_integer_gpu_ci
 
-      - name: Test C API
+      - name: Run signed integer multi-bit tests
         run: |
-          make test_c_api_gpu
-
-      - name: Run High Level API Tests
-        run: |
-          make test_high_level_api_gpu
+          BIG_TESTS_INSTANCE=TRUE make test_signed_integer_multi_bit_gpu_ci
 
   slack-notify:
     name: Slack Notification
     needs: [ setup-instance, cuda-tests-linux ]
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
-    if: ${{ !success() && !cancelled() }}
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' }}
     continue-on-error: true
     steps:
       - name: Send message
@@ -143,7 +171,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/gpu_signed_integer_tests.yml

@@ -0,0 +1,202 @@
+# Compile and test tfhe-cuda-backend signed integer on an AWS instance
+name: TFHE Cuda Backend - Signed integer tests
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_TESTS: TRUE
+  NIGHTLY_TESTS: FALSE
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+    types:
+      - opened
+      - synchronize
+      - labeled
+  schedule:
+    # Nightly tests @ 1AM after each work day
+    - cron: "0 1 * * MON-FRI"
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - '.github/workflows/gpu_signed_integer_tests.yml'
+              - Makefile
+              - scripts/**
+              - ci/**
+
+  setup-instance:
+    name: Setup instance (cuda-signed-integer-tests)
+    runs-on: ubuntu-latest
+    needs: should-run
+    if: (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true')
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: gpu-test
+
+  cuda-signed-integer-tests:
+    name: CUDA signed integer tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11 
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          persist-credentials: 'false'
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Should run nightly tests
+        if: github.event_name == 'schedule'
+        run: |
+          {
+            echo "FAST_TESTS=FALSE";
+            echo "NIGHTLY_TESTS=TRUE";
+          } >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run signed integer multi-bit tests
+        run: |
+          make test_signed_integer_multi_bit_gpu_ci
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-signed-integer-tests ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-signed-integer-tests.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-signed-integer-tests.result }}
+          SLACK_MESSAGE: "Base GPU tests finished with status: ${{ needs.cuda-signed-integer-tests.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-tests)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-signed-integer-tests ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-signed-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_unsigned_integer_h100_tests.yml

@@ -0,0 +1,188 @@
+# Test unsigned integers on an H100 VM on hyperstack
+name: TFHE Cuda Backend - Unsigned integer tests on H100
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+      types: [ labeled ]
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - Makefile
+              - '.github/workflows/gpu_unsigned_integer_tests.yml'
+              - scripts/**
+              - ci/**
+
+  setup-instance:
+    name: Setup instance (cuda-h100-tests)
+    needs: should-run
+    if: github.event_name != 'pull_request' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true') ||
+      (github.event.action == 'labeled' && github.event.label.name == 'approved' && needs.should-run.outputs.gpu_test == 'true')
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: single-h100
+
+  cuda-tests-linux:
+    name: CUDA H100 unsigned integer tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11 
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run unsigned integer tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_unsigned_integer_gpu_ci
+
+      - name: Run unsigned integer multi-bit tests
+        run: |
+          BIG_TESTS_INSTANCE=TRUE make test_unsigned_integer_multi_bit_gpu_ci
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
+          SLACK_MESSAGE: "Unsigned integer GPU H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-h100-tests)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-tests-linux ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_unsigned_integer_tests.yml

@@ -0,0 +1,199 @@
+# Compile and test tfhe-cuda-backend unsigned integer on an AWS instance
+name: TFHE Cuda Backend - Unsigned integer tests
+
+env:
+  CARGO_TERM_COLOR: always
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUSTFLAGS: "-C target-cpu=native"
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_TESTS: TRUE
+  NIGHTLY_TESTS: FALSE
+
+on:
+  # Allows you to run this workflow manually from the Actions tab as an alternative.
+  workflow_dispatch:
+  pull_request:
+    types:
+      - opened
+      - synchronize
+      - labeled
+  schedule:
+    # Nightly tests @ 1AM after each work day
+    - cron: "0 1 * * MON-FRI"
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: write
+    outputs:
+      gpu_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.gpu_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            gpu:
+              - tfhe/Cargo.toml
+              - tfhe/build.rs
+              - backends/tfhe-cuda-backend/**
+              - tfhe/src/core_crypto/gpu/**
+              - tfhe/src/integer/gpu/**
+              - tfhe/shortint/parameters/**
+              - tfhe/src/high_level_api/**
+              - tfhe/src/c_api/**
+              - 'tfhe/docs/**.md'
+              - '.github/workflows/gpu_unsigned_integer_tests.yml'
+              - Makefile
+              - scripts/**
+              - ci/**
+
+  setup-instance:
+    name: Setup instance (cuda-unsigned-integer-tests)
+    needs: should-run
+    if: (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      github.event_name == 'workflow_dispatch' ||
+      (github.event.action != 'labeled' && needs.should-run.outputs.gpu_test == 'true')
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: gpu-test
+
+  cuda-unsigned-integer-tests:
+    name: CUDA unsigned integer tests
+    needs: [ should-run, setup-instance ]
+    if: github.event_name != 'pull_request' ||
+      (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      fail-fast: false
+      # explicit include-based build matrix, of known valid options
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev 
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+
+      - name: Set up home
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install latest stable
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: stable
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc" >> "${GITHUB_ENV}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "HOME=/home/ubuntu";
+          } >> "${GITHUB_ENV}"
+
+      - name: Should run nightly tests
+        if: github.event_name == 'schedule'
+        run: |
+          {
+            echo "FAST_TESTS=FALSE";
+            echo "NIGHTLY_TESTS=TRUE";
+          } >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run unsigned integer multi-bit tests
+        run: |
+          make test_unsigned_integer_multi_bit_gpu_ci
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-unsigned-integer-tests ]
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-unsigned-integer-tests.result != 'skipped' }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-unsigned-integer-tests.result }}
+          SLACK_MESSAGE: "Unsigned integer GPU tests finished with status: ${{ needs.cuda-unsigned-integer-tests.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-tests)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-unsigned-integer-tests ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-unsigned-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/integer_benchmark.yml

@@ -1,130 +0,0 @@
-# Run integer benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Integer benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  PARSE_INTEGER_BENCH_CSV_FILE: tfhe_rs_integer_benches_${{ github.sha }}.csv
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  run-integer-benchmarks:
-    name: Execute integer benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Run benchmarks with AVX512
-        run: |
-          make FAST_BENCH=TRUE bench_integer
-
-      - name: Parse benchmarks to csv
-        run: |
-          make PARSE_INTEGER_BENCH_CSV_FILE=${{ env.PARSE_INTEGER_BENCH_CSV_FILE }} \
-            parse_integer_benches
-
-      - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_csv_integer
-          path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_integer
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-      - name: Slack Notification
-        if: ${{ failure() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Integer benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/integer_cpu_benchmark.yml

@@ -1,28 +1,20 @@
 # Run all integer benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Integer full benchmarks
+name: Integer benchmarks
 
 on:
   workflow_dispatch:
     inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
+      all_precisions:
+        description: "Run all precisions"
+        type: boolean
+        default: false
+
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
+    # Quarterly benchmarks will be triggered right before end of quarter, the 25th of the current month at 4a.m.
+    # These benchmarks are far longer to execute hence the reason to run them only four time a year.
+    - cron: '0 4 25 MAR,JUN,SEP,DEC *'
 
 env:
   CARGO_TERM_COLOR: always
@@ -30,21 +22,29 @@ env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_BENCH: TRUE
 
 jobs:
   prepare-matrix:
     name: Prepare operations matrix
     runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
       op_flavor: ${{ steps.set_op_flavor.outputs.op_flavor }}
     steps:
       - name: Weekly benchmarks
-        if: ${{ github.event.inputs.user_inputs == 'weekly_benchmarks' }}
+        if: github.event_name == 'workflow_dispatch' ||
+          github.event.schedule == '0 1 * * 6'
         run: |
           echo "OP_FLAVOR=[\"default\"]" >> "${GITHUB_ENV}"
 
       - name: Quarterly benchmarks
-        if: ${{ github.event.inputs.user_inputs == 'quarterly_benchmarks' }}
+        if: github.event.schedule == '0 4 25 MAR,JUN,SEP,DEC *'
         run: |
           echo "OP_FLAVOR=[\"default\", \"smart\", \"unchecked\", \"misc\"]" >> "${GITHUB_ENV}"
 
@@ -53,11 +53,31 @@ jobs:
          run: |
           echo "op_flavor=${{ toJSON(env.OP_FLAVOR) }}" >> "${GITHUB_OUTPUT}"
 
+  setup-instance:
+    name: Setup instance (integer-benchmarks)
+    needs: prepare-matrix
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
+
   integer-benchmarks:
     name: Execute integer benchmarks for all operations flavor
-    needs: prepare-matrix
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
+    needs: [ prepare-matrix, setup-instance ]
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     timeout-minutes: 1440  # 24 hours
     strategy:
@@ -66,15 +86,8 @@ jobs:
         command: [ integer, integer_multi_bit]
         op_flavor: ${{ fromJson(needs.prepare-matrix.outputs.op_flavor) }}
     steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -92,17 +105,22 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
           token: ${{ secrets.FHE_ACTIONS_TOKEN }}
 
+      - name: Should run benchmarks with all precisions
+        if: inputs.all_precisions
+        run: |
+          echo "FAST_BENCH=FALSE" >> "${GITHUB_ENV}"
+
       - name: Run benchmarks with AVX512
         run: |
           make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_${{ matrix.command }}
@@ -111,7 +129,7 @@ jobs:
         run: |
           python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
           --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
+          --hardware "hpc7a.96xlarge" \
           --project-version "${{ env.COMMIT_HASH }}" \
           --branch ${{ github.ref_name }} \
           --commit-date "${{ env.COMMIT_DATE }}" \
@@ -121,7 +139,7 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
           path: ${{ env.RESULTS_FILENAME }}
@@ -140,19 +158,34 @@ jobs:
           -d @${{ env.RESULTS_FILENAME }} \
           ${{ secrets.SLAB_URL }}
 
-  slack-notification:
-    name: Slack Notification
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ failure() }}
-    needs: integer-benchmarks
-    steps:
-      - name: Notify
+      - name: Slack Notification
+        if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "Integer full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+  teardown-instance:
+    name: Teardown instance (integer-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, integer-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (integer-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/integer_gpu_benchmark.yml

@@ -23,13 +23,14 @@ jobs:
   setup-instance:
     name: Setup instance (cuda-integer-benchmarks)
     runs-on: ubuntu-latest
-    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
+    if:  github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -52,22 +53,13 @@ jobs:
             gcc: 11
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-      CMAKE_VERSION: 3.29.1
+      CMAKE_VERSION: 3.29.6
     steps:
       # Mandatory on hyperstack since a bootable volume is not re-usable yet.
       - name: Install dependencies
         run: |
           sudo apt update
-          sudo apt install ca-certificates curl
-          sudo install -m 0755 -d /etc/apt/keyrings
-          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
-          sudo chmod a+r /etc/apt/keyrings/docker.asc
-          echo \
-          "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
-          $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
-          sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
-          sudo apt update
-          sudo apt install -y checkinstall zlib1g-dev libssl-dev docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
           wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
           tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
           cd cmake-${{ env.CMAKE_VERSION }}
@@ -76,7 +68,7 @@ jobs:
           sudo make install
 
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -94,7 +86,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
@@ -118,6 +110,10 @@ jobs:
             echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
           } >> "${GITHUB_ENV}"
 
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
       - name: Run benchmarks with AVX512
         run: |
           make FAST_BENCH=TRUE BENCH_OP_FLAVOR=default bench_integer_gpu
@@ -128,7 +124,7 @@ jobs:
             parse_integer_benches
 
       - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_csv_integer
           path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
@@ -148,13 +144,13 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_integer
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -177,8 +173,8 @@ jobs:
   slack-notify:
     name: Slack Notification
     needs: [ setup-instance, cuda-integer-benchmarks ]
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
-    if: ${{ !success() && !cancelled() }}
+    runs-on: ubuntu-latest
+    if: ${{ always() && needs.cuda-integer-benchmarks.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
       - name: Send message
@@ -195,7 +191,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/integer_gpu_full_benchmark.yml

@@ -22,13 +22,14 @@ jobs:
   setup-instance:
     name: Setup instance (cuda-integer-full-benchmarks)
     runs-on: ubuntu-latest
-    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -56,22 +57,13 @@ jobs:
             gcc: 11
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-      CMAKE_VERSION: 3.29.1
+      CMAKE_VERSION: 3.29.6
     steps:
       # Mandatory on hyperstack since a bootable volume is not re-usable yet.
       - name: Install dependencies
         run: |
           sudo apt update
-          sudo apt install ca-certificates curl
-          sudo install -m 0755 -d /etc/apt/keyrings
-          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
-          sudo chmod a+r /etc/apt/keyrings/docker.asc
-          echo \
-          "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
-          $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
-          sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
-          sudo apt update
-          sudo apt install -y checkinstall zlib1g-dev libssl-dev docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
           wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
           tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
           cd cmake-${{ env.CMAKE_VERSION }}
@@ -80,7 +72,7 @@ jobs:
           sudo make install
 
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -98,7 +90,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
@@ -123,12 +115,16 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
           token: ${{ secrets.FHE_ACTIONS_TOKEN }}
 
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
       - name: Run benchmarks with AVX512
         run: |
           make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_${{ matrix.command }}_gpu
@@ -148,7 +144,7 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
           path: ${{ env.RESULTS_FILENAME }}
@@ -170,7 +166,7 @@ jobs:
   slack-notify:
     name: Slack Notification
     needs: [ setup-instance, cuda-integer-full-benchmarks ]
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    runs-on: ubuntu-latest
     if: ${{ !success() && !cancelled() }}
     continue-on-error: true
     steps:
@@ -188,7 +184,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/integer_multi_bit_benchmark.yml

@@ -1,130 +0,0 @@
-# Run integer benchmarks with multi-bit cryptographic parameters on an AWS instance and return parsed results to Slab CI bot.
-name: Integer Multi-bit benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  PARSE_INTEGER_BENCH_CSV_FILE: tfhe_rs_integer_benches_${{ github.sha }}.csv
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  run-integer-benchmarks:
-    name: Execute integer multi-bit benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Run multi-bit benchmarks with AVX512
-        run: |
-          make FAST_BENCH=TRUE bench_integer_multi_bit
-
-      - name: Parse benchmarks to csv
-        run: |
-          make PARSE_INTEGER_BENCH_CSV_FILE=${{ env.PARSE_INTEGER_BENCH_CSV_FILE }} \
-            parse_integer_benches
-
-      - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_csv_integer
-          path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_integer
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-      - name: Slack Notification
-        if: ${{ failure() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Integer benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/integer_multi_bit_gpu_benchmark.yml

@@ -3,6 +3,16 @@ name: Integer GPU Multi-bit benchmarks
 
 on:
   workflow_dispatch:
+    inputs:
+      all_precisions:
+        description: "Run all precisions"
+        type: boolean
+        default: false
+      fast_default:
+        description: "Run only deduplicated default operations without scalar variants"
+        type: boolean
+        default: false
+
   schedule:
     # Weekly benchmarks will be triggered each Saturday at 1a.m.
     - cron: '0 1 * * 6'
@@ -18,18 +28,21 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_BENCH: TRUE
+  BENCH_OP_FLAVOR: default
 
 jobs:
   setup-instance:
     name: Setup instance (cuda-integer-multi-bit-benchmarks)
     runs-on: ubuntu-latest
-    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
       runner-name: ${{ steps.start-instance.outputs.label }}
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -53,22 +66,13 @@ jobs:
             gcc: 11
     env:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
-      CMAKE_VERSION: 3.29.1
+      CMAKE_VERSION: 3.29.6
     steps:
       # Mandatory on hyperstack since a bootable volume is not re-usable yet.
       - name: Install dependencies
         run: |
           sudo apt update
-          sudo apt install ca-certificates curl
-          sudo install -m 0755 -d /etc/apt/keyrings
-          sudo curl -fsSL https://download.docker.com/linux/ubuntu/gpg -o /etc/apt/keyrings/docker.asc
-          sudo chmod a+r /etc/apt/keyrings/docker.asc
-          echo \
-          "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/docker.asc] https://download.docker.com/linux/ubuntu \
-          $(. /etc/os-release && echo "$VERSION_CODENAME") stable" | \
-          sudo tee /etc/apt/sources.list.d/docker.list > /dev/null
-          sudo apt update
-          sudo apt install -y checkinstall zlib1g-dev libssl-dev docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
           wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
           tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
           cd cmake-${{ env.CMAKE_VERSION }}
@@ -77,7 +81,7 @@ jobs:
           sudo make install
 
       - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -95,7 +99,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
@@ -119,9 +123,23 @@ jobs:
             echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
           } >> "${GITHUB_ENV}"
 
+      - name: Should run benchmarks with all precisions
+        if: inputs.all_precisions
+        run: |
+          echo "FAST_BENCH=FALSE" >> "${GITHUB_ENV}"
+
+      - name: Should run fast subset benchmarks
+        if: inputs.fast_default
+        run: |
+          echo "BENCH_OP_FLAVOR=fast_default" >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
       - name: Run multi-bit benchmarks with AVX512
         run: |
-          make FAST_BENCH=TRUE BENCH_OP_FLAVOR=default bench_integer_multi_bit_gpu
+          make bench_unsigned_integer_multi_bit_gpu
 
       - name: Parse benchmarks to csv
         run: |
@@ -129,7 +147,7 @@ jobs:
             parse_integer_benches
 
       - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_csv_integer
           path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
@@ -149,13 +167,13 @@ jobs:
           --throughput
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_integer
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -179,7 +197,7 @@ jobs:
   slack-notify:
     name: Slack Notification
     needs: [ setup-instance, cuda-integer-multi-bit-benchmarks ]
-    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    runs-on: ubuntu-latest
     if: ${{ !success() && !cancelled() }}
     continue-on-error: true
     steps:
@@ -197,7 +215,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/integer_multi_bit_multi_gpu_benchmark.yml

@@ -0,0 +1,221 @@
+# Run 64-bit multi-bit integer benchmarks on an instance with CUDA and return parsed results to Slab CI bot.
+name: Integer multi GPU Multi-bit benchmarks
+
+on:
+  workflow_dispatch:
+    inputs:
+      all_precisions:
+        description: "Run all precisions"
+        type: boolean
+        default: false
+      fast_default:
+        description: "Run only deduplicated default operations without scalar variants"
+        type: boolean
+        default: false
+
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
+
+env:
+  CARGO_TERM_COLOR: always
+  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_BENCH: TRUE
+  BENCH_OP_FLAVOR: default
+
+jobs:
+  setup-instance:
+    name: Setup instance (cuda-integer-multi-bit-multi-gpu-benchmarks)
+    runs-on: ubuntu-latest
+    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      github.event_name == 'workflow_dispatch' }}
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: multi-h100
+
+  cuda-integer-multi-bit-multi-gpu-benchmarks:
+    name: Execute multi GPU integer multi-bit benchmarks
+    needs: setup-instance
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    timeout-minutes: 1440 # 24 hours
+    continue-on-error: true
+    strategy:
+      fail-fast: false
+      max-parallel: 1
+      matrix:
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
+      - name: Set up home
+        # "Install rust" step require root user to have a HOME directory which is not set.
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install rust
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: nightly
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CUDA_PATH=$CUDA_PATH";
+            echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH";
+            echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc";
+          } >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+          } >> "${GITHUB_ENV}"
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Should run benchmarks with all precisions
+        if: inputs.all_precisions
+        run: |
+          echo "FAST_BENCH=FALSE" >> "${GITHUB_ENV}"
+
+      - name: Should run fast subset benchmarks
+        if: inputs.fast_default
+        run: |
+          echo "BENCH_OP_FLAVOR=fast_default" >> "${GITHUB_ENV}"
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run multi-bit benchmarks with AVX512
+        run: |
+          make bench_unsigned_integer_multi_bit_gpu
+
+      - name: Parse results
+        run: |
+          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
+          --database tfhe_rs \
+          --hardware "n3-H100x8" \
+          --backend gpu \
+          --project-version "${{ env.COMMIT_HASH }}" \
+          --branch ${{ github.ref_name }} \
+          --commit-date "${{ env.COMMIT_DATE }}" \
+          --bench-date "${{ env.BENCH_DATE }}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --throughput
+
+      - name: Upload parsed results artifact
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
+        with:
+          name: ${{ github.sha }}_integer
+          path: ${{ env.RESULTS_FILENAME }}
+
+      - name: Send data to Slab
+        shell: bash
+        run: |
+          echo "Computing HMac on results file"
+          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
+          echo "Sending results to Slab..."
+          curl -v -k \
+          -H "Content-Type: application/json" \
+          -H "X-Slab-Repository: ${{ github.repository }}" \
+          -H "X-Slab-Command: store_data_v2" \
+          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
+          -d @${{ env.RESULTS_FILENAME }} \
+          ${{ secrets.SLAB_URL }}
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-integer-multi-bit-multi-gpu-benchmarks ]
+    runs-on: ubuntu-latest
+    if: ${{ !success() && !cancelled() }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-integer-multi-bit-multi-gpu-benchmarks.result }}
+          SLACK_MESSAGE: "Integer multi GPU multi-bit benchmarks finished with status: ${{ needs.cuda-integer-multi-bit-multi-gpu-benchmarks.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-integer-multi-bit-multi-gpu-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-integer-multi-bit-multi-gpu-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-integer-multi-bit-multi-gpu-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/integer_multi_gpu_full_benchmark.yml

@@ -0,0 +1,201 @@
+# Run all integer benchmarks on an instance with CUDA and return parsed results to Slab CI bot.
+name: Integer multi GPU full benchmarks
+
+on:
+  workflow_dispatch:
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
+
+env:
+  CARGO_TERM_COLOR: always
+  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+jobs:
+  setup-instance:
+    name: Setup instance (cuda-integer-full-multi-gpu-benchmarks)
+    runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: hyperstack
+          profile: multi-h100
+
+  cuda-integer-full-multi-gpu-benchmarks:
+    name: Execute multi GPU integer benchmarks for all operations flavor
+    needs: setup-instance
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    timeout-minutes: 1440 # 24 hours
+    continue-on-error: true
+    strategy:
+      fail-fast: false
+      max-parallel: 1
+      matrix:
+        command: [integer, integer_multi_bit]
+        op_flavor: [default, unchecked]
+        # explicit include-based build matrix, of known valid options
+        include:
+          - os: ubuntu-22.04
+            cuda: "12.2"
+            gcc: 11
+    env:
+      CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
+      CMAKE_VERSION: 3.29.6
+    steps:
+      # Mandatory on hyperstack since a bootable volume is not re-usable yet.
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y checkinstall zlib1g-dev libssl-dev
+          wget https://github.com/Kitware/CMake/releases/download/v${{ env.CMAKE_VERSION }}/cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          tar -zxvf cmake-${{ env.CMAKE_VERSION }}.tar.gz
+          cd cmake-${{ env.CMAKE_VERSION }}
+          ./bootstrap
+          make -j"$(nproc)"
+          sudo make install
+
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
+      - name: Set up home
+        # "Install rust" step require root user to have a HOME directory which is not set.
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install rust
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: nightly
+
+      - name: Export CUDA variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CUDA_PATH=$CUDA_PATH";
+            echo "LD_LIBRARY_PATH=$CUDA_PATH/lib:$LD_LIBRARY_PATH";
+            echo "CUDACXX=/usr/local/cuda-${{ matrix.cuda }}/bin/nvcc";
+          } >> "${GITHUB_ENV}"
+          echo "$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+
+      # Specify the correct host compilers
+      - name: Export gcc and g++ variables
+        if: ${{ !cancelled() }}
+        run: |
+          {
+            echo "CC=/usr/bin/gcc-${{ matrix.gcc }}";
+            echo "CXX=/usr/bin/g++-${{ matrix.gcc }}";
+            echo "CUDAHOSTCXX=/usr/bin/g++-${{ matrix.gcc }}";
+          } >> "${GITHUB_ENV}"
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Check device is detected
+        if: ${{ !cancelled() }}
+        run: nvidia-smi
+
+      - name: Run benchmarks with AVX512
+        run: |
+          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_${{ matrix.command }}_gpu
+
+      - name: Parse results
+        run: |
+          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
+          --database tfhe_rs \
+          --hardware "n3-H100x8" \
+          --backend gpu \
+          --project-version "${{ env.COMMIT_HASH }}" \
+          --branch ${{ github.ref_name }} \
+          --commit-date "${{ env.COMMIT_DATE }}" \
+          --bench-date "${{ env.BENCH_DATE }}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --throughput
+
+      - name: Upload parsed results artifact
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
+        with:
+          name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
+          path: ${{ env.RESULTS_FILENAME }}
+
+      - name: Send data to Slab
+        shell: bash
+        run: |
+          echo "Computing HMac on results file"
+          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
+          echo "Sending results to Slab..."
+          curl -v -k \
+          -H "Content-Type: application/json" \
+          -H "X-Slab-Repository: ${{ github.repository }}" \
+          -H "X-Slab-Command: store_data_v2" \
+          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
+          -d @${{ env.RESULTS_FILENAME }} \
+          ${{ secrets.SLAB_URL }}
+
+  slack-notify:
+    name: Slack Notification
+    needs: [ setup-instance, cuda-integer-full-multi-gpu-benchmarks ]
+    runs-on: ubuntu-latest
+    if: ${{ !success() && !cancelled() }}
+    continue-on-error: true
+    steps:
+      - name: Send message
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ needs.cuda-integer-full-multi-gpu-benchmarks.result }}
+          SLACK_MESSAGE: "Integer GPU full benchmarks finished with status: ${{ needs.cuda-integer-full-multi-gpu-benchmarks.result }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (cuda-integer-full-multi-gpu-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, cuda-integer-full-multi-gpu-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (cuda-integer-full-multi-gpu-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/m1_tests.yml

@@ -18,6 +18,9 @@ env:
   RUST_MIN_STACK: "8388608"
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
   FAST_TESTS: "TRUE"
+  # We clear the cache to reduce memory pressure because of the numerous processes of cargo
+  # nextest
+  TFHE_RS_CLEAR_IN_MEMORY_KEY_CACHE: "1"
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref }}
@@ -31,12 +34,12 @@ jobs:
     timeout-minutes: 720
 
     steps:
-      - uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+      - uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           persist-credentials: 'false'
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 

.github/workflows/make_release.yml

@@ -20,20 +20,72 @@ on:
         description: "Push node js package"
         type: boolean
         default: true
+      npm_latest_tag:
+        description: "Set NPM tag as latest"
+        type: boolean
+        default: false
 
 env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  NPM_TAG: ""
 
 jobs:
-  publish_release:
-    name: Publish Release
+  package:
     runs-on: ubuntu-latest
+    outputs:
+      hash: ${{ steps.hash.outputs.hash }}
     steps:
       - name: Checkout
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
+      - name: Prepare package
+        run: |
+          cargo package -p tfhe
+      - uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a # v4.3.6
+        with:
+          name: crate
+          path: target/package/*.crate
+      - name: generate hash
+        id: hash
+        run: cd target/package && echo "hash=$(sha256sum ./*.crate | base64 -w0)" >> "${GITHUB_OUTPUT}"
 
+  provenance:
+    if: ${{ !inputs.dry_run  }}
+    needs: [package]
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    permissions:
+      # Needed to detect the GitHub Actions environment
+      actions: read
+      # Needed to create the provenance via GitHub OIDC
+      id-token: write
+      # Needed to upload assets/artifacts
+      contents: write
+    with:
+      # SHA-256 hashes of the Crate package.
+      base64-subjects: ${{ needs.package.outputs.hash }}
+
+  publish_release:
+    name: Publish Release
+    needs: [package] # for comparing hashes
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      id-token: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+      - name: Create NPM version tag
+        if: ${{ inputs.npm_latest_tag }}
+        run: |
+          echo "NPM_TAG=latest" >> "${GITHUB_ENV}"
+      - name: Download artifact
+        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        with:
+          name: crate
+          path: target/package
       - name: Publish crate.io package
         if: ${{ inputs.push_to_crates }}
         env:
@@ -42,10 +94,26 @@ jobs:
         run: |
           cargo publish -p tfhe --token ${{ env.CRATES_TOKEN }} ${{ env.DRY_RUN }}
 
+      - name: Generate hash
+        id: published_hash
+        run: cd target/package && echo "pub_hash=$(sha256sum ./*.crate | base64 -w0)" >> "${GITHUB_OUTPUT}"
+
+      - name: Slack notification (hashes comparison)
+        if: ${{ needs.package.outputs.hash != steps.published_hash.outputs.pub_hash }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: failure
+          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+          SLACK_MESSAGE: "SLSA tfhe crate - hash comparison failure: (${{ env.ACTION_RUN_URL }})"
+          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
       - name: Build web package
         if: ${{ inputs.push_web_package }}
         run: |
-          make build_web_js_api
+          make build_web_js_api_parallel
 
       - name: Publish web package
         if: ${{ inputs.push_web_package }}
@@ -54,6 +122,8 @@ jobs:
           token: ${{ secrets.NPM_TOKEN }}
           package: tfhe/pkg/package.json
           dry-run: ${{ inputs.dry_run }}
+          tag: ${{ env.NPM_TAG }}
+          provenance: true
 
       - name: Build Node package
         if: ${{ inputs.push_node_package }}
@@ -70,6 +140,8 @@ jobs:
           token: ${{ secrets.NPM_TOKEN }}
           package: tfhe/pkg/package.json
           dry-run: ${{ inputs.dry_run }}
+          tag: ${{ env.NPM_TAG }}
+          provenance: true
 
       - name: Slack Notification
         if: ${{ failure() }}

.github/workflows/make_release_concrete_csprng.yml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 

.github/workflows/make_release_cuda.yml

@@ -29,7 +29,7 @@ jobs:
     steps:
       - name: Start instance
         id: start-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: start
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
@@ -54,7 +54,7 @@ jobs:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
     steps:
       - name: Checkout
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
@@ -63,7 +63,7 @@ jobs:
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: stable
 
@@ -112,7 +112,7 @@ jobs:
     steps:
       - name: Stop instance
         id: stop-instance
-        uses: zama-ai/slab-github-runner@1dced74825027fe3d481392163ed8fc56813fb5d
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
         with:
           mode: stop
           github-token: ${{ secrets.SLAB_ACTION_TOKEN }}

.github/workflows/make_release_zk_pok.yml

@@ -18,7 +18,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 

.github/workflows/parameters_check.yml

@@ -14,17 +14,17 @@ on:
 
 jobs:
   params-curves-security-check:
-    runs-on: ubuntu-latest
+    runs-on: large_ubuntu_16
     steps:
       - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
 
       - name: Checkout lattice-estimator
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: malb/lattice-estimator
           path: lattice_estimator
-          ref: '53508253629d3b5d31a2ad110e85dc69391ccb95'
+          ref: 'e80ec6bbbba212428b0e92d0467c18629cf9ed67'
 
       - name: Install Sage
         run: |

.github/workflows/shortint_benchmark.yml

@@ -1,128 +0,0 @@
-# Run shortint benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Shortint benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  run-shortint-benchmarks:
-    name: Execute shortint benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Run benchmarks with AVX512
-        run: |
-          make bench_shortint
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Measure key sizes
-        run: |
-          make measure_shortint_key_sizes
-
-      - name: Parse key sizes results
-        run: |
-          python3 ./ci/benchmark_parser.py tfhe/shortint_key_sizes.csv ${{ env.RESULTS_FILENAME }} \
-          --key-sizes \
-          --append-results
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_shortint
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-      - name: Slack Notification
-        if: ${{ failure() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Shortint benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/shortint_cpu_benchmark.yml

@@ -0,0 +1,193 @@
+# Run all shortint benchmarks on an AWS instance and return parsed results to Slab CI bot.
+name: Shortint full benchmarks
+
+on:
+  workflow_dispatch:
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
+    # Quarterly benchmarks will be triggered right before end of quarter, the 25th of the current month at 4a.m.
+    # These benchmarks are far longer to execute hence the reason to run them only four time a year.
+    - cron: '0 4 25 MAR,JUN,SEP,DEC *'
+
+
+env:
+  CARGO_TERM_COLOR: always
+  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+jobs:
+  prepare-matrix:
+    name: Prepare operations matrix
+    runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      op_flavor: ${{ steps.set_op_flavor.outputs.op_flavor }}
+    steps:
+      - name: Weekly benchmarks
+        if: github.event_name == 'workflow_dispatch' ||
+          github.event.schedule == '0 1 * * 6'
+        run: |
+          echo "OP_FLAVOR=[\"default\"]" >> "${GITHUB_ENV}"
+
+      - name: Quarterly benchmarks
+        if: github.event.schedule == '0 4 25 MAR,JUN,SEP,DEC *'
+        run: |
+          echo "OP_FLAVOR=[\"default\", \"smart\", \"unchecked\"]" >> "${GITHUB_ENV}"
+
+      - name: Set operation flavor output
+        id: set_op_flavor
+        run: |
+          echo "op_flavor=${{ toJSON(env.OP_FLAVOR) }}" >> "${GITHUB_OUTPUT}"
+
+  setup-instance:
+    name: Setup instance (shortint-benchmarks)
+    needs: prepare-matrix
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
+
+  shortint-benchmarks:
+    name: Execute shortint benchmarks for all operations flavor
+    needs: [ prepare-matrix, setup-instance ]
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    continue-on-error: true
+    strategy:
+      max-parallel: 1
+      matrix:
+        op_flavor: ${{ fromJson(needs.prepare-matrix.outputs.op_flavor) }}
+    steps:
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
+      - name: Set up home
+        # "Install rust" step require root user to have a HOME directory which is not set.
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install rust
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: nightly
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Run benchmarks with AVX512
+        run: |
+          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_shortint
+
+      - name: Parse results
+        run: |
+          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
+          COMMIT_HASH="$(git describe --tags --dirty)"
+          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
+          --database tfhe_rs \
+          --hardware "hpc7a.96xlarge" \
+          --project-version "${COMMIT_HASH}" \
+          --branch ${{ github.ref_name }} \
+          --commit-date "${COMMIT_DATE}" \
+          --bench-date "${{ env.BENCH_DATE }}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --throughput
+
+      # This small benchmark needs to be executed only once.
+      - name: Measure key sizes
+        if: matrix.op_flavor == 'default'
+        run: |
+          make measure_shortint_key_sizes
+
+      - name: Parse key sizes results
+        if: matrix.op_flavor == 'default'
+        run: |
+          python3 ./ci/benchmark_parser.py tfhe/shortint_key_sizes.csv ${{ env.RESULTS_FILENAME }} \
+          --key-sizes \
+          --append-results
+
+      - name: Upload parsed results artifact
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
+        with:
+          name: ${{ github.sha }}_shortint_${{ matrix.op_flavor }}
+          path: ${{ env.RESULTS_FILENAME }}
+
+      - name: Send data to Slab
+        shell: bash
+        run: |
+          echo "Computing HMac on results file"
+          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
+          echo "Sending results to Slab..."
+          curl -v -k \
+          -H "Content-Type: application/json" \
+          -H "X-Slab-Repository: ${{ github.repository }}" \
+          -H "X-Slab-Command: store_data_v2" \
+          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
+          -d @${{ env.RESULTS_FILENAME }} \
+          ${{ secrets.SLAB_URL }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Shortint full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (shortint-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, shortint-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (shortint-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/shortint_full_benchmark.yml

@@ -1,152 +0,0 @@
-# Run all shortint benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Shortint full benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      # This input is not used in this workflow but still mandatory since a calling workflow could
-      # use it. If a triggering command include a user_inputs field, then the triggered workflow
-      # must include this very input, otherwise the workflow won't be called.
-      # See start_full_benchmarks.yml as example.
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  shortint-benchmarks:
-    name: Execute shortint benchmarks for all operations flavor
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    strategy:
-      max-parallel: 1
-      matrix:
-        op_flavor: [ default, smart, unchecked ]
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Get benchmark details
-        run: |
-          {
-            echo "BENCH_DATE=$(date --iso-8601=seconds)";
-            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
-            echo "COMMIT_HASH=$(git describe --tags --dirty)";
-          } >> "${GITHUB_ENV}"
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Run benchmarks with AVX512
-        run: |
-          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_shortint
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      # This small benchmark needs to be executed only once.
-      - name: Measure key sizes
-        if: matrix.op_flavor == 'default'
-        run: |
-          make measure_shortint_key_sizes
-
-      - name: Parse key sizes results
-        if: matrix.op_flavor == 'default'
-        run: |
-          python3 ./ci/benchmark_parser.py tfhe/shortint_key_sizes.csv ${{ env.RESULTS_FILENAME }} \
-          --key-sizes \
-          --append-results
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_shortint_${{ matrix.op_flavor }}
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-  slack-notification:
-    name: Slack Notification
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ failure() }}
-    needs: shortint-benchmarks
-    steps:
-      - name: Notify
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Shortint full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/signed_integer_benchmark.yml

@@ -1,130 +0,0 @@
-# Run signed integer benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Signed Integer benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  PARSE_INTEGER_BENCH_CSV_FILE: tfhe_rs_integer_benches_${{ github.sha }}.csv
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  run-integer-benchmarks:
-    name: Execute signed integer benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Run benchmarks with AVX512
-        run: |
-          make FAST_BENCH=TRUE bench_signed_integer
-
-      - name: Parse benchmarks to csv
-        run: |
-          make PARSE_INTEGER_BENCH_CSV_FILE=${{ env.PARSE_INTEGER_BENCH_CSV_FILE }} \
-            parse_integer_benches
-
-      - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_csv_integer
-          path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_integer
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-      - name: Slack Notification
-        if: ${{ failure() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Signed integer benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/signed_integer_cpu_benchmark.yml

@@ -0,0 +1,191 @@
+# Run all signed integer benchmarks on an AWS instance and return parsed results to Slab CI bot.
+name: Signed Integer full benchmarks
+
+on:
+  workflow_dispatch:
+    inputs:
+      all_precisions:
+        description: "Run all precisions"
+        type: boolean
+        default: false
+
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
+    # Quarterly benchmarks will be triggered right before end of quarter, the 25th of the current month at 4a.m.
+    # These benchmarks are far longer to execute hence the reason to run them only four time a year.
+    - cron: '0 4 25 MAR,JUN,SEP,DEC *'
+
+env:
+  CARGO_TERM_COLOR: always
+  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  FAST_BENCH: TRUE
+
+jobs:
+  prepare-matrix:
+    name: Prepare operations matrix
+    runs-on: ubuntu-latest
+    if: github.event_name != 'schedule' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      op_flavor: ${{ steps.set_op_flavor.outputs.op_flavor }}
+    steps:
+      - name: Weekly benchmarks
+        if: github.event_name == 'workflow_dispatch' ||
+          github.event.schedule == '0 1 * * 6'
+        run: |
+          echo "OP_FLAVOR=[\"default\"]" >> "${GITHUB_ENV}"
+
+      - name: Quarterly benchmarks
+        if: github.event.schedule == '0 4 25 MAR,JUN,SEP,DEC *'
+        run: |
+          echo "OP_FLAVOR=[\"default\", \"unchecked\"]" >> "${GITHUB_ENV}"
+
+      - name: Set operation flavor output
+        id: set_op_flavor
+        run: |
+          echo "op_flavor=${{ toJSON(env.OP_FLAVOR) }}" >> "${GITHUB_OUTPUT}"
+
+  setup-instance:
+    name: Setup instance (signed-integer-benchmarks)
+    needs: prepare-matrix
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
+
+  signed-integer-benchmarks:
+    name: Execute signed integer benchmarks for all operations flavor
+    needs: [ prepare-matrix, setup-instance ]
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    concurrency:
+      group: ${{ github.workflow }}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    continue-on-error: true
+    timeout-minutes: 1440  # 24 hours
+    strategy:
+      max-parallel: 1
+      matrix:
+        command: [ integer, integer_multi_bit ]
+        op_flavor: [ default, unchecked ]
+    steps:
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
+      - name: Set up home
+        # "Install rust" step require root user to have a HOME directory which is not set.
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install rust
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: nightly
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Should run benchmarks with all precisions
+        if: inputs.all_precisions
+        run: |
+          echo "FAST_BENCH=FALSE" >> "${GITHUB_ENV}"
+
+      - name: Run benchmarks with AVX512
+        run: |
+          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_signed_${{ matrix.command }}
+
+      - name: Parse results
+        run: |
+          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
+          --database tfhe_rs \
+          --hardware "hpc7a.96xlarge" \
+          --project-version "${{ env.COMMIT_HASH }}" \
+          --branch ${{ github.ref_name }} \
+          --commit-date "${{ env.COMMIT_DATE }}" \
+          --bench-date "${{ env.BENCH_DATE }}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --throughput
+
+      - name: Upload parsed results artifact
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
+        with:
+          name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
+          path: ${{ env.RESULTS_FILENAME }}
+
+      - name: Send data to Slab
+        shell: bash
+        run: |
+          echo "Computing HMac on results file"
+          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
+          echo "Sending results to Slab..."
+          curl -v -k \
+          -H "Content-Type: application/json" \
+          -H "X-Slab-Repository: ${{ github.repository }}" \
+          -H "X-Slab-Command: store_data_v2" \
+          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
+          -d @${{ env.RESULTS_FILENAME }} \
+          ${{ secrets.SLAB_URL }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Signed integer full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (integer-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, signed-integer-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (signed-integer-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/signed_integer_full_benchmark.yml

@@ -1,136 +0,0 @@
-# Run all signed integer benchmarks on an AWS instance and return parsed results to Slab CI bot.
-name: Signed Integer full benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  integer-benchmarks:
-    name: Execute signed integer benchmarks for all operations flavor
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    continue-on-error: true
-    timeout-minutes: 1440  # 24 hours
-    strategy:
-      max-parallel: 1
-      matrix:
-        command: [ integer, integer_multi_bit ]
-        op_flavor: [ default, unchecked ]
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Get benchmark details
-        run: |
-          {
-            echo "BENCH_DATE=$(date --iso-8601=seconds)";
-            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
-            echo "COMMIT_HASH=$(git describe --tags --dirty)";
-          } >> "${GITHUB_ENV}"
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Run benchmarks with AVX512
-        run: |
-          make BENCH_OP_FLAVOR=${{ matrix.op_flavor }} bench_signed_${{ matrix.command }}
-
-      - name: Parse results
-        run: |
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${{ env.COMMIT_HASH }}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${{ env.COMMIT_DATE }}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-  slack-notification:
-    name: Slack Notification
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ failure() }}
-    needs: integer-benchmarks
-    steps:
-      - name: Notify
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Signed integer full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/signed_integer_multi_bit_benchmark.yml

