feat(hlapi): bind string [r]split_once

Bumps [actions/download-artifact](https://github.com/actions/download-artifact) from 4.1.9 to 4.2.1.
- [Release notes](https://github.com/actions/download-artifact/releases)
- [Commits](cc20338598...95815c38cf)

---
updated-dependencies:
- dependency-name: actions/download-artifact
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <support@github.com>

.github/workflows/aws_tfhe_backward_compat_tests.yml

@@ -11,6 +11,8 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -51,7 +53,7 @@ jobs:
     name: Backward compatibility tests
     needs: [ setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -62,18 +64,14 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         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 }}
+          PR_BRANCH: ${{ github.ref_name }}
         run: |
           echo "BACKWARD_COMPAT_DATA_BRANCH=${PR_BRANCH}" >> "${GITHUB_ENV}"
 
@@ -83,7 +81,23 @@ jobs:
           BRANCH="$(make backward_compat_branch)"
           echo "branch=${BRANCH}" >> "${GITHUB_OUTPUT}"
 
+      - name: Get backward compat branch head SHA
+        id: backward_compat_sha
+        env:
+          REPO_URL: "https://github.com/zama-ai/tfhe-backward-compat-data"
+        run: |
+          SHA=$(git ls-remote ${{ env.REPO_URL }} refs/heads/${{ steps.backward_compat_branch.outputs.branch }} | awk '{print $1}')
+          echo "sha=${SHA}" >> "${GITHUB_OUTPUT}"
+
+      - name: Retrieve data from cache
+        id: retrieve-data-cache
+        uses: actions/cache/restore@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
+        with:
+          path: tests/tfhe-backward-compat-data
+          key: ${{ steps.backward_compat_branch.outputs.branch }}_${{ steps.backward_compat_sha.outputs.sha }}
+
       - name: Clone test data
+        if: steps.retrieve-data-cache.outputs.cache-hit != 'true'
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
         with:
           persist-credentials: 'false'
@@ -96,13 +110,26 @@ jobs:
         run: |
           make test_backward_compatibility_ci
 
+      - name: Store data in cache
+        if: steps.retrieve-data-cache.outputs.cache-hit != 'true'
+        continue-on-error: true
+        uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
+        with:
+          path: tests/tfhe-backward-compat-data
+          key: ${{ steps.backward_compat_branch.outputs.branch }}_${{ steps.backward_compat_sha.outputs.sha }}
+
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Backward compatibility tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Backward compatibility tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (backward-compat-tests)
@@ -127,4 +154,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (backward-compat-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (backward-compat-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_fast_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -63,7 +65,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             dependencies:
@@ -158,7 +160,7 @@ jobs:
     name: Fast CPU tests
     needs: [ should-run, setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -169,7 +171,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -209,7 +211,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/restore@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         with:
           path: |
             ~/.nvm
@@ -222,7 +224,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |
@@ -264,13 +266,18 @@ jobs:
         run: |
           make test_zk
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() && env.SECRETS_AVAILABLE == 'true' }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Fast AWS tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Fast AWS tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (fast-tests)
@@ -295,4 +302,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (fast-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (fast-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_integer_tests.yml

@@ -10,6 +10,8 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
+  PULL_REQUEST_MD_LINK: ""
   # We clear the cache to reduce memory pressure because of the numerous processes of cargo
   # nextest
   TFHE_RS_CLEAR_IN_MEMORY_KEY_CACHE: "1"
@@ -50,7 +52,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             integer:
@@ -98,7 +100,7 @@ jobs:
     name: Unsigned integer tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -109,7 +111,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -134,13 +136,18 @@ jobs:
         run: |
           AVX512_SUPPORT=ON NO_BIG_PARAMS=${{ env.NO_BIG_PARAMS }} BIG_TESTS_INSTANCE=TRUE make test_unsigned_integer_ci
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Unsigned Integer tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Unsigned Integer tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (unsigned-integer-tests)
@@ -165,4 +172,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (unsigned-integer-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (unsigned-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_signed_integer_tests.yml

@@ -10,6 +10,8 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
+  PULL_REQUEST_MD_LINK: ""
   # We clear the cache to reduce memory pressure because of the numerous processes of cargo
   # nextest
   TFHE_RS_CLEAR_IN_MEMORY_KEY_CACHE: "1"
@@ -51,7 +53,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             integer:
@@ -99,7 +101,7 @@ jobs:
     name: Signed integer tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -110,7 +112,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -139,13 +141,18 @@ jobs:
         run: |
           AVX512_SUPPORT=ON NO_BIG_PARAMS=${{ env.NO_BIG_PARAMS }} BIG_TESTS_INSTANCE=TRUE make test_signed_integer_ci
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Signed Integer tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Signed Integer tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (signed-integer-tests)
@@ -170,4 +177,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (signed-integer-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (signed-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_tests.yml

@@ -10,7 +10,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -72,7 +74,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             dependencies:
@@ -169,7 +171,7 @@ jobs:
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     needs: [ should-run, setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}_${{github.event_name}}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -180,7 +182,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -246,13 +248,18 @@ jobs:
           make test_trivium
           make test_kreyvium
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "CPU tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "CPU tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cpu-tests)
@@ -277,4 +284,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cpu-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cpu-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/aws_tfhe_wasm_tests.yml

@@ -10,6 +10,8 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -52,7 +54,7 @@ jobs:
     name: WASM tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -63,7 +65,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -73,7 +75,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/restore@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         with:
           path: |
             ~/.nvm
@@ -86,7 +88,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |
@@ -115,13 +117,18 @@ jobs:
         run: |
           make test_zk_wasm_x86_compat_ci
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "WASM tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "WASM tests finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (wasm-tests)
@@ -146,4 +153,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (wasm-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (wasm-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/benchmark_boolean.yml