@@ -1,130 +0,0 @@
-# Run signed integer benchmarks with multi-bit cryptographic parameters on an AWS instance and return parsed results to Slab CI bot.
-name: Signed Integer Multi-bit benchmarks
-
-on:
-  workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-
-env:
-  CARGO_TERM_COLOR: always
-  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
-  PARSE_INTEGER_BENCH_CSV_FILE: tfhe_rs_integer_benches_${{ github.sha }}.csv
-  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-  RUST_BACKTRACE: "full"
-  RUST_MIN_STACK: "8388608"
-
-jobs:
-  run-integer-benchmarks:
-    name: Execute signed integer multi-bit benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
-    steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Set up home
-        # "Install rust" step require root user to have a HOME directory which is not set.
-        run: |
-          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
-
-      - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
-        with:
-          toolchain: nightly
-
-      - name: Run multi-bit benchmarks with AVX512
-        run: |
-          make FAST_BENCH=TRUE bench_signed_integer_multi_bit
-
-      - name: Parse benchmarks to csv
-        run: |
-          make PARSE_INTEGER_BENCH_CSV_FILE=${{ env.PARSE_INTEGER_BENCH_CSV_FILE }} \
-            parse_integer_benches
-
-      - name: Upload csv results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_csv_integer
-          path: ${{ env.PARSE_INTEGER_BENCH_CSV_FILE }}
-
-      - name: Parse results
-        run: |
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
-          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
-          --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
-          --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
-          --bench-date "${{ env.BENCH_DATE }}" \
-          --walk-subdirs \
-          --name-suffix avx512 \
-          --throughput
-
-      - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
-        with:
-          name: ${{ github.sha }}_integer
-          path: ${{ env.RESULTS_FILENAME }}
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Send data to Slab
-        shell: bash
-        run: |
-          echo "Computing HMac on results file"
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
-          echo "Sending results to Slab..."
-          curl -v -k \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: store_data_v2" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @${{ env.RESULTS_FILENAME }} \
-          ${{ secrets.SLAB_URL }}
-
-      - name: Slack Notification
-        if: ${{ failure() }}
-        continue-on-error: true
-        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
-        env:
-          SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "Signed integer benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/start_benchmarks.yml

@@ -1,130 +0,0 @@
-# Start all benchmark jobs on Slab CI bot.
-name: Start all benchmarks
-
-on:
-  push:
-    branches:
-      - "main"
-  workflow_dispatch:
-    inputs:
-      # The input name must be the name of the slab command to launch
-      boolean_bench:
-        description: "Run Boolean benches"
-        type: boolean
-        default: true
-      shortint_bench:
-        description: "Run shortint benches"
-        type: boolean
-        default: true
-      integer_bench:
-        description: "Run integer benches"
-        type: boolean
-        default: true
-      signed_integer_bench:
-        description: "Run signed integer benches"
-        type: boolean
-        default: true
-      integer_multi_bit_bench:
-        description: "Run integer multi bit benches"
-        type: boolean
-        default: true
-      signed_integer_multi_bit_bench:
-        description: "Run signed integer multi bit benches"
-        type: boolean
-        default: true
-      core_crypto_bench:
-        description: "Run core crypto benches"
-        type: boolean
-        default: true
-      wasm_client_bench:
-        description: "Run WASM client benches"
-        type: boolean
-        default: true
-
-jobs:
-  start-benchmarks:
-    if: ${{ (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
-    strategy:
-      matrix:
-        command: [ boolean_bench, shortint_bench,
-                   integer_bench, integer_multi_bit_bench,
-                   signed_integer_bench, signed_integer_multi_bit_bench,
-                   core_crypto_bench, wasm_client_bench ]
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Check for file changes
-        id: changed-files
-        uses: tj-actions/changed-files@03334d095e2739fa9ac4034ec16f66d5d01e9eba
-        with:
-          files_yaml: |
-            common_benches:
-              - toolchain.txt
-              - Makefile
-              - ci/slab.toml
-              - tfhe/Cargo.toml
-              - tfhe/src/core_crypto/**
-              - .github/workflows/start_benchmarks.yml
-            boolean_bench:
-              - tfhe/src/boolean/**
-              - tfhe/benches/boolean/**
-              - .github/workflows/boolean_benchmark.yml
-            shortint_bench:
-              - tfhe/src/shortint/**
-              - tfhe/benches/shortint/**
-              - .github/workflows/shortint_benchmark.yml
-            integer_bench:
-              - tfhe/src/shortint/**
-              - tfhe/src/integer/**
-              - tfhe/benches/integer/bench.rs
-              - .github/workflows/integer_benchmark.yml
-            integer_multi_bit_bench:
-              - tfhe/src/shortint/**
-              - tfhe/src/integer/**
-              - tfhe/benches/integer/bench.rs
-              - .github/workflows/integer_multi_bit_benchmark.yml
-            signed_integer_bench:
-              - tfhe/src/shortint/**
-              - tfhe/src/integer/**
-              - tfhe/benches/integer/signed_bench.rs
-              - .github/workflows/signed_integer_benchmark.yml
-            signed_integer_multi_bit_bench:
-              - tfhe/src/shortint/**
-              - tfhe/src/integer/**
-              - tfhe/benches/integer/signed_bench.rs
-              - .github/workflows/signed_integer_multi_bit_benchmark.yml
-            core_crypto_bench:
-              - tfhe/src/core_crypto/**
-              - tfhe/benches/core_crypto/**
-              - .github/workflows/core_crypto_benchmark.yml
-            wasm_client_bench:
-              - tfhe/web_wasm_parallel_tests/**
-              - .github/workflows/wasm_client_benchmark.yml
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Start AWS job in Slab
-        # If manually triggered check that the current bench has been requested
-        # Otherwise if it's on push check that files relevant to benchmarks have changed
-        if: (github.event_name == 'workflow_dispatch' && github.event.inputs[matrix.command] == 'true') || (github.event_name == 'push' && (steps.changed-files.outputs.common_benches_any_changed == 'true' || steps.changed-files.outputs[format('{0}_any_changed', matrix.command)] == 'true'))
-        shell: bash
-        run: |
-          echo -n '{"command": "${{ matrix.command }}", "git_ref": "${{ github.ref }}", "sha": "${{ github.sha }}"}' > command.json
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh command.json '${{ secrets.JOB_SECRET }}')"
-          curl -v -k \
-          --fail-with-body \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: start_aws" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @command.json \
-          ${{ secrets.SLAB_URL }}

.github/workflows/start_full_benchmarks.yml

@@ -1,66 +0,0 @@
-# Start all benchmark jobs, including full shortint and integer, on Slab CI bot.
-name: Start full suite benchmarks
-
-on:
-  schedule:
-    # Weekly benchmarks will be triggered each Saturday at 1a.m.
-    - cron: '0 1 * * 6'
-    # Quarterly benchmarks will be triggered right before end of quarter, the 25th of the current month at 4a.m.
-    # These benchmarks are far longer to execute hence the reason to run them only four time a year.
-    - cron: '0 4 25 MAR,JUN,SEP,DEC *'
-  workflow_dispatch:
-    inputs:
-      benchmark_type:
-        description: 'Benchmark type'
-        required: true
-        default: 'weekly'
-        type: choice
-        options:
-          - weekly
-          - quarterly
-
-jobs:
-  start-benchmarks:
-    if: ${{ (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') || github.event_name == 'workflow_dispatch' }}
-    strategy:
-      matrix:
-        command: [ boolean_bench, shortint_full_bench,
-                   integer_full_bench, signed_integer_full_bench,
-                   core_crypto_bench, wasm_client_bench ]
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout tfhe-rs
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          fetch-depth: 0
-
-      - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
-        with:
-          repository: zama-ai/slab
-          path: slab
-          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
-
-      - name: Set benchmarks type as weekly
-        if: (github.event_name == 'workflow_dispatch' && inputs.benchmark_type == 'weekly') || github.event.schedule == '0 1 * * 6'
-        run: |
-          echo "BENCH_TYPE=weekly_benchmarks" >> "${GITHUB_ENV}"
-
-      - name: Set benchmarks type as quarterly
-        if: (github.event_name == 'workflow_dispatch' && inputs.benchmark_type == 'quarterly') || github.event.schedule == '0 4 25 MAR,JUN,SEP,DEC *'
-        run: |
-          echo "BENCH_TYPE=quarterly_benchmarks" >> "${GITHUB_ENV}"
-
-      - name: Start AWS job in Slab
-        shell: bash
-        run: |
-          echo -n '{"command": "${{ matrix.command }}", "git_ref": "${{ github.ref }}", "sha": "${{ github.sha }}", "user_inputs": "${{ env.BENCH_TYPE }}"}' > command.json
-          SIGNATURE="$(slab/scripts/hmac_calculator.sh command.json '${{ secrets.JOB_SECRET }}')"
-          curl -v -k \
-          --fail-with-body \
-          -H "Content-Type: application/json" \
-          -H "X-Slab-Repository: ${{ github.repository }}" \
-          -H "X-Slab-Command: start_aws" \
-          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
-          -d @command.json \
-          ${{ secrets.SLAB_URL }}

.github/workflows/sync_on_push.yml

@@ -13,7 +13,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
       - name: git-sync

.github/workflows/wasm_client_benchmark.yml

@@ -1,32 +1,14 @@
-# Run WASM client benchmarks on an AWS instance and return parsed results to Slab CI bot.
+# Run WASM client benchmarks on an instance and return parsed results to Slab CI bot.
 name: WASM client benchmarks
 
 on:
   workflow_dispatch:
-    inputs:
-      instance_id:
-        description: "Instance ID"
-        type: string
-      instance_image_id:
-        description: "Instance AMI ID"
-        type: string
-      instance_type:
-        description: "Instance product type"
-        type: string
-      runner_name:
-        description: "Action runner name"
-        type: string
-      request_id:
-        description: "Slab request ID"
-        type: string
-      # This input is not used in this workflow but still mandatory since a calling workflow could
-      # use it. If a triggering command include a user_inputs field, then the triggered workflow
-      # must include this very input, otherwise the workflow won't be called.
-      # See start_full_benchmarks.yml as example.
-      user_inputs:
-        description: "Type of benchmarks to run"
-        type: string
-        default: "weekly_benchmarks"
+  push:
+    branches:
+      - main
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 1a.m.
+    - cron: '0 1 * * 6'
 
 env:
   CARGO_TERM_COLOR: always
@@ -34,56 +16,106 @@ env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
   RUST_BACKTRACE: "full"
   RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
 
 jobs:
-  run-wasm-client-benchmarks:
-    name: Execute WASM client benchmarks in EC2
-    runs-on: ${{ github.event.inputs.runner_name }}
-    if: ${{ !cancelled() }}
+  should-run:
+    runs-on: ubuntu-latest
+    if: github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs')
+    permissions:
+      pull-requests: write
+    outputs:
+      wasm_bench: ${{ steps.changed-files.outputs.wasm_bench_any_changed }}
     steps:
-      - name: Instance configuration used
-        run: |
-          echo "IDs: ${{ inputs.instance_id }}"
-          echo "AMI: ${{ inputs.instance_image_id }}"
-          echo "Type: ${{ inputs.instance_type }}"
-          echo "Request ID: ${{ inputs.request_id }}"
-
-      - name: Get benchmark date
-        run: |
-          echo "BENCH_DATE=$(date --iso-8601=seconds)" >> "${GITHUB_ENV}"
-
-      - name: Checkout tfhe-rs repo with tags
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           fetch-depth: 0
 
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            wasm_bench:
+              - tfhe/Cargo.toml
+              - concrete-csprng/**
+              - tfhe-zk-pok/**
+              - tfhe/src/**
+              - '!tfhe/src/c_api/**'
+              - tfhe/web_wasm_parallel_tests/**
+              - .github/workflows/wasm_client_benchmark.yml
+
+  setup-instance:
+    name: Setup instance (wasm-client-benchmarks)
+    if: github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      (github.event_name == 'push' && github.repository == 'zama-ai/tfhe-rs' && needs.should-run.outputs.wasm_bench)
+    needs: should-run
+    runs-on: ubuntu-latest
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: cpu-small
+
+  wasm-client-benchmarks:
+    name: Execute WASM client benchmarks
+    needs: setup-instance
+    if: needs.setup-instance.result != 'skipped'
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    steps:
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
       - name: Set up home
         # "Install rust" step require root user to have a HOME directory which is not set.
         run: |
           echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@d8352f6b1d2e870bc5716e7a6d9b65c4cc244a1a
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
         with:
           toolchain: nightly
 
       - name: Run benchmarks
         run: |
           make install_node
-          make ci_bench_web_js_api_parallel
+          make bench_web_js_api_parallel_ci
 
       - name: Parse results
         run: |
           make parse_wasm_benchmarks
-
-          COMMIT_DATE="$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})"
-          COMMIT_HASH="$(git describe --tags --dirty)"
           python3 ./ci/benchmark_parser.py tfhe/wasm_pk_gen.csv ${{ env.RESULTS_FILENAME }} \
           --database tfhe_rs \
-          --hardware ${{ inputs.instance_type }} \
-          --project-version "${COMMIT_HASH}" \
+          --hardware "m6i.4xlarge" \
+          --project-version "${{ env.COMMIT_HASH }}" \
           --branch ${{ github.ref_name }} \
-          --commit-date "${COMMIT_DATE}" \
+          --commit-date "${{ env.COMMIT_DATE }}" \
           --bench-date "${{ env.BENCH_DATE }}" \
           --key-gen
 
@@ -98,13 +130,13 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65462800fd760344b1a7b4382951275a0abb4808
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
         with:
           name: ${{ github.sha }}_wasm
           path: ${{ env.RESULTS_FILENAME }}
 
       - name: Checkout Slab repo
-        uses: actions/checkout@a5ac7e51b41094c92402da3b24376905380afc29
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
         with:
           repository: zama-ai/slab
           path: slab
@@ -130,8 +162,28 @@ jobs:
         uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "WASM benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+  teardown-instance:
+    name: Teardown instance (wasm-client-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, wasm-client-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (wasm-client-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/zk_pke_benchmark.yml

@@ -0,0 +1,197 @@
+# Run PKE Zero-Knowledge benchmarks on an instance and return parsed results to Slab CI bot.
+name: PKE ZK benchmarks
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - main
+  schedule:
+    # Weekly benchmarks will be triggered each Saturday at 3a.m.
+    - cron: '0 3 * * 6'
+env:
+  CARGO_TERM_COLOR: always
+  RESULTS_FILENAME: parsed_benchmark_results_${{ github.sha }}.json
+  PARSE_INTEGER_BENCH_CSV_FILE: tfhe_rs_integer_benches_${{ github.sha }}.csv
+  ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  RUST_BACKTRACE: "full"
+  RUST_MIN_STACK: "8388608"
+  SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
+  SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
+  SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
+  SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+
+jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    if: github.event_name == 'workflow_dispatch' ||
+      ((github.event_name == 'push' || github.event_name == 'schedule') && github.repository == 'zama-ai/tfhe-rs')
+    outputs:
+      zk_pok_changed: ${{ steps.changed-files.outputs.zk_pok_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@c65cd883420fd2eb864698a825fc4162dd94482c
+        with:
+          since_last_remote_commit: true
+          files_yaml: |
+            zk_pok:
+              - tfhe/Cargo.toml
+              - concrete-csprng/**
+              - tfhe-zk-pok/**
+              - tfhe/src/core_crypto/**
+              - tfhe/src/shortint/**
+              - tfhe/src/integer/**
+              - tfhe/src/zk.rs
+              - tfhe/benches/integer/zk_pke.rs
+              - .github/workflows/zk_pke_benchmark.yml
+
+  setup-instance:
+    name: Setup instance (pke-zk-benchmarks)
+    runs-on: ubuntu-latest
+    needs: should-run
+    if: github.event_name == 'workflow_dispatch' ||
+      (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs') ||
+      (github.event_name == 'push' &&
+      github.repository == 'zama-ai/tfhe-rs' &&
+      needs.should-run.outputs.zk_pok_changed == 'true')
+    outputs:
+      runner-name: ${{ steps.start-instance.outputs.label }}
+    steps:
+      - name: Start instance
+        id: start-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: start
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          backend: aws
+          profile: bench
+
+  pke-zk-benchmarks:
+    name: Execute PKE ZK benchmarks
+    if: needs.setup-instance.result != 'skipped'
+    needs: setup-instance
+    concurrency:
+      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.ref }}
+      cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    steps:
+      - name: Checkout tfhe-rs repo with tags
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          fetch-depth: 0
+
+      - name: Get benchmark details
+        run: |
+          {
+            echo "BENCH_DATE=$(date --iso-8601=seconds)";
+            echo "COMMIT_DATE=$(git --no-pager show -s --format=%cd --date=iso8601-strict ${{ github.sha }})";
+            echo "COMMIT_HASH=$(git describe --tags --dirty)";
+          } >> "${GITHUB_ENV}"
+
+      - name: Set up home
+        # "Install rust" step require root user to have a HOME directory which is not set.
+        run: |
+          echo "HOME=/home/ubuntu" >> "${GITHUB_ENV}"
+
+      - name: Install rust
+        uses: dtolnay/rust-toolchain@7b1c307e0dcbda6122208f10795a713336a9b35a
+        with:
+          toolchain: nightly
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Run benchmarks with AVX512
+        run: |
+          make bench_integer_zk
+
+      - name: Parse results
+        run: |
+          python3 ./ci/benchmark_parser.py target/criterion ${{ env.RESULTS_FILENAME }} \
+          --database tfhe_rs \
+          --hardware "hpc7a.96xlarge" \
+          --backend cpu \
+          --project-version "${{ env.COMMIT_HASH }}" \
+          --branch ${{ github.ref_name }} \
+          --commit-date "${{ env.COMMIT_DATE }}" \
+          --bench-date "${{ env.BENCH_DATE }}" \
+          --walk-subdirs \
+          --name-suffix avx512 \
+          --throughput
+
+      - name: Parse CRS sizes results
+        run: |
+          python3 ./ci/benchmark_parser.py tfhe/pke_zk_crs_sizes.csv ${{ env.RESULTS_FILENAME }} \
+          --key-sizes \
+          --append-results
+
+      - name: Upload parsed results artifact
+        uses: actions/upload-artifact@834a144ee995460fba8ed112a2fc961b36a5ec5a
+        with:
+          name: ${{ github.sha }}_integer_zk
+          path: ${{ env.RESULTS_FILENAME }}
+
+      - name: Checkout Slab repo
+        uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332
+        with:
+          repository: zama-ai/slab
+          path: slab
+          token: ${{ secrets.FHE_ACTIONS_TOKEN }}
+
+      - name: Send data to Slab
+        shell: bash
+        run: |
+          echo "Computing HMac on results file"
+          SIGNATURE="$(slab/scripts/hmac_calculator.sh ${{ env.RESULTS_FILENAME }} '${{ secrets.JOB_SECRET }}')"
+          echo "Sending results to Slab..."
+          curl -v -k \
+          -H "Content-Type: application/json" \
+          -H "X-Slab-Repository: ${{ github.repository }}" \
+          -H "X-Slab-Command: store_data_v2" \
+          -H "X-Hub-Signature-256: sha256=${SIGNATURE}" \
+          -d @${{ env.RESULTS_FILENAME }} \
+          ${{ secrets.SLAB_URL }}
+
+      - name: Slack Notification
+        if: ${{ !success() && !cancelled() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "PKE ZK benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
+
+  teardown-instance:
+    name: Teardown instance (pke-zk-benchmarks)
+    if: ${{ always() && needs.setup-instance.result != 'skipped' }}
+    needs: [ setup-instance, pke-zk-benchmarks ]
+    runs-on: ubuntu-latest
+    steps:
+      - name: Stop instance
+        id: stop-instance
+        uses: zama-ai/slab-github-runner@447a2d0fd2d1a9d647aa0d0723a6e9255372f261
+        with:
+          mode: stop
+          github-token: ${{ secrets.SLAB_ACTION_TOKEN }}
+          slab-url: ${{ secrets.SLAB_BASE_URL }}
+          job-secret: ${{ secrets.JOB_SECRET }}
+          label: ${{ needs.setup-instance.outputs.runner-name }}
+
+      - name: Slack Notification
+        if: ${{ failure() }}
+        continue-on-error: true
+        uses: rtCamp/action-slack-notify@4e5fb42d249be6a45a298f3c9543b111b02f7907
+        env:
+          SLACK_COLOR: ${{ job.status }}
+          SLACK_MESSAGE: "Instance teardown (pke-zk-benchmarks) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.gitignore

@@ -7,6 +7,7 @@ target/
 # In case of symlinked keys
 /keys
 
+**/*.rmeta
 **/Cargo.lock
 **/*.bin
 
@@ -22,3 +23,9 @@ dieharder_run.log
 
 # Cuda local build
 backends/tfhe-cuda-backend/cuda/cmake-build-debug/
+
+# WASM tests
+tfhe/web_wasm_parallel_tests/server.PID
+
+# Dir used for backward compatibility test data
+tfhe/tfhe-backward-compat-data/

Cargo.toml

@@ -7,6 +7,14 @@ members = [
     "apps/trivium",
     "concrete-csprng",
     "backends/tfhe-cuda-backend",
+    "utils/tfhe-versionable",
+    "utils/tfhe-versionable-derive",
+]
+
+exclude = [
+    "tfhe/backward_compatibility_tests",
+    "utils/cargo-tfhe-lints-inner",
+    "utils/cargo-tfhe-lints"
 ]
 
 [profile.bench]

Makefile

@@ -16,19 +16,15 @@ GEN_KEY_CACHE_COVERAGE_ONLY?=FALSE
 PARSE_INTEGER_BENCH_CSV_FILE?=tfhe_rs_integer_benches.csv
 FAST_TESTS?=FALSE
 FAST_BENCH?=FALSE
+NIGHTLY_TESTS?=FALSE
 BENCH_OP_FLAVOR?=DEFAULT
-NODE_VERSION=20
+NODE_VERSION=22.4
 FORWARD_COMPAT?=OFF
-# sed: -n, do not print input stream, -e means a script/expression
-# 1,/version/ indicates from the first line, to the line matching version at the start of the line
-# p indicates to print, so we keep only the start of the Cargo.toml until we hit the first version
-# entry which should be the version of tfhe
-TFHE_CURRENT_VERSION:=\
-$(shell sed -n -e '1,/^version/p' tfhe/Cargo.toml | \
-grep '^version[[:space:]]*=' | cut -d '=' -f 2 | xargs)
-# Cargo has a hard time distinguishing between our package from the workspace and a package that
-# could be a dependency, so we build an unambiguous spec here
-TFHE_SPEC:=tfhe@$(TFHE_CURRENT_VERSION)
+BACKWARD_COMPAT_DATA_URL=https://github.com/zama-ai/tfhe-backward-compat-data.git
+BACKWARD_COMPAT_DATA_BRANCH?=v0.1
+BACKWARD_COMPAT_DATA_PROJECT=tfhe-backward-compat-data
+BACKWARD_COMPAT_DATA_DIR=$(BACKWARD_COMPAT_DATA_PROJECT)
+TFHE_SPEC:=tfhe
 # This is done to avoid forgetting it, we still precise the RUSTFLAGS in the commands to be able to
 # copy paste the command in the terminal and change them if required without forgetting the flags
 export RUSTFLAGS?=-C target-cpu=native
@@ -115,7 +111,7 @@ install_cargo_nextest: install_rs_build_toolchain
 .PHONY: install_wasm_pack # Install wasm-pack to build JS packages
 install_wasm_pack: install_rs_build_toolchain
 	@wasm-pack --version > /dev/null 2>&1 || \
-	cargo $(CARGO_RS_BUILD_TOOLCHAIN) install wasm-pack || \
+	cargo $(CARGO_RS_BUILD_TOOLCHAIN) install --locked wasm-pack@0.13.0 || \
 	( echo "Unable to install cargo wasm-pack, unknown error." && exit 1 )
 
 .PHONY: install_node # Install last version of NodeJS via nvm
@@ -145,6 +141,11 @@ install_tarpaulin: install_rs_build_toolchain
 	cargo $(CARGO_RS_BUILD_TOOLCHAIN) install cargo-tarpaulin --locked || \
 	( echo "Unable to install cargo tarpaulin, unknown error." && exit 1 )
 
+.PHONY: install_tfhe_lints # Install custom tfhe-rs lints
+install_tfhe_lints:
+	(cd utils/cargo-tfhe-lints-inner && cargo install --path .) && \
+	cd utils/cargo-tfhe-lints && cargo install --path .
+
 .PHONY: check_linelint_installed # Check if linelint newline linter is installed
 check_linelint_installed:
 	@printf "\n" | linelint - > /dev/null 2>&1 || \
@@ -264,6 +265,17 @@ clippy: install_rs_check_toolchain
 		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
+.PHONY: clippy_rustdoc # Run clippy lints on doctests enabling the boolean, shortint, integer and zk-pok
+clippy_rustdoc: install_rs_check_toolchain
+	if [[ "$(OS)" != "Linux" && "$(OS)" != "Darwin" ]]; then \
+		echo "WARNING: skipped clippy_rustdoc, unsupported OS $(OS)"; \
+		exit 0; \
+	fi && \
+	CLIPPYFLAGS="-D warnings" RUSTDOCFLAGS="--no-run --nocapture --test-builder ./scripts/clippy_driver.sh -Z unstable-options" \
+		cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" test --doc \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,zk-pok,pbs-stats \
+		-p $(TFHE_SPEC)
+
 .PHONY: clippy_c_api # Run clippy lints enabling the boolean, shortint and the C API
 clippy_c_api: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
@@ -273,7 +285,7 @@ clippy_c_api: install_rs_check_toolchain
 .PHONY: clippy_js_wasm_api # Run clippy lints enabling the boolean, shortint, integer and the js wasm API
 clippy_js_wasm_api: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api \
+		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: clippy_tasks # Run clippy lints on helper tasks crate.
@@ -289,7 +301,7 @@ clippy_trivium: install_rs_check_toolchain
 .PHONY: clippy_all_targets # Run clippy lints on all targets (benches, examples, etc.)
 clippy_all_targets: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,zk-pok-experimental \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,zk-pok \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: clippy_concrete_csprng # Run clippy lints on concrete-csprng
@@ -304,18 +316,23 @@ clippy_zk_pok: install_rs_check_toolchain
 		-p tfhe-zk-pok -- --no-deps -D warnings
 
 .PHONY: clippy_all # Run all clippy targets
-clippy_all: clippy clippy_boolean clippy_shortint clippy_integer clippy_all_targets clippy_c_api \
-clippy_js_wasm_api clippy_tasks clippy_core clippy_concrete_csprng clippy_zk_pok clippy_trivium
+clippy_all: clippy_rustdoc clippy clippy_boolean clippy_shortint clippy_integer clippy_all_targets \
+clippy_c_api clippy_js_wasm_api clippy_tasks clippy_core clippy_concrete_csprng clippy_zk_pok clippy_trivium
 
 .PHONY: clippy_fast # Run main clippy targets
-clippy_fast: clippy clippy_all_targets clippy_c_api clippy_js_wasm_api clippy_tasks clippy_core \
-clippy_concrete_csprng
+clippy_fast: clippy_rustdoc clippy clippy_all_targets clippy_c_api clippy_js_wasm_api clippy_tasks \
+clippy_core clippy_concrete_csprng
 
 .PHONY: clippy_cuda_backend # Run clippy lints on the tfhe-cuda-backend
 clippy_cuda_backend: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
 		-p tfhe-cuda-backend -- --no-deps -D warnings
 
+.PHONY: tfhe_lints # Run custom tfhe-rs lints
+tfhe_lints: install_tfhe_lints
+	cd tfhe && RUSTFLAGS="$(RUSTFLAGS)" cargo tfhe-lints \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer -- -D warnings
+
 .PHONY: build_core # Build core_crypto without experimental features
 build_core: install_rs_build_toolchain install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) build --profile $(CARGO_PROFILE) \
@@ -368,21 +385,21 @@ symlink_c_libs_without_fingerprint:
 .PHONY: build_c_api # Build the C API for boolean, shortint and integer
 build_c_api: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) build --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok-experimental,$(FORWARD_COMPAT_FEATURE) \
+		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,$(FORWARD_COMPAT_FEATURE) \
 		-p $(TFHE_SPEC)
 	@"$(MAKE)" symlink_c_libs_without_fingerprint
 
 .PHONY: build_c_api_gpu # Build the C API for boolean, shortint and integer
 build_c_api_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) build --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok-experimental,gpu \
+		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,gpu \
 		-p $(TFHE_SPEC)
 	@"$(MAKE)" symlink_c_libs_without_fingerprint
 
 .PHONY: build_c_api_experimental_deterministic_fft # Build the C API for boolean, shortint and integer with experimental deterministic FFT
 build_c_api_experimental_deterministic_fft: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) build --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok-experimental,experimental-force_fft_algo_dif4,$(FORWARD_COMPAT_FEATURE) \
+		--features=$(TARGET_ARCH_FEATURE),boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,experimental-force_fft_algo_dif4,$(FORWARD_COMPAT_FEATURE) \
 		-p $(TFHE_SPEC)
 	@"$(MAKE)" symlink_c_libs_without_fingerprint
 
@@ -391,7 +408,7 @@ build_web_js_api: install_rs_build_toolchain install_wasm_pack
 	cd tfhe && \
 	RUSTFLAGS="$(WASM_RUSTFLAGS)" rustup run "$(RS_BUILD_TOOLCHAIN)" \
 		wasm-pack build --release --target=web \
-		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok-experimental
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok
 
 .PHONY: build_web_js_api_parallel # Build the js API targeting the web browser with parallelism support
 build_web_js_api_parallel: install_rs_check_toolchain install_wasm_pack
@@ -399,15 +416,16 @@ build_web_js_api_parallel: install_rs_check_toolchain install_wasm_pack
 	rustup component add rust-src --toolchain $(RS_CHECK_TOOLCHAIN) && \
 	RUSTFLAGS="$(WASM_RUSTFLAGS) -C target-feature=+atomics,+bulk-memory,+mutable-globals" rustup run $(RS_CHECK_TOOLCHAIN) \
 		wasm-pack build --release --target=web \
-		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,parallel-wasm-api,zk-pok-experimental \
-		-Z build-std=panic_abort,std
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,parallel-wasm-api,zk-pok \
+		-Z build-std=panic_abort,std && \
+	find pkg/snippets -type f -iname workerHelpers.worker.js -exec sed -i "s|from '..\/..\/..\/';|from '..\/..\/..\/tfhe.js';|" {} \;
 
 .PHONY: build_node_js_api # Build the js API targeting nodejs
 build_node_js_api: install_rs_build_toolchain install_wasm_pack
 	cd tfhe && \
 	RUSTFLAGS="$(WASM_RUSTFLAGS)" rustup run "$(RS_BUILD_TOOLCHAIN)" \
 		wasm-pack build --release --target=nodejs \
-		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok-experimental
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok
 
 .PHONY: build_concrete_csprng # Build concrete_csprng
 build_concrete_csprng: install_rs_build_toolchain
@@ -417,10 +435,10 @@ build_concrete_csprng: install_rs_build_toolchain
 .PHONY: test_core_crypto # Run the tests of the core_crypto module including experimental ones
 test_core_crypto: install_rs_build_toolchain install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),experimental,zk-pok-experimental -p $(TFHE_SPEC) -- core_crypto::
+		--features=$(TARGET_ARCH_FEATURE),experimental,zk-pok -p $(TFHE_SPEC) -- core_crypto::
 	@if [[ "$(AVX512_SUPPORT)" == "ON" ]]; then \
 		RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
-			--features=$(TARGET_ARCH_FEATURE),experimental,zk-pok-experimental,$(AVX512_FEATURE) -p $(TFHE_SPEC) -- core_crypto::; \
+			--features=$(TARGET_ARCH_FEATURE),experimental,zk-pok,$(AVX512_FEATURE) -p $(TFHE_SPEC) -- core_crypto::; \
 	fi
 
 .PHONY: test_core_crypto_cov # Run the tests of the core_crypto module with code coverage
@@ -443,8 +461,8 @@ test_cuda_backend:
 	mkdir -p "$(TFHECUDA_BUILD)" && \
 		cd "$(TFHECUDA_BUILD)" && \
 		cmake .. -DCMAKE_BUILD_TYPE=Release -DTFHE_CUDA_BACKEND_BUILD_TESTS=ON && \
-		make -j "$(CPU_COUNT)" && \
-		make test
+		"$(MAKE)" -j "$(CPU_COUNT)" && \
+		"$(MAKE)" test
 
 .PHONY: test_gpu # Run the tests of the core_crypto module including experimental on the gpu backend
 test_gpu: test_core_crypto_gpu test_integer_gpu test_cuda_backend
@@ -459,10 +477,64 @@ test_core_crypto_gpu: install_rs_build_toolchain
 .PHONY: test_integer_gpu # Run the tests of the integer module including experimental on the gpu backend
 test_integer_gpu: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key::
+		--features=$(TARGET_ARCH_FEATURE),integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key:: --test-threads=6
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --doc --profile $(CARGO_PROFILE) \
 		--features=$(TARGET_ARCH_FEATURE),integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key::
 
+.PHONY: test_integer_gpu_ci # Run the tests for integer ci on gpu backend
+test_integer_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --backend "gpu" \
+		--tfhe-package "$(TFHE_SPEC)"
+
+.PHONY: test_unsigned_integer_gpu_ci # Run the tests for unsigned integer ci on gpu backend
+test_unsigned_integer_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --backend "gpu" \
+		--unsigned-only --tfhe-package "$(TFHE_SPEC)"
+
+.PHONY: test_signed_integer_gpu_ci # Run the tests for signed integer ci on gpu backend
+test_signed_integer_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --backend "gpu" \
+		--signed-only --tfhe-package "$(TFHE_SPEC)"
+
+.PHONY: test_integer_multi_bit_gpu_ci # Run the tests for integer ci on gpu backend running only multibit tests
+test_integer_multi_bit_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --backend "gpu" \
+		--tfhe-package "$(TFHE_SPEC)"
+
+.PHONY: test_unsigned_integer_multi_bit_gpu_ci # Run the tests for unsigned integer ci on gpu backend running only multibit tests
+test_unsigned_integer_multi_bit_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --backend "gpu" \
+		--unsigned-only --tfhe-package "$(TFHE_SPEC)"
+
+.PHONY: test_signed_integer_multi_bit_gpu_ci # Run the tests for signed integer ci on gpu backend running only multibit tests
+test_signed_integer_multi_bit_gpu_ci: install_rs_check_toolchain install_cargo_nextest
+	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
+	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
+		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
+		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --backend "gpu" \
+		--signed-only --tfhe-package "$(TFHE_SPEC)"
+
 .PHONY: test_boolean # Run the tests of the boolean module
 test_boolean: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
@@ -525,6 +597,7 @@ test_shortint_cov: install_rs_check_toolchain install_tarpaulin
 test_integer_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --avx512-support "$(AVX512_SUPPORT)" \
 		--tfhe-package "$(TFHE_SPEC)"
@@ -533,6 +606,7 @@ test_integer_ci: install_rs_check_toolchain install_cargo_nextest
 test_unsigned_integer_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --avx512-support "$(AVX512_SUPPORT)" \
 		--unsigned-only --tfhe-package "$(TFHE_SPEC)"
@@ -541,6 +615,7 @@ test_unsigned_integer_ci: install_rs_check_toolchain install_cargo_nextest
 test_signed_integer_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --avx512-support "$(AVX512_SUPPORT)" \
 		--signed-only --tfhe-package "$(TFHE_SPEC)"
@@ -549,22 +624,25 @@ test_signed_integer_ci: install_rs_check_toolchain install_cargo_nextest
 test_integer_multi_bit_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --avx512-support "$(AVX512_SUPPORT)" \
 		--tfhe-package "$(TFHE_SPEC)"
 
-.PHONY: test_unsigned_integer_multi_bit_ci # Run the tests for nsigned integer ci running only multibit tests
+.PHONY: test_unsigned_integer_multi_bit_ci # Run the tests for unsigned integer ci running only multibit tests
 test_unsigned_integer_multi_bit_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --avx512-support "$(AVX512_SUPPORT)" \
 		--unsigned-only --tfhe-package "$(TFHE_SPEC)"
 
-.PHONY: test_signed_integer_multi_bit_ci # Run the tests for nsigned integer ci running only multibit tests
+.PHONY: test_signed_integer_multi_bit_ci # Run the tests for signed integer ci running only multibit tests
 test_signed_integer_multi_bit_ci: install_rs_check_toolchain install_cargo_nextest
 	BIG_TESTS_INSTANCE="$(BIG_TESTS_INSTANCE)" \
 	FAST_TESTS="$(FAST_TESTS)" \
+	NIGHTLY_TESTS="$(NIGHTLY_TESTS)" \
 		./scripts/integer-tests.sh --rust-toolchain $(CARGO_RS_CHECK_TOOLCHAIN) \
 		--cargo-profile "$(CARGO_PROFILE)" --multi-bit --avx512-support "$(AVX512_SUPPORT)" \
 		--signed-only --tfhe-package "$(TFHE_SPEC)"
@@ -591,7 +669,7 @@ test_integer_cov: install_rs_check_toolchain install_tarpaulin
 .PHONY: test_high_level_api # Run all the tests for high_level_api
 test_high_level_api: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,zk-pok-experimental -p $(TFHE_SPEC) \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,zk-pok -p $(TFHE_SPEC) \
 		-- high_level_api::
 
 test_high_level_api_gpu: install_rs_build_toolchain install_cargo_nextest
@@ -602,14 +680,14 @@ test_high_level_api_gpu: install_rs_build_toolchain install_cargo_nextest
 .PHONY: test_user_doc # Run tests from the .md documentation
 test_user_doc: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) --doc \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,pbs-stats,zk-pok-experimental \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,pbs-stats,zk-pok \
 		-p $(TFHE_SPEC) \
 		-- test_user_docs::
 
 .PHONY: test_user_doc_gpu # Run tests for GPU from the .md documentation
 test_user_doc_gpu: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) --doc \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,gpu,zk-pok-experimental -p $(TFHE_SPEC) \
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,internal-keycache,gpu,zk-pok -p $(TFHE_SPEC) \
 		-- test_user_docs::
 
 .PHONY: test_fhe_strings # Run tests for fhe_strings example
@@ -648,18 +726,38 @@ test_concrete_csprng: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
 		--features=$(TARGET_ARCH_FEATURE) -p concrete-csprng
 
-.PHONY: test_zk_pok # Run tfhe-zk-pok-experimental tests
+.PHONY: test_zk_pok # Run tfhe-zk-pok tests
 test_zk_pok: install_rs_build_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
 		-p tfhe-zk-pok
 
+.PHONY: test_versionable # Run tests for tfhe-versionable subcrate
+test_versionable: install_rs_build_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
+		-p tfhe-versionable
+
+# The backward compat data repo holds historical binary data but also rust code to generate and load them.
+# Here we use the "patch" functionality of Cargo to make sure the repo used for the data is the same as the one used for the code.
+.PHONY: test_backward_compatibility_ci
+test_backward_compatibility_ci: install_rs_build_toolchain
+	TFHE_BACKWARD_COMPAT_DATA_DIR="$(BACKWARD_COMPAT_DATA_DIR)" RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
+		--config "patch.'$(BACKWARD_COMPAT_DATA_URL)'.$(BACKWARD_COMPAT_DATA_PROJECT).path=\"tfhe/$(BACKWARD_COMPAT_DATA_DIR)\"" \
+		--features=$(TARGET_ARCH_FEATURE),shortint,integer -p $(TFHE_SPEC) test_backward_compatibility -- --nocapture
+
+.PHONY: test_backward_compatibility # Same as test_backward_compatibility_ci but tries to clone the data repo first if needed
+test_backward_compatibility: tfhe/$(BACKWARD_COMPAT_DATA_DIR) test_backward_compatibility_ci
+
+.PHONY: backward_compat_branch # Prints the required backward compatibility branch
+backward_compat_branch:
+	@echo "$(BACKWARD_COMPAT_DATA_BRANCH)"
+
 .PHONY: doc # Build rust doc
 doc: install_rs_check_toolchain
 	@# Even though we are not in docs.rs, this allows to "just" build the doc
 	DOCS_RS=1 \
 	RUSTDOCFLAGS="--html-in-header katex-header.html" \
 	cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" doc \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,gpu,internal-keycache,experimental --no-deps -p $(TFHE_SPEC)
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,gpu,internal-keycache,experimental,zk-pok --no-deps -p $(TFHE_SPEC)
 
 .PHONY: docs # Build rust doc alias for doc
 docs: doc
@@ -670,7 +768,7 @@ lint_doc: install_rs_check_toolchain
 	DOCS_RS=1 \
 	RUSTDOCFLAGS="--html-in-header katex-header.html -Dwarnings" \
 	cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" doc \
-		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,gpu,internal-keycache,experimental -p $(TFHE_SPEC) --no-deps
+		--features=$(TARGET_ARCH_FEATURE),boolean,shortint,integer,gpu,internal-keycache,experimental,zk-pok -p $(TFHE_SPEC) --no-deps
 
 .PHONY: lint_docs # Build rust doc with linting enabled alias for lint_doc
 lint_docs: lint_doc
@@ -715,7 +813,7 @@ check_compile_tests_benches_gpu: install_rs_build_toolchain
 	mkdir -p "$(TFHECUDA_BUILD)" && \
 		cd "$(TFHECUDA_BUILD)" && \
 		cmake .. -DCMAKE_BUILD_TYPE=Debug -DTFHE_CUDA_BACKEND_BUILD_TESTS=ON -DTFHE_CUDA_BACKEND_BUILD_BENCHMARKS=ON && \
-		make -j "$(CPU_COUNT)"
+		"$(MAKE)" -j "$(CPU_COUNT)"
 
 .PHONY: build_nodejs_test_docker # Build a docker image with tools to run nodejs tests for wasm API
 build_nodejs_test_docker:
@@ -735,14 +833,14 @@ test_nodejs_wasm_api_in_docker: build_nodejs_test_docker
 
 .PHONY: test_nodejs_wasm_api # Run tests for the nodejs on wasm API
 test_nodejs_wasm_api: build_node_js_api
-	cd tfhe && node --test js_on_wasm_tests
+	cd tfhe/js_on_wasm_tests && npm run test
 
 .PHONY: test_web_js_api_parallel # Run tests for the web wasm api
 test_web_js_api_parallel: build_web_js_api_parallel
 	$(MAKE) -C tfhe/web_wasm_parallel_tests test
 
-.PHONY: ci_test_web_js_api_parallel # Run tests for the web wasm api
-ci_test_web_js_api_parallel: build_web_js_api_parallel
+.PHONY: test_web_js_api_parallel_ci # Run tests for the web wasm api
+test_web_js_api_parallel_ci: build_web_js_api_parallel
 	source ~/.nvm/nvm.sh && \
 	nvm install $(NODE_VERSION) && \
 	nvm use $(NODE_VERSION) && \
@@ -809,6 +907,22 @@ bench_integer_multi_bit_gpu: install_rs_check_toolchain
 	--bench integer-bench \
 	--features=$(TARGET_ARCH_FEATURE),integer,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
 
+.PHONY: bench_unsigned_integer_multi_bit_gpu # Run benchmarks for unsigned integer on GPU backend using multi-bit parameters
+bench_unsigned_integer_multi_bit_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=MULTI_BIT \
+	__TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench integer-bench \
+	--features=$(TARGET_ARCH_FEATURE),integer,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) -- ::unsigned
+
+.PHONY: bench_integer_zk # Run benchmarks for integer encryption with ZK proofs
+bench_integer_zk: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench zk-pke-bench \
+	--features=$(TARGET_ARCH_FEATURE),integer,internal-keycache,zk-pok,nightly-avx512 \
+	-p $(TFHE_SPEC) --
+
 .PHONY: bench_shortint # Run benchmarks for shortint
 bench_shortint: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) \
@@ -816,16 +930,12 @@ bench_shortint: install_rs_check_toolchain
 	--bench shortint-bench \
 	--features=$(TARGET_ARCH_FEATURE),shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
-.PHONY: bench_oprf # Run benchmarks for shortint
-bench_oprf: install_rs_check_toolchain
+.PHONY: bench_shortint_oprf # Run benchmarks for shortint
+bench_shortint_oprf: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench oprf-shortint-bench \
 	--features=$(TARGET_ARCH_FEATURE),shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
-	RUSTFLAGS="$(RUSTFLAGS)" \
-	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench oprf-integer-bench \
-	--features=$(TARGET_ARCH_FEATURE),integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
 .PHONY: bench_shortint_multi_bit # Run benchmarks for shortint using multi-bit parameters
 bench_shortint_multi_bit: install_rs_check_toolchain
@@ -847,9 +957,15 @@ bench_pbs: install_rs_check_toolchain
 	--bench pbs-bench \
 	--features=$(TARGET_ARCH_FEATURE),boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
+.PHONY: bench_pbs128 # Run benchmarks for PBS using FFT 128 bits
+bench_pbs128: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench pbs128-bench \
+	--features=$(TARGET_ARCH_FEATURE),boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
+
 .PHONY: bench_pbs_gpu # Run benchmarks for PBS on GPU backend
 bench_pbs_gpu: install_rs_check_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_FAST_BENCH=$(FAST_BENCH) cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench pbs-bench \
 	--features=$(TARGET_ARCH_FEATURE),boolean,shortint,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
@@ -869,10 +985,10 @@ bench_ks_gpu: install_rs_check_toolchain
 bench_web_js_api_parallel: build_web_js_api_parallel
 	$(MAKE) -C tfhe/web_wasm_parallel_tests bench
 
-.PHONY: ci_bench_web_js_api_parallel # Run benchmarks for the web wasm api
-ci_bench_web_js_api_parallel: build_web_js_api_parallel
+.PHONY: bench_web_js_api_parallel_ci # Run benchmarks for the web wasm api
+bench_web_js_api_parallel_ci: build_web_js_api_parallel
 	source ~/.nvm/nvm.sh && \
-	nvm use node && \
+	nvm use $(NODE_VERSION) && \
 	$(MAKE) -C tfhe/web_wasm_parallel_tests bench-ci
 
 #
@@ -929,6 +1045,12 @@ write_params_to_file: install_rs_check_toolchain
 	--example write_params_to_file \
 	--features=$(TARGET_ARCH_FEATURE),boolean,shortint,internal-keycache
 
+.PHONY: clone_backward_compat_data # Clone the data repo needed for backward compatibility tests
+clone_backward_compat_data:
+	./scripts/clone_backward_compat_data.sh $(BACKWARD_COMPAT_DATA_URL) $(BACKWARD_COMPAT_DATA_BRANCH) tfhe/$(BACKWARD_COMPAT_DATA_DIR)
+
+tfhe/$(BACKWARD_COMPAT_DATA_DIR): clone_backward_compat_data
+
 #
 # Real use case examples
 #
@@ -955,7 +1077,7 @@ sha256_bool: install_rs_check_toolchain
 
 .PHONY: pcc # pcc stands for pre commit checks (except GPU)
 pcc: no_tfhe_typo no_dbg_log check_fmt lint_doc check_md_docs_are_tested check_intra_md_links \
-clippy_all check_compile_tests
+clippy_all tfhe_lints check_compile_tests
 
 .PHONY: pcc_gpu # pcc stands for pre commit checks for GPU compilation
 pcc_gpu: clippy_gpu clippy_cuda_backend check_compile_tests_benches_gpu

README.md

@@ -50,7 +50,7 @@ production-ready library for all the advanced features of TFHE.
 <br></br>
 
 ## Table of Contents
-- **[Getting Started](#getting-started)**
+- **[Getting started](#getting-started)**
    - [Cargo.toml configuration](#cargotoml-configuration)
    - [A simple example](#a-simple-example)
 - **[Resources](#resources)**
@@ -65,7 +65,7 @@ production-ready library for all the advanced features of TFHE.
 - **[Support](#support)**
 <br></br>
 
-## Getting Started
+## Getting started
 
 ### Cargo.toml configuration
 To use the latest version of `TFHE-rs` in your project, you first need to add it as a dependency in your `Cargo.toml`:
@@ -198,7 +198,7 @@ Full, comprehensive documentation is available here: [https://docs.zama.ai/tfhe-
 
 ### Disclaimers
 
-#### Security Estimation
+#### Security estimation
 
 Security estimations are done using the
 [Lattice Estimator](https://github.com/malb/lattice-estimator)
@@ -206,13 +206,13 @@ with `red_cost_model = reduction.RC.BDGL16`.
 
 When a new update is published in the Lattice Estimator, we update parameters accordingly.
 
-### Security Model
+### Security model
 
-The default parameters for the TFHE-rs library are chosen considering the IND-CPA security model, and are selected with a bootstrapping failure probability fixed at p_error = $2^{-40}$. In particular, it is assumed that the results of decrypted computations are not shared by the secret key owner with any third parties, as such an action can lead to leakage of the secret encryption key. If you are designing an application where decryptions must be shared, you will need to craft custom encryption parameters which are chosen in consideration of the IND-CPA^D security model [1]. 
+The default parameters for the TFHE-rs library are chosen considering the IND-CPA security model, and are selected with a bootstrapping failure probability fixed at p_error = $2^{-64}$. In particular, it is assumed that the results of decrypted computations are not shared by the secret key owner with any third parties, as such an action can lead to leakage of the secret encryption key. If you are designing an application where decryptions must be shared, you will need to craft custom encryption parameters which are chosen in consideration of the IND-CPA^D security model [1]. 
 
 [1] Li, Baiyu, et al. "Securing approximate homomorphic encryption using differential privacy." Annual International Cryptology Conference. Cham: Springer Nature Switzerland, 2022. https://eprint.iacr.org/2022/816.pdf
 
-#### Side-Channel Attacks
+#### Side-channel attacks
 
 Mitigation for side-channel attacks has not yet been implemented in TFHE-rs,
 and will be released in upcoming versions.
@@ -241,7 +241,23 @@ Becoming an approved contributor involves signing our Contributor License Agreem
 <br></br>
 
 ### License
-This software is distributed under the **BSD-3-Clause-Clear** license. If you have any questions, please contact us at hello@zama.ai.
+This software is distributed under the **BSD-3-Clause-Clear** license. Read [this](LICENSE) for more details.
+
+#### FAQ
+**Is Zama’s technology free to use?**
+>Zama’s libraries are free to use under the BSD 3-Clause Clear license only for development, research, prototyping, and experimentation purposes. However, for any commercial use of Zama's open source code, companies must purchase Zama’s commercial patent license.
+>
+>Everything we do is open source and we are very transparent on what it means for our users, you can read more about how we monetize our open source products at Zama in [this blogpost](https://www.zama.ai/post/open-source).
+
+**What do I need to do if I want to use Zama’s technology for commercial purposes?**
+>To commercially use Zama’s technology you need to be granted Zama’s patent license. Please contact us hello@zama.ai for more information.
+
+**Do you file IP on your technology?**
+>Yes, all Zama’s technologies are patented.
+
+**Can you customize a solution for my specific use case?**
+>We are open to collaborating and advancing the FHE space with our partners. If you have specific needs, please email us at hello@zama.ai.
+
 <p align="right">
   <a href="#about" > ↑ Back to top </a> 
 </p>

apps/trivium/benches/kreyvium_byte.rs

@@ -4,9 +4,8 @@ use tfhe::{generate_keys, ConfigBuilder, FheUint64, FheUint8};
 use tfhe_trivium::{KreyviumStreamByte, TransCiphering};
 
 pub fn kreyvium_byte_gen(c: &mut Criterion) {
-    let config = ConfigBuilder::default()
-        .enable_function_evaluation()
-        .build();
+    let config = ConfigBuilder::default().build();
+
     let (client_key, server_key) = generate_keys(config);
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();
@@ -33,9 +32,8 @@ pub fn kreyvium_byte_gen(c: &mut Criterion) {
 }
 
 pub fn kreyvium_byte_trans(c: &mut Criterion) {
-    let config = ConfigBuilder::default()
-        .enable_function_evaluation()
-        .build();
+    let config = ConfigBuilder::default().build();
+
     let (client_key, server_key) = generate_keys(config);
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();
@@ -63,9 +61,8 @@ pub fn kreyvium_byte_trans(c: &mut Criterion) {
 }
 
 pub fn kreyvium_byte_warmup(c: &mut Criterion) {
-    let config = ConfigBuilder::default()
-        .enable_function_evaluation()
-        .build();
+    let config = ConfigBuilder::default().build();
+
     let (client_key, server_key) = generate_keys(config);
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();

apps/trivium/benches/kreyvium_shortint.rs

@@ -13,7 +13,7 @@ pub fn kreyvium_shortint_warmup(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );
@@ -63,7 +63,7 @@ pub fn kreyvium_shortint_gen(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );
@@ -108,7 +108,7 @@ pub fn kreyvium_shortint_trans(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );

apps/trivium/benches/trivium_shortint.rs

@@ -13,7 +13,7 @@ pub fn trivium_shortint_warmup(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );
@@ -63,7 +63,7 @@ pub fn trivium_shortint_gen(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );
@@ -108,7 +108,7 @@ pub fn trivium_shortint_trans(c: &mut Criterion) {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );

apps/trivium/src/kreyvium/kreyvium_byte.rs

@@ -119,7 +119,7 @@ impl KreyviumStreamByte<FheUint8> {
         }
 
         // Key and iv are stored in reverse in their shift registers
-        let mut key = key_bytes.map(|b| b.map(|x| (x as u8).reverse_bits() as u64));
+        let mut key = key_bytes.map(|b| b.reverse_bits());
         let mut iv = iv_bytes.map(|x| FheUint8::encrypt_trivial(x.reverse_bits()));
         key.reverse();
         iv.reverse();

apps/trivium/src/kreyvium/test.rs

@@ -224,7 +224,7 @@ fn kreyvium_test_shortint_long() {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );
@@ -299,9 +299,8 @@ fn kreyvium_test_clear_byte() {
 
 #[test]
 fn kreyvium_test_byte_long() {
-    let config = ConfigBuilder::default()
-        .enable_function_evaluation()
-        .build();
+    let config = ConfigBuilder::default().build();
+
     let (client_key, server_key) = generate_keys(config);
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();
@@ -338,9 +337,8 @@ fn kreyvium_test_byte_long() {
 
 #[test]
 fn kreyvium_test_fhe_byte_transciphering_long() {
-    let config = ConfigBuilder::default()
-        .enable_function_evaluation()
-        .build();
+    let config = ConfigBuilder::default().build();
+
     let (client_key, server_key) = generate_keys(config);
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();

apps/trivium/src/trivium/test.rs

@@ -360,7 +360,7 @@ fn trivium_test_shortint_long() {
     let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(PARAM_MESSAGE_1_CARRY_1_KS_PBS);
 
     let ksk = KeySwitchingKey::new(
-        (&client_key, &server_key),
+        (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
         PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS,
     );

backends/tfhe-cuda-backend/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tfhe-cuda-backend"
-version = "0.2.0"
+version = "0.4.0-alpha.0"
 edition = "2021"
 authors = ["Zama team"]
 license = "BSD-3-Clause-Clear"
@@ -14,6 +14,3 @@ keywords = ["fully", "homomorphic", "encryption", "fhe", "cryptography"]
 [build-dependencies]
 cmake = { version = "0.1" }
 pkg-config = { version = "0.3" }
-
-[dependencies]
-thiserror = "1.0"

backends/tfhe-cuda-backend/build.rs

@@ -8,7 +8,24 @@ fn main() {
         }
     }
 
+    // This is a workaround to the current nightly toolchain (2024-06-27 which started with
+    // toolchain 2024-05-05) build issue
+    // Essentially if cbindgen is running, a wrong argument ends up forwarded to the cuda backend
+    // "make" command during macro expansions for TFHE-rs C API, crashing make for make < 4.4 and
+    // thus crashing the build
+    // On the other hand, this speeds up C API build greatly given we don't have macro expansions
+    // in the CUDA backend so this skips the second compilation of TFHE-rs for macro inspection by
+    // cbindgen
+    if std::env::var("_CBINDGEN_IS_RUNNING").is_ok() {
+        return;
+    }
+
     println!("Build tfhe-cuda-backend");
+    println!("cargo::rerun-if-changed=cuda/include");
+    println!("cargo::rerun-if-changed=cuda/src");
+    println!("cargo::rerun-if-changed=cuda/tests_and_benchmarks");
+    println!("cargo::rerun-if-changed=cuda/CMakeLists.txt");
+    println!("cargo::rerun-if-changed=src");
     if env::consts::OS == "linux" {
         let output = Command::new("./get_os_name.sh").output().unwrap();
         let distribution = String::from_utf8(output.stdout).unwrap();

backends/tfhe-cuda-backend/cuda/include/ciphertext.h

@@ -1,6 +1,7 @@
 #ifndef CUDA_CIPHERTEXT_H
 #define CUDA_CIPHERTEXT_H
 
+#include "device.h"
 #include <cstdint>
 
 extern "C" {
@@ -14,5 +15,11 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu_64(void *stream,
                                                   void *dest, void *src,
                                                   uint32_t number_of_cts,
                                                   uint32_t lwe_dimension);
+
+void cuda_glwe_sample_extract_64(void *stream, uint32_t gpu_index,
+                                 void *lwe_array_out, void *glwe_array_in,
+                                 uint32_t *nth_array, uint32_t num_glwes,
+                                 uint32_t glwe_dimension,
+                                 uint32_t polynomial_size);
 };
 #endif

backends/tfhe-cuda-backend/cuda/include/device.h

@@ -6,6 +6,7 @@
 #include <cstdlib>
 #include <cstring>
 #include <cuda_runtime.h>
+#include <vector>
 
 #define synchronize_threads_in_block() __syncthreads()
 extern "C" {
@@ -63,14 +64,8 @@ void cuda_drop(void *ptr, uint32_t gpu_index);
 void cuda_drop_async(void *ptr, cudaStream_t stream, uint32_t gpu_index);
 
 int cuda_get_max_shared_memory(uint32_t gpu_index);
-
-void cuda_stream_add_callback(cudaStream_t stream, uint32_t gpu_index,
-                              cudaStreamCallback_t callback, void *user_data);
 }
 
-void host_free_on_stream_callback(cudaStream_t stream, cudaError_t status,
-                                  void *host_pointer);
-
 template <typename Torus>
 void cuda_set_value_async(cudaStream_t stream, uint32_t gpu_index,
                           Torus *d_array, Torus value, Torus n);

backends/tfhe-cuda-backend/cuda/include/helper.h

@@ -1,10 +0,0 @@
-#ifndef HELPER_H
-#define HELPER_H
-
-extern "C" {
-int cuda_setup_multi_gpu();
-}
-
-void multi_gpu_checks(uint32_t gpu_count);
-
-#endif

backends/tfhe-cuda-backend/cuda/include/helper_multi_gpu.h

@@ -0,0 +1,34 @@
+#ifndef HELPER_MULTI_GPU_H
+#define HELPER_MULTI_GPU_H
+#include <mutex>
+#include <variant>
+#include <vector>
+
+extern std::mutex m;
+extern bool p2p_enabled;
+
+extern "C" {
+int cuda_setup_multi_gpu();
+}
+
+// Define a variant type that can be either a vector or a single pointer
+template <typename Torus>
+using LweArrayVariant = std::variant<std::vector<Torus *>, Torus *>;
+
+// Macro to define the visitor logic using std::holds_alternative for vectors
+#define GET_VARIANT_ELEMENT(variant, index)                                    \
+  [&] {                                                                        \
+    if (std::holds_alternative<std::vector<Torus *>>(variant)) {               \
+      return std::get<std::vector<Torus *>>(variant)[index];                   \
+    } else {                                                                   \
+      return std::get<Torus *>(variant);                                       \
+    }                                                                          \
+  }()
+
+int get_active_gpu_count(int num_inputs, int gpu_count);
+
+int get_num_inputs_on_gpu(int total_num_inputs, int gpu_index, int gpu_count);
+
+int get_gpu_offset(int total_num_inputs, int gpu_index, int gpu_count);
+
+#endif

backends/tfhe-cuda-backend/cuda/include/programmable_bootstrap.h

@@ -26,14 +26,12 @@ void cuda_convert_lwe_programmable_bootstrap_key_64(
 void scratch_cuda_programmable_bootstrap_amortized_32(
     void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 void scratch_cuda_programmable_bootstrap_amortized_64(
     void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 void cuda_programmable_bootstrap_amortized_lwe_ciphertext_vector_32(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
@@ -41,8 +39,7 @@ void cuda_programmable_bootstrap_amortized_lwe_ciphertext_vector_32(
     void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
     int8_t *pbs_buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t base_log, uint32_t level_count,
-    uint32_t num_samples, uint32_t num_luts, uint32_t lwe_idx,
-    uint32_t max_shared_memory);
+    uint32_t num_samples);
 
 void cuda_programmable_bootstrap_amortized_lwe_ciphertext_vector_64(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
@@ -50,8 +47,7 @@ void cuda_programmable_bootstrap_amortized_lwe_ciphertext_vector_64(
     void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
     int8_t *pbs_buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t base_log, uint32_t level_count,
-    uint32_t num_samples, uint32_t num_luts, uint32_t lwe_idx,
-    uint32_t max_shared_memory);
+    uint32_t num_samples);
 
 void cleanup_cuda_programmable_bootstrap_amortized(void *stream,
                                                    uint32_t gpu_index,
@@ -60,14 +56,12 @@ void cleanup_cuda_programmable_bootstrap_amortized(void *stream,
 void scratch_cuda_programmable_bootstrap_32(
     void *stream, uint32_t gpu_index, int8_t **buffer, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 void scratch_cuda_programmable_bootstrap_64(
     void *stream, uint32_t gpu_index, int8_t **buffer, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 void cuda_programmable_bootstrap_lwe_ciphertext_vector_32(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
@@ -75,8 +69,7 @@ void cuda_programmable_bootstrap_lwe_ciphertext_vector_32(
     void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
     int8_t *buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t base_log, uint32_t level_count,
-    uint32_t num_samples, uint32_t num_luts, uint32_t lwe_idx,
-    uint32_t max_shared_memory);
+    uint32_t num_samples);
 
 void cuda_programmable_bootstrap_lwe_ciphertext_vector_64(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
@@ -84,44 +77,41 @@ void cuda_programmable_bootstrap_lwe_ciphertext_vector_64(
     void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
     int8_t *buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t base_log, uint32_t level_count,
-    uint32_t num_samples, uint32_t num_luts, uint32_t lwe_idx,
-    uint32_t max_shared_memory);
+    uint32_t num_samples);
 
 void cleanup_cuda_programmable_bootstrap(void *stream, uint32_t gpu_index,
                                          int8_t **pbs_buffer);
 
 uint64_t get_buffer_size_programmable_bootstrap_amortized_64(
     uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory);
+    uint32_t input_lwe_ciphertext_count);
 
 uint64_t get_buffer_size_programmable_bootstrap_64(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory);
+    uint32_t input_lwe_ciphertext_count);
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_programmable_bootstrap_step_one(
+uint64_t get_buffer_size_full_sm_programmable_bootstrap_step_one(
     uint32_t polynomial_size) {
   return sizeof(Torus) * polynomial_size +      // accumulator_rotated
          sizeof(double2) * polynomial_size / 2; // accumulator fft
 }
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_programmable_bootstrap_step_two(
+uint64_t get_buffer_size_full_sm_programmable_bootstrap_step_two(
     uint32_t polynomial_size) {
   return sizeof(Torus) * polynomial_size +      // accumulator
          sizeof(double2) * polynomial_size / 2; // accumulator fft
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
+uint64_t
 get_buffer_size_partial_sm_programmable_bootstrap(uint32_t polynomial_size) {
   return sizeof(double2) * polynomial_size / 2; // accumulator fft
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
+uint64_t
 get_buffer_size_full_sm_programmable_bootstrap_tbc(uint32_t polynomial_size) {
   return sizeof(Torus) * polynomial_size +      // accumulator_rotated
          sizeof(Torus) * polynomial_size +      // accumulator
@@ -129,21 +119,19 @@ get_buffer_size_full_sm_programmable_bootstrap_tbc(uint32_t polynomial_size) {
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_partial_sm_programmable_bootstrap_tbc(
+uint64_t get_buffer_size_partial_sm_programmable_bootstrap_tbc(
     uint32_t polynomial_size) {
   return sizeof(double2) * polynomial_size / 2; // accumulator fft mask & body
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap(
+uint64_t get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap(
     uint32_t polynomial_size) {
   return sizeof(double2) * polynomial_size / 2; // tbc
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
+uint64_t
 get_buffer_size_full_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
   return sizeof(Torus) * polynomial_size +      // accumulator_rotated
          sizeof(Torus) * polynomial_size +      // accumulator
@@ -151,15 +139,14 @@ get_buffer_size_full_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
 }
 
 template <typename Torus>
-__host__ __device__ uint64_t
+uint64_t
 get_buffer_size_partial_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
   return sizeof(double2) * polynomial_size / 2; // accumulator fft mask & body
 }
 
 template <typename Torus>
-__host__ bool
-supports_distributed_shared_memory_on_classic_programmable_bootstrap(
-    uint32_t polynomial_size, uint32_t max_shared_memory);
+bool supports_distributed_shared_memory_on_classic_programmable_bootstrap(
+    uint32_t polynomial_size);
 
 template <typename Torus, PBS_TYPE pbs_type> struct pbs_buffer;
 
@@ -178,7 +165,7 @@ template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::CLASSICAL> {
 
     this->pbs_variant = pbs_variant;
 
-    auto max_shared_memory = cuda_get_max_shared_memory(gpu_index);
+    auto max_shared_memory = cuda_get_max_shared_memory(0);
 
     if (allocate_gpu_memory) {
       switch (pbs_variant) {
@@ -255,7 +242,7 @@ template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::CLASSICAL> {
 
         bool supports_dsm =
             supports_distributed_shared_memory_on_classic_programmable_bootstrap<
-                Torus>(polynomial_size, max_shared_memory);
+                Torus>(polynomial_size);
 
         uint64_t full_sm =
             get_buffer_size_full_sm_programmable_bootstrap_tbc<Torus>(
@@ -314,10 +301,10 @@ template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::CLASSICAL> {
 };
 
 template <typename Torus>
-__host__ __device__ uint64_t get_buffer_size_programmable_bootstrap_cg(
+uint64_t get_buffer_size_programmable_bootstrap_cg(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory) {
-
+    uint32_t input_lwe_ciphertext_count) {
+  int max_shared_memory = cuda_get_max_shared_memory(0);
   uint64_t full_sm =
       get_buffer_size_full_sm_programmable_bootstrap_cg<Torus>(polynomial_size);
   uint64_t partial_sm =
@@ -343,8 +330,7 @@ template <typename Torus>
 bool has_support_to_cuda_programmable_bootstrap_cg(uint32_t glwe_dimension,
                                                    uint32_t polynomial_size,
                                                    uint32_t level_count,
-                                                   uint32_t num_samples,
-                                                   uint32_t max_shared_memory);
+                                                   uint32_t num_samples);
 
 template <typename Torus>
 void cuda_programmable_bootstrap_cg_lwe_ciphertext_vector(
@@ -353,8 +339,7 @@ void cuda_programmable_bootstrap_cg_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, double2 *bootstrapping_key,
     pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
-    uint32_t level_count, uint32_t num_samples, uint32_t num_luts,
-    uint32_t lwe_idx, uint32_t max_shared_memory);
+    uint32_t level_count, uint32_t num_samples);
 
 template <typename Torus>
 void cuda_programmable_bootstrap_lwe_ciphertext_vector(
@@ -363,8 +348,7 @@ void cuda_programmable_bootstrap_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, double2 *bootstrapping_key,
     pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
-    uint32_t level_count, uint32_t num_samples, uint32_t num_luts,
-    uint32_t lwe_idx, uint32_t max_shared_memory);
+    uint32_t level_count, uint32_t num_samples);
 
 #if (CUDA_ARCH >= 900)
 template <typename Torus>
@@ -374,43 +358,44 @@ void cuda_programmable_bootstrap_tbc_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, double2 *bootstrapping_key,
     pbs_buffer<Torus, CLASSICAL> *buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t base_log,
-    uint32_t level_count, uint32_t num_samples, uint32_t num_luts,
-    uint32_t lwe_idx, uint32_t max_shared_memory);
+    uint32_t level_count, uint32_t num_samples);
 
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_programmable_bootstrap_tbc(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **pbs_buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 #endif
 
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_programmable_bootstrap_cg(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **pbs_buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_programmable_bootstrap(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, CLASSICAL> **buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 template <typename Torus>
 bool has_support_to_cuda_programmable_bootstrap_tbc(uint32_t num_samples,
                                                     uint32_t glwe_dimension,
                                                     uint32_t polynomial_size,
-                                                    uint32_t level_count,
-                                                    uint32_t max_shared_memory);
+                                                    uint32_t level_count);
 
 #ifdef __CUDACC__
 __device__ inline int get_start_ith_ggsw(int i, uint32_t polynomial_size,
                                          int glwe_dimension,
                                          uint32_t level_count);
 
+template <typename T>
+__device__ const T *get_ith_mask_kth_block(const T *ptr, int i, int k,
+                                           int level, uint32_t polynomial_size,
+                                           int glwe_dimension,
+                                           uint32_t level_count);
+
 template <typename T>
 __device__ T *get_ith_mask_kth_block(T *ptr, int i, int k, int level,
                                      uint32_t polynomial_size,
@@ -422,8 +407,8 @@ __device__ T *get_ith_body_kth_block(T *ptr, int i, int k, int level,
                                      int glwe_dimension, uint32_t level_count);
 
 template <typename T>
-__device__ T *get_multi_bit_ith_lwe_gth_group_kth_block(
-    T *ptr, int g, int i, int k, int level, uint32_t grouping_factor,
+__device__ const T *get_multi_bit_ith_lwe_gth_group_kth_block(
+    const T *ptr, int g, int i, int k, int level, uint32_t grouping_factor,
     uint32_t polynomial_size, uint32_t glwe_dimension, uint32_t level_count);
 
 #endif

backends/tfhe-cuda-backend/cuda/include/programmable_bootstrap_multibit.h

@@ -8,7 +8,7 @@ extern "C" {
 
 bool has_support_to_cuda_programmable_bootstrap_cg_multi_bit(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t num_samples, uint32_t max_shared_memory);
+    uint32_t num_samples);
 
 void cuda_convert_lwe_multi_bit_programmable_bootstrap_key_64(
     void *stream, uint32_t gpu_index, void *dest, void *src,
@@ -19,8 +19,7 @@ void scratch_cuda_multi_bit_programmable_bootstrap_64(
     void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
     uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
     uint32_t level_count, uint32_t grouping_factor,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t chunk_size = 0);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 void cuda_multi_bit_programmable_bootstrap_lwe_ciphertext_vector_64(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
@@ -28,24 +27,7 @@ void cuda_multi_bit_programmable_bootstrap_lwe_ciphertext_vector_64(
     void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
     int8_t *buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
     uint32_t polynomial_size, uint32_t grouping_factor, uint32_t base_log,
-    uint32_t level_count, uint32_t num_samples, uint32_t num_luts,
-    uint32_t lwe_idx, uint32_t max_shared_memory, uint32_t lwe_chunk_size = 0);
-
-void scratch_cuda_generic_multi_bit_programmable_bootstrap_64(
-    void *stream, uint32_t gpu_index, int8_t **pbs_buffer,
-    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t level_count, uint32_t grouping_factor,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t lwe_chunk_size = 0);
-
-void cuda_generic_multi_bit_programmable_bootstrap_lwe_ciphertext_vector_64(
-    void *stream, uint32_t gpu_index, void *lwe_array_out,
-    void *lwe_output_indexes, void *lut_vector, void *lut_vector_indexes,
-    void *lwe_array_in, void *lwe_input_indexes, void *bootstrapping_key,
-    int8_t *pbs_buffer, uint32_t lwe_dimension, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t grouping_factor, uint32_t base_log,
-    uint32_t level_count, uint32_t num_samples, uint32_t num_luts,
-    uint32_t lwe_idx, uint32_t max_shared_memory, uint32_t lwe_chunk_size = 0);
+    uint32_t level_count, uint32_t num_samples);
 
 void cleanup_cuda_multi_bit_programmable_bootstrap(void *stream,
                                                    uint32_t gpu_index,
@@ -53,23 +35,21 @@ void cleanup_cuda_multi_bit_programmable_bootstrap(void *stream,
 }
 
 template <typename Torus>
-__host__ bool
-supports_distributed_shared_memory_on_multibit_programmable_bootstrap(
-    uint32_t polynomial_size, uint32_t max_shared_memory);
+bool supports_distributed_shared_memory_on_multibit_programmable_bootstrap(
+    uint32_t polynomial_size);
 
 template <typename Torus>
 bool has_support_to_cuda_programmable_bootstrap_tbc_multi_bit(
     uint32_t num_samples, uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t level_count, uint32_t max_shared_memory);
+    uint32_t level_count);
 
 #if CUDA_ARCH >= 900
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_tbc_multi_bit_programmable_bootstrap(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, MULTI_BIT> **buffer,
     uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
     uint32_t level_count, uint32_t grouping_factor,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t lwe_chunk_size);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 template <typename Torus>
 void cuda_tbc_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
@@ -78,25 +58,14 @@ void cuda_tbc_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, Torus *bootstrapping_key,
     pbs_buffer<Torus, MULTI_BIT> *pbs_buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
-    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
-    uint32_t num_luts, uint32_t lwe_idx, uint32_t max_shared_memory,
-    uint32_t lwe_chunk_size);
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples);
 #endif
 
-template <typename Torus, typename STorus>
-void scratch_cuda_cg_multi_bit_programmable_bootstrap(
-    void *stream, uint32_t gpu_index, pbs_buffer<Torus, MULTI_BIT> **pbs_buffer,
-    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t level_count, uint32_t grouping_factor,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t lwe_chunk_size = 0);
-
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_cg_multi_bit_programmable_bootstrap(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, MULTI_BIT> **pbs_buffer,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t lwe_chunk_size = 0);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 template <typename Torus>
 void cuda_cg_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
@@ -105,17 +74,14 @@ void cuda_cg_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, Torus *bootstrapping_key,
     pbs_buffer<Torus, MULTI_BIT> *pbs_buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
-    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
-    uint32_t num_luts, uint32_t lwe_idx, uint32_t max_shared_memory,
-    uint32_t lwe_chunk_size = 0);
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples);
 
-template <typename Torus, typename STorus>
+template <typename Torus>
 void scratch_cuda_multi_bit_programmable_bootstrap(
     void *stream, uint32_t gpu_index, pbs_buffer<Torus, MULTI_BIT> **pbs_buffer,
     uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
     uint32_t level_count, uint32_t grouping_factor,
-    uint32_t input_lwe_ciphertext_count, uint32_t max_shared_memory,
-    bool allocate_gpu_memory, uint32_t lwe_chunk_size = 0);
+    uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
 template <typename Torus>
 void cuda_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
@@ -124,45 +90,34 @@ void cuda_multi_bit_programmable_bootstrap_lwe_ciphertext_vector(
     Torus *lwe_array_in, Torus *lwe_input_indexes, Torus *bootstrapping_key,
     pbs_buffer<Torus, MULTI_BIT> *pbs_buffer, uint32_t lwe_dimension,
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t grouping_factor,
-    uint32_t base_log, uint32_t level_count, uint32_t num_samples,
-    uint32_t num_luts, uint32_t lwe_idx, uint32_t max_shared_memory,
-    uint32_t lwe_chunk_size = 0);
+    uint32_t base_log, uint32_t level_count, uint32_t num_samples);
 
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_multibit_programmable_bootstrap_keybundle(
+uint64_t get_buffer_size_full_sm_multibit_programmable_bootstrap_keybundle(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_multibit_programmable_bootstrap_step_one(
+uint64_t get_buffer_size_full_sm_multibit_programmable_bootstrap_step_one(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_multibit_programmable_bootstrap_step_two(
+uint64_t get_buffer_size_full_sm_multibit_programmable_bootstrap_step_two(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_partial_sm_multibit_programmable_bootstrap_step_one(
+uint64_t get_buffer_size_partial_sm_multibit_programmable_bootstrap_step_one(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_cg_multibit_programmable_bootstrap(
+uint64_t get_buffer_size_full_sm_cg_multibit_programmable_bootstrap(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_partial_sm_cg_multibit_programmable_bootstrap(
+uint64_t get_buffer_size_partial_sm_cg_multibit_programmable_bootstrap(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_sm_dsm_plus_tbc_multibit_programmable_bootstrap(
+uint64_t get_buffer_size_sm_dsm_plus_tbc_multibit_programmable_bootstrap(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_partial_sm_tbc_multibit_programmable_bootstrap(
+uint64_t get_buffer_size_partial_sm_tbc_multibit_programmable_bootstrap(
     uint32_t polynomial_size);
 template <typename Torus>
-__host__ __device__ uint64_t
-get_buffer_size_full_sm_tbc_multibit_programmable_bootstrap(
+uint64_t get_buffer_size_full_sm_tbc_multibit_programmable_bootstrap(
     uint32_t polynomial_size);
 
 template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::MULTI_BIT> {
@@ -332,8 +287,7 @@ template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::MULTI_BIT> {
 };
 
 template <typename Torus, class params>
-__host__ uint32_t get_lwe_chunk_size(uint32_t gpu_index, uint32_t max_num_pbs,
-                                     uint32_t polynomial_size,
-                                     uint32_t max_shared_memory);
+uint32_t get_lwe_chunk_size(uint32_t gpu_index, uint32_t max_num_pbs,
+                            uint32_t polynomial_size);
 
 #endif // CUDA_MULTI_BIT_H

backends/tfhe-cuda-backend/cuda/src/CMakeLists.txt

@@ -11,7 +11,7 @@ set(SOURCES
     ${CMAKE_SOURCE_DIR}/${INCLUDE_DIR}/linear_algebra.h
     ${CMAKE_SOURCE_DIR}/${INCLUDE_DIR}/shifts.h
     ${CMAKE_SOURCE_DIR}/${INCLUDE_DIR}/vertical_packing.h
-    ${CMAKE_SOURCE_DIR}/${INCLUDE_DIR}/helper.h)
+    ${CMAKE_SOURCE_DIR}/${INCLUDE_DIR}/helper_multi_gpu.h)
 file(GLOB_RECURSE SOURCES "*.cu")
 add_library(tfhe_cuda_backend STATIC ${SOURCES})
 set_target_properties(tfhe_cuda_backend PROPERTIES CUDA_SEPARABLE_COMPILATION ON CUDA_RESOLVE_DEVICE_SYMBOLS ON)

backends/tfhe-cuda-backend/cuda/src/crypto/ciphertext.cu

@@ -1,4 +1,5 @@
 #include "ciphertext.cuh"
+#include "polynomial/parameters.cuh"
 
 void cuda_convert_lwe_ciphertext_vector_to_gpu_64(void *stream,
                                                   uint32_t gpu_index,
@@ -19,3 +20,58 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu_64(void *stream,
       static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)dest,
       (uint64_t *)src, number_of_cts, lwe_dimension);
 }
+
+void cuda_glwe_sample_extract_64(void *stream, uint32_t gpu_index,
+                                 void *lwe_array_out, void *glwe_array_in,
+                                 uint32_t *nth_array, uint32_t num_glwes,
+                                 uint32_t glwe_dimension,
+                                 uint32_t polynomial_size) {
+
+  switch (polynomial_size) {
+  case 256:
+    host_sample_extract<uint64_t, AmortizedDegree<256>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 512:
+    host_sample_extract<uint64_t, AmortizedDegree<512>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 1024:
+    host_sample_extract<uint64_t, AmortizedDegree<1024>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 2048:
+    host_sample_extract<uint64_t, AmortizedDegree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 4096:
+    host_sample_extract<uint64_t, AmortizedDegree<4096>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 8192:
+    host_sample_extract<uint64_t, AmortizedDegree<8192>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  case 16384:
+    host_sample_extract<uint64_t, AmortizedDegree<16384>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
+        (uint64_t *)glwe_array_in, (uint32_t *)nth_array, num_glwes,
+        glwe_dimension);
+    break;
+  default:
+    PANIC("Cuda error: unsupported polynomial size. Supported "
+          "N's are powers of two in the interval [256..16384].")
+  }
+}

backends/tfhe-cuda-backend/cuda/src/crypto/ciphertext.cuh

@@ -3,6 +3,7 @@
 
 #include "ciphertext.h"
 #include "device.h"
+#include "polynomial/functions.cuh"
 #include <cstdint>
 
 template <typename T>
@@ -25,4 +26,39 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu(cudaStream_t stream,
   cuda_memcpy_async_to_cpu(dest, src, size, stream, gpu_index);
 }
 
+template <typename Torus, class params>
+__global__ void sample_extract(Torus *lwe_array_out, Torus *glwe_array_in,
+                               uint32_t *nth_array, uint32_t glwe_dimension) {
+
+  const int input_id = blockIdx.x;
+
+  const int glwe_input_size = (glwe_dimension + 1) * params::degree;
+  const int lwe_output_size = glwe_dimension * params::degree + 1;
+
+  auto lwe_out = lwe_array_out + input_id * lwe_output_size;
+
+  // We assume each GLWE will store the first polynomial_size inputs
+  uint32_t nth_per_glwe = params::degree;
+  auto glwe_in = glwe_array_in + (input_id / nth_per_glwe) * glwe_input_size;
+
+  auto nth = nth_array[input_id];
+
+  sample_extract_mask<Torus, params>(lwe_out, glwe_in, glwe_dimension, nth);
+  sample_extract_body<Torus, params>(lwe_out, glwe_in, glwe_dimension, nth);
+}
+
+template <typename Torus, class params>
+__host__ void host_sample_extract(cudaStream_t stream, uint32_t gpu_index,
+                                  Torus *lwe_array_out, Torus *glwe_array_in,
+                                  uint32_t *nth_array, uint32_t num_glwes,
+                                  uint32_t glwe_dimension) {
+  cudaSetDevice(gpu_index);
+
+  dim3 grid(num_glwes);
+  dim3 thds(params::degree / params::opt);
+  sample_extract<Torus, params><<<grid, thds, 0, stream>>>(
+      lwe_array_out, glwe_array_in, nth_array, glwe_dimension);
+  check_cuda_error(cudaGetLastError());
+}
+
 #endif

backends/tfhe-cuda-backend/cuda/src/crypto/keyswitch.cuh

@@ -3,8 +3,11 @@
 
 #include "device.h"
 #include "gadget.cuh"
+#include "helper_multi_gpu.h"
+#include "polynomial/functions.cuh"
 #include "polynomial/polynomial_math.cuh"
 #include "torus.cuh"
+#include "utils/kernel_dimensions.cuh"
 #include <thread>
 #include <vector>
 
@@ -31,43 +34,69 @@ __device__ Torus *get_ith_block(Torus *ksk, int i, int level,
  * scaling factor) under key s2 instead of s1, with an increased noise
  *
  */
+// Each thread in x are used to calculate one output.
+// threads in y are used to paralelize the lwe_dimension_in loop.
+// shared memory is used to store intermediate results of the reduction.
 template <typename Torus>
 __global__ void
-keyswitch(Torus *lwe_array_out, Torus *lwe_output_indexes, Torus *lwe_array_in,
-          Torus *lwe_input_indexes, Torus *ksk, uint32_t lwe_dimension_in,
+keyswitch(Torus *lwe_array_out, const Torus *__restrict__ lwe_output_indexes,
+          const Torus *__restrict__ lwe_array_in,
+          const Torus *__restrict__ lwe_input_indexes,
+          const Torus *__restrict__ ksk, uint32_t lwe_dimension_in,
           uint32_t lwe_dimension_out, uint32_t base_log, uint32_t level_count) {
-  int tid = threadIdx.x;
+  const int tid = threadIdx.x + blockIdx.x * blockDim.x;
+  const int shmem_index = threadIdx.x + threadIdx.y * blockDim.x;
+
   extern __shared__ int8_t sharedmem[];
+  Torus *lwe_acc_out = (Torus *)sharedmem;
+  auto block_lwe_array_out = get_chunk(
+      lwe_array_out, lwe_output_indexes[blockIdx.y], lwe_dimension_out + 1);
+
   if (tid <= lwe_dimension_out) {
-    Torus *local_lwe_array_out = (Torus *)sharedmem;
+
+    Torus local_lwe_out = 0;
     auto block_lwe_array_in = get_chunk(
-        lwe_array_in, lwe_input_indexes[blockIdx.x], lwe_dimension_in + 1);
-    auto block_lwe_array_out = get_chunk(
-        lwe_array_out, lwe_output_indexes[blockIdx.x], lwe_dimension_out + 1);
-    local_lwe_array_out[tid] = 0;
+        lwe_array_in, lwe_input_indexes[blockIdx.y], lwe_dimension_in + 1);
 