@@ -43,7 +43,7 @@ jobs:
     needs: setup-instance
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     steps:
@@ -63,7 +63,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -94,7 +94,7 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_boolean
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_core_crypto.yml

@@ -43,7 +43,7 @@ jobs:
     needs: setup-instance
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     steps:
       - name: Checkout tfhe-rs repo with tags
@@ -62,12 +62,13 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
       - name: Run benchmarks with AVX512
         run: |
+          make bench_ks_pbs
           make bench_pbs
           make bench_pbs128
           make bench_ks
@@ -85,7 +86,7 @@ jobs:
           --walk-subdirs
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_erc20.yml

@@ -43,7 +43,7 @@ jobs:
     needs: setup-instance
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     timeout-minutes: 720  # 12 hours
@@ -64,7 +64,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -99,7 +99,7 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_erc20
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_gpu.yml

@@ -26,6 +26,7 @@ on:
           - integer_multi_bit
           - integer_compression
           - pbs
+          - pbs128
           - ks
       op_flavor:
         description: "Operations set to run"
@@ -68,7 +69,7 @@ jobs:
   run-benchmarks:
     name: Run benchmarks
     needs: parse-inputs
-    uses: ./.github/workflows/benchmark_gpu_integer_common.yml
+    uses: ./.github/workflows/benchmark_gpu_common.yml
     with:
       profile: ${{ needs.parse-inputs.outputs.profile }}
       hardware_name: ${{ needs.parse-inputs.outputs.hardware_name }}

.github/workflows/benchmark_gpu_4090.yml

@@ -29,7 +29,7 @@ jobs:
       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
+      group: ${{ github.workflow_ref }}_cuda_integer_bench
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ["self-hosted", "4090-desktop"]
     timeout-minutes: 1440 # 24 hours
@@ -51,7 +51,7 @@ jobs:
           echo "FAST_BENCH=TRUE" >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -80,7 +80,7 @@ jobs:
           --walk-subdirs
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_integer_multi_bit_gpu_default
           path: ${{ env.RESULTS_FILENAME }}
@@ -97,14 +97,14 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Integer RTX 4090 full benchmarks finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Integer RTX 4090 full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
 
   cuda-core-crypto-benchmarks:
     name: Cuda core crypto benchmarks  (RTX 4090)
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'schedule' || contains(github.event.label.name, '4090_bench') }}
     needs: cuda-integer-benchmarks
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}_cuda_core_crypto_bench
+      group: ${{ github.workflow_ref }}_cuda_core_crypto_bench
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ["self-hosted", "4090-desktop"]
     timeout-minutes: 1440 # 24 hours
@@ -126,7 +126,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -157,7 +157,7 @@ jobs:
       
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}
@@ -182,7 +182,7 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Core crypto RTX 4090 full benchmarks finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Core crypto RTX 4090 full benchmarks finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
 
   remove_github_label:
     name: Remove 4090 bench label

.github/workflows/benchmark_gpu_common.yml

@@ -1,4 +1,4 @@
-# Run integer benchmarks on CUDA instance and return parsed results to Slab CI bot.
+# Run benchmarks on CUDA instance and return parsed results to Slab CI bot.
 name: Cuda benchmarks - common
 
 on:
@@ -114,10 +114,16 @@ jobs:
     needs: prepare-matrix
     runs-on: ubuntu-latest
     outputs:
-      runner-name: ${{ steps.start-instance.outputs.label }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
-      - name: Start instance
-        id: start-instance
+      - name: Start remote instance
+        id: start-remote-instance
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -127,9 +133,43 @@ jobs:
           backend: ${{ inputs.backend }}
           profile: ${{ inputs.profile }}
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: steps.start-remote-instance.outcome == 'failure' &&
+          inputs.profile == 'single-h100'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
+  # Install dependencies only once since cuda-benchmarks uses a matrix strategy, thus running multiple times.
+  install-dependencies:
+    name: Install dependencies
+    needs: [ setup-instance ]
+    runs-on: ${{ needs.setup-instance.outputs.runner-name }}
+    strategy:
+      matrix:
+        # explicit include-based build matrix, of known valid options
+        include:
+          - cuda: "12.2"
+            gcc: 11
+    steps:
+      - name: Checkout tfhe-rs repo
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
+        with:
+          fetch-depth: 0
+          persist-credentials: 'false'
+          token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
+
+      - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
+        uses: ./.github/actions/gpu_setup
+        with:
+          cuda-version: ${{ matrix.cuda }}
+          gcc-version: ${{ matrix.gcc }}
+
   cuda-benchmarks:
     name: Cuda benchmarks (${{ inputs.profile }})
-    needs: [ prepare-matrix, setup-instance ]
+    needs: [ prepare-matrix, setup-instance, install-dependencies ]
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     timeout-minutes: 1440 # 24 hours
     continue-on-error: true
@@ -140,10 +180,8 @@ jobs:
         command: ${{ fromJSON(needs.prepare-matrix.outputs.command) }}
         op_flavor: ${{ fromJSON(needs.prepare-matrix.outputs.op_flavor) }}
         bench_type: ${{ fromJSON(needs.prepare-matrix.outputs.bench_type) }}
-        # explicit include-based build matrix, of known valid options
         include:
-          - os: ubuntu-22.04
-            cuda: "12.2"
+          - cuda: "12.2"
             gcc: 11
     steps:
       - name: Checkout tfhe-rs repo with tags
@@ -153,12 +191,6 @@ jobs:
           persist-credentials