-    if (tid == lwe_dimension_out) {
-      local_lwe_array_out[lwe_dimension_out] =
-          block_lwe_array_in[lwe_dimension_in];
+    if (tid == lwe_dimension_out && threadIdx.y == 0) {
+      local_lwe_out = block_lwe_array_in[lwe_dimension_in];
     }
+    const Torus mask_mod_b = (1ll << base_log) - 1ll;
 
-    for (int i = 0; i < lwe_dimension_in; i++) {
+    const int pack_size = (lwe_dimension_in + blockDim.y - 1) / blockDim.y;
+    const int start_i = pack_size * threadIdx.y;
+    const int end_i = SEL(lwe_dimension_in, pack_size * (threadIdx.y + 1),
+                          pack_size * (threadIdx.y + 1) <= lwe_dimension_in);
+
+    // This loop distribution seems to benefit the global mem reads
+    for (int i = start_i; i < end_i; i++) {
       Torus a_i = round_to_closest_multiple(block_lwe_array_in[i], base_log,
                                             level_count);
       Torus state = a_i >> (sizeof(Torus) * 8 - base_log * level_count);
-      Torus mask_mod_b = (1ll << base_log) - 1ll;
+
       for (int j = 0; j < level_count; j++) {
         auto ksk_block =
             get_ith_block(ksk, i, j, lwe_dimension_out, level_count);
         Torus decomposed = decompose_one<Torus>(state, mask_mod_b, base_log);
-        local_lwe_array_out[tid] -= (Torus)ksk_block[tid] * decomposed;
+        local_lwe_out -= (Torus)ksk_block[tid] * decomposed;
       }
     }
-    block_lwe_array_out[tid] = local_lwe_array_out[tid];
+
+    lwe_acc_out[shmem_index] = local_lwe_out;
+  }
+
+  if (tid <= lwe_dimension_out) {
+    for (int offset = blockDim.y / 2; offset > 0 && threadIdx.y < offset;
+         offset /= 2) {
+      __syncthreads();
+      lwe_acc_out[shmem_index] +=
+          lwe_acc_out[shmem_index + offset * blockDim.x];
+    }
+    if (threadIdx.y == 0)
+      block_lwe_array_out[tid] = lwe_acc_out[shmem_index];
   }
 }
 
-/// assume lwe_array_in in the gpu
 template <typename Torus>
 __host__ void cuda_keyswitch_lwe_ciphertext_vector(
     cudaStream_t stream, uint32_t gpu_index, Torus *lwe_array_out,
@@ -76,15 +105,16 @@ __host__ void cuda_keyswitch_lwe_ciphertext_vector(
     uint32_t base_log, uint32_t level_count, uint32_t num_samples) {
 
   cudaSetDevice(gpu_index);
-  constexpr int ideal_threads = 1024;
-  if (lwe_dimension_out + 1 > ideal_threads)
-    PANIC("Cuda error (keyswitch): lwe dimension size out should be greater "
-          "or equal to the number of threads per block")
 
-  int lwe_size = lwe_dimension_out + 1;
-  int shared_mem = sizeof(Torus) * lwe_size;
-  dim3 grid(num_samples, 1, 1);
-  dim3 threads(ideal_threads, 1, 1);
+  constexpr int num_threads_y = 32;
+  int num_blocks, num_threads_x;
+
+  getNumBlocksAndThreads2D(lwe_dimension_out + 1, 512, num_threads_y,
+                           num_blocks, num_threads_x);
+
+  int shared_mem = sizeof(Torus) * num_threads_y * num_threads_x;
+  dim3 grid(num_blocks, num_samples, 1);
+  dim3 threads(num_threads_x, num_threads_y, 1);
 
   keyswitch<Torus><<<grid, threads, shared_mem, stream>>>(
       lwe_array_out, lwe_output_indexes, lwe_array_in, lwe_input_indexes, ksk,
@@ -92,4 +122,36 @@ __host__ void cuda_keyswitch_lwe_ciphertext_vector(
   check_cuda_error(cudaGetLastError());
 }
 
+template <typename Torus>
+void execute_keyswitch_async(cudaStream_t *streams, uint32_t *gpu_indexes,
+                             uint32_t gpu_count,
+                             const LweArrayVariant<Torus> &lwe_array_out,
+                             const LweArrayVariant<Torus> &lwe_output_indexes,
+                             const LweArrayVariant<Torus> &lwe_array_in,
+                             const LweArrayVariant<Torus> &lwe_input_indexes,
+                             Torus **ksks, uint32_t lwe_dimension_in,
+                             uint32_t lwe_dimension_out, uint32_t base_log,
+                             uint32_t level_count, uint32_t num_samples) {
+
+  /// If the number of radix blocks is lower than the number of GPUs, not all
+  /// GPUs will be active and there will be 1 input per GPU
+  for (uint i = 0; i < gpu_count; i++) {
+    int num_samples_on_gpu = get_num_inputs_on_gpu(num_samples, i, gpu_count);
+
+    Torus *current_lwe_array_out = GET_VARIANT_ELEMENT(lwe_array_out, i);
+    Torus *current_lwe_output_indexes =
+        GET_VARIANT_ELEMENT(lwe_output_indexes, i);
+    Torus *current_lwe_array_in = GET_VARIANT_ELEMENT(lwe_array_in, i);
+    Torus *current_lwe_input_indexes =
+        GET_VARIANT_ELEMENT(lwe_input_indexes, i);
+
+    // Compute Keyswitch
+    cuda_keyswitch_lwe_ciphertext_vector<Torus>(
+        streams[i], gpu_indexes[i], current_lwe_array_out,
+        current_lwe_output_indexes, current_lwe_array_in,
+        current_lwe_input_indexes, ksks[i], lwe_dimension_in, lwe_dimension_out,
+        base_log, level_count, num_samples_on_gpu);
+  }
+}
+
 #endif

backends/tfhe-cuda-backend/cuda/src/crypto/torus.cuh

@@ -39,36 +39,19 @@ __device__ inline T round_to_closest_multiple(T x, uint32_t base_log,
 }
 
 template <typename T>
-__device__ __forceinline__ void rescale_torus_element(T element, T &output,
-                                                      uint32_t log_shift) {
-  output =
-      round((double)element / (double(std::numeric_limits<T>::max()) + 1.0) *
-            (double)log_shift);
+__device__ __forceinline__ void modulus_switch(T input, T &output,
+                                               uint32_t log_modulus) {
+  constexpr uint32_t BITS = sizeof(T) * 8;
+
+  output = input + (((T)1) << (BITS - log_modulus - 1));
+  output >>= (BITS - log_modulus);
 }
 
 template <typename T>
-__device__ __forceinline__ T rescale_torus_element(T element,
-                                                   uint32_t log_shift) {
-  return round((double)element / (double(std::numeric_limits<T>::max()) + 1.0) *
-               (double)log_shift);
+__device__ __forceinline__ T modulus_switch(T input, uint32_t log_modulus) {
+  T output;
+  modulus_switch(input, output, log_modulus);
+  return output;
 }
 
-template <>
-__device__ __forceinline__ void
-rescale_torus_element<uint32_t>(uint32_t element, uint32_t &output,
-                                uint32_t log_shift) {
-  output =
-      round(__uint2double_rn(element) /
-            (__uint2double_rn(std::numeric_limits<uint32_t>::max()) + 1.0) *
-            __uint2double_rn(log_shift));
-}
-
-template <>
-__device__ __forceinline__ void
-rescale_torus_element<uint64_t>(uint64_t element, uint64_t &output,
-                                uint32_t log_shift) {
-  output = round(__ull2double_rn(element) /
-                 (__ull2double_rn(std::numeric_limits<uint64_t>::max()) + 1.0) *
-                 __uint2double_rn(log_shift));
-}
 #endif // CNCRT_TORUS_H

backends/tfhe-cuda-backend/cuda/src/device.cu

@@ -6,7 +6,7 @@
 cudaStream_t cuda_create_stream(uint32_t gpu_index) {
   check_cuda_error(cudaSetDevice(gpu_index));
   cudaStream_t stream;
-  check_cuda_error(cudaStreamCreate(&stream));
+  check_cuda_error(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
   return stream;
 }
 
@@ -47,9 +47,7 @@ void *cuda_malloc_async(uint64_t size, cudaStream_t stream,
       &support_async_alloc, cudaDevAttrMemoryPoolsSupported, gpu_index));
 
   if (support_async_alloc) {
-    cuda_synchronize_stream(stream, gpu_index);
     check_cuda_error(cudaMallocAsync((void **)&ptr, size, stream));
-    cuda_synchronize_stream(stream, gpu_index);
   } else {
     check_cuda_error(cudaMalloc((void **)&ptr, size));
   }
@@ -121,21 +119,22 @@ void cuda_memcpy_async_gpu_to_gpu(void *dest, void *src, uint64_t size,
     return;
   cudaPointerAttributes attr_dest;
   check_cuda_error(cudaPointerGetAttributes(&attr_dest, dest));
-  if (attr_dest.device != gpu_index && attr_dest.type != cudaMemoryTypeDevice) {
+  if (attr_dest.type != cudaMemoryTypeDevice) {
     PANIC("Cuda error: invalid dest device pointer in copy from GPU to GPU.")
   }
   cudaPointerAttributes attr_src;
   check_cuda_error(cudaPointerGetAttributes(&attr_src, src));
-  if (attr_src.device != gpu_index && attr_src.type != cudaMemoryTypeDevice) {
+  if (attr_src.type != cudaMemoryTypeDevice) {
     PANIC("Cuda error: invalid src device pointer in copy from GPU to GPU.")
   }
-  if (attr_src.device != attr_dest.device) {
-    PANIC("Cuda error: different devices specified in copy from GPU to GPU.")
-  }
-
   check_cuda_error(cudaSetDevice(gpu_index));
-  check_cuda_error(
-      cudaMemcpyAsync(dest, src, size, cudaMemcpyDeviceToDevice, stream));
+  if (attr_src.device == attr_dest.device) {
+    check_cuda_error(
+        cudaMemcpyAsync(dest, src, size, cudaMemcpyDeviceToDevice, stream));
+  } else {
+    check_cuda_error(cudaMemcpyPeerAsync(dest, attr_dest.device, src,
+                                         attr_src.device, size, stream));
+  }
 }
 
 /// Synchronizes device
@@ -167,19 +166,21 @@ __global__ void cuda_set_value_kernel(Torus *array, Torus value, Torus n) {
 template <typename Torus>
 void cuda_set_value_async(cudaStream_t stream, uint32_t gpu_index,
                           Torus *d_array, Torus value, Torus n) {
-  cudaPointerAttributes attr;
-  check_cuda_error(cudaPointerGetAttributes(&attr, d_array));
-  if (attr.type != cudaMemoryTypeDevice) {
-    PANIC("Cuda error: invalid dest device pointer in cuda set value.")
-  }
-  check_cuda_error(cudaSetDevice(gpu_index));
-  int block_size = 256;
-  int num_blocks = (n + block_size - 1) / block_size;
+  if (n > 0) {
+    cudaPointerAttributes attr;
+    check_cuda_error(cudaPointerGetAttributes(&attr, d_array));
+    if (attr.type != cudaMemoryTypeDevice) {
+      PANIC("Cuda error: invalid dest device pointer in cuda set value.")
+    }
+    check_cuda_error(cudaSetDevice(gpu_index));
+    int block_size = 256;
+    int num_blocks = (n + block_size - 1) / block_size;
 
-  // Launch the kernel
-  cuda_set_value_kernel<<<num_blocks, block_size, 0, stream>>>(d_array, value,
-                                                               n);
-  check_cuda_error(cudaGetLastError());
+    // Launch the kernel
+    cuda_set_value_kernel<<<num_blocks, block_size, 0, stream>>>(d_array, value,
+                                                                 n);
+    check_cuda_error(cudaGetLastError());
+  }
 }
 
 /// Explicitly instantiate cuda_set_value_async for 32 and 64 bits
@@ -242,22 +243,18 @@ void cuda_drop_async(void *ptr, cudaStream_t stream, uint32_t gpu_index) {
 
 /// Get the maximum size for the shared memory
 int cuda_get_max_shared_memory(uint32_t gpu_index) {
-  check_cuda_error(cudaSetDevice(gpu_index));
   int max_shared_memory = 0;
   cudaDeviceGetAttribute(&max_shared_memory, cudaDevAttrMaxSharedMemoryPerBlock,
                          gpu_index);
   check_cuda_error(cudaGetLastError());
+#if CUDA_ARCH == 900
+  max_shared_memory = 226000;
+#elif CUDA_ARCH == 890
+  max_shared_memory = 127000;
+#elif CUDA_ARCH == 800
+  max_shared_memory = 163000;
+#elif CUDA_ARCH == 700
+  max_shared_memory = 95000;
+#endif
   return max_shared_memory;
 }
-
-void cuda_stream_add_callback(cudaStream_t stream, uint32_t gpu_index,
-                              cudaStreamCallback_t callback, void *user_data) {
-
-  check_cuda_error(cudaSetDevice(gpu_index));
-  check_cuda_error(cudaStreamAddCallback(stream, callback, user_data, 0));
-}
-
-void host_free_on_stream_callback(cudaStream_t stream, cudaError_t status,
-                                  void *host_pointer) {
-  free(host_pointer);
-}

backends/tfhe-cuda-backend/cuda/src/fft/bnsmfft.cuh

@@ -294,9 +294,6 @@ template <class params> __device__ void NSMFFT_direct(double2 *A) {
     __syncthreads();
   }
 
-  // compressed size = 8192 is actual polynomial size = 16384.
-  // from this size, twiddles can't fit in constant memory,
-  // so from here, butterfly operation access device memory.
   if constexpr (params::degree >= 8192) {
     // level 13
     tid = threadIdx.x;
@@ -307,7 +304,7 @@ template <class params> __device__ void NSMFFT_direct(double2 *A) {
            (tid & (params::degree / 8192 - 1));
       i2 = i1 + params::degree / 8192;
 
-      w = negtwiddles13[twid_id];
+      w = negtwiddles[twid_id + 4096];
       u = A[i1];
       v = A[i2] * w;
 
@@ -351,10 +348,6 @@ template <class params> __device__ void NSMFFT_inverse(double2 *A) {
   // mapping in backward fft is reversed
   // butterfly operation is started from last level
 
-  // compressed size = 8192 is actual polynomial size = 16384.
-  // twiddles for this size can't fit in constant memory so
-  // butterfly operation for this level access device memory to fetch
-  // twiddles
   if constexpr (params::degree >= 8192) {
     // level 13
     tid = threadIdx.x;
@@ -365,7 +358,7 @@ template <class params> __device__ void NSMFFT_inverse(double2 *A) {
            (tid & (params::degree / 8192 - 1));
       i2 = i1 + params::degree / 8192;
 
-      w = negtwiddles13[twid_id];
+      w = negtwiddles[twid_id + 4096];
       u = A[i1] - A[i2];
 
       A[i1] += A[i2];

backends/tfhe-cuda-backend/cuda/src/fft/twiddles.cu

@@ -1,6 +1,6 @@
 #include "cuComplex.h"
 
-__constant__ double2 negtwiddles[4096] = {
+__device__ double2 negtwiddles[8192] = {
     {0, 0},
     {0.707106781186547461715008466854, 0.707106781186547572737310929369},
     {0.92387953251128673848313610506, 0.382683432365089781779232680492},
@@ -4096,9 +4096,7 @@ __constant__ double2 negtwiddles[4096] = {
     {0.70791982920081630847874976098, 0.706292797233758484765075991163},
     {-0.706292797233758484765075991163, 0.70791982920081630847874976098},
     {0.00115048533711384847431913325266, 0.99999933819152553304832053982},
-    {-0.99999933819152553304832053982, 0.00115048533711384847431913325266}};
-
-__device__ double2 negtwiddles13[4096] = {
+    {-0.99999933819152553304832053982, 0.00115048533711384847431913325266},
     {0.999999981616429334252416083473, 0.000191747597310703291528452552051},
     {-0.000191747597310703291528452552051, 0.999999981616429334252416083473},
     {0.706971182161065359039753275283, 0.707242354213734603085583785287},

backends/tfhe-cuda-backend/cuda/src/fft/twiddles.cuh

@@ -2,12 +2,7 @@
 #define GPU_BOOTSTRAP_TWIDDLES_CUH
 
 /*
- * 'negtwiddles' are stored in constant memory for faster access times
- * because of it's limited size, only twiddles for up to 2^12 polynomial size
- * can be stored there, twiddles for 2^13 are stored in device memory
- * 'negtwiddles13'
+ * 'negtwiddles' are stored in device memory to profit caching
  */
-
-extern __constant__ double2 negtwiddles[4096];
-extern __device__ double2 negtwiddles13[4096];
+extern __device__ double2 negtwiddles[8192];
 #endif

backends/tfhe-cuda-backend/cuda/src/integer/addition.cu

@@ -0,0 +1,49 @@
+#include "integer/addition.cuh"
+
+void scratch_cuda_signed_overflowing_add_or_sub_radix_ciphertext_kb_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t num_blocks, int8_t signed_operation,
+    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
+    bool allocate_gpu_memory) {
+
+  SIGNED_OPERATION op = (signed_operation == 1) ? SIGNED_OPERATION::ADDITION
+                                                : SIGNED_OPERATION::SUBTRACTION;
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          big_lwe_dimension, small_lwe_dimension, ks_level,
+                          ks_base_log, pbs_level, pbs_base_log, grouping_factor,
+                          message_modulus, carry_modulus);
+
+  scratch_cuda_integer_signed_overflowing_add_or_sub_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      (int_signed_overflowing_add_or_sub_memory<uint64_t> **)mem_ptr,
+      num_blocks, op, params, allocate_gpu_memory);
+}
+
+void cuda_signed_overflowing_add_or_sub_radix_ciphertext_kb_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, void *lhs,
+    void *rhs, void *overflowed, int8_t signed_operation, int8_t *mem_ptr,
+    void **bsks, void **ksks, uint32_t num_blocks) {
+
+  auto mem = (int_signed_overflowing_add_or_sub_memory<uint64_t> *)mem_ptr;
+  SIGNED_OPERATION op = (signed_operation == 1) ? SIGNED_OPERATION::ADDITION
+                                                : SIGNED_OPERATION::SUBTRACTION;
+
+  host_integer_signed_overflowing_add_or_sub_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(lhs), static_cast<uint64_t *>(rhs),
+      static_cast<uint64_t *>(overflowed), op, bsks, (uint64_t **)(ksks), mem,
+      num_blocks);
+}
+
+void cleanup_signed_overflowing_add_or_sub(void **streams,
+                                           uint32_t *gpu_indexes,
+                                           uint32_t gpu_count,
+                                           int8_t **mem_ptr_void) {
+  int_signed_overflowing_add_or_sub_memory<uint64_t> *mem_ptr =
+      (int_signed_overflowing_add_or_sub_memory<uint64_t> *)(*mem_ptr_void);
+
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
+}

backends/tfhe-cuda-backend/cuda/src/integer/addition.cuh

@@ -0,0 +1,137 @@
+#ifndef TFHE_RS_ADDITION_CUH
+#define TFHE_RS_ADDITION_CUH
+
+#include "crypto/keyswitch.cuh"
+#include "device.h"
+#include "integer.h"
+#include "integer/comparison.cuh"
+#include "integer/integer.cuh"
+#include "integer/negation.cuh"
+#include "integer/scalar_shifts.cuh"
+#include "linear_algebra.h"
+#include "programmable_bootstrap.h"
+#include "utils/helper.cuh"
+#include "utils/kernel_dimensions.cuh"
+#include <fstream>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <vector>
+
+template <typename Torus>
+void host_resolve_signed_overflow(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *result, Torus *last_block_inner_propagation,
+    Torus *last_block_input_carry, Torus *last_block_output_carry,
+    int_resolve_signed_overflow_memory<Torus> *mem, void **bsks, Torus **ksks) {
+
+  auto x = mem->x;
+
+  Torus *d_clears =
+      (Torus *)cuda_malloc_async(sizeof(Torus), streams[0], gpu_indexes[0]);
+
+  cuda_set_value_async<Torus>(streams[0], gpu_indexes[0], d_clears, 2, 1);
+
+  // replace with host function call
+  cuda_mult_lwe_ciphertext_vector_cleartext_vector_64(
+      streams[0], gpu_indexes[0], x, last_block_output_carry, d_clears,
+      mem->params.big_lwe_dimension, 1);
+
+  host_addition(streams[0], gpu_indexes[0], last_block_inner_propagation,
+                last_block_inner_propagation, x, mem->params.big_lwe_dimension,
+                1);
+  host_addition(streams[0], gpu_indexes[0], last_block_inner_propagation,
+                last_block_inner_propagation, last_block_input_carry,
+                mem->params.big_lwe_dimension, 1);
+
+  host_apply_univariate_lut_kb<Torus>(streams, gpu_indexes, gpu_count, result,
+                                      last_block_inner_propagation,
+                                      mem->resolve_overflow_lut, ksks, bsks, 1);
+
+  cuda_drop_async(d_clears, streams[0], gpu_indexes[0]);
+}
+
+template <typename Torus>
+__host__ void scratch_cuda_integer_signed_overflowing_add_or_sub_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_signed_overflowing_add_or_sub_memory<Torus> **mem_ptr,
+    uint32_t num_blocks, SIGNED_OPERATION op, int_radix_params params,
+    bool allocate_gpu_memory) {
+
+  *mem_ptr = new int_signed_overflowing_add_or_sub_memory<Torus>(
+      streams, gpu_indexes, gpu_count, params, num_blocks, op,
+      allocate_gpu_memory);
+}
+
+/*
+ * Addition - signed_operation = 1
+ * Subtraction - signed_operation = -1
+ */
+template <typename Torus>
+__host__ void host_integer_signed_overflowing_add_or_sub_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *lhs, Torus *rhs, Torus *overflowed, SIGNED_OPERATION op, void **bsks,
+    uint64_t **ksks,
+    int_signed_overflowing_add_or_sub_memory<uint64_t> *mem_ptr,
+    uint32_t num_blocks) {
+
+  auto radix_params = mem_ptr->params;
+
+  uint32_t big_lwe_dimension = radix_params.big_lwe_dimension;
+  uint32_t big_lwe_size = big_lwe_dimension + 1;
+  uint32_t big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
+
+  assert(radix_params.message_modulus >= 4 && radix_params.carry_modulus >= 4);
+
+  auto result = mem_ptr->result;
+  auto neg_rhs = mem_ptr->neg_rhs;
+  auto input_carries = mem_ptr->input_carries;
+  auto output_carry = mem_ptr->output_carry;
+  auto last_block_inner_propagation = mem_ptr->last_block_inner_propagation;
+
+  cuda_memcpy_async_gpu_to_gpu(result, lhs, num_blocks * big_lwe_size_bytes,
+                               streams[0], gpu_indexes[0]);
+
+  // phase 1
+  if (op == SIGNED_OPERATION::ADDITION) {
+    host_addition(streams[0], gpu_indexes[0], result, lhs, rhs,
+                  big_lwe_dimension, num_blocks);
+  } else {
+    host_integer_radix_negation(
+        streams, gpu_indexes, gpu_count, neg_rhs, rhs, big_lwe_dimension,
+        num_blocks, radix_params.message_modulus, radix_params.carry_modulus);
+    host_addition(streams[0], gpu_indexes[0], result, lhs, neg_rhs,
+                  big_lwe_dimension, num_blocks);
+  }
+
+  // phase 2
+  for (uint j = 0; j < gpu_count; j++) {
+    cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+  }
+
+  host_propagate_single_carry(mem_ptr->sub_streams_1, gpu_indexes, gpu_count,
+                              result, output_carry, input_carries,
+                              mem_ptr->scp_mem, bsks, ksks, num_blocks);
+  host_generate_last_block_inner_propagation(
+      mem_ptr->sub_streams_2, gpu_indexes, gpu_count,
+      last_block_inner_propagation, &lhs[(num_blocks - 1) * big_lwe_size],
+      &rhs[(num_blocks - 1) * big_lwe_size], mem_ptr->las_block_prop_mem, bsks,
+      ksks);
+
+  for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+    cuda_synchronize_stream(mem_ptr->sub_streams_1[j], gpu_indexes[j]);
+    cuda_synchronize_stream(mem_ptr->sub_streams_2[j], gpu_indexes[j]);
+  }
+
+  // phase 3
+  auto input_carry = &input_carries[(num_blocks - 1) * big_lwe_size];
+
+  host_resolve_signed_overflow(
+      streams, gpu_indexes, gpu_count, overflowed, last_block_inner_propagation,
+      input_carry, output_carry, mem_ptr->resolve_overflow_mem, bsks, ksks);
+
+  cuda_memcpy_async_gpu_to_gpu(lhs, result, num_blocks * big_lwe_size_bytes,
+                               streams[0], gpu_indexes[0]);
+}
+
+#endif // TFHE_RS_ADDITION_CUH

backends/tfhe-cuda-backend/cuda/src/integer/bitwise_ops.cu

@@ -1,13 +1,13 @@
 #include "integer/bitwise_ops.cuh"
 
 void scratch_cuda_integer_radix_bitop_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, BITOP_TYPE op_type,
-    bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t lwe_ciphertext_count,
+    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
+    BITOP_TYPE op_type, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -15,7 +15,7 @@ void scratch_cuda_integer_radix_bitop_kb_64(
                           message_modulus, carry_modulus);
 
   scratch_cuda_integer_radix_bitop_kb<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_bitop_buffer<uint64_t> **)mem_ptr, lwe_ciphertext_count, params,
       op_type, allocate_gpu_memory);
 }
@@ -23,34 +23,21 @@ void scratch_cuda_integer_radix_bitop_kb_64(
 void cuda_bitop_integer_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *lwe_array_out, void *lwe_array_1, void *lwe_array_2, int8_t *mem_ptr,
-    void *bsk, void *ksk, uint32_t lwe_ciphertext_count) {
+    void **bsks, void **ksks, uint32_t lwe_ciphertext_count) {
 
   host_integer_radix_bitop_kb<uint64_t>(
       (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       static_cast<uint64_t *>(lwe_array_out),
       static_cast<uint64_t *>(lwe_array_1),
       static_cast<uint64_t *>(lwe_array_2),
-      (int_bitop_buffer<uint64_t> *)mem_ptr, bsk, static_cast<uint64_t *>(ksk),
+      (int_bitop_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks),
       lwe_ciphertext_count);
 }
 
-void cuda_bitnot_integer_radix_ciphertext_kb_64(
-    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void *lwe_array_in, int8_t *mem_ptr, void *bsk,
-    void *ksk, uint32_t lwe_ciphertext_count) {
-
-  host_integer_radix_bitnot_kb<uint64_t>(
-      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-      static_cast<uint64_t *>(lwe_array_out),
-      static_cast<uint64_t *>(lwe_array_in),
-      (int_bitop_buffer<uint64_t> *)mem_ptr, bsk, static_cast<uint64_t *>(ksk),
-      lwe_ciphertext_count);
-}
-
-void cleanup_cuda_integer_bitop(void *stream, uint32_t gpu_index,
-                                int8_t **mem_ptr_void) {
+void cleanup_cuda_integer_bitop(void **streams, uint32_t *gpu_indexes,
+                                uint32_t gpu_count, int8_t **mem_ptr_void) {
 
   int_bitop_buffer<uint64_t> *mem_ptr =
       (int_bitop_buffer<uint64_t> *)(*mem_ptr_void);
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/bitwise_ops.cuh

@@ -16,38 +16,25 @@ __host__ void
 host_integer_radix_bitop_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                             uint32_t gpu_count, Torus *lwe_array_out,
                             Torus *lwe_array_1, Torus *lwe_array_2,
-                            int_bitop_buffer<Torus> *mem_ptr, void *bsk,
-                            Torus *ksk, uint32_t num_radix_blocks) {
+                            int_bitop_buffer<Torus> *mem_ptr, void **bsks,
+                            Torus **ksks, uint32_t num_radix_blocks) {
 
   auto lut = mem_ptr->lut;
 
   integer_radix_apply_bivariate_lookup_table_kb<Torus>(
       streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_1, lwe_array_2,
-      bsk, ksk, num_radix_blocks, lut, lut->params.message_modulus);
-}
-
-template <typename Torus>
-__host__ void host_integer_radix_bitnot_kb(
-    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    Torus *lwe_array_out, Torus *lwe_array_in, int_bitop_buffer<Torus> *mem_ptr,
-    void *bsk, Torus *ksk, uint32_t num_radix_blocks) {
-
-  auto lut = mem_ptr->lut;
-
-  integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in, bsk, ksk,
-      num_radix_blocks, lut);
+      bsks, ksks, num_radix_blocks, lut, lut->params.message_modulus);
 }
 
 template <typename Torus>
 __host__ void scratch_cuda_integer_radix_bitop_kb(
-    cudaStream_t stream, uint32_t gpu_index, int_bitop_buffer<Torus> **mem_ptr,
-    uint32_t num_radix_blocks, int_radix_params params, BITOP_TYPE op,
-    bool allocate_gpu_memory) {
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_bitop_buffer<Torus> **mem_ptr, uint32_t num_radix_blocks,
+    int_radix_params params, BITOP_TYPE op, bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
-  *mem_ptr = new int_bitop_buffer<Torus>(stream, gpu_index, op, params,
-                                         num_radix_blocks, allocate_gpu_memory);
+  *mem_ptr =
+      new int_bitop_buffer<Torus>(streams, gpu_indexes, gpu_count, op, params,
+                                  num_radix_blocks, allocate_gpu_memory);
 }
 
 #endif

backends/tfhe-cuda-backend/cuda/src/integer/cmux.cu

@@ -1,12 +1,13 @@
 #include "integer/cmux.cuh"
 
 void scratch_cuda_integer_radix_cmux_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t lwe_ciphertext_count, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t lwe_ciphertext_count,
+    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
+    bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -17,7 +18,7 @@ void scratch_cuda_integer_radix_cmux_kb_64(
       [](uint64_t x) -> uint64_t { return x == 1; };
 
   scratch_cuda_integer_radix_cmux_kb(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_cmux_buffer<uint64_t> **)mem_ptr, predicate_lut_f,
       lwe_ciphertext_count, params, allocate_gpu_memory);
 }
@@ -25,7 +26,7 @@ void scratch_cuda_integer_radix_cmux_kb_64(
 void cuda_cmux_integer_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *lwe_array_out, void *lwe_condition, void *lwe_array_true,
-    void *lwe_array_false, int8_t *mem_ptr, void *bsk, void *ksk,
+    void *lwe_array_false, int8_t *mem_ptr, void **bsks, void **ksks,
     uint32_t lwe_ciphertext_count) {
 
   host_integer_radix_cmux_kb<uint64_t>(
@@ -34,15 +35,16 @@ void cuda_cmux_integer_radix_ciphertext_kb_64(
       static_cast<uint64_t *>(lwe_condition),
       static_cast<uint64_t *>(lwe_array_true),
       static_cast<uint64_t *>(lwe_array_false),
-      (int_cmux_buffer<uint64_t> *)mem_ptr, bsk, static_cast<uint64_t *>(ksk),
+      (int_cmux_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks),
 
       lwe_ciphertext_count);
 }
 
-void cleanup_cuda_integer_radix_cmux(void *stream, uint32_t gpu_index,
+void cleanup_cuda_integer_radix_cmux(void **streams, uint32_t *gpu_indexes,
+                                     uint32_t gpu_count,
                                      int8_t **mem_ptr_void) {
 
   int_cmux_buffer<uint64_t> *mem_ptr =
       (int_cmux_buffer<uint64_t> *)(*mem_ptr_void);
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/cmux.cuh

@@ -2,15 +2,14 @@
 #define CUDA_INTEGER_CMUX_CUH
 
 #include "integer.cuh"
-#include <omp.h>
 
 template <typename Torus>
 __host__ void zero_out_if(cudaStream_t *streams, uint32_t *gpu_indexes,
                           uint32_t gpu_count, Torus *lwe_array_out,
                           Torus *lwe_array_input, Torus *lwe_condition,
                           int_zero_out_if_buffer<Torus> *mem_ptr,
-                          int_radix_lut<Torus> *predicate, void *bsk,
-                          Torus *ksk, uint32_t num_radix_blocks) {
+                          int_radix_lut<Torus> *predicate, void **bsks,
+                          Torus **ksks, uint32_t num_radix_blocks) {
   cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
 
@@ -36,46 +35,41 @@ __host__ void zero_out_if(cudaStream_t *streams, uint32_t *gpu_indexes,
   }
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, tmp_lwe_array_input, bsk,
-      ksk, num_radix_blocks, predicate);
+      streams, gpu_indexes, gpu_count, lwe_array_out, tmp_lwe_array_input, bsks,
+      ksks, num_radix_blocks, predicate);
 }
 
 template <typename Torus>
 __host__ void host_integer_radix_cmux_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_condition, Torus *lwe_array_true,
-    Torus *lwe_array_false, int_cmux_buffer<Torus> *mem_ptr, void *bsk,
-    Torus *ksk, uint32_t num_radix_blocks) {
+    Torus *lwe_array_false, int_cmux_buffer<Torus> *mem_ptr, void **bsks,
+    Torus **ksks, uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
 
   // Since our CPU threads will be working on different streams we shall assert
   // the work in the main stream is completed
-  cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-  auto true_stream = mem_ptr->zero_if_true_buffer->local_stream;
-  auto false_stream = mem_ptr->zero_if_false_buffer->local_stream;
-
-#pragma omp parallel sections
-  {
-    // Both sections may be executed in parallel
-#pragma omp section
-    {
-      auto mem_true = mem_ptr->zero_if_true_buffer;
-      zero_out_if(&true_stream, gpu_indexes, gpu_count, mem_ptr->tmp_true_ct,
-                  lwe_array_true, lwe_condition, mem_true,
-                  mem_ptr->inverted_predicate_lut, bsk, ksk, num_radix_blocks);
-    }
-#pragma omp section
-    {
-      auto mem_false = mem_ptr->zero_if_false_buffer;
-      zero_out_if(&false_stream, gpu_indexes, gpu_count, mem_ptr->tmp_false_ct,
-                  lwe_array_false, lwe_condition, mem_false,
-                  mem_ptr->predicate_lut, bsk, ksk, num_radix_blocks);
-    }
+  auto true_streams = mem_ptr->zero_if_true_buffer->true_streams;
+  auto false_streams = mem_ptr->zero_if_false_buffer->false_streams;
+  for (uint j = 0; j < gpu_count; j++) {
+    cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+  }
+
+  auto mem_true = mem_ptr->zero_if_true_buffer;
+  zero_out_if(true_streams, gpu_indexes, gpu_count, mem_ptr->tmp_true_ct,
+              lwe_array_true, lwe_condition, mem_true,
+              mem_ptr->inverted_predicate_lut, bsks, ksks, num_radix_blocks);
+  auto mem_false = mem_ptr->zero_if_false_buffer;
+  zero_out_if(false_streams, gpu_indexes, gpu_count, mem_ptr->tmp_false_ct,
+              lwe_array_false, lwe_condition, mem_false, mem_ptr->predicate_lut,
+              bsks, ksks, num_radix_blocks);
+  for (uint j = 0; j < mem_ptr->zero_if_true_buffer->active_gpu_count; j++) {
+    cuda_synchronize_stream(true_streams[j], gpu_indexes[j]);
+  }
+  for (uint j = 0; j < mem_ptr->zero_if_false_buffer->active_gpu_count; j++) {
+    cuda_synchronize_stream(false_streams[j], gpu_indexes[j]);
   }
-  cuda_synchronize_stream(true_stream, gpu_indexes[0]);
-  cuda_synchronize_stream(false_stream, gpu_indexes[0]);
 
   // If the condition was true, true_ct will have kept its value and false_ct
   // will be 0 If the condition was false, true_ct will be 0 and false_ct will
@@ -86,19 +80,19 @@ __host__ void host_integer_radix_cmux_kb(
                 num_radix_blocks);
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, added_cts, bsk, ksk,
+      streams, gpu_indexes, gpu_count, lwe_array_out, added_cts, bsks, ksks,
       num_radix_blocks, mem_ptr->message_extract_lut);
 }
 
 template <typename Torus>
 __host__ void scratch_cuda_integer_radix_cmux_kb(
-    cudaStream_t stream, uint32_t gpu_index, int_cmux_buffer<Torus> **mem_ptr,
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_cmux_buffer<Torus> **mem_ptr,
     std::function<Torus(Torus)> predicate_lut_f, uint32_t num_radix_blocks,
     int_radix_params params, bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
-  *mem_ptr =
-      new int_cmux_buffer<Torus>(stream, gpu_index, predicate_lut_f, params,
-                                 num_radix_blocks, allocate_gpu_memory);
+  *mem_ptr = new int_cmux_buffer<Torus>(streams, gpu_indexes, gpu_count,
+                                        predicate_lut_f, params,
+                                        num_radix_blocks, allocate_gpu_memory);
 }
 #endif

backends/tfhe-cuda-backend/cuda/src/integer/comparison.cu

@@ -1,13 +1,13 @@
 #include "integer/comparison.cuh"
 
 void scratch_cuda_integer_radix_comparison_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t num_radix_blocks, uint32_t message_modulus, uint32_t carry_modulus,
-    PBS_TYPE pbs_type, COMPARISON_TYPE op_type, bool is_signed,
-    bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t num_radix_blocks,
+    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
+    COMPARISON_TYPE op_type, bool is_signed, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -18,7 +18,7 @@ void scratch_cuda_integer_radix_comparison_kb_64(
   case EQ:
   case NE:
     scratch_cuda_integer_radix_comparison_check_kb<uint64_t>(
-        static_cast<cudaStream_t>(stream), gpu_index,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         (int_comparison_buffer<uint64_t> **)mem_ptr, num_radix_blocks, params,
         op_type, false, allocate_gpu_memory);
     break;
@@ -29,7 +29,7 @@ void scratch_cuda_integer_radix_comparison_kb_64(
   case MAX:
   case MIN:
     scratch_cuda_integer_radix_comparison_check_kb<uint64_t>(
-        static_cast<cudaStream_t>(stream), gpu_index,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         (int_comparison_buffer<uint64_t> **)mem_ptr, num_radix_blocks, params,
         op_type, is_signed, allocate_gpu_memory);
     break;
@@ -39,7 +39,7 @@ void scratch_cuda_integer_radix_comparison_kb_64(
 void cuda_comparison_integer_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *lwe_array_out, void *lwe_array_1, void *lwe_array_2, int8_t *mem_ptr,
-    void *bsk, void *ksk, uint32_t num_radix_blocks) {
+    void **bsks, void **ksks, uint32_t num_radix_blocks) {
 
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;
@@ -50,8 +50,8 @@ void cuda_comparison_integer_radix_ciphertext_kb_64(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_1),
-        static_cast<uint64_t *>(lwe_array_2), buffer, bsk,
-        static_cast<uint64_t *>(ksk), num_radix_blocks);
+        static_cast<uint64_t *>(lwe_array_2), buffer, bsks, (uint64_t **)(ksks),
+        num_radix_blocks);
     break;
   case GT:
   case GE:
@@ -62,7 +62,7 @@ void cuda_comparison_integer_radix_ciphertext_kb_64(
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_1),
         static_cast<uint64_t *>(lwe_array_2), buffer,
-        buffer->diff_buffer->operator_f, bsk, static_cast<uint64_t *>(ksk),
+        buffer->diff_buffer->operator_f, bsks, (uint64_t **)(ksks),
         num_radix_blocks);
     break;
   case MAX:
@@ -71,18 +71,19 @@ void cuda_comparison_integer_radix_ciphertext_kb_64(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_1),
-        static_cast<uint64_t *>(lwe_array_2), buffer, bsk,
-        static_cast<uint64_t *>(ksk), num_radix_blocks);
+        static_cast<uint64_t *>(lwe_array_2), buffer, bsks, (uint64_t **)(ksks),
+        num_radix_blocks);
     break;
   default:
     PANIC("Cuda error: integer operation not supported")
   }
 }
 
-void cleanup_cuda_integer_comparison(void *stream, uint32_t gpu_index,
+void cleanup_cuda_integer_comparison(void **streams, uint32_t *gpu_indexes,
+                                     uint32_t gpu_count,
                                      int8_t **mem_ptr_void) {
 
   int_comparison_buffer<uint64_t> *mem_ptr =
       (int_comparison_buffer<uint64_t> *)(*mem_ptr_void);
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/comparison.cuh

@@ -56,12 +56,11 @@ __host__ void accumulate_all_blocks(cudaStream_t stream, uint32_t gpu_index,
  *
  */
 template <typename Torus>
-__host__ void
-are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
-                               uint32_t gpu_count, Torus *lwe_array_out,
-                               Torus *lwe_array_in,
-                               int_comparison_buffer<Torus> *mem_ptr, void *bsk,
-                               Torus *ksk, uint32_t num_radix_blocks) {
+__host__ void are_all_comparisons_block_true(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *lwe_array_out, Torus *lwe_array_in,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
+    uint32_t num_radix_blocks) {
 
   cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
@@ -94,6 +93,8 @@ are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
     // as in the worst case we will be adding `max_value` ones
     auto input_blocks = tmp_out;
     auto accumulator = are_all_block_true_buffer->tmp_block_accumulated;
+    auto is_equal_to_num_blocks_map =
+        &are_all_block_true_buffer->is_equal_to_lut_map;
     for (int i = 0; i < num_chunks; i++) {
       accumulate_all_blocks(streams[0], gpu_indexes[0], accumulator,
                             input_blocks, big_lwe_dimension, chunk_length);
@@ -103,8 +104,6 @@ are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
       input_blocks += (big_lwe_dimension + 1) * chunk_length;
     }
     accumulator = are_all_block_true_buffer->tmp_block_accumulated;
-    auto is_equal_to_num_blocks_map =
-        &are_all_block_true_buffer->is_equal_to_lut_map;
 
     // Selects a LUT
     int_radix_lut<Torus> *lut;
@@ -119,7 +118,7 @@ are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
       } else {
         // LUT needs to be computed
         auto new_lut =
-            new int_radix_lut<Torus>(streams[0], gpu_indexes[0], params,
+            new int_radix_lut<Torus>(streams, gpu_indexes, gpu_count, params,
                                      max_value, num_radix_blocks, true);
 
         auto is_equal_to_num_blocks_lut_f = [max_value,
@@ -127,10 +126,12 @@ are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
           return (x & max_value) == chunk_length;
         };
         generate_device_accumulator<Torus>(
-            streams[0], gpu_indexes[0], new_lut->lut, glwe_dimension,
-            polynomial_size, message_modulus, carry_modulus,
+            streams[0], gpu_indexes[0], new_lut->get_lut(gpu_indexes[0], 0),
+            glwe_dimension, polynomial_size, message_modulus, carry_modulus,
             is_equal_to_num_blocks_lut_f);
 
+        new_lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
+
         (*is_equal_to_num_blocks_map)[chunk_length] = new_lut;
         lut = new_lut;
       }
@@ -140,12 +141,12 @@ are_all_comparisons_block_true(cudaStream_t *streams, uint32_t *gpu_indexes,
     if (remaining_blocks == 1) {
       // In the last iteration we copy the output to the final address
       integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsk, ksk,
-          1, lut);
+          streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsks,
+          ksks, 1, lut);
       return;
     } else {
       integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, tmp_out, accumulator, bsk, ksk,
+          streams, gpu_indexes, gpu_count, tmp_out, accumulator, bsks, ksks,
           num_chunks, lut);
     }
   }
@@ -161,7 +162,7 @@ template <typename Torus>
 __host__ void is_at_least_one_comparisons_block_true(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t num_radix_blocks) {
 
   cudaSetDevice(gpu_indexes[0]);
@@ -207,13 +208,13 @@ __host__ void is_at_least_one_comparisons_block_true(
     if (remaining_blocks == 1) {
       // In the last iteration we copy the output to the final address
       integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsk, ksk,
-          1, lut);
+          streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsks,
+          ksks, 1, lut);
       return;
     } else {
       integer_radix_apply_univariate_lookup_table_kb<Torus>(
           streams, gpu_indexes, gpu_count, mem_ptr->tmp_lwe_array_out,
-          accumulator, bsk, ksk, num_chunks, lut);
+          accumulator, bsks, ksks, num_chunks, lut);
     }
   }
 }
@@ -241,10 +242,9 @@ template <typename Torus>
 __host__ void host_compare_with_zero_equality(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     int32_t num_radix_blocks, int_radix_lut<Torus> *zero_comparison) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -293,27 +293,26 @@ __host__ void host_compare_with_zero_equality(
   }
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, sum, sum, bsk, ksk, num_sum_blocks,
+      streams, gpu_indexes, gpu_count, sum, sum, bsks, ksks, num_sum_blocks,
       zero_comparison);
   are_all_comparisons_block_true(streams, gpu_indexes, gpu_count, lwe_array_out,
-                                 sum, mem_ptr, bsk, ksk, num_sum_blocks);
+                                 sum, mem_ptr, bsks, ksks, num_sum_blocks);
 }
 
 template <typename Torus>
 __host__ void host_integer_radix_equality_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_1, Torus *lwe_array_2,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto eq_buffer = mem_ptr->eq_buffer;
 
   // Applies the LUT for the comparison operation
   auto comparisons = mem_ptr->tmp_block_comparisons;
   integer_radix_apply_bivariate_lookup_table_kb(
       streams, gpu_indexes, gpu_count, comparisons, lwe_array_1, lwe_array_2,
-      bsk, ksk, num_radix_blocks, eq_buffer->operator_lut,
+      bsks, ksks, num_radix_blocks, eq_buffer->operator_lut,
       eq_buffer->operator_lut->params.message_modulus);
 
   // This takes a Vec of blocks, where each block is either 0 or 1.
@@ -321,7 +320,7 @@ __host__ void host_integer_radix_equality_check_kb(
   // It returns a block encrypting 1 if all input blocks are 1
   // otherwise the block encrypts 0
   are_all_comparisons_block_true(streams, gpu_indexes, gpu_count, lwe_array_out,
-                                 comparisons, mem_ptr, bsk, ksk,
+                                 comparisons, mem_ptr, bsks, ksks,
                                  num_radix_blocks);
 }
 
@@ -330,10 +329,9 @@ __host__ void
 compare_radix_blocks_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                         uint32_t gpu_count, Torus *lwe_array_out,
                         Torus *lwe_array_left, Torus *lwe_array_right,
-                        int_comparison_buffer<Torus> *mem_ptr, void *bsk,
-                        Torus *ksk, uint32_t num_radix_blocks) {
+                        int_comparison_buffer<Torus> *mem_ptr, void **bsks,
+                        Torus **ksks, uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -360,7 +358,7 @@ compare_radix_blocks_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
   // Apply LUT to compare to 0
   auto is_non_zero_lut = mem_ptr->eq_buffer->is_non_zero_lut;
   integer_radix_apply_univariate_lookup_table_kb(
-      streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_out, bsk, ksk,
+      streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_out, bsks, ksks,
       num_radix_blocks, is_non_zero_lut);
 
   // Add one
@@ -380,10 +378,9 @@ tree_sign_reduction(cudaStream_t *streams, uint32_t *gpu_indexes,
                     uint32_t gpu_count, Torus *lwe_array_out,
                     Torus *lwe_block_comparisons,
                     int_tree_sign_reduction_buffer<Torus> *tree_buffer,
-                    std::function<Torus(Torus)> sign_handler_f, void *bsk,
-                    Torus *ksk, uint32_t num_radix_blocks) {
+                    std::function<Torus(Torus)> sign_handler_f, void **bsks,
+                    Torus **ksks, uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = tree_buffer->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto glwe_dimension = params.glwe_dimension;
@@ -413,7 +410,7 @@ tree_sign_reduction(cudaStream_t *streams, uint32_t *gpu_indexes,
                 partial_block_count, 4);
 
     integer_radix_apply_univariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, x, y, bsk, ksk,
+        streams, gpu_indexes, gpu_count, x, y, bsks, ksks,
         partial_block_count >> 1, inner_tree_leaf);
 
     if ((partial_block_count % 2) != 0) {
@@ -451,13 +448,15 @@ tree_sign_reduction(cudaStream_t *streams, uint32_t *gpu_indexes,
     y = x;
     f = sign_handler_f;
   }
-  generate_device_accumulator<Torus>(streams[0], gpu_indexes[0], last_lut->lut,
-                                     glwe_dimension, polynomial_size,
-                                     message_modulus, carry_modulus, f);
+  generate_device_accumulator<Torus>(
+      streams[0], gpu_indexes[0], last_lut->get_lut(gpu_indexes[0], 0),
+      glwe_dimension, polynomial_size, message_modulus, carry_modulus, f);
+  last_lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
   // Last leaf
-  integer_radix_apply_univariate_lookup_table_kb(
-      streams, gpu_indexes, gpu_count, lwe_array_out, y, bsk, ksk, 1, last_lut);
+  integer_radix_apply_univariate_lookup_table_kb(streams, gpu_indexes,
+                                                 gpu_count, lwe_array_out, y,
+                                                 bsks, ksks, 1, last_lut);
 }
 
 template <typename Torus>
@@ -465,10 +464,9 @@ __host__ void host_integer_radix_difference_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_left, Torus *lwe_array_right,
     int_comparison_buffer<Torus> *mem_ptr,
-    std::function<Torus(Torus)> reduction_lut_f, void *bsk, Torus *ksk,
+    std::function<Torus(Torus)> reduction_lut_f, void **bsks, Torus **ksks,
     uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto diff_buffer = mem_ptr->diff_buffer;
 
   auto params = mem_ptr->params;
@@ -500,10 +498,10 @@ __host__ void host_integer_radix_difference_check_kb(
     // Clean noise
     auto identity_lut = mem_ptr->identity_lut;
     integer_radix_apply_univariate_lookup_table_kb(
-        streams, gpu_indexes, gpu_count, packed_left, packed_left, bsk, ksk,
+        streams, gpu_indexes, gpu_count, packed_left, packed_left, bsks, ksks,
         packed_num_radix_blocks, identity_lut);
     integer_radix_apply_univariate_lookup_table_kb(
-        streams, gpu_indexes, gpu_count, packed_right, packed_right, bsk, ksk,
+        streams, gpu_indexes, gpu_count, packed_right, packed_right, bsks, ksks,
         packed_num_radix_blocks, identity_lut);
 
     lhs = packed_left;
@@ -520,14 +518,15 @@ __host__ void host_integer_radix_difference_check_kb(
     // Compare packed blocks, or simply the total number of radix blocks in the
     // inputs
     compare_radix_blocks_kb(streams, gpu_indexes, gpu_count, comparisons, lhs,
-                            rhs, mem_ptr, bsk, ksk, packed_num_radix_blocks);
+                            rhs, mem_ptr, bsks, ksks, packed_num_radix_blocks);
     num_comparisons = packed_num_radix_blocks;
   } else {
     // Packing is possible
     if (carry_modulus >= message_modulus) {
       // Compare (num_radix_blocks - 2) / 2 packed blocks
       compare_radix_blocks_kb(streams, gpu_indexes, gpu_count, comparisons, lhs,
-                              rhs, mem_ptr, bsk, ksk, packed_num_radix_blocks);
+                              rhs, mem_ptr, bsks, ksks,
+                              packed_num_radix_blocks);
 
       // Compare the last block before the sign block separately
       auto identity_lut = mem_ptr->identity_lut;
@@ -538,37 +537,37 @@ __host__ void host_integer_radix_difference_check_kb(
           packed_num_radix_blocks * big_lwe_size;
       integer_radix_apply_univariate_lookup_table_kb(
           streams, gpu_indexes, gpu_count, last_left_block_before_sign_block,
-          lwe_array_left + (num_radix_blocks - 2) * big_lwe_size, bsk, ksk, 1,
+          lwe_array_left + (num_radix_blocks - 2) * big_lwe_size, bsks, ksks, 1,
           identity_lut);
       integer_radix_apply_univariate_lookup_table_kb(
           streams, gpu_indexes, gpu_count, last_right_block_before_sign_block,
-          lwe_array_right + (num_radix_blocks - 2) * big_lwe_size, bsk, ksk, 1,
-          identity_lut);
+          lwe_array_right + (num_radix_blocks - 2) * big_lwe_size, bsks, ksks,
+          1, identity_lut);
       compare_radix_blocks_kb(
           streams, gpu_indexes, gpu_count,
           comparisons + packed_num_radix_blocks * big_lwe_size,
           last_left_block_before_sign_block, last_right_block_before_sign_block,
-          mem_ptr, bsk, ksk, 1);
+          mem_ptr, bsks, ksks, 1);
       // Compare the sign block separately
       integer_radix_apply_bivariate_lookup_table_kb(
           streams, gpu_indexes, gpu_count,
           comparisons + (packed_num_radix_blocks + 1) * big_lwe_size,
           lwe_array_left + (num_radix_blocks - 1) * big_lwe_size,
-          lwe_array_right + (num_radix_blocks - 1) * big_lwe_size, bsk, ksk, 1,
-          mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
+          lwe_array_right + (num_radix_blocks - 1) * big_lwe_size, bsks, ksks,
+          1, mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
       num_comparisons = packed_num_radix_blocks + 2;
 
     } else {
       compare_radix_blocks_kb(streams, gpu_indexes, gpu_count, comparisons,
-                              lwe_array_left, lwe_array_right, mem_ptr, bsk,
-                              ksk, num_radix_blocks - 1);
+                              lwe_array_left, lwe_array_right, mem_ptr, bsks,
+                              ksks, num_radix_blocks - 1);
       // Compare the sign block separately
       integer_radix_apply_bivariate_lookup_table_kb(
           streams, gpu_indexes, gpu_count,
           comparisons + (num_radix_blocks - 1) * big_lwe_size,
           lwe_array_left + (num_radix_blocks - 1) * big_lwe_size,
-          lwe_array_right + (num_radix_blocks - 1) * big_lwe_size, bsk, ksk, 1,
-          mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
+          lwe_array_right + (num_radix_blocks - 1) * big_lwe_size, bsks, ksks,
+          1, mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
       num_comparisons = num_radix_blocks;
     }
   }
@@ -578,20 +577,19 @@ __host__ void host_integer_radix_difference_check_kb(
   // final sign
   tree_sign_reduction(streams, gpu_indexes, gpu_count, lwe_array_out,
                       comparisons, mem_ptr->diff_buffer->tree_buffer,
-                      reduction_lut_f, bsk, ksk, num_comparisons);
+                      reduction_lut_f, bsks, ksks, num_comparisons);
 }
 
 template <typename Torus>
 __host__ void scratch_cuda_integer_radix_comparison_check_kb(
-    cudaStream_t stream, uint32_t gpu_index,
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     int_comparison_buffer<Torus> **mem_ptr, uint32_t num_radix_blocks,
     int_radix_params params, COMPARISON_TYPE op, bool is_signed,
     bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
-  *mem_ptr = new int_comparison_buffer<Torus>(stream, gpu_index, op, params,
-                                              num_radix_blocks, is_signed,
-                                              allocate_gpu_memory);
+  *mem_ptr = new int_comparison_buffer<Torus>(streams, gpu_indexes, gpu_count,
+                                              op, params, num_radix_blocks,
+                                              is_signed, allocate_gpu_memory);
 }
 
 template <typename Torus>
@@ -599,20 +597,19 @@ __host__ void
 host_integer_radix_maxmin_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                              uint32_t gpu_count, Torus *lwe_array_out,
                              Torus *lwe_array_left, Torus *lwe_array_right,
-                             int_comparison_buffer<Torus> *mem_ptr, void *bsk,
-                             Torus *ksk, uint32_t total_num_radix_blocks) {
+                             int_comparison_buffer<Torus> *mem_ptr, void **bsks,
+                             Torus **ksks, uint32_t total_num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   // Compute the sign
   host_integer_radix_difference_check_kb(
       streams, gpu_indexes, gpu_count, mem_ptr->tmp_lwe_array_out,
-      lwe_array_left, lwe_array_right, mem_ptr, mem_ptr->identity_lut_f, bsk,
-      ksk, total_num_radix_blocks);
+      lwe_array_left, lwe_array_right, mem_ptr, mem_ptr->identity_lut_f, bsks,
+      ksks, total_num_radix_blocks);
 
   // Selector
   host_integer_radix_cmux_kb(streams, gpu_indexes, gpu_count, lwe_array_out,
                              mem_ptr->tmp_lwe_array_out, lwe_array_left,
-                             lwe_array_right, mem_ptr->cmux_buffer, bsk, ksk,
+                             lwe_array_right, mem_ptr->cmux_buffer, bsks, ksks,
                              total_num_radix_blocks);
 }
 

backends/tfhe-cuda-backend/cuda/src/integer/div_rem.cu

@@ -1,12 +1,12 @@
 #include "integer/div_rem.cuh"
 
 void scratch_cuda_integer_div_rem_radix_ciphertext_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t num_blocks, uint32_t message_modulus, uint32_t carry_modulus,
-    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t num_blocks, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -14,72 +14,29 @@ void scratch_cuda_integer_div_rem_radix_ciphertext_kb_64(
                           message_modulus, carry_modulus);
 
   scratch_cuda_integer_div_rem_kb<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_div_rem_memory<uint64_t> **)mem_ptr, num_blocks, params,
       allocate_gpu_memory);
 }
 
 void cuda_integer_div_rem_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, void *quotient,
-    void *remainder, void *numerator, void *divisor, int8_t *mem_ptr, void *bsk,
-    void *ksk, uint32_t num_blocks) {
+    void *remainder, void *numerator, void *divisor, int8_t *mem_ptr,
+    void **bsks, void **ksks, uint32_t num_blocks) {
 
   auto mem = (int_div_rem_memory<uint64_t> *)mem_ptr;
 
-  switch (mem->params.polynomial_size) {
-  case 512:
-    host_integer_div_rem_kb<uint64_t, Degree<512>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 1024:
-
-    host_integer_div_rem_kb<uint64_t, Degree<1024>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 2048:
-    host_integer_div_rem_kb<uint64_t, Degree<2048>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 4096:
-    host_integer_div_rem_kb<uint64_t, Degree<4096>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 8192:
-    host_integer_div_rem_kb<uint64_t, Degree<8192>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 16384:
-    host_integer_div_rem_kb<uint64_t, Degree<16384>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
-        static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
-        bsk, static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  default:
-    PANIC("Cuda error (integer div_rem): unsupported polynomial size. "
-          "Only N = 512, 1024, 2048, 4096, 8192, 16384 is supported")
-  }
+  host_integer_div_rem_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(quotient), static_cast<uint64_t *>(remainder),
+      static_cast<uint64_t *>(numerator), static_cast<uint64_t *>(divisor),
+      bsks, (uint64_t **)(ksks), mem, num_blocks);
 }
 
-void cleanup_cuda_integer_div_rem(void *stream, uint32_t gpu_index,
-                                  int8_t **mem_ptr_void) {
+void cleanup_cuda_integer_div_rem(void **streams, uint32_t *gpu_indexes,
+                                  uint32_t gpu_count, int8_t **mem_ptr_void) {
   int_div_rem_memory<uint64_t> *mem_ptr =
       (int_div_rem_memory<uint64_t> *)(*mem_ptr_void);
 
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/div_rem.cuh

@@ -14,7 +14,6 @@
 #include "utils/kernel_dimensions.cuh"
 #include <fstream>
 #include <iostream>
-#include <omp.h>
 #include <sstream>
 #include <string>
 #include <vector>
@@ -31,17 +30,13 @@ template <typename Torus> struct lwe_ciphertext_list {
   int_radix_params params;
 
   size_t big_lwe_size;
-  size_t radix_size;
   size_t big_lwe_size_bytes;
-  size_t radix_size_bytes;
   size_t big_lwe_dimension;
 
   lwe_ciphertext_list(Torus *src, int_radix_params params, size_t max_blocks)
       : data(src), params(params), max_blocks(max_blocks) {
     big_lwe_size = params.big_lwe_dimension + 1;
     big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
-    radix_size = max_blocks * big_lwe_size;
-    radix_size_bytes = radix_size * sizeof(Torus);
     big_lwe_dimension = params.big_lwe_dimension;
     len = max_blocks;
   }
@@ -164,22 +159,21 @@ template <typename Torus> struct lwe_ciphertext_list {
 };
 
 template <typename Torus>
-__host__ void
-scratch_cuda_integer_div_rem_kb(cudaStream_t stream, uint32_t gpu_index,
-                                int_div_rem_memory<Torus> **mem_ptr,
-                                uint32_t num_blocks, int_radix_params params,
-                                bool allocate_gpu_memory) {
+__host__ void scratch_cuda_integer_div_rem_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_div_rem_memory<Torus> **mem_ptr, uint32_t num_blocks,
+    int_radix_params params, bool allocate_gpu_memory) {
 
-  *mem_ptr = new int_div_rem_memory<Torus>(stream, gpu_index, params,
-                                           num_blocks, allocate_gpu_memory);
+  *mem_ptr = new int_div_rem_memory<Torus>(
+      streams, gpu_indexes, gpu_count, params, num_blocks, allocate_gpu_memory);
 }
 
-template <typename Torus, class params>
+template <typename Torus>
 __host__ void
 host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                         uint32_t gpu_count, Torus *quotient, Torus *remainder,
-                        Torus *numerator, Torus *divisor, void *bsk,
-                        uint64_t *ksk, int_div_rem_memory<uint64_t> *mem_ptr,
+                        Torus *numerator, Torus *divisor, void **bsks,
+                        uint64_t **ksks, int_div_rem_memory<uint64_t> *mem_ptr,
                         uint32_t num_blocks) {
 
   auto radix_params = mem_ptr->params;
@@ -290,7 +284,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
 
           integer_radix_apply_univariate_lookup_table_kb(
               streams, gpu_indexes, gpu_count, interesting_divisor.last_block(),
-              interesting_divisor.last_block(), bsk, ksk, 1,
+              interesting_divisor.last_block(), bsks, ksks, 1,
               mem_ptr->masking_luts_1[shifted_mask]);
         }; // trim_last_interesting_divisor_bits
 
@@ -318,7 +312,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
 
           integer_radix_apply_univariate_lookup_table_kb(
               streams, gpu_indexes, gpu_count, divisor_ms_blocks.first_block(),
-              divisor_ms_blocks.first_block(), bsk, ksk, 1,
+              divisor_ms_blocks.first_block(), bsks, ksks, 1,
               mem_ptr->masking_luts_2[shifted_mask]);
         }; // trim_first_divisor_ms_bits
 
@@ -342,7 +336,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
 
           host_integer_radix_logical_scalar_shift_kb_inplace(
               streams, gpu_indexes, gpu_count, interesting_remainder1.data, 1,
-              mem_ptr->shift_mem_1, bsk, ksk, interesting_remainder1.len);
+              mem_ptr->shift_mem_1, bsks, ksks, interesting_remainder1.len);
 
           tmp_radix.clone_from(interesting_remainder1, 0,
                                interesting_remainder1.len - 1, streams[0],
@@ -371,41 +365,30 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
         [&](cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count) {
           host_integer_radix_logical_scalar_shift_kb_inplace(
               streams, gpu_indexes, gpu_count, interesting_remainder2.data, 1,
-              mem_ptr->shift_mem_2, bsk, ksk, interesting_remainder2.len);
+              mem_ptr->shift_mem_2, bsks, ksks, interesting_remainder2.len);
         }; // left_shift_interesting_remainder2
 
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-#pragma omp parallel sections
-    {
-#pragma omp section
-      {
-        // interesting_divisor
-        trim_last_interesting_divisor_bits(&mem_ptr->sub_stream_1,
-                                           &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // divisor_ms_blocks
-        trim_first_divisor_ms_bits(&mem_ptr->sub_stream_2, &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // interesting_remainder1
-        // numerator_block_stack
-        left_shift_interesting_remainder1(&mem_ptr->sub_stream_3,
-                                          &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // interesting_remainder2
-        left_shift_interesting_remainder2(&mem_ptr->sub_stream_4,
-                                          &gpu_indexes[0], 1);
-      }
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+    // interesting_divisor
+    trim_last_interesting_divisor_bits(mem_ptr->sub_streams_1, gpu_indexes,
+                                       gpu_count);
+    // divisor_ms_blocks
+    trim_first_divisor_ms_bits(mem_ptr->sub_streams_2, gpu_indexes, gpu_count);
+    // interesting_remainder1
+    // numerator_block_stack
+    left_shift_interesting_remainder1(mem_ptr->sub_streams_3, gpu_indexes,
+                                      gpu_count);
+    // interesting_remainder2
+    left_shift_interesting_remainder2(mem_ptr->sub_streams_4, gpu_indexes,
+                                      gpu_count);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(mem_ptr->sub_streams_1[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_2[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_3[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_4[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(mem_ptr->sub_stream_1, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_2, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_3, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_4, gpu_indexes[0]);
 
     // if interesting_remainder1 != 0 -> interesting_remainder2 == 0
     // if interesting_remainder1 == 0 -> interesting_remainder2 != 0
@@ -435,10 +418,10 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
     //  `subtraction_overflowed` - single ciphertext
     auto do_overflowing_sub = [&](cudaStream_t *streams, uint32_t *gpu_indexes,
                                   uint32_t gpu_count) {
-      host_integer_overflowing_sub_kb<Torus, params>(
+      host_integer_overflowing_sub_kb<Torus>(
           streams, gpu_indexes, gpu_count, new_remainder.data,
           subtraction_overflowed.data, merged_interesting_remainder.data,
-          interesting_divisor.data, bsk, ksk, mem_ptr->overflow_sub_mem,
+          interesting_divisor.data, bsks, ksks, mem_ptr->overflow_sub_mem,
           merged_interesting_remainder.len);
     };
 
@@ -458,7 +441,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
         // So we can skip some stuff
         host_compare_with_zero_equality(
             streams, gpu_indexes, gpu_count, tmp_1.data, trivial_blocks.data,
-            mem_ptr->comparison_buffer, bsk, ksk, trivial_blocks.len,
+            mem_ptr->comparison_buffer, bsks, ksks, trivial_blocks.len,
             mem_ptr->comparison_buffer->eq_buffer->is_non_zero_lut);
 
         tmp_1.len =
@@ -467,7 +450,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
         is_at_least_one_comparisons_block_true(
             streams, gpu_indexes, gpu_count,
             at_least_one_upper_block_is_non_zero.data, tmp_1.data,
-            mem_ptr->comparison_buffer, bsk, ksk, tmp_1.len);
+            mem_ptr->comparison_buffer, bsks, ksks, tmp_1.len);
       }
     };
 
@@ -480,36 +463,28 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
           integer_radix_apply_univariate_lookup_table_kb(
               streams, gpu_indexes, gpu_count,
               cleaned_merged_interesting_remainder.data,
-              cleaned_merged_interesting_remainder.data, bsk, ksk,
+              cleaned_merged_interesting_remainder.data, bsks, ksks,
               cleaned_merged_interesting_remainder.len,
               mem_ptr->message_extract_lut_1);
         };
 
     // phase 2
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-#pragma omp parallel sections
-    {
-#pragma omp section
-      {
-        // new_remainder
-        // subtraction_overflowed
-        do_overflowing_sub(&mem_ptr->sub_stream_1, &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // at_least_one_upper_block_is_non_zero
-        check_divisor_upper_blocks(&mem_ptr->sub_stream_2, &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // cleaned_merged_interesting_remainder
-        create_clean_version_of_merged_remainder(&mem_ptr->sub_stream_3,
-                                                 &gpu_indexes[0], 1);
-      }
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+    // new_remainder
+    // subtraction_overflowed
+    do_overflowing_sub(mem_ptr->sub_streams_1, gpu_indexes, gpu_count);
+    // at_least_one_upper_block_is_non_zero
+    check_divisor_upper_blocks(mem_ptr->sub_streams_2, gpu_indexes, gpu_count);
+    // cleaned_merged_interesting_remainder
+    create_clean_version_of_merged_remainder(mem_ptr->sub_streams_3,
+                                             gpu_indexes, gpu_count);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(mem_ptr->sub_streams_1[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_2[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_3[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(mem_ptr->sub_stream_1, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_2, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_3, gpu_indexes[0]);
 
     host_addition(streams[0], gpu_indexes[0], overflow_sum.data,
                   subtraction_overflowed.data,
@@ -528,7 +503,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
               streams, gpu_indexes, gpu_count,
               cleaned_merged_interesting_remainder.data,
               cleaned_merged_interesting_remainder.data,
-              overflow_sum_radix.data, bsk, ksk,
+              overflow_sum_radix.data, bsks, ksks,
               cleaned_merged_interesting_remainder.len,
               mem_ptr->zero_out_if_overflow_did_not_happen[factor_lut_id],
               factor);
@@ -538,7 +513,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
         [&](cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count) {
           integer_radix_apply_bivariate_lookup_table_kb<Torus>(
               streams, gpu_indexes, gpu_count, new_remainder.data,
-              new_remainder.data, overflow_sum_radix.data, bsk, ksk,
+              new_remainder.data, overflow_sum_radix.data, bsks, ksks,
               new_remainder.len,
               mem_ptr->zero_out_if_overflow_happened[factor_lut_id], factor);
         };
@@ -548,7 +523,7 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
       integer_radix_apply_bivariate_lookup_table_kb<Torus>(
           streams, gpu_indexes, gpu_count, did_not_overflow.data,
           subtraction_overflowed.data,
-          at_least_one_upper_block_is_non_zero.data, bsk, ksk, 1,
+          at_least_one_upper_block_is_non_zero.data, bsks, ksks, 1,
           mem_ptr->merge_overflow_flags_luts[pos_in_block],
           mem_ptr->merge_overflow_flags_luts[pos_in_block]
               ->params.message_modulus);
@@ -559,30 +534,22 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                     did_not_overflow.data, radix_params.big_lwe_dimension, 1);
     };
 
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-#pragma omp parallel sections
-    {
-#pragma omp section
-      {
-        // cleaned_merged_interesting_remainder
-        conditionally_zero_out_merged_interesting_remainder(
-            &mem_ptr->sub_stream_1, &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // new_remainder
-        conditionally_zero_out_merged_new_remainder(&mem_ptr->sub_stream_2,
-                                                    &gpu_indexes[0], 1);
-      }
-#pragma omp section
-      {
-        // quotient
-        set_quotient_bit(&mem_ptr->sub_stream_3, &gpu_indexes[0], 1);
-      }
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+    // cleaned_merged_interesting_remainder
+    conditionally_zero_out_merged_interesting_remainder(mem_ptr->sub_streams_1,
+                                                        gpu_indexes, gpu_count);
+    // new_remainder
+    conditionally_zero_out_merged_new_remainder(mem_ptr->sub_streams_2,
+                                                gpu_indexes, gpu_count);
+    // quotient
+    set_quotient_bit(mem_ptr->sub_streams_3, gpu_indexes, gpu_count);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(mem_ptr->sub_streams_1[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_2[j], gpu_indexes[j]);
+      cuda_synchronize_stream(mem_ptr->sub_streams_3[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(mem_ptr->sub_stream_1, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_2, gpu_indexes[0]);
-    cuda_synchronize_stream(mem_ptr->sub_stream_3, gpu_indexes[0]);
 
     assert(first_trivial_block - 1 == cleaned_merged_interesting_remainder.len);
     assert(first_trivial_block - 1 == new_remainder.len);
@@ -601,24 +568,19 @@ host_integer_div_rem_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                 remainder2.data, radix_params.big_lwe_dimension,
                 remainder1.len);
 
-  cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-#pragma omp parallel sections
-  {
-#pragma omp section
-    {
-      integer_radix_apply_univariate_lookup_table_kb(
-          &mem_ptr->sub_stream_1, &gpu_indexes[0], 1, remainder, remainder, bsk,
-          ksk, num_blocks, mem_ptr->message_extract_lut_1);
-    }
-#pragma omp section
-    {
-      integer_radix_apply_univariate_lookup_table_kb(
-          &mem_ptr->sub_stream_2, &gpu_indexes[0], 1, quotient, quotient, bsk,
-          ksk, num_blocks, mem_ptr->message_extract_lut_2);
-    }
+  for (uint j = 0; j < gpu_count; j++) {
+    cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+  }
+  integer_radix_apply_univariate_lookup_table_kb(
+      mem_ptr->sub_streams_1, gpu_indexes, gpu_count, remainder, remainder,
+      bsks, ksks, num_blocks, mem_ptr->message_extract_lut_1);
+  integer_radix_apply_univariate_lookup_table_kb(
+      mem_ptr->sub_streams_2, gpu_indexes, gpu_count, quotient, quotient, bsks,
+      ksks, num_blocks, mem_ptr->message_extract_lut_2);
+  for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+    cuda_synchronize_stream(mem_ptr->sub_streams_1[j], gpu_indexes[j]);
+    cuda_synchronize_stream(mem_ptr->sub_streams_2[j], gpu_indexes[j]);
   }
-  cuda_synchronize_stream(mem_ptr->sub_stream_1, gpu_indexes[0]);
-  cuda_synchronize_stream(mem_ptr->sub_stream_2, gpu_indexes[0]);
 }
 
 #endif // TFHE_RS_DIV_REM_CUH

backends/tfhe-cuda-backend/cuda/src/integer/integer.cu

@@ -1,127 +1,52 @@
 #include "integer/integer.cuh"
 #include <linear_algebra.h>
 
-void cuda_full_propagation_64_inplace(
-    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    void *input_blocks, int8_t *mem_ptr, void *ksk, void *bsk,
-    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t ks_base_log, uint32_t ks_level, uint32_t pbs_base_log,
-    uint32_t pbs_level, uint32_t grouping_factor, uint32_t num_blocks) {
+void cuda_full_propagation_64_inplace(void **streams, uint32_t *gpu_indexes,
+                                      uint32_t gpu_count, void *input_blocks,
+                                      int8_t *mem_ptr, void **ksks, void **bsks,
+                                      uint32_t num_blocks) {
 
-  switch (polynomial_size) {
-  case 256:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<256>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 512:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<512>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 1024:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<1024>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 2048:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<2048>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 4096:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<4096>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 8192:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<8192>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  case 16384:
-    host_full_propagate_inplace<uint64_t, int64_t, AmortizedDegree<16384>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(input_blocks),
-        (int_fullprop_buffer<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk),
-        bsk, lwe_dimension, glwe_dimension, polynomial_size, ks_base_log,
-        ks_level, pbs_base_log, pbs_level, grouping_factor, num_blocks);
-    break;
-  default:
-    PANIC("Cuda error (full propagation inplace): unsupported polynomial size. "
-          "Supported N's are powers of two"
-          " in the interval [256..16384].")
-  }
+  int_fullprop_buffer<uint64_t> *buffer =
+      (int_fullprop_buffer<uint64_t> *)mem_ptr;
+
+  host_full_propagate_inplace<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(input_blocks), buffer, (uint64_t **)(ksks), bsks,
+      num_blocks);
 }
 
 void scratch_cuda_full_propagation_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t lwe_dimension,
-    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
-    uint32_t grouping_factor, uint32_t input_lwe_ciphertext_count,
-    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
-    bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t ks_level, uint32_t ks_base_log, uint32_t pbs_level,
+    uint32_t pbs_base_log, uint32_t grouping_factor, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          glwe_dimension * polynomial_size, lwe_dimension,
+                          ks_level, ks_base_log, pbs_level, pbs_base_log,
+                          grouping_factor, message_modulus, carry_modulus);
 
   scratch_cuda_full_propagation<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
-      (int_fullprop_buffer<uint64_t> **)mem_ptr, lwe_dimension, glwe_dimension,
-      polynomial_size, level_count, grouping_factor, input_lwe_ciphertext_count,
-      message_modulus, carry_modulus, pbs_type, allocate_gpu_memory);
+      (cudaStream_t *)streams, gpu_indexes, gpu_count,
+      (int_fullprop_buffer<uint64_t> **)mem_ptr, params, allocate_gpu_memory);
 }
 
-void cleanup_cuda_full_propagation(void *stream, uint32_t gpu_index,
-                                   int8_t **mem_ptr_void) {
+void cleanup_cuda_full_propagation(void **streams, uint32_t *gpu_indexes,
+                                   uint32_t gpu_count, int8_t **mem_ptr_void) {
 
   int_fullprop_buffer<uint64_t> *mem_ptr =
       (int_fullprop_buffer<uint64_t> *)(*mem_ptr_void);
-  auto s = static_cast<cudaStream_t>(stream);
 
-  cuda_drop_async(mem_ptr->lut_buffer, s, gpu_index);
-  cuda_drop_async(mem_ptr->lut_indexes, s, gpu_index);
-
-  cuda_drop_async(mem_ptr->lwe_indexes, s, gpu_index);
-
-  cuda_drop_async(mem_ptr->tmp_small_lwe_vector, s, gpu_index);
-  cuda_drop_async(mem_ptr->tmp_big_lwe_vector, s, gpu_index);
-
-  switch (mem_ptr->pbs_type) {
-  case CLASSICAL: {
-    auto x = (pbs_buffer<uint64_t, CLASSICAL> *)(mem_ptr->pbs_buffer);
-    x->release(s, gpu_index);
-  } break;
-  case MULTI_BIT: {
-    auto x = (pbs_buffer<uint64_t, MULTI_BIT> *)(mem_ptr->pbs_buffer);
-    x->release(s, gpu_index);
-  } break;
-  default:
-    PANIC("Cuda error (PBS): unsupported implementation variant.")
-  }
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }
 
 void scratch_cuda_propagate_single_carry_kb_64_inplace(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t num_blocks, uint32_t message_modulus, uint32_t carry_modulus,
-    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t num_blocks, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -129,35 +54,49 @@ void scratch_cuda_propagate_single_carry_kb_64_inplace(
                           message_modulus, carry_modulus);
 
   scratch_cuda_propagate_single_carry_kb_inplace(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_sc_prop_memory<uint64_t> **)mem_ptr, num_blocks, params,
       allocate_gpu_memory);
 }
 
 void cuda_propagate_single_carry_kb_64_inplace(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, void *lwe_array,
-    int8_t *mem_ptr, void *bsk, void *ksk, uint32_t num_blocks) {
+    void *carry_out, int8_t *mem_ptr, void **bsks, void **ksks,
+    uint32_t num_blocks) {
   host_propagate_single_carry<uint64_t>(
       (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-      static_cast<uint64_t *>(lwe_array),
-      (int_sc_prop_memory<uint64_t> *)mem_ptr, bsk,
-      static_cast<uint64_t *>(ksk), num_blocks);
+      static_cast<uint64_t *>(lwe_array), static_cast<uint64_t *>(carry_out),
+      nullptr, (int_sc_prop_memory<uint64_t> *)mem_ptr, bsks,
+      (uint64_t **)(ksks), num_blocks);
 }
 
-void cleanup_cuda_propagate_single_carry(void *stream, uint32_t gpu_index,
+void cuda_propagate_single_carry_get_input_carries_kb_64_inplace(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, void *lwe_array,
+    void *carry_out, void *input_carries, int8_t *mem_ptr, void **bsks,
+    void **ksks, uint32_t num_blocks) {
+  host_propagate_single_carry<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(lwe_array), static_cast<uint64_t *>(carry_out),
+      static_cast<uint64_t *>(input_carries),
+      (int_sc_prop_memory<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks),
+      num_blocks);
+}
+
+void cleanup_cuda_propagate_single_carry(void **streams, uint32_t *gpu_indexes,
+                                         uint32_t gpu_count,
                                          int8_t **mem_ptr_void) {
   int_sc_prop_memory<uint64_t> *mem_ptr =
       (int_sc_prop_memory<uint64_t> *)(*mem_ptr_void);
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }
 
 void scratch_cuda_apply_univariate_lut_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, void *input_lut,
-    uint32_t lwe_dimension, uint32_t glwe_dimension, uint32_t polynomial_size,
-    uint32_t ks_level, uint32_t ks_base_log, uint32_t pbs_level,
-    uint32_t pbs_base_log, uint32_t grouping_factor, uint32_t num_radix_blocks,
-    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
-    bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    void *input_lut, uint32_t lwe_dimension, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t num_radix_blocks, uint32_t message_modulus, uint32_t carry_modulus,
+    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           glwe_dimension * polynomial_size, lwe_dimension,
@@ -165,7 +104,7 @@ void scratch_cuda_apply_univariate_lut_kb_64(
                           grouping_factor, message_modulus, carry_modulus);
 
   scratch_cuda_apply_univariate_lut_kb<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_radix_lut<uint64_t> **)mem_ptr, static_cast<uint64_t *>(input_lut),
       num_radix_blocks, params, allocate_gpu_memory);
 }
@@ -173,19 +112,116 @@ void scratch_cuda_apply_univariate_lut_kb_64(
 void cuda_apply_univariate_lut_kb_64(void **streams, uint32_t *gpu_indexes,
                                      uint32_t gpu_count, void *output_radix_lwe,
                                      void *input_radix_lwe, int8_t *mem_ptr,
-                                     void *ksk, void *bsk,
+                                     void **ksks, void **bsks,
                                      uint32_t num_blocks) {
 
   host_apply_univariate_lut_kb<uint64_t>(
       (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       static_cast<uint64_t *>(output_radix_lwe),
       static_cast<uint64_t *>(input_radix_lwe),
-      (int_radix_lut<uint64_t> *)mem_ptr, static_cast<uint64_t *>(ksk), bsk,
+      (int_radix_lut<uint64_t> *)mem_ptr, (uint64_t **)(ksks), bsks,
       num_blocks);
 }
 
-void cleanup_cuda_apply_univariate_lut_kb_64(void *stream, uint32_t gpu_index,
+void cleanup_cuda_apply_univariate_lut_kb_64(void **streams,
+                                             uint32_t *gpu_indexes,
+                                             uint32_t gpu_count,
                                              int8_t **mem_ptr_void) {
   int_radix_lut<uint64_t> *mem_ptr = (int_radix_lut<uint64_t> *)(*mem_ptr_void);
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
+}
+
+void scratch_cuda_apply_bivariate_lut_kb_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    void *input_lut, uint32_t lwe_dimension, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t num_radix_blocks, uint32_t message_modulus, uint32_t carry_modulus,
+    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          glwe_dimension * polynomial_size, lwe_dimension,
+                          ks_level, ks_base_log, pbs_level, pbs_base_log,
+                          grouping_factor, message_modulus, carry_modulus);
+
+  scratch_cuda_apply_bivariate_lut_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      (int_radix_lut<uint64_t> **)mem_ptr, static_cast<uint64_t *>(input_lut),
+      num_radix_blocks, params, allocate_gpu_memory);
+}
+
+void cuda_apply_bivariate_lut_kb_64(void **streams, uint32_t *gpu_indexes,
+                                    uint32_t gpu_count, void *output_radix_lwe,
+                                    void *input_radix_lwe_1,
+                                    void *input_radix_lwe_2, int8_t *mem_ptr,
+                                    void **ksks, void **bsks,
+                                    uint32_t num_blocks, uint32_t shift) {
+
+  host_apply_bivariate_lut_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(output_radix_lwe),
+      static_cast<uint64_t *>(input_radix_lwe_1),
+      static_cast<uint64_t *>(input_radix_lwe_2),
+      (int_radix_lut<uint64_t> *)mem_ptr, (uint64_t **)(ksks), bsks, num_blocks,
+      shift);
+}
+
+void cleanup_cuda_apply_bivariate_lut_kb_64(void **streams,
+                                            uint32_t *gpu_indexes,
+                                            uint32_t gpu_count,
+                                            int8_t **mem_ptr_void) {
+  int_radix_lut<uint64_t> *mem_ptr = (int_radix_lut<uint64_t> *)(*mem_ptr_void);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
+}
+
+void scratch_cuda_integer_compute_prefix_sum_hillis_steele_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    void *input_lut, uint32_t lwe_dimension, uint32_t glwe_dimension,
+    uint32_t polynomial_size, uint32_t ks_level, uint32_t ks_base_log,
+    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t num_radix_blocks, uint32_t message_modulus, uint32_t carry_modulus,
+    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+
+  int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
+                          glwe_dimension * polynomial_size, lwe_dimension,
+                          ks_level, ks_base_log, pbs_level, pbs_base_log,
+                          grouping_factor, message_modulus, carry_modulus);
+
+  scratch_cuda_apply_bivariate_lut_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      (int_radix_lut<uint64_t> **)mem_ptr, static_cast<uint64_t *>(input_lut),
+      num_radix_blocks, params, allocate_gpu_memory);
+}
+
+void cuda_integer_compute_prefix_sum_hillis_steele_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    void *output_radix_lwe, void *input_radix_lwe, int8_t *mem_ptr, void **ksks,
+    void **bsks, uint32_t num_blocks, uint32_t shift) {
+
+  int_radix_params params = ((int_radix_lut<uint64_t> *)mem_ptr)->params;
+
+  host_compute_prefix_sum_hillis_steele<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(output_radix_lwe),
+      static_cast<uint64_t *>(input_radix_lwe), params,
+      (int_radix_lut<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks),
+      num_blocks);
+}
+
+void cleanup_cuda_integer_compute_prefix_sum_hillis_steele_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int8_t **mem_ptr_void) {
+  int_radix_lut<uint64_t> *mem_ptr = (int_radix_lut<uint64_t> *)(*mem_ptr_void);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
+}
+
+void cuda_integer_reverse_blocks_64_inplace(void **streams,
+                                            uint32_t *gpu_indexes,
+                                            uint32_t gpu_count, void *lwe_array,
+                                            uint32_t num_blocks,
+                                            uint32_t lwe_size) {
+
+  host_radix_blocks_reverse_inplace<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes,
+      static_cast<uint64_t *>(lwe_array), num_blocks, lwe_size);
 }

backends/tfhe-cuda-backend/cuda/src/integer/integer.cuh

@@ -3,6 +3,7 @@
 
 #include "crypto/keyswitch.cuh"
 #include "device.h"
+#include "helper_multi_gpu.h"
 #include "integer.h"
 #include "integer/scalar_addition.cuh"
 #include "linear_algebra.h"
@@ -10,6 +11,7 @@
 #include "polynomial/functions.cuh"
 #include "programmable_bootstrap.h"
 #include "utils/helper.cuh"
+#include "utils/helper_multi_gpu.cuh"
 #include "utils/kernel_dimensions.cuh"
 #include <functional>
 
@@ -20,18 +22,19 @@ template <typename Torus>
 __global__ void radix_blocks_rotate_right(Torus *dst, Torus *src,
                                           uint32_t value, uint32_t blocks_count,
                                           uint32_t lwe_size) {
-  value %= blocks_count;
-
   size_t tid = threadIdx.x;
-  size_t src_block_id = blockIdx.x;
-  size_t dst_block_id = (src_block_id + value) % blocks_count;
-  size_t stride = blockDim.x;
+  if (tid < lwe_size) {
+    value %= blocks_count;
+    size_t src_block_id = blockIdx.x;
+    size_t dst_block_id = (src_block_id + value) % blocks_count;
+    size_t stride = blockDim.x;
 
-  auto cur_src_block = &src[src_block_id * lwe_size];
-  auto cur_dst_block = &dst[dst_block_id * lwe_size];
+    auto cur_src_block = &src[src_block_id * lwe_size];
+    auto cur_dst_block = &dst[dst_block_id * lwe_size];
 
-  for (size_t i = tid; i < lwe_size; i += stride) {
-    cur_dst_block[i] = cur_src_block[i];
+    for (size_t i = tid; i < lwe_size; i += stride) {
+      cur_dst_block[i] = cur_src_block[i];
+    }
   }
 }
 
@@ -42,25 +45,28 @@ template <typename Torus>
 __global__ void radix_blocks_rotate_left(Torus *dst, Torus *src, uint32_t value,
                                          uint32_t blocks_count,
                                          uint32_t lwe_size) {
-  value %= blocks_count;
-  size_t src_block_id = blockIdx.x;
-
   size_t tid = threadIdx.x;
-  size_t dst_block_id = (src_block_id >= value)
-                            ? src_block_id - value
-                            : src_block_id - value + blocks_count;
-  size_t stride = blockDim.x;
+  if (tid < lwe_size) {
+    value %= blocks_count;
+    size_t src_block_id = blockIdx.x;
 
-  auto cur_src_block = &src[src_block_id * lwe_size];
-  auto cur_dst_block = &dst[dst_block_id * lwe_size];
+    size_t dst_block_id = (src_block_id >= value)
+                              ? src_block_id - value
+                              : src_block_id - value + blocks_count;
+    size_t stride = blockDim.x;
 
-  for (size_t i = tid; i < lwe_size; i += stride) {
-    cur_dst_block[i] = cur_src_block[i];
+    auto cur_src_block = &src[src_block_id * lwe_size];
+    auto cur_dst_block = &dst[dst_block_id * lwe_size];
+
+    for (size_t i = tid; i < lwe_size; i += stride) {
+      cur_dst_block[i] = cur_src_block[i];
+    }
   }
 }
 
 // rotate radix ciphertext right with specific value
 // calculation is not inplace, so `dst` and `src` must not be the same
+// one block is responsible to process single lwe ciphertext
 template <typename Torus>
 __host__ void
 host_radix_blocks_rotate_right(cudaStream_t *streams, uint32_t *gpu_indexes,
@@ -72,7 +78,7 @@ host_radix_blocks_rotate_right(cudaStream_t *streams, uint32_t *gpu_indexes,
           "pointers should be different");
   }
   cudaSetDevice(gpu_indexes[0]);
-  radix_blocks_rotate_right<<<blocks_count, 256, 0, streams[0]>>>(
+  radix_blocks_rotate_right<<<blocks_count, 1024, 0, streams[0]>>>(
       dst, src, value, blocks_count, lwe_size);
 }
 
@@ -89,10 +95,39 @@ host_radix_blocks_rotate_left(cudaStream_t *streams, uint32_t *gpu_indexes,
           "pointers should be different");
   }
   cudaSetDevice(gpu_indexes[0]);
-  radix_blocks_rotate_left<<<blocks_count, 256, 0, streams[0]>>>(
+  radix_blocks_rotate_left<<<blocks_count, 1024, 0, streams[0]>>>(
       dst, src, value, blocks_count, lwe_size);
 }
 
+// reverse the blocks in a list
+// each cuda block swaps a couple of blocks
+template <typename Torus>
+__global__ void radix_blocks_reverse_lwe_inplace(Torus *src,
+                                                 uint32_t blocks_count,
+                                                 uint32_t lwe_size) {
+
+  size_t idx = blockIdx.x;
+  size_t rev_idx = blocks_count - 1 - idx;
+
+  for (int j = threadIdx.x; j < lwe_size; j += blockDim.x) {
+    Torus back_element = src[rev_idx * lwe_size + j];
+    Torus front_element = src[idx * lwe_size + j];
+    src[idx * lwe_size + j] = back_element;
+    src[rev_idx * lwe_size + j] = front_element;
+  }
+}
+
+template <typename Torus>
+__host__ void
+host_radix_blocks_reverse_inplace(cudaStream_t *streams, uint32_t *gpu_indexes,
+                                  Torus *src, uint32_t blocks_count,
+                                  uint32_t lwe_size) {
+  cudaSetDevice(gpu_indexes[0]);
+  int num_blocks = blocks_count / 2, num_threads = 1024;
+  radix_blocks_reverse_lwe_inplace<<<num_blocks, num_threads, 0, streams[0]>>>(
+      src, blocks_count, lwe_size);
+}
+
 // polynomial_size threads
 template <typename Torus>
 __global__ void
@@ -138,7 +173,7 @@ __host__ void pack_bivariate_blocks(cudaStream_t *streams,
 template <typename Torus>
 __host__ void integer_radix_apply_univariate_lookup_table_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    Torus *lwe_array_out, Torus *lwe_array_in, void *bsk, Torus *ksk,
+    Torus *lwe_array_out, Torus *lwe_array_in, void **bsks, Torus **ksks,
     uint32_t num_radix_blocks, int_radix_lut<Torus> *lut) {
   // apply_lookup_table
   auto params = lut->params;
@@ -153,30 +188,77 @@ __host__ void integer_radix_apply_univariate_lookup_table_kb(
   auto polynomial_size = params.polynomial_size;
   auto grouping_factor = params.grouping_factor;
 
-  // Compute Keyswitch-PBS
-  cuda_keyswitch_lwe_ciphertext_vector(
-      streams[0], gpu_indexes[0], lut->tmp_lwe_after_ks,
-      lut->lwe_trivial_indexes, lwe_array_in, lut->lwe_indexes_in, ksk,
-      big_lwe_dimension, small_lwe_dimension, ks_base_log, ks_level,
-      num_radix_blocks);
+  /// For multi GPU execution we create vectors of pointers for inputs and
+  /// outputs
+  std::vector<Torus *> lwe_array_in_vec = lut->lwe_array_in_vec;
+  std::vector<Torus *> lwe_after_ks_vec = lut->lwe_after_ks_vec;
+  std::vector<Torus *> lwe_after_pbs_vec = lut->lwe_after_pbs_vec;
+  std::vector<Torus *> lwe_trivial_indexes_vec = lut->lwe_trivial_indexes_vec;
 
-  execute_pbs<Torus>(streams, gpu_indexes, gpu_count, lwe_array_out,
-                     lut->lwe_indexes_out, lut->lut, lut->lut_indexes,
-                     lut->tmp_lwe_after_ks, lut->lwe_trivial_indexes, bsk,
-                     lut->buffer, glwe_dimension, small_lwe_dimension,
-                     polynomial_size, pbs_base_log, pbs_level, grouping_factor,
-                     num_radix_blocks, 1, 0,
-                     cuda_get_max_shared_memory(gpu_indexes[0]), pbs_type);
+  auto active_gpu_count = get_active_gpu_count(num_radix_blocks, gpu_count);
+  if (active_gpu_count == 1) {
+    execute_keyswitch_async<Torus>(streams, gpu_indexes, 1, lwe_after_ks_vec[0],
+                                   lwe_trivial_indexes_vec[0], lwe_array_in,
+                                   lut->lwe_indexes_in, ksks, big_lwe_dimension,
+                                   small_lwe_dimension, ks_base_log, ks_level,
+                                   num_radix_blocks);
+
+    /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+    /// dimension to a big LWE dimension
+    execute_pbs_async<Torus>(
+        streams, gpu_indexes, 1, lwe_array_out, lut->lwe_indexes_out,
+        lut->lut_vec, lut->lut_indexes_vec, lwe_after_ks_vec[0],
+        lwe_trivial_indexes_vec[0], bsks, lut->buffer, glwe_dimension,
+        small_lwe_dimension, polynomial_size, pbs_base_log, pbs_level,
+        grouping_factor, num_radix_blocks, pbs_type);
+  } else {
+    /// Make sure all data that should be on GPU 0 is indeed there
+    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
+
+    /// With multiple GPUs we push to the vectors on each GPU then when we
+    /// gather data to GPU 0 we can copy back to the original indexing
+    multi_gpu_scatter_lwe_async<Torus>(
+        streams, gpu_indexes, active_gpu_count, lwe_array_in_vec, lwe_array_in,
+        lut->h_lwe_indexes_in, lut->using_trivial_lwe_indexes, num_radix_blocks,
+        big_lwe_dimension + 1);
+
+    /// Apply KS to go from a big LWE dimension to a small LWE dimension
+    execute_keyswitch_async<Torus>(streams, gpu_indexes, active_gpu_count,
+                                   lwe_after_ks_vec, lwe_trivial_indexes_vec,
+                                   lwe_array_in_vec, lwe_trivial_indexes_vec,
+                                   ksks, big_lwe_dimension, small_lwe_dimension,
+                                   ks_base_log, ks_level, num_radix_blocks);
+
+    /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+    /// dimension to a big LWE dimension
+    execute_pbs_async<Torus>(
+        streams, gpu_indexes, active_gpu_count, lwe_after_pbs_vec,
+        lwe_trivial_indexes_vec, lut->lut_vec, lut->lut_indexes_vec,
+        lwe_after_ks_vec, lwe_trivial_indexes_vec, bsks, lut->buffer,
+        glwe_dimension, small_lwe_dimension, polynomial_size, pbs_base_log,
+        pbs_level, grouping_factor, num_radix_blocks, pbs_type);
+
+    /// Copy data back to GPU 0 and release vecs
+    multi_gpu_gather_lwe_async<Torus>(streams, gpu_indexes, active_gpu_count,
+                                      lwe_array_out, lwe_after_pbs_vec,
+                                      lut->h_lwe_indexes_out,
+                                      lut->using_trivial_lwe_indexes,
+                                      num_radix_blocks, big_lwe_dimension + 1);
+
+    /// Synchronize all GPUs
+    for (uint i = 0; i < active_gpu_count; i++) {
+      cuda_synchronize_stream(streams[i], gpu_indexes[i]);
+    }
+  }
 }
 
 template <typename Torus>
 __host__ void integer_radix_apply_bivariate_lookup_table_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    Torus *lwe_array_out, Torus *lwe_array_1, Torus *lwe_array_2, void *bsk,
-    Torus *ksk, uint32_t num_radix_blocks, int_radix_lut<Torus> *lut,
+    Torus *lwe_array_out, Torus *lwe_array_1, Torus *lwe_array_2, void **bsks,
+    Torus **ksks, uint32_t num_radix_blocks, int_radix_lut<Torus> *lut,
     uint32_t shift) {
-  cudaSetDevice(gpu_indexes[0]);
-  // apply_lookup_table_bivariate
+
   auto params = lut->params;
   auto pbs_type = params.pbs_type;
   auto big_lwe_dimension = params.big_lwe_dimension;
@@ -188,7 +270,6 @@ __host__ void integer_radix_apply_bivariate_lookup_table_kb(
   auto glwe_dimension = params.glwe_dimension;
   auto polynomial_size = params.polynomial_size;
   auto grouping_factor = params.grouping_factor;
-  auto message_modulus = params.message_modulus;
 
   // Left message is shifted
   auto lwe_array_pbs_in = lut->tmp_lwe_before_ks;
@@ -198,20 +279,64 @@ __host__ void integer_radix_apply_bivariate_lookup_table_kb(
                         num_radix_blocks);
   check_cuda_error(cudaGetLastError());
 
-  // Apply LUT
-  cuda_keyswitch_lwe_ciphertext_vector(
-      streams[0], gpu_indexes[0], lut->tmp_lwe_after_ks,
-      lut->lwe_trivial_indexes, lwe_array_pbs_in, lut->lwe_trivial_indexes, ksk,
-      big_lwe_dimension, small_lwe_dimension, ks_base_log, ks_level,
-      num_radix_blocks);
+  /// For multi GPU execution we create vectors of pointers for inputs and
+  /// outputs
+  std::vector<Torus *> lwe_array_in_vec = lut->lwe_array_in_vec;
+  std::vector<Torus *> lwe_after_ks_vec = lut->lwe_after_ks_vec;
+  std::vector<Torus *> lwe_after_pbs_vec = lut->lwe_after_pbs_vec;
+  std::vector<Torus *> lwe_trivial_indexes_vec = lut->lwe_trivial_indexes_vec;
 
-  execute_pbs<Torus>(streams, gpu_indexes, gpu_count, lwe_array_out,
-                     lut->lwe_indexes_out, lut->lut, lut->lut_indexes,
-                     lut->tmp_lwe_after_ks, lut->lwe_trivial_indexes, bsk,
-                     lut->buffer, glwe_dimension, small_lwe_dimension,
-                     polynomial_size, pbs_base_log, pbs_level, grouping_factor,
-                     num_radix_blocks, 1, 0,
-                     cuda_get_max_shared_memory(gpu_indexes[0]), pbs_type);
+  auto active_gpu_count = get_active_gpu_count(num_radix_blocks, gpu_count);
+  if (active_gpu_count == 1) {
+    execute_keyswitch_async<Torus>(streams, gpu_indexes, 1, lwe_after_ks_vec[0],
+                                   lwe_trivial_indexes_vec[0], lwe_array_pbs_in,
+                                   lut->lwe_indexes_in, ksks, big_lwe_dimension,
+                                   small_lwe_dimension, ks_base_log, ks_level,
+                                   num_radix_blocks);
+
+    /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+    /// dimension to a big LWE dimension
+    execute_pbs_async<Torus>(
+        streams, gpu_indexes, 1, lwe_array_out, lut->lwe_indexes_out,
+        lut->lut_vec, lut->lut_indexes_vec, lwe_after_ks_vec[0],
+        lwe_trivial_indexes_vec[0], bsks, lut->buffer, glwe_dimension,
+        small_lwe_dimension, polynomial_size, pbs_base_log, pbs_level,
+        grouping_factor, num_radix_blocks, pbs_type);
+  } else {
+    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
+    multi_gpu_scatter_lwe_async<Torus>(
+        streams, gpu_indexes, active_gpu_count, lwe_array_in_vec,
+        lwe_array_pbs_in, lut->h_lwe_indexes_in, lut->using_trivial_lwe_indexes,
+        num_radix_blocks, big_lwe_dimension + 1);
+
+    /// Apply KS to go from a big LWE dimension to a small LWE dimension
+    execute_keyswitch_async<Torus>(streams, gpu_indexes, active_gpu_count,
+                                   lwe_after_ks_vec, lwe_trivial_indexes_vec,
+                                   lwe_array_in_vec, lwe_trivial_indexes_vec,
+                                   ksks, big_lwe_dimension, small_lwe_dimension,
+                                   ks_base_log, ks_level, num_radix_blocks);
+
+    /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+    /// dimension to a big LWE dimension
+    execute_pbs_async<Torus>(
+        streams, gpu_indexes, active_gpu_count, lwe_after_pbs_vec,
+        lwe_trivial_indexes_vec, lut->lut_vec, lut->lut_indexes_vec,
+        lwe_after_ks_vec, lwe_trivial_indexes_vec, bsks, lut->buffer,
+        glwe_dimension, small_lwe_dimension, polynomial_size, pbs_base_log,
+        pbs_level, grouping_factor, num_radix_blocks, pbs_type);
+
+    /// Copy data back to GPU 0 and release vecs
+    multi_gpu_gather_lwe_async<Torus>(streams, gpu_indexes, active_gpu_count,
+                                      lwe_array_out, lwe_after_pbs_vec,
+                                      lut->h_lwe_indexes_out,
+                                      lut->using_trivial_lwe_indexes,
+                                      num_radix_blocks, big_lwe_dimension + 1);
+
+    /// Synchronize all GPUs
+    for (uint i = 0; i < active_gpu_count; i++) {
+      cuda_synchronize_stream(streams[i], gpu_indexes[i]);
+    }
+  }
 }
 
 // Rotates the slice in-place such that the first mid elements of the slice move
@@ -308,7 +433,6 @@ void generate_device_accumulator_bivariate(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t message_modulus,
     uint32_t carry_modulus, std::function<Torus(Torus, Torus)> f) {
 
-  cudaSetDevice(gpu_index);
   // host lut
   Torus *h_lut =
       (Torus *)malloc((glwe_dimension + 1) * polynomial_size * sizeof(Torus));
@@ -317,15 +441,15 @@ void generate_device_accumulator_bivariate(
   generate_lookup_table_bivariate<Torus>(h_lut, glwe_dimension, polynomial_size,
                                          message_modulus, carry_modulus, f);
 
-  // copy host lut and lut_indexes to device
+  // copy host lut and lut_indexes_vec to device
+  cuda_synchronize_stream(stream, gpu_index);
   cuda_memcpy_async_to_gpu(acc_bivariate, h_lut,
                            (glwe_dimension + 1) * polynomial_size *
                                sizeof(Torus),
                            stream, gpu_index);
 
-  // Release memory when possible
-  cuda_stream_add_callback(stream, gpu_index, host_free_on_stream_callback,
-                           h_lut);
+  cuda_synchronize_stream(stream, gpu_index);
+  free(h_lut);
 }
 
 /*
@@ -341,7 +465,6 @@ void generate_device_accumulator_bivariate_with_factor(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t message_modulus,
     uint32_t carry_modulus, std::function<Torus(Torus, Torus)> f, int factor) {
 
-  cudaSetDevice(gpu_index);
   // host lut
   Torus *h_lut =
       (Torus *)malloc((glwe_dimension + 1) * polynomial_size * sizeof(Torus));
@@ -351,15 +474,15 @@ void generate_device_accumulator_bivariate_with_factor(
       h_lut, glwe_dimension, polynomial_size, message_modulus, carry_modulus, f,
       factor);
 
-  // copy host lut and lut_indexes to device
+  cuda_synchronize_stream(stream, gpu_index);
+  // copy host lut and lut_indexes_vec to device
   cuda_memcpy_async_to_gpu(acc_bivariate, h_lut,
                            (glwe_dimension + 1) * polynomial_size *
                                sizeof(Torus),
                            stream, gpu_index);
 
-  // Release memory when possible
-  cuda_stream_add_callback(stream, gpu_index, host_free_on_stream_callback,
-                           h_lut);
+  cuda_synchronize_stream(stream, gpu_index);
+  free(h_lut);
 }
 
 /*
@@ -377,7 +500,6 @@ void generate_device_accumulator(cudaStream_t stream, uint32_t gpu_index,
                                  uint32_t carry_modulus,
                                  std::function<Torus(Torus)> f) {
 
-  cudaSetDevice(gpu_index);
   // host lut
   Torus *h_lut =
       (Torus *)malloc((glwe_dimension + 1) * polynomial_size * sizeof(Torus));
@@ -386,32 +508,70 @@ void generate_device_accumulator(cudaStream_t stream, uint32_t gpu_index,
   generate_lookup_table<Torus>(h_lut, glwe_dimension, polynomial_size,
                                message_modulus, carry_modulus, f);
 
-  // copy host lut and lut_indexes to device
+  cuda_synchronize_stream(stream, gpu_index);
+  // copy host lut and lut_indexes_vec to device
   cuda_memcpy_async_to_gpu(
       acc, h_lut, (glwe_dimension + 1) * polynomial_size * sizeof(Torus),
       stream, gpu_index);
 
-  // Release memory when possible
-  cuda_stream_add_callback(stream, gpu_index, host_free_on_stream_callback,
-                           h_lut);
+  cuda_synchronize_stream(stream, gpu_index);
+  free(h_lut);
 }
 
 template <typename Torus>
 void scratch_cuda_propagate_single_carry_kb_inplace(
-    cudaStream_t stream, uint32_t gpu_index,
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     int_sc_prop_memory<Torus> **mem_ptr, uint32_t num_radix_blocks,
     int_radix_params params, bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
-  *mem_ptr = new int_sc_prop_memory<Torus>(
-      stream, gpu_index, params, num_radix_blocks, allocate_gpu_memory);
+  *mem_ptr =
+      new int_sc_prop_memory<Torus>(streams, gpu_indexes, gpu_count, params,
+                                    num_radix_blocks, allocate_gpu_memory);
+}
+
+template <typename Torus>
+void host_compute_prefix_sum_hillis_steele(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *step_output, Torus *generates_or_propagates, int_radix_params params,
+    int_radix_lut<Torus> *luts, void **bsks, Torus **ksks,
+    uint32_t num_blocks) {
+
+  auto glwe_dimension = params.glwe_dimension;
+  auto polynomial_size = params.polynomial_size;
+  auto big_lwe_size = glwe_dimension * polynomial_size + 1;
+  auto big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
+
+  int num_steps = ceil(log2((double)num_blocks));
+  int space = 1;
+  cuda_memcpy_async_gpu_to_gpu(step_output, generates_or_propagates,
+                               big_lwe_size_bytes * num_blocks, streams[0],
+                               gpu_indexes[0]);
+
+  for (int step = 0; step < num_steps; step++) {
+    if (space > num_blocks - 1)
+      PANIC("Cuda error: step output is going out of bounds in Hillis Steele "
+            "propagation")
+    auto cur_blocks = &step_output[space * big_lwe_size];
+    auto prev_blocks = generates_or_propagates;
+    int cur_total_blocks = num_blocks - space;
+
+    integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+        streams, gpu_indexes, gpu_count, cur_blocks, cur_blocks, prev_blocks,
+        bsks, ksks, cur_total_blocks, luts, luts->params.message_modulus);
+
+    cuda_memcpy_async_gpu_to_gpu(
+        &generates_or_propagates[space * big_lwe_size], cur_blocks,
+        big_lwe_size_bytes * cur_total_blocks, streams[0], gpu_indexes[0]);
+    space *= 2;
+  }
 }
 
 template <typename Torus>
 void host_propagate_single_carry(cudaStream_t *streams, uint32_t *gpu_indexes,
                                  uint32_t gpu_count, Torus *lwe_array,
-                                 int_sc_prop_memory<Torus> *mem, void *bsk,
-                                 Torus *ksk, uint32_t num_blocks) {
+                                 Torus *carry_out, Torus *input_carries,
+                                 int_sc_prop_memory<Torus> *mem, void **bsks,
+                                 Torus **ksks, uint32_t num_blocks) {
   auto params = mem->params;
   auto glwe_dimension = params.glwe_dimension;
   auto polynomial_size = params.polynomial_size;
@@ -426,58 +586,58 @@ void host_propagate_single_carry(cudaStream_t *streams, uint32_t *gpu_indexes,
   auto message_acc = mem->message_acc;
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, generates_or_propagates, lwe_array, bsk,
-      ksk, num_blocks, luts_array);
+      streams, gpu_indexes, gpu_count, generates_or_propagates, lwe_array, bsks,
+      ksks, num_blocks, luts_array);
 
   // compute prefix sum with hillis&steele
+  host_compute_prefix_sum_hillis_steele(
+      streams, gpu_indexes, gpu_count, step_output, generates_or_propagates,
+      params, luts_carry_propagation_sum, bsks, ksks, num_blocks);
 
-  int num_steps = ceil(log2((double)num_blocks));
-  int space = 1;
-  cudaSetDevice(gpu_indexes[0]);
-  cuda_memcpy_async_gpu_to_gpu(step_output, generates_or_propagates,
-                               big_lwe_size_bytes * num_blocks, streams[0],
-                               gpu_indexes[0]);
-
-  for (int step = 0; step < num_steps; step++) {
-    auto cur_blocks = &step_output[space * big_lwe_size];
-    auto prev_blocks = generates_or_propagates;
-    int cur_total_blocks = num_blocks - space;
-
-    integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, cur_blocks, cur_blocks, prev_blocks,
-        bsk, ksk, cur_total_blocks, luts_carry_propagation_sum,
-        luts_carry_propagation_sum->params.message_modulus);
-
-    cudaSetDevice(gpu_indexes[0]);
-    cuda_memcpy_async_gpu_to_gpu(
-        &generates_or_propagates[space * big_lwe_size], cur_blocks,
-        big_lwe_size_bytes * cur_total_blocks, streams[0], gpu_indexes[0]);
-    space *= 2;
-  }
-
-  cudaSetDevice(gpu_indexes[0]);
   host_radix_blocks_rotate_right(streams, gpu_indexes, gpu_count, step_output,
                                  generates_or_propagates, 1, num_blocks,
                                  big_lwe_size);
+  if (carry_out != nullptr) {
+    cuda_memcpy_async_gpu_to_gpu(carry_out, step_output, big_lwe_size_bytes,
+                                 streams[0], gpu_indexes[0]);
+  }
   cuda_memset_async(step_output, 0, big_lwe_size_bytes, streams[0],
                     gpu_indexes[0]);
 
+  if (input_carries != nullptr) {
+    cuda_memcpy_async_gpu_to_gpu(input_carries, step_output,
+                                 big_lwe_size_bytes * num_blocks, streams[0],
+                                 gpu_indexes[0]);
+  }
+
   host_addition(streams[0], gpu_indexes[0], lwe_array, lwe_array, step_output,
                 glwe_dimension * polynomial_size, num_blocks);
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array, lwe_array, bsk, ksk,
+      streams, gpu_indexes, gpu_count, lwe_array, lwe_array, bsks, ksks,
       num_blocks, message_acc);
 }
 
+template <typename Torus>
+void host_generate_last_block_inner_propagation(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *last_block_inner_propagation, Torus *lhs, Torus *rhs,
+    int_last_block_inner_propagate_memory<Torus> *mem, void **bsks,
+    Torus **ksks) {
+
+  integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+      streams, gpu_indexes, gpu_count, last_block_inner_propagation, lhs, rhs,
+      bsks, ksks, 1, mem->last_block_inner_propagation_lut,
+      mem->params.message_modulus);
+}
+
 template <typename Torus>
 void host_propagate_single_sub_borrow(cudaStream_t *streams,
                                       uint32_t *gpu_indexes, uint32_t gpu_count,
                                       Torus *overflowed, Torus *lwe_array,
-                                      int_single_borrow_prop_memory<Torus> *mem,
-                                      void *bsk, Torus *ksk,
+                                      int_overflowing_sub_memory<Torus> *mem,
+                                      void **bsks, Torus **ksks,
                                       uint32_t num_blocks) {
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem->params;
   auto glwe_dimension = params.glwe_dimension;
   auto polynomial_size = params.polynomial_size;
@@ -492,31 +652,13 @@ void host_propagate_single_sub_borrow(cudaStream_t *streams,
   auto message_acc = mem->message_acc;
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, generates_or_propagates, lwe_array, bsk,
-      ksk, num_blocks, luts_array);
+      streams, gpu_indexes, gpu_count, generates_or_propagates, lwe_array, bsks,
+      ksks, num_blocks, luts_array);
 
   // compute prefix sum with hillis&steele
-  int num_steps = ceil(log2((double)num_blocks));
-  int space = 1;
-  cuda_memcpy_async_gpu_to_gpu(step_output, generates_or_propagates,
-                               big_lwe_size_bytes * num_blocks, streams[0],
-                               gpu_indexes[0]);
-
-  for (int step = 0; step < num_steps; step++) {
-    auto cur_blocks = &step_output[space * big_lwe_size];
-    auto prev_blocks = generates_or_propagates;
-    int cur_total_blocks = num_blocks - space;
-
-    integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, cur_blocks, cur_blocks, prev_blocks,
-        bsk, ksk, cur_total_blocks, luts_carry_propagation_sum,
-        luts_carry_propagation_sum->params.message_modulus);
-
-    cuda_memcpy_async_gpu_to_gpu(
-        &generates_or_propagates[space * big_lwe_size], cur_blocks,
-        big_lwe_size_bytes * cur_total_blocks, streams[0], gpu_indexes[0]);
-    space *= 2;
-  }
+  host_compute_prefix_sum_hillis_steele<Torus>(
+      streams, gpu_indexes, gpu_count, step_output, generates_or_propagates,
+      params, luts_carry_propagation_sum, bsks, ksks, num_blocks);
 
   cuda_memcpy_async_gpu_to_gpu(
       overflowed, &generates_or_propagates[big_lwe_size * (num_blocks - 1)],
@@ -532,54 +674,52 @@ void host_propagate_single_sub_borrow(cudaStream_t *streams,
                    step_output, glwe_dimension * polynomial_size, num_blocks);
 
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array, lwe_array, bsk, ksk,
+      streams, gpu_indexes, gpu_count, lwe_array, lwe_array, bsks, ksks,
       num_blocks, message_acc);
 }
 
 /*
  * input_blocks: input radix ciphertext propagation will happen inplace
  * acc_message_carry: list of two lut s, [(message_acc), (carry_acc)]
- * lut_indexes_message_carry: lut_indexes for message and carry, should always
- * be  {0, 1} small_lwe_vector: output of keyswitch should have size = 2 *
- * (lwe_dimension + 1) * sizeof(Torus) big_lwe_vector: output of pbs should have
- *     size = 2 * (glwe_dimension * polynomial_size + 1) * sizeof(Torus)
+ * lut_indexes_message_carry: lut_indexes_vec for message and carry, should
+ * always be  {0, 1} small_lwe_vector: output of keyswitch should have size = 2
+ * * (lwe_dimension + 1) * sizeof(Torus) big_lwe_vector: output of pbs should
+ * have size = 2 * (glwe_dimension * polynomial_size + 1) * sizeof(Torus)
  */
-template <typename Torus, typename STorus, class params>
+template <typename Torus>
 void host_full_propagate_inplace(cudaStream_t *streams, uint32_t *gpu_indexes,
                                  uint32_t gpu_count, Torus *input_blocks,
                                  int_fullprop_buffer<Torus> *mem_ptr,
-                                 Torus *ksk, void *bsk, uint32_t lwe_dimension,
-                                 uint32_t glwe_dimension,
-                                 uint32_t polynomial_size, uint32_t ks_base_log,
-                                 uint32_t ks_level, uint32_t pbs_base_log,
-                                 uint32_t pbs_level, uint32_t grouping_factor,
+                                 Torus **ksks, void **bsks,
                                  uint32_t num_blocks) {
+  auto params = mem_ptr->lut->params;
 
-  cudaSetDevice(gpu_indexes[0]);
-  int big_lwe_size = (glwe_dimension * polynomial_size + 1);
-  int small_lwe_size = (lwe_dimension + 1);
+  int big_lwe_size = (params.glwe_dimension * params.polynomial_size + 1);
+  int small_lwe_size = (params.small_lwe_dimension + 1);
 
   for (int i = 0; i < num_blocks; i++) {
     auto cur_input_block = &input_blocks[i * big_lwe_size];
 
-    cuda_keyswitch_lwe_ciphertext_vector<Torus>(
-        streams[0], gpu_indexes[0], mem_ptr->tmp_small_lwe_vector,
-        mem_ptr->lwe_indexes, cur_input_block, mem_ptr->lwe_indexes, ksk,
-        polynomial_size * glwe_dimension, lwe_dimension, ks_base_log, ks_level,
-        1);
+    /// Since the keyswitch is done on one input only, use only 1 GPU
+    execute_keyswitch_async<Torus>(
+        streams, gpu_indexes, 1, mem_ptr->tmp_small_lwe_vector,
+        mem_ptr->lut->lwe_trivial_indexes, cur_input_block,
+        mem_ptr->lut->lwe_trivial_indexes, ksks, params.big_lwe_dimension,
+        params.small_lwe_dimension, params.ks_base_log, params.ks_level, 1);
 
     cuda_memcpy_async_gpu_to_gpu(&mem_ptr->tmp_small_lwe_vector[small_lwe_size],
                                  mem_ptr->tmp_small_lwe_vector,
                                  small_lwe_size * sizeof(Torus), streams[0],
                                  gpu_indexes[0]);
 
-    execute_pbs<Torus>(
+    execute_pbs_async<Torus>(
         streams, gpu_indexes, 1, mem_ptr->tmp_big_lwe_vector,
-        mem_ptr->lwe_indexes, mem_ptr->lut_buffer, mem_ptr->lut_indexes,
-        mem_ptr->tmp_small_lwe_vector, mem_ptr->lwe_indexes, bsk,
-        mem_ptr->pbs_buffer, glwe_dimension, lwe_dimension, polynomial_size,
-        pbs_base_log, pbs_level, grouping_factor, 2, 2, 0,
-        cuda_get_max_shared_memory(gpu_indexes[0]), mem_ptr->pbs_type);
+        mem_ptr->lut->lwe_trivial_indexes, mem_ptr->lut->lut_vec,
+        mem_ptr->lut->lut_indexes_vec, mem_ptr->tmp_small_lwe_vector,
+        mem_ptr->lut->lwe_trivial_indexes, bsks, mem_ptr->lut->buffer,
+        params.glwe_dimension, params.small_lwe_dimension,
+        params.polynomial_size, params.pbs_base_log, params.pbs_level,
+        params.grouping_factor, 2, params.pbs_type);
 
     cuda_memcpy_async_gpu_to_gpu(cur_input_block, mem_ptr->tmp_big_lwe_vector,
                                  big_lwe_size * sizeof(Torus), streams[0],
@@ -590,108 +730,20 @@ void host_full_propagate_inplace(cudaStream_t *streams, uint32_t *gpu_indexes,
       host_addition(streams[0], gpu_indexes[0], next_input_block,
                     next_input_block,
                     &mem_ptr->tmp_big_lwe_vector[big_lwe_size],
-                    glwe_dimension * polynomial_size, 1);
+                    params.big_lwe_dimension, 1);
     }
   }
 }
 
 template <typename Torus>
-void scratch_cuda_full_propagation(
-    cudaStream_t stream, uint32_t gpu_index,
-    int_fullprop_buffer<Torus> **mem_ptr, uint32_t lwe_dimension,
-    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t pbs_level,
-    uint32_t grouping_factor, uint32_t num_radix_blocks,
-    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
-    bool allocate_gpu_memory) {
+void scratch_cuda_full_propagation(cudaStream_t *streams, uint32_t *gpu_indexes,
+                                   uint32_t gpu_count,
+                                   int_fullprop_buffer<Torus> **mem_ptr,
+                                   int_radix_params params,
+                                   bool allocate_gpu_memory) {
 
-  int8_t *pbs_buffer;
-  execute_scratch_pbs<Torus>(
-      stream, gpu_index, &pbs_buffer, glwe_dimension, lwe_dimension,
-      polynomial_size, pbs_level, grouping_factor, num_radix_blocks,
-      cuda_get_max_shared_memory(gpu_index), pbs_type, allocate_gpu_memory);
-
-  // LUT
-  Torus *lut_buffer;
-  if (allocate_gpu_memory) {
-    // LUT is used as a trivial encryption, so we only allocate memory for the
-    // body
-    Torus lut_buffer_size =
-        2 * (glwe_dimension + 1) * polynomial_size * sizeof(Torus);
-
-    lut_buffer = (Torus *)cuda_malloc_async(lut_buffer_size, stream, gpu_index);
-
-    // LUTs
-    auto lut_f_message = [message_modulus](Torus x) -> Torus {
-      return x % message_modulus;
-    };
-    auto lut_f_carry = [message_modulus](Torus x) -> Torus {
-      return x / message_modulus;
-    };
-
-    //
-    Torus *lut_buffer_message = lut_buffer;
-    Torus *lut_buffer_carry =
-        lut_buffer + (glwe_dimension + 1) * polynomial_size;
-
-    generate_device_accumulator<Torus>(
-        stream, gpu_index, lut_buffer_message, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, lut_f_message);
-
-    generate_device_accumulator<Torus>(
-        stream, gpu_index, lut_buffer_carry, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, lut_f_carry);
-  }
-
-  Torus *lut_indexes;
-  if (allocate_gpu_memory) {
-    lut_indexes =
-        (Torus *)cuda_malloc_async(2 * sizeof(Torus), stream, gpu_index);
-
-    Torus h_lut_indexes[2] = {0, 1};
-    cuda_memcpy_async_to_gpu(lut_indexes, h_lut_indexes, 2 * sizeof(Torus),
-                             stream, gpu_index);
-  }
-
-  Torus *lwe_indexes;
-  if (allocate_gpu_memory) {
-    Torus lwe_indexes_size = num_radix_blocks * sizeof(Torus);
-
-    lwe_indexes =
-        (Torus *)cuda_malloc_async(lwe_indexes_size, stream, gpu_index);
-    Torus *h_lwe_indexes = (Torus *)malloc(lwe_indexes_size);
-    for (int i = 0; i < num_radix_blocks; i++)
-      h_lwe_indexes[i] = i;
-    cuda_memcpy_async_to_gpu(lwe_indexes, h_lwe_indexes, lwe_indexes_size,
-                             stream, gpu_index);
-    cuda_stream_add_callback(stream, gpu_index, host_free_on_stream_callback,
-                             h_lwe_indexes);
-  }
-
-  // Temporary arrays
-  Torus *small_lwe_vector;
-  Torus *big_lwe_vector;
-  if (allocate_gpu_memory) {
-    Torus small_vector_size = 2 * (lwe_dimension + 1) * sizeof(Torus);
-    Torus big_vector_size =
-        2 * (glwe_dimension * polynomial_size + 1) * sizeof(Torus);
-
-    small_lwe_vector =
-        (Torus *)cuda_malloc_async(small_vector_size, stream, gpu_index);
-    big_lwe_vector =
-        (Torus *)cuda_malloc_async(big_vector_size, stream, gpu_index);
-  }
-
-  *mem_ptr = new int_fullprop_buffer<Torus>;
-
-  (*mem_ptr)->pbs_type = pbs_type;
-  (*mem_ptr)->pbs_buffer = pbs_buffer;
-
-  (*mem_ptr)->lut_buffer = lut_buffer;
-  (*mem_ptr)->lut_indexes = lut_indexes;
-  (*mem_ptr)->lwe_indexes = lwe_indexes;
-
-  (*mem_ptr)->tmp_small_lwe_vector = small_lwe_vector;
-  (*mem_ptr)->tmp_big_lwe_vector = big_lwe_vector;
+  *mem_ptr = new int_fullprop_buffer<Torus>(streams, gpu_indexes, gpu_count,
+                                            params, allocate_gpu_memory);
 }
 
 // (lwe_dimension+1) threads
@@ -736,11 +788,12 @@ __host__ void pack_blocks(cudaStream_t stream, uint32_t gpu_index,
                           Torus *lwe_array_out, Torus *lwe_array_in,
                           uint32_t lwe_dimension, uint32_t num_radix_blocks,
                           uint32_t factor) {
+  if (num_radix_blocks == 0)
+    return;
   cudaSetDevice(gpu_index);
-
   int num_blocks = 0, num_threads = 0;
   int num_entries = (lwe_dimension + 1);
-  getNumBlocksAndThreads(num_entries, 512, num_blocks, num_threads);
+  getNumBlocksAndThreads(num_entries, 1024, num_blocks, num_threads);
   device_pack_blocks<<<num_blocks, num_threads, 0, stream>>>(
       lwe_array_out, lwe_array_in, lwe_dimension, num_radix_blocks, factor);
 }
@@ -800,22 +853,22 @@ create_trivial_radix(cudaStream_t stream, uint32_t gpu_index,
 template <typename Torus>
 __host__ void extract_n_bits(cudaStream_t *streams, uint32_t *gpu_indexes,
                              uint32_t gpu_count, Torus *lwe_array_out,
-                             Torus *lwe_array_in, void *bsk, Torus *ksk,
+                             Torus *lwe_array_in, void **bsks, Torus **ksks,
                              uint32_t num_radix_blocks, uint32_t bits_per_block,
                              int_bit_extract_luts_buffer<Torus> *bit_extract) {
 
   integer_radix_apply_univariate_lookup_table_kb(
-      streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in, bsk, ksk,
+      streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in, bsks, ksks,
       num_radix_blocks * bits_per_block, bit_extract->lut);
 }
 
 template <typename Torus>
-__host__ void reduce_signs(cudaStream_t *streams, uint32_t *gpu_indexes,
-                           uint32_t gpu_count, Torus *signs_array_out,
-                           Torus *signs_array_in,
-                           int_comparison_buffer<Torus> *mem_ptr,
-                           std::function<Torus(Torus)> sign_handler_f,
-                           void *bsk, Torus *ksk, uint32_t num_sign_blocks) {
+__host__ void
+reduce_signs(cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+             Torus *signs_array_out, Torus *signs_array_in,
+             int_comparison_buffer<Torus> *mem_ptr,
+             std::function<Torus(Torus)> sign_handler_f, void **bsks,
+             Torus **ksks, uint32_t num_sign_blocks) {
 
   cudaSetDevice(gpu_indexes[0]);
   auto diff_buffer = mem_ptr->diff_buffer;
@@ -845,14 +898,16 @@ __host__ void reduce_signs(cudaStream_t *streams, uint32_t *gpu_indexes,
   if (num_sign_blocks > 2) {
     auto lut = diff_buffer->reduce_signs_lut;
     generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, reduce_two_orderings_function);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        reduce_two_orderings_function);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     while (num_sign_blocks > 2) {
       pack_blocks(streams[0], gpu_indexes[0], signs_b, signs_a,
                   big_lwe_dimension, num_sign_blocks, 4);
       integer_radix_apply_univariate_lookup_table_kb(
-          streams, gpu_indexes, gpu_count, signs_a, signs_b, bsk, ksk,
+          streams, gpu_indexes, gpu_count, signs_a, signs_b, bsks, ksks,
           num_sign_blocks / 2, lut);
 
       auto last_block_signs_b =
@@ -877,14 +932,16 @@ __host__ void reduce_signs(cudaStream_t *streams, uint32_t *gpu_indexes,
 
     auto lut = diff_buffer->reduce_signs_lut;
     generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, final_lut_f);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        final_lut_f);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     pack_blocks(streams[0], gpu_indexes[0], signs_b, signs_a, big_lwe_dimension,
                 2, 4);
     integer_radix_apply_univariate_lookup_table_kb(streams, gpu_indexes,
                                                    gpu_count, signs_array_out,
-                                                   signs_b, bsk, ksk, 1, lut);
+                                                   signs_b, bsks, ksks, 1, lut);
 
   } else {
 
@@ -895,40 +952,74 @@ __host__ void reduce_signs(cudaStream_t *streams, uint32_t *gpu_indexes,
 
     auto lut = mem_ptr->diff_buffer->reduce_signs_lut;
     generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, final_lut_f);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        final_lut_f);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     integer_radix_apply_univariate_lookup_table_kb(streams, gpu_indexes,
                                                    gpu_count, signs_array_out,
-                                                   signs_a, bsk, ksk, 1, lut);
+                                                   signs_a, bsks, ksks, 1, lut);
   }
 }
 
 template <typename Torus>
 void scratch_cuda_apply_univariate_lut_kb(
-    cudaStream_t stream, uint32_t gpu_index, int_radix_lut<Torus> **mem_ptr,
-    Torus *input_lut, uint32_t num_radix_blocks, int_radix_params params,
-    bool allocate_gpu_memory) {
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_radix_lut<Torus> **mem_ptr, Torus *input_lut, uint32_t num_radix_blocks,
+    int_radix_params params, bool allocate_gpu_memory) {
 
-  *mem_ptr = new int_radix_lut<Torus>(stream, gpu_index, params, 1,
-                                      num_radix_blocks, allocate_gpu_memory);
-  cuda_memcpy_async_to_gpu((*mem_ptr)->lut, input_lut,
+  *mem_ptr = new int_radix_lut<Torus>(streams, gpu_indexes, gpu_count, params,
+                                      1, num_radix_blocks, allocate_gpu_memory);
+  // It is safe to do this copy on GPU 0, because all LUTs always reside on GPU
+  // 0
+  cuda_memcpy_async_to_gpu((*mem_ptr)->get_lut(gpu_indexes[0], 0), input_lut,
                            (params.glwe_dimension + 1) *
                                params.polynomial_size * sizeof(Torus),
-                           stream, gpu_index);
+                           streams[0], gpu_indexes[0]);
+  (*mem_ptr)->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 }
 
 template <typename Torus>
 void host_apply_univariate_lut_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
                                   uint32_t gpu_count, Torus *radix_lwe_out,
                                   Torus *radix_lwe_in,
-                                  int_radix_lut<Torus> *mem, Torus *ksk,
-                                  void *bsk, uint32_t num_blocks) {
+                                  int_radix_lut<Torus> *mem, Torus **ksks,
+                                  void **bsks, uint32_t num_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, radix_lwe_out, radix_lwe_in, bsk, ksk,
+      streams, gpu_indexes, gpu_count, radix_lwe_out, radix_lwe_in, bsks, ksks,
       num_blocks, mem);
 }
 
+template <typename Torus>
+void scratch_cuda_apply_bivariate_lut_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_radix_lut<Torus> **mem_ptr, Torus *input_lut, uint32_t num_radix_blocks,
+    int_radix_params params, bool allocate_gpu_memory) {
+
+  *mem_ptr = new int_radix_lut<Torus>(streams, gpu_indexes, gpu_count, params,
+                                      1, num_radix_blocks, allocate_gpu_memory);
+  // It is safe to do this copy on GPU 0, because all LUTs always reside on GPU
+  // 0
+  cuda_memcpy_async_to_gpu((*mem_ptr)->get_lut(gpu_indexes[0], 0), input_lut,
+                           (params.glwe_dimension + 1) *
+                               params.polynomial_size * sizeof(Torus),
+                           streams[0], gpu_indexes[0]);
+  (*mem_ptr)->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
+}
+
+template <typename Torus>
+void host_apply_bivariate_lut_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
+                                 uint32_t gpu_count, Torus *radix_lwe_out,
+                                 Torus *radix_lwe_in_1, Torus *radix_lwe_in_2,
+                                 int_radix_lut<Torus> *mem, Torus **ksks,
+                                 void **bsks, uint32_t num_blocks,
+                                 uint32_t shift) {
+
+  integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+      streams, gpu_indexes, gpu_count, radix_lwe_out, radix_lwe_in_1,
+      radix_lwe_in_2, bsks, ksks, num_blocks, mem, shift);
+}
+
 #endif // TFHE_RS_INTERNAL_INTEGER_CUH

backends/tfhe-cuda-backend/cuda/src/integer/multiplication.cu

@@ -66,12 +66,12 @@ void generate_ids_update_degrees(int *terms_degree, size_t *h_lwe_idx_in,
  * the integer radix multiplication in keyswitch->bootstrap order.
  */
 void scratch_cuda_integer_mult_radix_ciphertext_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr,
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
     uint32_t message_modulus, uint32_t carry_modulus, uint32_t glwe_dimension,
     uint32_t lwe_dimension, uint32_t polynomial_size, uint32_t pbs_base_log,
     uint32_t pbs_level, uint32_t ks_base_log, uint32_t ks_level,
     uint32_t grouping_factor, uint32_t num_radix_blocks, PBS_TYPE pbs_type,
-    uint32_t max_shared_memory, bool allocate_gpu_memory) {
+    bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           polynomial_size * glwe_dimension, lwe_dimension,
@@ -87,7 +87,7 @@ void scratch_cuda_integer_mult_radix_ciphertext_kb_64(
   case 8192:
   case 16384:
     scratch_cuda_integer_mult_radix_ciphertext_kb<uint64_t>(
-        static_cast<cudaStream_t>(stream), gpu_index,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         (int_mul_memory<uint64_t> **)mem_ptr, num_radix_blocks, params,
         allocate_gpu_memory);
     break;
@@ -123,76 +123,69 @@ void scratch_cuda_integer_mult_radix_ciphertext_kb_64(
  * - 'num_blocks' is the number of big lwe ciphertext blocks inside radix
  * ciphertext
  * - 'pbs_type' selects which PBS implementation should be used
- * - 'max_shared_memory' maximum shared memory per cuda block
  */
 void cuda_integer_mult_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    void *radix_lwe_out, void *radix_lwe_left, void *radix_lwe_right, void *bsk,
-    void *ksk, int8_t *mem_ptr, uint32_t polynomial_size, uint32_t num_blocks) {
+    void *radix_lwe_out, void *radix_lwe_left, void *radix_lwe_right,
+    void **bsks, void **ksks, int8_t *mem_ptr, uint32_t polynomial_size,
+    uint32_t num_blocks) {
 
   switch (polynomial_size) {
   case 256:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<256>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<256>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 512:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<512>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<512>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 1024:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<1024>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<1024>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 2048:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<2048>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<2048>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 4096:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<4096>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<4096>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 8192:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<8192>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<8192>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 16384:
-    host_integer_mult_radix_kb<uint64_t, int64_t, AmortizedDegree<16384>>(
+    host_integer_mult_radix_kb<uint64_t, AmortizedDegree<16384>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
         static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), (int_mul_memory<uint64_t> *)mem_ptr,
-        num_blocks);
+        static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks),
+        (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   default:
     PANIC("Cuda error (integer multiplication): unsupported polynomial size. "
@@ -200,37 +193,38 @@ void cuda_integer_mult_radix_ciphertext_kb_64(
   }
 }
 
-void cleanup_cuda_integer_mult(void *stream, uint32_t gpu_index,
-                               int8_t **mem_ptr_void) {
+void cleanup_cuda_integer_mult(void **streams, uint32_t *gpu_indexes,
+                               uint32_t gpu_count, int8_t **mem_ptr_void) {
 
   int_mul_memory<uint64_t> *mem_ptr =
       (int_mul_memory<uint64_t> *)(*mem_ptr_void);
 
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }
 
-void scratch_cuda_integer_radix_sum_ciphertexts_vec_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t lwe_dimension, uint32_t ks_level,
-    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
-    uint32_t grouping_factor, uint32_t num_blocks_in_radix,
-    uint32_t max_num_radix_in_vec, uint32_t message_modulus,
-    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+void scratch_cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t lwe_dimension,
+    uint32_t ks_level, uint32_t ks_base_log, uint32_t pbs_level,
+    uint32_t pbs_base_log, uint32_t grouping_factor,
+    uint32_t num_blocks_in_radix, uint32_t max_num_radix_in_vec,
+    uint32_t message_modulus, uint32_t carry_modulus, PBS_TYPE pbs_type,
+    bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           glwe_dimension * polynomial_size, lwe_dimension,
                           ks_level, ks_base_log, pbs_level, pbs_base_log,
                           grouping_factor, message_modulus, carry_modulus);
-  scratch_cuda_integer_sum_ciphertexts_vec_kb<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
+  scratch_cuda_integer_partial_sum_ciphertexts_vec_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_sum_ciphertexts_vec_memory<uint64_t> **)mem_ptr, num_blocks_in_radix,
       max_num_radix_in_vec, params, allocate_gpu_memory);
 }
 
-void cuda_integer_radix_sum_ciphertexts_vec_kb_64(
+void cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *radix_lwe_out, void *radix_lwe_vec, uint32_t num_radix_in_vec,
-    int8_t *mem_ptr, void *bsk, void *ksk, uint32_t num_blocks_in_radix) {
+    int8_t *mem_ptr, void **bsks, void **ksks, uint32_t num_blocks_in_radix) {
 
   auto mem = (int_sum_ciphertexts_vec_memory<uint64_t> *)mem_ptr;
 
@@ -243,52 +237,51 @@ void cuda_integer_radix_sum_ciphertexts_vec_kb_64(
 
   switch (mem->params.polynomial_size) {
   case 512:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<512>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<512>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   case 1024:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<1024>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t,
+                                                AmortizedDegree<1024>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   case 2048:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<2048>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t,
+                                                AmortizedDegree<2048>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   case 4096:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<4096>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t,
+                                                AmortizedDegree<4096>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   case 8192:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<8192>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t,
+                                                AmortizedDegree<8192>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   case 16384:
-    host_integer_sum_ciphertexts_vec_kb<uint64_t, AmortizedDegree<16384>>(
+    host_integer_partial_sum_ciphertexts_vec_kb<uint64_t,
+                                                AmortizedDegree<16384>>(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks_in_radix,
-        num_radix_in_vec);
+        static_cast<uint64_t *>(radix_lwe_vec), terms_degree, bsks,
+        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec);
     break;
   default:
     PANIC("Cuda error (integer multiplication): unsupported polynomial size. "
@@ -298,11 +291,11 @@ void cuda_integer_radix_sum_ciphertexts_vec_kb_64(
   free(terms_degree);
 }
 
-void cleanup_cuda_integer_radix_sum_ciphertexts_vec(void *stream,
-                                                    uint32_t gpu_index,
-                                                    int8_t **mem_ptr_void) {
+void cleanup_cuda_integer_radix_partial_sum_ciphertexts_vec(
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int8_t **mem_ptr_void) {
   int_sum_ciphertexts_vec_memory<uint64_t> *mem_ptr =
       (int_sum_ciphertexts_vec_memory<uint64_t> *)(*mem_ptr_void);
 
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/multiplication.cuh

@@ -8,11 +8,13 @@
 
 #include "crypto/keyswitch.cuh"
 #include "device.h"
+#include "helper_multi_gpu.h"
 #include "integer.h"
 #include "integer/integer.cuh"
 #include "linear_algebra.h"
 #include "programmable_bootstrap.h"
 #include "utils/helper.cuh"
+#include "utils/helper_multi_gpu.cuh"
 #include "utils/kernel_dimensions.cuh"
 #include <fstream>
 #include <iostream>
@@ -91,15 +93,11 @@ all_shifted_lhs_rhs(Torus *radix_lwe_left, Torus *lsb_ciphertext,
   }
 }
 
-template <typename Torus, sharedMemDegree SMD>
+template <typename Torus>
 __global__ void tree_add_chunks(Torus *result_blocks, Torus *input_blocks,
                                 uint32_t chunk_size, uint32_t block_size,
                                 uint32_t num_blocks) {
 
-  extern __shared__ int8_t sharedmem[];
-
-  Torus *result = (Torus *)sharedmem;
-
   size_t stride = blockDim.x;
   size_t chunk_id = blockIdx.x;
   size_t chunk_elem_size = chunk_size * num_blocks * block_size;
@@ -107,10 +105,7 @@ __global__ void tree_add_chunks(Torus *result_blocks, Torus *input_blocks,
   auto src_chunk = &input_blocks[chunk_id * chunk_elem_size];
   auto dst_radix = &result_blocks[chunk_id * radix_elem_size];
   size_t block_stride = blockIdx.y * block_size;
-  auto dst_block = &dst_radix[block_stride];
-
-  if constexpr (SMD == NOSM)
-    result = dst_block;
+  auto result = &dst_radix[block_stride];
 
   // init shared mem with first radix of chunk
   size_t tid = threadIdx.x;
@@ -125,18 +120,12 @@ __global__ void tree_add_chunks(Torus *result_blocks, Torus *input_blocks,
       result[i] += cur_src_radix[block_stride + i];
     }
   }
-
-  // put result from shared mem to global mem
-  if constexpr (SMD == FULLSM)
-    for (int i = tid; i < block_size; i += stride)
-      dst_block[i] = result[i];
 }
 
 template <typename Torus, class params>
 __global__ void fill_radix_from_lsb_msb(Torus *result_blocks, Torus *lsb_blocks,
                                         Torus *msb_blocks,
                                         uint32_t glwe_dimension,
-                                        uint32_t lsb_count, uint32_t msb_count,
                                         uint32_t num_blocks) {
   size_t big_lwe_dimension = glwe_dimension * params::degree + 1;
   size_t big_lwe_id = blockIdx.x;
@@ -180,41 +169,25 @@ __global__ void fill_radix_from_lsb_msb(Torus *result_blocks, Torus *lsb_blocks,
   }
 }
 template <typename Torus>
-__host__ void scratch_cuda_integer_sum_ciphertexts_vec_kb(
-    cudaStream_t stream, uint32_t gpu_index,
+__host__ void scratch_cuda_integer_partial_sum_ciphertexts_vec_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     int_sum_ciphertexts_vec_memory<Torus> **mem_ptr,
     uint32_t num_blocks_in_radix, uint32_t max_num_radix_in_vec,
     int_radix_params params, bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
-  size_t sm_size = (params.big_lwe_dimension + 1) * sizeof(Torus);
-  if (sm_size < cuda_get_max_shared_memory(gpu_index)) {
-    check_cuda_error(cudaFuncSetAttribute(
-        tree_add_chunks<Torus, FULLSM>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize, sm_size));
-    cudaFuncSetCacheConfig(tree_add_chunks<Torus, FULLSM>,
-                           cudaFuncCachePreferShared);
-    check_cuda_error(cudaGetLastError());
-  } else {
-    check_cuda_error(
-        cudaFuncSetAttribute(tree_add_chunks<Torus, NOSM>,
-                             cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
-    cudaFuncSetCacheConfig(tree_add_chunks<Torus, NOSM>, cudaFuncCachePreferL1);
-    check_cuda_error(cudaGetLastError());
-  }
   *mem_ptr = new int_sum_ciphertexts_vec_memory<Torus>(
-      stream, gpu_index, params, num_blocks_in_radix, max_num_radix_in_vec,
-      allocate_gpu_memory);
+      streams, gpu_indexes, gpu_count, params, num_blocks_in_radix,
+      max_num_radix_in_vec, allocate_gpu_memory);
 }
 
 template <typename Torus, class params>
-__host__ void host_integer_sum_ciphertexts_vec_kb(
+__host__ void host_integer_partial_sum_ciphertexts_vec_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
-    Torus *radix_lwe_out, Torus *terms, int *terms_degree, void *bsk,
-    uint64_t *ksk, int_sum_ciphertexts_vec_memory<uint64_t> *mem_ptr,
-    uint32_t num_blocks_in_radix, uint32_t num_radix_in_vec) {
+    Torus *radix_lwe_out, Torus *terms, int *terms_degree, void **bsks,
+    uint64_t **ksks, int_sum_ciphertexts_vec_memory<uint64_t> *mem_ptr,
+    uint32_t num_blocks_in_radix, uint32_t num_radix_in_vec,
+    int_radix_lut<Torus> *reused_lut = nullptr) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto new_blocks = mem_ptr->new_blocks;
   auto old_blocks = mem_ptr->old_blocks;
   auto small_lwe_vector = mem_ptr->small_lwe_vector;
@@ -225,11 +198,12 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
   auto message_modulus = mem_ptr->params.message_modulus;
   auto carry_modulus = mem_ptr->params.carry_modulus;
   auto num_blocks = num_blocks_in_radix;
-  auto big_lwe_size = mem_ptr->params.big_lwe_dimension + 1;
+  auto big_lwe_dimension = mem_ptr->params.big_lwe_dimension;
+  auto big_lwe_size = big_lwe_dimension + 1;
   auto glwe_dimension = mem_ptr->params.glwe_dimension;
   auto polynomial_size = mem_ptr->params.polynomial_size;
-  auto lwe_dimension = mem_ptr->params.small_lwe_dimension;
-  auto big_lwe_dimension = mem_ptr->params.big_lwe_dimension;
+  auto small_lwe_dimension = mem_ptr->params.small_lwe_dimension;
+  auto small_lwe_size = small_lwe_dimension + 1;
 
   if (old_blocks != terms) {
     cuda_memcpy_async_gpu_to_gpu(old_blocks, terms,
@@ -248,7 +222,48 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
   int32_t h_smart_copy_in[r * num_blocks];
   int32_t h_smart_copy_out[r * num_blocks];
 
-  auto max_shared_memory = cuda_get_max_shared_memory(gpu_indexes[0]);
+  /// Here it is important to query the default max shared memory on device 0
+  /// instead of cuda_get_max_shared_memory,
+  /// to avoid bugs with tree_add_chunks trying to use too much shared memory
+  int max_shared_memory = 0;
+  check_cuda_error(cudaDeviceGetAttribute(
+      &max_shared_memory, cudaDevAttrMaxSharedMemoryPerBlock, 0));
+
+  // create lut object for message and carry
+  // we allocate luts_message_carry in the host function (instead of scratch)
+  // to reduce average memory consumption
+  int_radix_lut<Torus> *luts_message_carry;
+  size_t ch_amount = r / chunk_size;
+  if (!ch_amount)
+    ch_amount++;
+  if (reused_lut == nullptr) {
+    luts_message_carry = new int_radix_lut<Torus>(
+        streams, gpu_indexes, gpu_count, mem_ptr->params, 2,
+        2 * ch_amount * num_blocks, true);
+  } else {
+    luts_message_carry = new int_radix_lut<Torus>(
+        streams, gpu_indexes, gpu_count, mem_ptr->params, 2,
+        2 * ch_amount * num_blocks, reused_lut);
+  }
+  auto message_acc = luts_message_carry->get_lut(gpu_indexes[0], 0);
+  auto carry_acc = luts_message_carry->get_lut(gpu_indexes[0], 1);
+
+  // define functions for each accumulator
+  auto lut_f_message = [message_modulus](Torus x) -> Torus {
+    return x % message_modulus;
+  };
+  auto lut_f_carry = [message_modulus](Torus x) -> Torus {
+    return x / message_modulus;
+  };
+
+  // generate accumulators
+  generate_device_accumulator<Torus>(
+      streams[0], gpu_indexes[0], message_acc, glwe_dimension, polynomial_size,
+      message_modulus, carry_modulus, lut_f_message);
+  generate_device_accumulator<Torus>(
+      streams[0], gpu_indexes[0], carry_acc, glwe_dimension, polynomial_size,
+      message_modulus, carry_modulus, lut_f_carry);
+  luts_message_carry->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
   while (r > 2) {
     size_t cur_total_blocks = r * num_blocks;
@@ -258,12 +273,9 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
     dim3 add_grid(ch_amount, num_blocks, 1);
     size_t sm_size = big_lwe_size * sizeof(Torus);
 
-    if (sm_size < max_shared_memory)
-      tree_add_chunks<Torus, FULLSM><<<add_grid, 512, sm_size, streams[0]>>>(
-          new_blocks, old_blocks, min(r, chunk_size), big_lwe_size, num_blocks);
-    else
-      tree_add_chunks<Torus, NOSM><<<add_grid, 512, 0, streams[0]>>>(
-          new_blocks, old_blocks, min(r, chunk_size), big_lwe_size, num_blocks);
+    cudaSetDevice(gpu_indexes[0]);
+    tree_add_chunks<Torus><<<add_grid, 512, 0, streams[0]>>>(
+        new_blocks, old_blocks, min(r, chunk_size), big_lwe_size, num_blocks);
 
     check_cuda_error(cudaGetLastError());
 
@@ -276,47 +288,22 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
         terms_degree, h_lwe_idx_in, h_lwe_idx_out, h_smart_copy_in,
         h_smart_copy_out, ch_amount, r, num_blocks, chunk_size, message_max,
         total_count, message_count, carry_count, sm_copy_count);
-
-    // create lut object for message and carry
-    // we allocate luts_message_carry in the host function (instead of scratch)
-    // to reduce average memory consumption
-    auto luts_message_carry = new int_radix_lut<Torus>(
-        streams[0], gpu_indexes[0], mem_ptr->params, 2, total_count, true);
-
-    auto message_acc = luts_message_carry->get_lut(0);
-    auto carry_acc = luts_message_carry->get_lut(1);
-
-    // define functions for each accumulator
-    auto lut_f_message = [message_modulus](Torus x) -> Torus {
-      return x % message_modulus;
-    };
-    auto lut_f_carry = [message_modulus](Torus x) -> Torus {
-      return x / message_modulus;
-    };
-
-    // generate accumulators
-    generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], message_acc, glwe_dimension,
-        polynomial_size, message_modulus, carry_modulus, lut_f_message);
-    generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], carry_acc, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, lut_f_carry);
-
+    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
     auto lwe_indexes_in = luts_message_carry->lwe_indexes_in;
     auto lwe_indexes_out = luts_message_carry->lwe_indexes_out;
+    luts_message_carry->set_lwe_indexes(streams[0], gpu_indexes[0],
+                                        h_lwe_idx_in, h_lwe_idx_out);
 
-    size_t copy_size = total_count * sizeof(Torus);
-    cuda_memcpy_async_to_gpu(lwe_indexes_in, h_lwe_idx_in, copy_size,
-                             streams[0], gpu_indexes[0]);
-    cuda_memcpy_async_to_gpu(lwe_indexes_out, h_lwe_idx_out, copy_size,
-                             streams[0], gpu_indexes[0]);
-    copy_size = sm_copy_count * sizeof(int32_t);
+    size_t copy_size = sm_copy_count * sizeof(int32_t);
     cuda_memcpy_async_to_gpu(d_smart_copy_in, h_smart_copy_in, copy_size,
                              streams[0], gpu_indexes[0]);
     cuda_memcpy_async_to_gpu(d_smart_copy_out, h_smart_copy_out, copy_size,
                              streams[0], gpu_indexes[0]);
 
-    smart_copy<<<sm_copy_count, 256, 0, streams[0]>>>(
+    // inside d_smart_copy_in there are only -1 values
+    // it's fine to call smart_copy with same pointer
+    // as source and destination
+    smart_copy<<<sm_copy_count, 1024, 0, streams[0]>>>(
         new_blocks, new_blocks, d_smart_copy_out, d_smart_copy_in,
         big_lwe_size);
     check_cuda_error(cudaGetLastError());
@@ -324,24 +311,102 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
     if (carry_count > 0)
       cuda_set_value_async<Torus>(
           streams[0], gpu_indexes[0],
-          luts_message_carry->get_lut_indexes(message_count), 1, carry_count);
+          luts_message_carry->get_lut_indexes(gpu_indexes[0], message_count), 1,
+          carry_count);
 
-    cuda_keyswitch_lwe_ciphertext_vector(
-        streams[0], gpu_indexes[0], small_lwe_vector, lwe_indexes_in,
-        new_blocks, lwe_indexes_in, ksk, polynomial_size * glwe_dimension,
-        lwe_dimension, mem_ptr->params.ks_base_log, mem_ptr->params.ks_level,
-        message_count);
+    luts_message_carry->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
-    execute_pbs<Torus>(streams, gpu_indexes, gpu_count, new_blocks,
-                       lwe_indexes_out, luts_message_carry->lut,
-                       luts_message_carry->lut_indexes, small_lwe_vector,
-                       lwe_indexes_in, bsk, luts_message_carry->buffer,
-                       glwe_dimension, lwe_dimension, polynomial_size,
-                       mem_ptr->params.pbs_base_log, mem_ptr->params.pbs_level,
-                       mem_ptr->params.grouping_factor, total_count, 2, 0,
-                       max_shared_memory, mem_ptr->params.pbs_type);
+    /// For multi GPU execution we create vectors of pointers for inputs and
+    /// outputs
+    std::vector<Torus *> new_blocks_vec = luts_message_carry->lwe_array_in_vec;
+    std::vector<Torus *> small_lwe_vector_vec =
+        luts_message_carry->lwe_after_ks_vec;
+    std::vector<Torus *> lwe_after_pbs_vec =
+        luts_message_carry->lwe_after_pbs_vec;
+    std::vector<Torus *> lwe_trivial_indexes_vec =
+        luts_message_carry->lwe_trivial_indexes_vec;
 
-    luts_message_carry->release(streams[0], gpu_indexes[0]);
+    auto active_gpu_count = get_active_gpu_count(total_count, gpu_count);
+    if (active_gpu_count == 1) {
+      /// Apply KS to go from a big LWE dimension to a small LWE dimension
+      /// After this keyswitch execution, we need to synchronize the streams
+      /// because the keyswitch and PBS do not operate on the same number of
+      /// inputs
+      execute_keyswitch_async<Torus>(
+          streams, gpu_indexes, 1, small_lwe_vector, lwe_indexes_in, new_blocks,
+          lwe_indexes_in, ksks, polynomial_size * glwe_dimension,
+          small_lwe_dimension, mem_ptr->params.ks_base_log,
+          mem_ptr->params.ks_level, message_count);
+
+      /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+      /// dimension to a big LWE dimension
+      execute_pbs_async<Torus>(
+          streams, gpu_indexes, 1, new_blocks, lwe_indexes_out,
+          luts_message_carry->lut_vec, luts_message_carry->lut_indexes_vec,
+          small_lwe_vector, lwe_indexes_in, bsks, luts_message_carry->buffer,
+          glwe_dimension, small_lwe_dimension, polynomial_size,
+          mem_ptr->params.pbs_base_log, mem_ptr->params.pbs_level,
+          mem_ptr->params.grouping_factor, total_count,
+          mem_ptr->params.pbs_type);
+    } else {
+      cuda_synchronize_stream(streams[0], gpu_indexes[0]);
+
+      multi_gpu_scatter_lwe_async<Torus>(
+          streams, gpu_indexes, active_gpu_count, new_blocks_vec, new_blocks,
+          luts_message_carry->h_lwe_indexes_in,
+          luts_message_carry->using_trivial_lwe_indexes, message_count,
+          big_lwe_size);
+
+      /// Apply KS to go from a big LWE dimension to a small LWE dimension
+      /// After this keyswitch execution, we need to synchronize the streams
+      /// because the keyswitch and PBS do not operate on the same number of
+      /// inputs
+      execute_keyswitch_async<Torus>(
+          streams, gpu_indexes, active_gpu_count, small_lwe_vector_vec,
+          lwe_trivial_indexes_vec, new_blocks_vec, lwe_trivial_indexes_vec,
+          ksks, big_lwe_dimension, small_lwe_dimension,
+          mem_ptr->params.ks_base_log, mem_ptr->params.ks_level, total_count);
+
+      /// Copy data back to GPU 0, rebuild the lwe array, and scatter again on a
+      /// different configuration
+      multi_gpu_gather_lwe_async<Torus>(
+          streams, gpu_indexes, gpu_count, small_lwe_vector,
+          small_lwe_vector_vec, luts_message_carry->h_lwe_indexes_in,
+          luts_message_carry->using_trivial_lwe_indexes, message_count,
+          small_lwe_size);
+      /// Synchronize all GPUs
+      for (uint i = 0; i < active_gpu_count; i++) {
+        cuda_synchronize_stream(streams[i], gpu_indexes[i]);
+      }
+
+      multi_gpu_scatter_lwe_async<Torus>(
+          streams, gpu_indexes, gpu_count, small_lwe_vector_vec,
+          small_lwe_vector, luts_message_carry->h_lwe_indexes_in,
+          luts_message_carry->using_trivial_lwe_indexes, total_count,
+          small_lwe_size);
+
+      /// Apply PBS to apply a LUT, reduce the noise and go from a small LWE
+      /// dimension to a big LWE dimension
+      execute_pbs_async<Torus>(
+          streams, gpu_indexes, active_gpu_count, lwe_after_pbs_vec,
+          lwe_trivial_indexes_vec, luts_message_carry->lut_vec,
+          luts_message_carry->lut_indexes_vec, small_lwe_vector_vec,
+          lwe_trivial_indexes_vec, bsks, luts_message_carry->buffer,
+          glwe_dimension, small_lwe_dimension, polynomial_size,
+          mem_ptr->params.pbs_base_log, mem_ptr->params.pbs_level,
+          mem_ptr->params.grouping_factor, total_count,
+          mem_ptr->params.pbs_type);
+
+      multi_gpu_gather_lwe_async<Torus>(
+          streams, gpu_indexes, active_gpu_count, new_blocks, lwe_after_pbs_vec,
+          luts_message_carry->h_lwe_indexes_out,
+          luts_message_carry->using_trivial_lwe_indexes, total_count,
+          big_lwe_size);
+      /// Synchronize all GPUs
+      for (uint i = 0; i < active_gpu_count; i++) {
+        cuda_synchronize_stream(streams[i], gpu_indexes[i]);
+      }
+    }
 
     int rem_blocks = (r > chunk_size) ? r % chunk_size * num_blocks : 0;
     int new_blocks_created = 2 * ch_amount * num_blocks;
@@ -354,24 +419,21 @@ __host__ void host_integer_sum_ciphertexts_vec_kb(
     std::swap(new_blocks, old_blocks);
     r = (new_blocks_created + rem_blocks) / num_blocks;
   }
+  luts_message_carry->release(streams, gpu_indexes, gpu_count);
+  delete (luts_message_carry);
 
   host_addition(streams[0], gpu_indexes[0], radix_lwe_out, old_blocks,
                 &old_blocks[num_blocks * big_lwe_size], big_lwe_dimension,
                 num_blocks);
-
-  host_propagate_single_carry<Torus>(streams, gpu_indexes, gpu_count,
-                                     radix_lwe_out, mem_ptr->scp_mem, bsk, ksk,
-                                     num_blocks);
 }
 
-template <typename Torus, typename STorus, class params>
+template <typename Torus, class params>
 __host__ void host_integer_mult_radix_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     uint64_t *radix_lwe_out, uint64_t *radix_lwe_left,
-    uint64_t *radix_lwe_right, void *bsk, uint64_t *ksk,
+    uint64_t *radix_lwe_right, void **bsks, uint64_t **ksks,
     int_mul_memory<Torus> *mem_ptr, uint32_t num_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto glwe_dimension = mem_ptr->params.glwe_dimension;
   auto polynomial_size = mem_ptr->params.polynomial_size;
   auto lwe_dimension = mem_ptr->params.small_lwe_dimension;
@@ -438,6 +500,7 @@ __host__ void host_integer_mult_radix_kb(
   dim3 grid(lsb_vector_block_count, 1, 1);
   dim3 thds(params::degree / params::opt, 1, 1);
 
+  cudaSetDevice(gpu_indexes[0]);
   all_shifted_lhs_rhs<Torus, params><<<grid, thds, 0, streams[0]>>>(
       radix_lwe_left, vector_result_lsb, vector_result_msb, radix_lwe_right,
       vector_lsb_rhs, vector_msb_rhs, num_blocks);
@@ -445,18 +508,18 @@ __host__ void host_integer_mult_radix_kb(
 
   integer_radix_apply_bivariate_lookup_table_kb<Torus>(
       streams, gpu_indexes, gpu_count, block_mul_res, block_mul_res,
-      vector_result_sb, bsk, ksk, total_block_count, luts_array,
+      vector_result_sb, bsks, ksks, total_block_count, luts_array,
       luts_array->params.message_modulus);
 
   vector_result_lsb = &block_mul_res[0];
   vector_result_msb = &block_mul_res[lsb_vector_block_count *
                                      (polynomial_size * glwe_dimension + 1)];
 
+  cudaSetDevice(gpu_indexes[0]);
   fill_radix_from_lsb_msb<Torus, params>
       <<<num_blocks * num_blocks, params::degree / params::opt, 0,
          streams[0]>>>(vector_result_sb, vector_result_lsb, vector_result_msb,
-                       glwe_dimension, lsb_vector_block_count,
-                       msb_vector_block_count, num_blocks);
+                       glwe_dimension, num_blocks);
   check_cuda_error(cudaGetLastError());
 
   int terms_degree[2 * num_blocks * num_blocks];
@@ -472,35 +535,23 @@ __host__ void host_integer_mult_radix_kb(
     terms_degree_msb[i] = (b_id > r_id) ? message_modulus - 2 : 0;
   }
 
-  host_integer_sum_ciphertexts_vec_kb<Torus, params>(
+  host_integer_partial_sum_ciphertexts_vec_kb<Torus, params>(
       streams, gpu_indexes, gpu_count, radix_lwe_out, vector_result_sb,
-      terms_degree, bsk, ksk, mem_ptr->sum_ciphertexts_mem, num_blocks,
-      2 * num_blocks);
+      terms_degree, bsks, ksks, mem_ptr->sum_ciphertexts_mem, num_blocks,
+      2 * num_blocks, mem_ptr->luts_array);
+
+  auto scp_mem_ptr = mem_ptr->sum_ciphertexts_mem->scp_mem;
+  host_propagate_single_carry<Torus>(streams, gpu_indexes, gpu_count,
+                                     radix_lwe_out, nullptr, nullptr,
+                                     scp_mem_ptr, bsks, ksks, num_blocks);
 }
 
 template <typename Torus>
 __host__ void scratch_cuda_integer_mult_radix_ciphertext_kb(
-    cudaStream_t stream, uint32_t gpu_index, int_mul_memory<Torus> **mem_ptr,
-    uint32_t num_radix_blocks, int_radix_params params,
-    bool allocate_gpu_memory) {
-  cudaSetDevice(gpu_index);
-  size_t sm_size = (params.big_lwe_dimension + 1) * sizeof(Torus);
-  if (sm_size < cuda_get_max_shared_memory(gpu_index)) {
-    check_cuda_error(cudaFuncSetAttribute(
-        tree_add_chunks<Torus, FULLSM>,
-        cudaFuncAttributeMaxDynamicSharedMemorySize, sm_size));
-    cudaFuncSetCacheConfig(tree_add_chunks<Torus, FULLSM>,
-                           cudaFuncCachePreferShared);
-    check_cuda_error(cudaGetLastError());
-  } else {
-    check_cuda_error(
-        cudaFuncSetAttribute(tree_add_chunks<Torus, NOSM>,
-                             cudaFuncAttributeMaxDynamicSharedMemorySize, 0));
-    cudaFuncSetCacheConfig(tree_add_chunks<Torus, NOSM>, cudaFuncCachePreferL1);
-    check_cuda_error(cudaGetLastError());
-  }
-
-  *mem_ptr = new int_mul_memory<Torus>(stream, gpu_index, params,
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    int_mul_memory<Torus> **mem_ptr, uint32_t num_radix_blocks,
+    int_radix_params params, bool allocate_gpu_memory) {
+  *mem_ptr = new int_mul_memory<Torus>(streams, gpu_indexes, gpu_count, params,
                                        num_radix_blocks, allocate_gpu_memory);
 }
 

backends/tfhe-cuda-backend/cuda/src/integer/negation.cu

@@ -12,12 +12,12 @@ void cuda_negate_integer_radix_ciphertext_64_inplace(
 }
 
 void scratch_cuda_integer_radix_overflowing_sub_kb_64(
-    void *stream, uint32_t gpu_index, int8_t **mem_ptr, uint32_t glwe_dimension,
-    uint32_t polynomial_size, uint32_t big_lwe_dimension,
-    uint32_t small_lwe_dimension, uint32_t ks_level, uint32_t ks_base_log,
-    uint32_t pbs_level, uint32_t pbs_base_log, uint32_t grouping_factor,
-    uint32_t num_blocks, uint32_t message_modulus, uint32_t carry_modulus,
-    PBS_TYPE pbs_type, bool allocate_gpu_memory) {
+    void **streams, uint32_t *gpu_indexes, uint32_t gpu_count, int8_t **mem_ptr,
+    uint32_t glwe_dimension, uint32_t polynomial_size,
+    uint32_t big_lwe_dimension, uint32_t small_lwe_dimension, uint32_t ks_level,
+    uint32_t ks_base_log, uint32_t pbs_level, uint32_t pbs_base_log,
+    uint32_t grouping_factor, uint32_t num_blocks, uint32_t message_modulus,
+    uint32_t carry_modulus, PBS_TYPE pbs_type, bool allocate_gpu_memory) {
 
   int_radix_params params(pbs_type, glwe_dimension, polynomial_size,
                           big_lwe_dimension, small_lwe_dimension, ks_level,
@@ -25,7 +25,7 @@ void scratch_cuda_integer_radix_overflowing_sub_kb_64(
                           message_modulus, carry_modulus);
 
   scratch_cuda_integer_overflowing_sub_kb<uint64_t>(
-      static_cast<cudaStream_t>(stream), gpu_index,
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
       (int_overflowing_sub_memory<uint64_t> **)mem_ptr, num_blocks, params,
       allocate_gpu_memory);
 }
@@ -33,77 +33,26 @@ void scratch_cuda_integer_radix_overflowing_sub_kb_64(
 void cuda_integer_radix_overflowing_sub_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *radix_lwe_out, void *radix_lwe_overflowed, void *radix_lwe_left,
-    void *radix_lwe_right, int8_t *mem_ptr, void *bsk, void *ksk,
+    void *radix_lwe_right, int8_t *mem_ptr, void **bsks, void **ksks,
     uint32_t num_blocks) {
 
   auto mem = (int_overflowing_sub_memory<uint64_t> *)mem_ptr;
 
-  switch (mem->params.polynomial_size) {
-  case 512:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<512>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 1024:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<1024>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 2048:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<2048>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 4096:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<4096>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 8192:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<8192>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  case 16384:
-    host_integer_overflowing_sub_kb<uint64_t, AmortizedDegree<16384>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<uint64_t *>(radix_lwe_overflowed),
-        static_cast<uint64_t *>(radix_lwe_left),
-        static_cast<uint64_t *>(radix_lwe_right), bsk,
-        static_cast<uint64_t *>(ksk), mem, num_blocks);
-    break;
-  default:
-    PANIC("Cuda error (integer overflowing sub): unsupported polynomial size. "
-          "Only N = 512, 1024, 2048, 4096, 8192, 16384 is supported")
-  }
+  host_integer_overflowing_sub_kb<uint64_t>(
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
+      static_cast<uint64_t *>(radix_lwe_out),
+      static_cast<uint64_t *>(radix_lwe_overflowed),
+      static_cast<uint64_t *>(radix_lwe_left),
+      static_cast<uint64_t *>(radix_lwe_right), bsks, (uint64_t **)(ksks), mem,
+      num_blocks);
 }
 
-void cleanup_cuda_integer_radix_overflowing_sub(void *stream,
-                                                uint32_t gpu_index,
+void cleanup_cuda_integer_radix_overflowing_sub(void **streams,
+                                                uint32_t *gpu_indexes,
+                                                uint32_t gpu_count,
                                                 int8_t **mem_ptr_void) {
   int_overflowing_sub_memory<uint64_t> *mem_ptr =
       (int_overflowing_sub_memory<uint64_t> *)(*mem_ptr_void);
 
-  mem_ptr->release(static_cast<cudaStream_t>(stream), gpu_index);
+  mem_ptr->release((cudaStream_t *)(streams), gpu_indexes, gpu_count);
 }

backends/tfhe-cuda-backend/cuda/src/integer/negation.cuh

@@ -90,20 +90,19 @@ host_integer_radix_negation(cudaStream_t *streams, uint32_t *gpu_indexes,
 
 template <typename Torus>
 __host__ void scratch_cuda_integer_overflowing_sub_kb(
-    cudaStream_t stream, uint32_t gpu_index,
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     int_overflowing_sub_memory<Torus> **mem_ptr, uint32_t num_blocks,
     int_radix_params params, bool allocate_gpu_memory) {
 
-  cudaSetDevice(gpu_index);
   *mem_ptr = new int_overflowing_sub_memory<Torus>(
-      stream, gpu_index, params, num_blocks, allocate_gpu_memory);
+      streams, gpu_indexes, gpu_count, params, num_blocks, allocate_gpu_memory);
 }
 
-template <typename Torus, class params>
+template <typename Torus>
 __host__ void host_integer_overflowing_sub_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *radix_lwe_out, Torus *radix_lwe_overflowed, Torus *radix_lwe_left,
-    Torus *radix_lwe_right, void *bsk, uint64_t *ksk,
+    Torus *radix_lwe_right, void **bsks, uint64_t **ksks,
     int_overflowing_sub_memory<uint64_t> *mem_ptr, uint32_t num_blocks) {
 
   auto radix_params = mem_ptr->params;
@@ -114,9 +113,9 @@ __host__ void host_integer_overflowing_sub_kb(
       radix_params.message_modulus, radix_params.carry_modulus,
       radix_params.message_modulus - 1);
 
-  host_propagate_single_sub_borrow<Torus>(
-      streams, gpu_indexes, gpu_count, radix_lwe_overflowed, radix_lwe_out,
-      mem_ptr->borrow_prop_mem, bsk, ksk, num_blocks);
+  host_propagate_single_sub_borrow<Torus>(streams, gpu_indexes, gpu_count,
+                                          radix_lwe_overflowed, radix_lwe_out,
+                                          mem_ptr, bsks, ksks, num_blocks);
 }
 
 #endif

backends/tfhe-cuda-backend/cuda/src/integer/scalar_bitops.cu

@@ -3,7 +3,7 @@
 void cuda_scalar_bitop_integer_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *lwe_array_out, void *lwe_array_input, void *clear_blocks,
-    uint32_t num_clear_blocks, int8_t *mem_ptr, void *bsk, void *ksk,
+    uint32_t num_clear_blocks, int8_t *mem_ptr, void **bsks, void **ksks,
     uint32_t lwe_ciphertext_count, BITOP_TYPE op) {
 
   host_integer_radix_scalar_bitop_kb<uint64_t>(
@@ -11,6 +11,6 @@ void cuda_scalar_bitop_integer_radix_ciphertext_kb_64(
       static_cast<uint64_t *>(lwe_array_out),
       static_cast<uint64_t *>(lwe_array_input),
       static_cast<uint64_t *>(clear_blocks), num_clear_blocks,
-      (int_bitop_buffer<uint64_t> *)mem_ptr, bsk, static_cast<uint64_t *>(ksk),
+      (int_bitop_buffer<uint64_t> *)mem_ptr, bsks, (uint64_t **)(ksks),
       lwe_ciphertext_count, op);
 }

backends/tfhe-cuda-backend/cuda/src/integer/scalar_bitops.cuh

@@ -8,10 +8,9 @@ template <typename Torus>
 __host__ void host_integer_radix_scalar_bitop_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_input, Torus *clear_blocks,
-    uint32_t num_clear_blocks, int_bitop_buffer<Torus> *mem_ptr, void *bsk,
-    Torus *ksk, uint32_t num_radix_blocks, BITOP_TYPE op) {
+    uint32_t num_clear_blocks, int_bitop_buffer<Torus> *mem_ptr, void **bsks,
+    Torus **ksks, uint32_t num_radix_blocks, BITOP_TYPE op) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto lut = mem_ptr->lut;
   auto params = lut->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
@@ -31,13 +30,14 @@ __host__ void host_integer_radix_scalar_bitop_kb(
   } else {
     // We have all possible LUTs pre-computed and we use the decomposed scalar
     // as index to recover the right one
-    cuda_memcpy_async_gpu_to_gpu(lut->lut_indexes, clear_blocks,
-                                 num_clear_blocks * sizeof(Torus), streams[0],
-                                 gpu_indexes[0]);
+    cuda_memcpy_async_gpu_to_gpu(lut->get_lut_indexes(gpu_indexes[0], 0),
+                                 clear_blocks, num_clear_blocks * sizeof(Torus),
+                                 streams[0], gpu_indexes[0]);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     integer_radix_apply_univariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_input, bsk,
-        ksk, num_clear_blocks, lut);
+        streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_input, bsks,
+        ksks, num_clear_blocks, lut);
 
     if (op == SCALAR_BITAND && num_clear_blocks < num_radix_blocks) {
       auto lwe_array_out_block = lwe_array_out + num_clear_blocks * lwe_size;

backends/tfhe-cuda-backend/cuda/src/integer/scalar_comparison.cu

@@ -3,7 +3,7 @@
 void cuda_scalar_comparison_integer_radix_ciphertext_kb_64(
     void **streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     void *lwe_array_out, void *lwe_array_in, void *scalar_blocks,
-    int8_t *mem_ptr, void *bsk, void *ksk, uint32_t lwe_ciphertext_count,
+    int8_t *mem_ptr, void **bsks, void **ksks, uint32_t lwe_ciphertext_count,
     uint32_t num_scalar_blocks) {
 
   int_comparison_buffer<uint64_t> *buffer =
@@ -15,8 +15,8 @@ void cuda_scalar_comparison_integer_radix_ciphertext_kb_64(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_in),
-        static_cast<uint64_t *>(scalar_blocks), buffer, bsk,
-        static_cast<uint64_t *>(ksk), lwe_ciphertext_count, num_scalar_blocks);
+        static_cast<uint64_t *>(scalar_blocks), buffer, bsks,
+        (uint64_t **)(ksks), lwe_ciphertext_count, num_scalar_blocks);
     break;
   case GT:
   case GE:
@@ -27,7 +27,7 @@ void cuda_scalar_comparison_integer_radix_ciphertext_kb_64(
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_in),
         static_cast<uint64_t *>(scalar_blocks), buffer,
-        buffer->diff_buffer->operator_f, bsk, static_cast<uint64_t *>(ksk),
+        buffer->diff_buffer->operator_f, bsks, (uint64_t **)(ksks),
         lwe_ciphertext_count, num_scalar_blocks);
     break;
   case MAX:
@@ -36,8 +36,8 @@ void cuda_scalar_comparison_integer_radix_ciphertext_kb_64(
         (cudaStream_t *)(streams), gpu_indexes, gpu_count,
         static_cast<uint64_t *>(lwe_array_out),
         static_cast<uint64_t *>(lwe_array_in),
-        static_cast<uint64_t *>(scalar_blocks), buffer, bsk,
-        static_cast<uint64_t *>(ksk), lwe_ciphertext_count, num_scalar_blocks);
+        static_cast<uint64_t *>(scalar_blocks), buffer, bsks,
+        (uint64_t **)(ksks), lwe_ciphertext_count, num_scalar_blocks);
     break;
   default:
     PANIC("Cuda error: integer operation not supported")

backends/tfhe-cuda-backend/cuda/src/integer/scalar_comparison.cuh

@@ -2,17 +2,15 @@
 #define CUDA_INTEGER_SCALAR_COMPARISON_OPS_CUH
 
 #include "integer/comparison.cuh"
-#include <omp.h>
 
 template <typename Torus>
 __host__ void integer_radix_unsigned_scalar_difference_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
     int_comparison_buffer<Torus> *mem_ptr,
-    std::function<Torus(Torus)> sign_handler_f, void *bsk, Torus *ksk,
+    std::function<Torus(Torus)> sign_handler_f, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto glwe_dimension = params.glwe_dimension;
@@ -49,7 +47,7 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
     // means scalar is zero
     host_compare_with_zero_equality(streams, gpu_indexes, gpu_count,
                                     mem_ptr->tmp_lwe_array_out, lwe_array_in,
-                                    mem_ptr, bsk, ksk, total_num_radix_blocks,
+                                    mem_ptr, bsks, ksks, total_num_radix_blocks,
                                     mem_ptr->is_zero_lut);
 
     auto scalar_last_leaf_lut_f = [sign_handler_f](Torus x) -> Torus {
@@ -60,12 +58,14 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
 
     auto lut = mem_ptr->diff_buffer->tree_buffer->tree_last_leaf_scalar_lut;
     generate_device_accumulator<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, scalar_last_leaf_lut_f);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        scalar_last_leaf_lut_f);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     integer_radix_apply_univariate_lookup_table_kb<Torus>(
         streams, gpu_indexes, gpu_count, lwe_array_out,
-        mem_ptr->tmp_lwe_array_out, bsk, ksk, 1, lut);
+        mem_ptr->tmp_lwe_array_out, bsks, ksks, 1, lut);
 
   } else if (total_num_scalar_blocks < total_num_radix_blocks) {
     // We have to handle both part of the work described above
@@ -79,58 +79,53 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
     auto lwe_array_lsb_out = mem_ptr->tmp_lwe_array_out;
     auto lwe_array_msb_out = lwe_array_lsb_out + big_lwe_size;
 
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-    auto lsb_stream = mem_ptr->lsb_stream;
-    auto msb_stream = mem_ptr->msb_stream;
+    auto lsb_streams = mem_ptr->lsb_streams;
+    auto msb_streams = mem_ptr->msb_streams;
 
-#pragma omp parallel sections
-    {
-      // Both sections may be executed in parallel
-#pragma omp section
-      {
-        //////////////
-        // lsb
-        Torus *lhs = diff_buffer->tmp_packed_left;
-        Torus *rhs = diff_buffer->tmp_packed_right;
-
-        pack_blocks(lsb_stream, gpu_indexes[0], lhs, lwe_array_in,
-                    big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
-        pack_blocks(lsb_stream, gpu_indexes[0], rhs, scalar_blocks, 0,
-                    total_num_scalar_blocks, message_modulus);
-
-        // From this point we have half number of blocks
-        num_lsb_radix_blocks /= 2;
-        num_lsb_radix_blocks += (total_num_scalar_blocks % 2);
-
-        // comparisons will be assigned
-        // - 0 if lhs < rhs
-        // - 1 if lhs == rhs
-        // - 2 if lhs > rhs
-
-        auto comparisons = mem_ptr->tmp_block_comparisons;
-        scalar_compare_radix_blocks_kb(&lsb_stream, &gpu_indexes[0], 1,
-                                       comparisons, lhs, rhs, mem_ptr, bsk, ksk,
-                                       num_lsb_radix_blocks);
-
-        // Reduces a vec containing radix blocks that encrypts a sign
-        // (inferior, equal, superior) to one single radix block containing the
-        // final sign
-        tree_sign_reduction(&lsb_stream, &gpu_indexes[0], 1, lwe_array_lsb_out,
-                            comparisons, mem_ptr->diff_buffer->tree_buffer,
-                            mem_ptr->identity_lut_f, bsk, ksk,
-                            num_lsb_radix_blocks);
-      }
-#pragma omp section
-      {
-        //////////////
-        // msb
-        host_compare_with_zero_equality(
-            &msb_stream, &gpu_indexes[0], 1, lwe_array_msb_out, msb, mem_ptr,
-            bsk, ksk, num_msb_radix_blocks, mem_ptr->is_zero_lut);
-      }
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+
+    //////////////
+    // lsb
+    Torus *lhs = diff_buffer->tmp_packed_left;
+    Torus *rhs = diff_buffer->tmp_packed_right;
+
+    pack_blocks(lsb_streams[0], gpu_indexes[0], lhs, lwe_array_in,
+                big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
+    pack_blocks(lsb_streams[0], gpu_indexes[0], rhs, scalar_blocks, 0,
+                total_num_scalar_blocks, message_modulus);
+
+    // From this point we have half number of blocks
+    num_lsb_radix_blocks /= 2;
+    num_lsb_radix_blocks += (total_num_scalar_blocks % 2);
+
+    // comparisons will be assigned
+    // - 0 if lhs < rhs
+    // - 1 if lhs == rhs
+    // - 2 if lhs > rhs
+
+    auto comparisons = mem_ptr->tmp_block_comparisons;
+    scalar_compare_radix_blocks_kb(lsb_streams, gpu_indexes, gpu_count,
+                                   comparisons, lhs, rhs, mem_ptr, bsks, ksks,
+                                   num_lsb_radix_blocks);
+
+    // Reduces a vec containing radix blocks that encrypts a sign
+    // (inferior, equal, superior) to one single radix block containing the
+    // final sign
+    tree_sign_reduction(lsb_streams, gpu_indexes, gpu_count, lwe_array_lsb_out,
+                        comparisons, mem_ptr->diff_buffer->tree_buffer,
+                        mem_ptr->identity_lut_f, bsks, ksks,
+                        num_lsb_radix_blocks);
+    //////////////
+    // msb
+    host_compare_with_zero_equality(msb_streams, gpu_indexes, gpu_count,
+                                    lwe_array_msb_out, msb, mem_ptr, bsks, ksks,
+                                    num_msb_radix_blocks, mem_ptr->is_zero_lut);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(lsb_streams[j], gpu_indexes[j]);
+      cuda_synchronize_stream(msb_streams[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(lsb_stream, gpu_indexes[0]);
-    cuda_synchronize_stream(msb_stream, gpu_indexes[0]);
 
     //////////////
     // Reduce the two blocks into one final
@@ -145,12 +140,14 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
 
     auto lut = diff_buffer->tree_buffer->tree_last_leaf_scalar_lut;
     generate_device_accumulator_bivariate<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, scalar_bivariate_last_leaf_lut_f);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        scalar_bivariate_last_leaf_lut_f);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     integer_radix_apply_bivariate_lookup_table_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_lsb_out,
-        lwe_array_msb_out, bsk, ksk, 1, lut, lut->params.message_modulus);
+        lwe_array_msb_out, bsks, ksks, 1, lut, lut->params.message_modulus);
 
   } else {
     // We only have to do the regular comparison
@@ -177,7 +174,7 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
     // - 2 if lhs > rhs
     auto comparisons = mem_ptr->tmp_lwe_array_out;
     scalar_compare_radix_blocks_kb(streams, gpu_indexes, gpu_count, comparisons,
-                                   lhs, rhs, mem_ptr, bsk, ksk,
+                                   lhs, rhs, mem_ptr, bsks, ksks,
                                    num_lsb_radix_blocks);
 
     // Reduces a vec containing radix blocks that encrypts a sign
@@ -185,7 +182,7 @@ __host__ void integer_radix_unsigned_scalar_difference_check_kb(
     // final sign
     tree_sign_reduction(streams, gpu_indexes, gpu_count, lwe_array_out,
                         comparisons, mem_ptr->diff_buffer->tree_buffer,
-                        sign_handler_f, bsk, ksk, num_lsb_radix_blocks);
+                        sign_handler_f, bsks, ksks, num_lsb_radix_blocks);
   }
 }
 
@@ -194,10 +191,9 @@ __host__ void integer_radix_signed_scalar_difference_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
     int_comparison_buffer<Torus> *mem_ptr,
-    std::function<Torus(Torus)> sign_handler_f, void *bsk, Torus *ksk,
+    std::function<Torus(Torus)> sign_handler_f, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto glwe_dimension = params.glwe_dimension;
@@ -235,7 +231,7 @@ __host__ void integer_radix_signed_scalar_difference_check_kb(
     Torus *are_all_msb_zeros = mem_ptr->tmp_lwe_array_out;
     host_compare_with_zero_equality(
         streams, gpu_indexes, gpu_count, are_all_msb_zeros, lwe_array_in,
-        mem_ptr, bsk, ksk, total_num_radix_blocks, mem_ptr->is_zero_lut);
+        mem_ptr, bsks, ksks, total_num_radix_blocks, mem_ptr->is_zero_lut);
     Torus *sign_block =
         lwe_array_in + (total_num_radix_blocks - 1) * big_lwe_size;
 
@@ -276,12 +272,14 @@ __host__ void integer_radix_signed_scalar_difference_check_kb(
 
     auto lut = mem_ptr->diff_buffer->tree_buffer->tree_last_leaf_scalar_lut;
     generate_device_accumulator_bivariate<Torus>(
-        streams[0], gpu_indexes[0], lut->lut, glwe_dimension, polynomial_size,
-        message_modulus, carry_modulus, scalar_bivariate_last_leaf_lut_f);
+        streams[0], gpu_indexes[0], lut->get_lut(gpu_indexes[0], 0),
+        glwe_dimension, polynomial_size, message_modulus, carry_modulus,
+        scalar_bivariate_last_leaf_lut_f);
+    lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
 
     integer_radix_apply_bivariate_lookup_table_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, are_all_msb_zeros,
-        sign_block, bsk, ksk, 1, lut, lut->params.message_modulus);
+        sign_block, bsks, ksks, 1, lut, lut->params.message_modulus);
 
   } else if (total_num_scalar_blocks < total_num_radix_blocks) {
     // We have to handle both part of the work described above
@@ -295,101 +293,97 @@ __host__ void integer_radix_signed_scalar_difference_check_kb(
     auto lwe_array_lsb_out = mem_ptr->tmp_lwe_array_out;
     auto lwe_array_msb_out = lwe_array_lsb_out + big_lwe_size;
 
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-    auto lsb_stream = mem_ptr->lsb_stream;
-    auto msb_stream = mem_ptr->msb_stream;
-
-#pragma omp parallel sections
-    {
-      // Both sections may be executed in parallel
-#pragma omp section
-      {
-        //////////////
-        // lsb
-        Torus *lhs = diff_buffer->tmp_packed_left;
-        Torus *rhs = diff_buffer->tmp_packed_right;
-
-        pack_blocks(lsb_stream, gpu_indexes[0], lhs, lwe_array_in,
-                    big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
-        pack_blocks(lsb_stream, gpu_indexes[0], rhs, scalar_blocks, 0,
-                    total_num_scalar_blocks, message_modulus);
-
-        // From this point we have half number of blocks
-        num_lsb_radix_blocks /= 2;
-        num_lsb_radix_blocks += (total_num_scalar_blocks % 2);
-
-        // comparisons will be assigned
-        // - 0 if lhs < rhs
-        // - 1 if lhs == rhs
-        // - 2 if lhs > rhs
-
-        auto comparisons = mem_ptr->tmp_block_comparisons;
-        scalar_compare_radix_blocks_kb(&lsb_stream, &gpu_indexes[0], 1,
-                                       comparisons, lhs, rhs, mem_ptr, bsk, ksk,
-                                       num_lsb_radix_blocks);
-
-        // Reduces a vec containing radix blocks that encrypts a sign
-        // (inferior, equal, superior) to one single radix block containing the
-        // final sign
-        tree_sign_reduction(&lsb_stream, &gpu_indexes[0], 1, lwe_array_lsb_out,
-                            comparisons, mem_ptr->diff_buffer->tree_buffer,
-                            mem_ptr->identity_lut_f, bsk, ksk,
-                            num_lsb_radix_blocks);
-      }
-#pragma omp section
-      {
-        //////////////
-        // msb
-        // We remove the last block (which is the sign)
-        Torus *are_all_msb_zeros = lwe_array_msb_out;
-        host_compare_with_zero_equality(
-            &msb_stream, &gpu_indexes[0], 1, are_all_msb_zeros, msb, mem_ptr,
-            bsk, ksk, num_msb_radix_blocks, mem_ptr->is_zero_lut);
-
-        auto sign_bit_pos = (int)log2(message_modulus) - 1;
-
-        auto lut_f = [mem_ptr, sign_bit_pos](Torus sign_block,
-                                             Torus msb_are_zeros) {
-          bool sign_bit_is_set = (sign_block >> sign_bit_pos) == 1;
-          CMP_ORDERING sign_block_ordering;
-          if (sign_bit_is_set) {
-            sign_block_ordering = CMP_ORDERING::IS_INFERIOR;
-          } else if (sign_block != 0) {
-            sign_block_ordering = CMP_ORDERING::IS_SUPERIOR;
-          } else {
-            sign_block_ordering = CMP_ORDERING::IS_EQUAL;
-          }
-
-          CMP_ORDERING msb_ordering;
-          if (msb_are_zeros == 1)
-            msb_ordering = CMP_ORDERING::IS_EQUAL;
-          else
-            msb_ordering = CMP_ORDERING::IS_SUPERIOR;
-
-          return mem_ptr->diff_buffer->tree_buffer->block_selector_f(
-              sign_block_ordering, msb_ordering);
-        };
-
-        auto signed_msb_lut = mem_ptr->signed_msb_lut;
-        generate_device_accumulator_bivariate<Torus>(
-            msb_stream, gpu_indexes[0], signed_msb_lut->lut,
-            params.glwe_dimension, params.polynomial_size,
-            params.message_modulus, params.carry_modulus, lut_f);
-
-        Torus *sign_block = msb + (num_msb_radix_blocks - 1) * big_lwe_size;
-        integer_radix_apply_bivariate_lookup_table_kb(
-            &msb_stream, &gpu_indexes[0], 1, lwe_array_msb_out, sign_block,
-            are_all_msb_zeros, bsk, ksk, 1, signed_msb_lut,
-            signed_msb_lut->params.message_modulus);
-      }
+    auto lsb_streams = mem_ptr->lsb_streams;
+    auto msb_streams = mem_ptr->msb_streams;
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+
+    //////////////
+    // lsb
+    Torus *lhs = diff_buffer->tmp_packed_left;
+    Torus *rhs = diff_buffer->tmp_packed_right;
+
+    pack_blocks(lsb_streams[0], gpu_indexes[0], lhs, lwe_array_in,
+                big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
+    pack_blocks(lsb_streams[0], gpu_indexes[0], rhs, scalar_blocks, 0,
+                total_num_scalar_blocks, message_modulus);
+
+    // From this point we have half number of blocks
+    num_lsb_radix_blocks /= 2;
+    num_lsb_radix_blocks += (total_num_scalar_blocks % 2);
+
+    // comparisons will be assigned
+    // - 0 if lhs < rhs
+    // - 1 if lhs == rhs
+    // - 2 if lhs > rhs
+
+    auto comparisons = mem_ptr->tmp_block_comparisons;
+    scalar_compare_radix_blocks_kb(lsb_streams, gpu_indexes, gpu_count,
+                                   comparisons, lhs, rhs, mem_ptr, bsks, ksks,
+                                   num_lsb_radix_blocks);
+
+    // Reduces a vec containing radix blocks that encrypts a sign
+    // (inferior, equal, superior) to one single radix block containing the
+    // final sign
+    tree_sign_reduction(lsb_streams, gpu_indexes, gpu_count, lwe_array_lsb_out,
+                        comparisons, mem_ptr->diff_buffer->tree_buffer,
+                        mem_ptr->identity_lut_f, bsks, ksks,
+                        num_lsb_radix_blocks);
+    //////////////
+    // msb
+    // We remove the last block (which is the sign)
+    Torus *are_all_msb_zeros = lwe_array_msb_out;
+    host_compare_with_zero_equality(msb_streams, gpu_indexes, gpu_count,
+                                    are_all_msb_zeros, msb, mem_ptr, bsks, ksks,
+                                    num_msb_radix_blocks, mem_ptr->is_zero_lut);
+
+    auto sign_bit_pos = (int)log2(message_modulus) - 1;
+
+    auto lut_f = [mem_ptr, sign_bit_pos](Torus sign_block,
+                                         Torus msb_are_zeros) {
+      bool sign_bit_is_set = (sign_block >> sign_bit_pos) == 1;
+      CMP_ORDERING sign_block_ordering;
+      if (sign_bit_is_set) {
+        sign_block_ordering = CMP_ORDERING::IS_INFERIOR;
+      } else if (sign_block != 0) {
+        sign_block_ordering = CMP_ORDERING::IS_SUPERIOR;
+      } else {
+        sign_block_ordering = CMP_ORDERING::IS_EQUAL;
+      }
+
+      CMP_ORDERING msb_ordering;
+      if (msb_are_zeros == 1)
+        msb_ordering = CMP_ORDERING::IS_EQUAL;
+      else
+        msb_ordering = CMP_ORDERING::IS_SUPERIOR;
+
+      return mem_ptr->diff_buffer->tree_buffer->block_selector_f(
+          sign_block_ordering, msb_ordering);
+    };
+
+    auto signed_msb_lut = mem_ptr->signed_msb_lut;
+    generate_device_accumulator_bivariate<Torus>(
+        msb_streams[0], gpu_indexes[0],
+        signed_msb_lut->get_lut(gpu_indexes[0], 0), params.glwe_dimension,
+        params.polynomial_size, params.message_modulus, params.carry_modulus,
+        lut_f);
+    signed_msb_lut->broadcast_lut(streams, gpu_indexes, gpu_indexes[0]);
+
+    Torus *sign_block = msb + (num_msb_radix_blocks - 1) * big_lwe_size;
+    integer_radix_apply_bivariate_lookup_table_kb(
+        msb_streams, gpu_indexes, gpu_count, lwe_array_msb_out, sign_block,
+        are_all_msb_zeros, bsks, ksks, 1, signed_msb_lut,
+        signed_msb_lut->params.message_modulus);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(lsb_streams[j], gpu_indexes[j]);
+      cuda_synchronize_stream(msb_streams[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(lsb_stream, gpu_indexes[0]);
-    cuda_synchronize_stream(msb_stream, gpu_indexes[0]);
 
     //////////////
     // Reduce the two blocks into one final
     reduce_signs(streams, gpu_indexes, gpu_count, lwe_array_out,
-                 lwe_array_lsb_out, mem_ptr, sign_handler_f, bsk, ksk, 2);
+                 lwe_array_lsb_out, mem_ptr, sign_handler_f, bsks, ksks, 2);
 
   } else {
     // We only have to do the regular comparison
@@ -397,64 +391,56 @@ __host__ void integer_radix_signed_scalar_difference_check_kb(
     // total_num_radix_blocks == total_num_scalar_blocks
     uint32_t num_lsb_radix_blocks = total_num_radix_blocks;
 
-    cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-    auto lsb_stream = mem_ptr->lsb_stream;
-    auto msb_stream = mem_ptr->msb_stream;
+    for (uint j = 0; j < gpu_count; j++) {
+      cuda_synchronize_stream(streams[j], gpu_indexes[j]);
+    }
+    auto lsb_streams = mem_ptr->lsb_streams;
+    auto msb_streams = mem_ptr->msb_streams;
 
     auto lwe_array_ct_out = mem_ptr->tmp_lwe_array_out;
     auto lwe_array_sign_out =
         lwe_array_ct_out + (num_lsb_radix_blocks / 2) * big_lwe_size;
-#pragma omp parallel sections
-    {
-      // Both sections may be executed in parallel
-#pragma omp section
-      {
-        Torus *lhs = diff_buffer->tmp_packed_left;
-        Torus *rhs = diff_buffer->tmp_packed_right;
+    Torus *lhs = diff_buffer->tmp_packed_left;
+    Torus *rhs = diff_buffer->tmp_packed_right;
 
-        pack_blocks(lsb_stream, gpu_indexes[0], lhs, lwe_array_in,
-                    big_lwe_dimension, num_lsb_radix_blocks - 1,
-                    message_modulus);
-        pack_blocks(lsb_stream, gpu_indexes[0], rhs, scalar_blocks, 0,
-                    num_lsb_radix_blocks - 1, message_modulus);
+    pack_blocks(lsb_streams[0], gpu_indexes[0], lhs, lwe_array_in,
+                big_lwe_dimension, num_lsb_radix_blocks - 1, message_modulus);
+    pack_blocks(lsb_streams[0], gpu_indexes[0], rhs, scalar_blocks, 0,
+                num_lsb_radix_blocks - 1, message_modulus);
 
-        // From this point we have half number of blocks
-        num_lsb_radix_blocks /= 2;
+    // From this point we have half number of blocks
+    num_lsb_radix_blocks /= 2;
 
-        // comparisons will be assigned
-        // - 0 if lhs < rhs
-        // - 1 if lhs == rhs
-        // - 2 if lhs > rhs
-        scalar_compare_radix_blocks_kb(&lsb_stream, &gpu_indexes[0], 1,
-                                       lwe_array_ct_out, lhs, rhs, mem_ptr, bsk,
-                                       ksk, num_lsb_radix_blocks);
-      }
-#pragma omp section
-      {
-        Torus *encrypted_sign_block =
-            lwe_array_in + (total_num_radix_blocks - 1) * big_lwe_size;
-        Torus *scalar_sign_block =
-            scalar_blocks + (total_num_scalar_blocks - 1);
+    // comparisons will be assigned
+    // - 0 if lhs < rhs
+    // - 1 if lhs == rhs
+    // - 2 if lhs > rhs
+    scalar_compare_radix_blocks_kb(lsb_streams, gpu_indexes, gpu_count,
+                                   lwe_array_ct_out, lhs, rhs, mem_ptr, bsks,
+                                   ksks, num_lsb_radix_blocks);
+    Torus *encrypted_sign_block =
+        lwe_array_in + (total_num_radix_blocks - 1) * big_lwe_size;
+    Torus *scalar_sign_block = scalar_blocks + (total_num_scalar_blocks - 1);
 
-        auto trivial_sign_block = mem_ptr->tmp_trivial_sign_block;
-        create_trivial_radix(msb_stream, gpu_indexes[0], trivial_sign_block,
-                             scalar_sign_block, big_lwe_dimension, 1, 1,
-                             message_modulus, carry_modulus);
+    auto trivial_sign_block = mem_ptr->tmp_trivial_sign_block;
+    create_trivial_radix(msb_streams[0], gpu_indexes[0], trivial_sign_block,
+                         scalar_sign_block, big_lwe_dimension, 1, 1,
+                         message_modulus, carry_modulus);
 
-        integer_radix_apply_bivariate_lookup_table_kb(
-            &msb_stream, &gpu_indexes[0], 1, lwe_array_sign_out,
-            encrypted_sign_block, trivial_sign_block, bsk, ksk, 1,
-            mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
-      }
+    integer_radix_apply_bivariate_lookup_table_kb(
+        msb_streams, gpu_indexes, gpu_count, lwe_array_sign_out,
+        encrypted_sign_block, trivial_sign_block, bsks, ksks, 1,
+        mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
+    for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+      cuda_synchronize_stream(lsb_streams[j], gpu_indexes[j]);
+      cuda_synchronize_stream(msb_streams[j], gpu_indexes[j]);
     }
-    cuda_synchronize_stream(lsb_stream, gpu_indexes[0]);
-    cuda_synchronize_stream(msb_stream, gpu_indexes[0]);
 
     // Reduces a vec containing radix blocks that encrypts a sign
     // (inferior, equal, superior) to one single radix block containing the
     // final sign
     reduce_signs(streams, gpu_indexes, gpu_count, lwe_array_out,
-                 lwe_array_ct_out, mem_ptr, sign_handler_f, bsk, ksk,
+                 lwe_array_ct_out, mem_ptr, sign_handler_f, bsks, ksks,
                  num_lsb_radix_blocks + 1);
   }
 }
@@ -463,7 +449,7 @@ template <typename Torus>
 __host__ void integer_radix_signed_scalar_maxmin_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
   cudaSetDevice(gpu_indexes[0]);
@@ -475,7 +461,7 @@ __host__ void integer_radix_signed_scalar_maxmin_kb(
   auto sign = mem_ptr->tmp_lwe_array_out;
   integer_radix_signed_scalar_difference_check_kb(
       streams, gpu_indexes, gpu_count, sign, lwe_array_in, scalar_blocks,
-      mem_ptr, mem_ptr->identity_lut_f, bsk, ksk, total_num_radix_blocks,
+      mem_ptr, mem_ptr->identity_lut_f, bsks, ksks, total_num_radix_blocks,
       total_num_scalar_blocks);
 
   // There is no optimized CMUX for scalars, so we convert to a trivial
@@ -490,9 +476,10 @@ __host__ void integer_radix_signed_scalar_maxmin_kb(
 
   // Selector
   // CMUX for Max or Min
-  host_integer_radix_cmux_kb(
-      streams, gpu_indexes, gpu_count, lwe_array_out, sign, lwe_array_left,
-      lwe_array_right, mem_ptr->cmux_buffer, bsk, ksk, total_num_radix_blocks);
+  host_integer_radix_cmux_kb(streams, gpu_indexes, gpu_count, lwe_array_out,
+                             sign, lwe_array_left, lwe_array_right,
+                             mem_ptr->cmux_buffer, bsks, ksks,
+                             total_num_radix_blocks);
 }
 
 template <typename Torus>
@@ -500,19 +487,19 @@ __host__ void host_integer_radix_scalar_difference_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
     int_comparison_buffer<Torus> *mem_ptr,
-    std::function<Torus(Torus)> sign_handler_f, void *bsk, Torus *ksk,
+    std::function<Torus(Torus)> sign_handler_f, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
   if (mem_ptr->is_signed) {
     // is signed and scalar is positive
     integer_radix_signed_scalar_difference_check_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in,
-        scalar_blocks, mem_ptr, sign_handler_f, bsk, ksk,
+        scalar_blocks, mem_ptr, sign_handler_f, bsks, ksks,
         total_num_radix_blocks, total_num_scalar_blocks);
   } else {
     integer_radix_unsigned_scalar_difference_check_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in,
-        scalar_blocks, mem_ptr, sign_handler_f, bsk, ksk,
+        scalar_blocks, mem_ptr, sign_handler_f, bsks, ksks,
         total_num_radix_blocks, total_num_scalar_blocks);
   }
 }
@@ -521,32 +508,32 @@ template <typename Torus>
 __host__ void host_integer_radix_signed_scalar_maxmin_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
   if (mem_ptr->is_signed) {
     // is signed and scalar is positive
     integer_radix_signed_scalar_maxmin_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in,
-        scalar_blocks, mem_ptr, bsk, ksk, total_num_radix_blocks,
+        scalar_blocks, mem_ptr, bsks, ksks, total_num_radix_blocks,
         total_num_scalar_blocks);
   } else {
     integer_radix_unsigned_scalar_maxmin_kb(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_in,
-        scalar_blocks, mem_ptr, bsk, ksk, total_num_radix_blocks,
+        scalar_blocks, mem_ptr, bsks, ksks, total_num_radix_blocks,
         total_num_scalar_blocks);
   }
 }
 
 template <typename Torus>
-__host__ void
-scalar_compare_radix_blocks_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
-                               uint32_t gpu_count, Torus *lwe_array_out,
-                               Torus *lwe_array_in, Torus *scalar_blocks,
-                               int_comparison_buffer<Torus> *mem_ptr, void *bsk,
-                               Torus *ksk, uint32_t num_radix_blocks) {
+__host__ void scalar_compare_radix_blocks_kb(
+    cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
+    Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
+    uint32_t num_radix_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
+  if (num_radix_blocks == 0)
+    return;
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -579,8 +566,8 @@ scalar_compare_radix_blocks_kb(cudaStream_t *streams, uint32_t *gpu_indexes,
   // Apply LUT to compare to 0
   auto sign_lut = mem_ptr->eq_buffer->is_non_zero_lut;
   integer_radix_apply_univariate_lookup_table_kb(
-      streams, gpu_indexes, gpu_count, lwe_array_out, subtracted_blocks, bsk,
-      ksk, num_radix_blocks, sign_lut);
+      streams, gpu_indexes, gpu_count, lwe_array_out, subtracted_blocks, bsks,
+      ksks, num_radix_blocks, sign_lut);
 
   // Add one
   // Here Lhs can have the following values: (-1) % (message modulus * carry
@@ -594,10 +581,9 @@ template <typename Torus>
 __host__ void host_integer_radix_scalar_maxmin_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t total_num_radix_blocks, uint32_t total_num_scalar_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
 
   // Calculates the difference sign between the ciphertext and the scalar
@@ -607,7 +593,7 @@ __host__ void host_integer_radix_scalar_maxmin_kb(
   auto sign = mem_ptr->tmp_lwe_array_out;
   host_integer_radix_scalar_difference_check_kb(
       streams, gpu_indexes, gpu_count, sign, lwe_array_in, scalar_blocks,
-      mem_ptr, mem_ptr->identity_lut_f, bsk, ksk, total_num_radix_blocks,
+      mem_ptr, mem_ptr->identity_lut_f, bsks, ksks, total_num_radix_blocks,
       total_num_scalar_blocks);
 
   // There is no optimized CMUX for scalars, so we convert to a trivial
@@ -624,7 +610,7 @@ __host__ void host_integer_radix_scalar_maxmin_kb(
   // CMUX for Max or Min
   host_integer_radix_cmux_kb(streams, gpu_indexes, gpu_count, lwe_array_out,
                              mem_ptr->tmp_lwe_array_out, lwe_array_left,
-                             lwe_array_right, mem_ptr->cmux_buffer, bsk, ksk,
+                             lwe_array_right, mem_ptr->cmux_buffer, bsks, ksks,
                              total_num_radix_blocks);
 }
 
@@ -632,10 +618,9 @@ template <typename Torus>
 __host__ void host_integer_radix_scalar_equality_check_kb(
     cudaStream_t *streams, uint32_t *gpu_indexes, uint32_t gpu_count,
     Torus *lwe_array_out, Torus *lwe_array_in, Torus *scalar_blocks,
-    int_comparison_buffer<Torus> *mem_ptr, void *bsk, Torus *ksk,
+    int_comparison_buffer<Torus> *mem_ptr, void **bsks, Torus **ksks,
     uint32_t num_radix_blocks, uint32_t num_scalar_blocks) {
 
-  cudaSetDevice(gpu_indexes[0]);
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -662,74 +647,69 @@ __host__ void host_integer_radix_scalar_equality_check_kb(
   auto lwe_array_msb_out =
       lwe_array_lsb_out + big_lwe_size * num_halved_lsb_radix_blocks;
 
-  cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-
-  auto lsb_stream = mem_ptr->lsb_stream;
-  auto msb_stream = mem_ptr->msb_stream;
-
-#pragma omp parallel sections
-  {
-    // Both sections may be executed in parallel
-#pragma omp section
-    {
-      if (num_halved_scalar_blocks > 0) {
-        auto packed_blocks = mem_ptr->tmp_packed_input;
-        auto packed_scalar =
-            packed_blocks + big_lwe_size * num_halved_lsb_radix_blocks;
-
-        pack_blocks(lsb_stream, gpu_indexes[0], packed_blocks, lsb,
-                    big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
-        pack_blocks(lsb_stream, gpu_indexes[0], packed_scalar, scalar_blocks, 0,
-                    num_scalar_blocks, message_modulus);
-
-        cuda_memcpy_async_gpu_to_gpu(scalar_comparison_luts->lut_indexes,
-                                     packed_scalar,
-                                     num_halved_scalar_blocks * sizeof(Torus),
-                                     lsb_stream, gpu_indexes[0]);
-
-        integer_radix_apply_univariate_lookup_table_kb(
-            &lsb_stream, &gpu_indexes[0], 1, lwe_array_lsb_out, packed_blocks,
-            bsk, ksk, num_halved_lsb_radix_blocks, scalar_comparison_luts);
-      }
-    }
-#pragma omp section
-    {
-      //////////////
-      // msb
-      if (num_msb_radix_blocks > 0) {
-        int_radix_lut<Torus> *msb_lut;
-        switch (mem_ptr->op) {
-        case COMPARISON_TYPE::EQ:
-          msb_lut = mem_ptr->is_zero_lut;
-          break;
-        case COMPARISON_TYPE::NE:
-          msb_lut = mem_ptr->eq_buffer->is_non_zero_lut;
-          break;
-        default:
-          PANIC("Cuda error: integer operation not supported")
-        }
-
-        host_compare_with_zero_equality(&msb_stream, &gpu_indexes[0], 1,
-                                        lwe_array_msb_out, msb, mem_ptr, bsk,
-                                        ksk, num_msb_radix_blocks, msb_lut);
-      }
-    }
+  for (uint j = 0; j < gpu_count; j++) {
+    cuda_synchronize_stream(streams[j], gpu_indexes[j]);
   }
 
-  cuda_synchronize_stream(lsb_stream, gpu_indexes[0]);
-  cuda_synchronize_stream(msb_stream, gpu_indexes[0]);
+  auto lsb_streams = mem_ptr->lsb_streams;
+  auto msb_streams = mem_ptr->msb_streams;
+
+  if (num_halved_scalar_blocks > 0) {
+    auto packed_blocks = mem_ptr->tmp_packed_input;
+    auto packed_scalar =
+        packed_blocks + big_lwe_size * num_halved_lsb_radix_blocks;
+
+    pack_blocks(lsb_streams[0], gpu_indexes[0], packed_blocks, lsb,
+                big_lwe_dimension, num_lsb_radix_blocks, message_modulus);
+    pack_blocks(lsb_streams[0], gpu_indexes[0], packed_scalar, scalar_blocks, 0,
+                num_scalar_blocks, message_modulus);
+
+    cuda_memcpy_async_gpu_to_gpu(
+        scalar_comparison_luts->get_lut_indexes(gpu_indexes[0], 0),
+        packed_scalar, num_halved_scalar_blocks * sizeof(Torus), lsb_streams[0],
+        gpu_indexes[0]);
+    scalar_comparison_luts->broadcast_lut(lsb_streams, gpu_indexes, 0);
+
+    integer_radix_apply_univariate_lookup_table_kb(
+        lsb_streams, gpu_indexes, gpu_count, lwe_array_lsb_out, packed_blocks,
+        bsks, ksks, num_halved_lsb_radix_blocks, scalar_comparison_luts);
+  }
+  //////////////
+  // msb
+  if (num_msb_radix_blocks > 0) {
+    int_radix_lut<Torus> *msb_lut;
+    switch (mem_ptr->op) {
+    case COMPARISON_TYPE::EQ:
+      msb_lut = mem_ptr->is_zero_lut;
+      break;
+    case COMPARISON_TYPE::NE:
+      msb_lut = mem_ptr->eq_buffer->is_non_zero_lut;
+      break;
+    default:
+      PANIC("Cuda error: integer operation not supported")
+    }
+
+    host_compare_with_zero_equality(msb_streams, gpu_indexes, gpu_count,
+                                    lwe_array_msb_out, msb, mem_ptr, bsks, ksks,
+                                    num_msb_radix_blocks, msb_lut);
+  }
+
+  for (uint j = 0; j < mem_ptr->active_gpu_count; j++) {
+    cuda_synchronize_stream(lsb_streams[j], gpu_indexes[j]);
+    cuda_synchronize_stream(msb_streams[j], gpu_indexes[j]);
+  }
 
   switch (mem_ptr->op) {
   case COMPARISON_TYPE::EQ:
     are_all_comparisons_block_true(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_lsb_out,
-        mem_ptr, bsk, ksk,
+        mem_ptr, bsks, ksks,
         num_halved_scalar_blocks + (num_msb_radix_blocks > 0));
     break;
   case COMPARISON_TYPE::NE:
     is_at_least_one_comparisons_block_true(
         streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_lsb_out,
-        mem_ptr, bsk, ksk,
+        mem_ptr, bsks, ksks,
         num_halved_scalar_blocks + (num_msb_radix_blocks > 0));
     break;
   default: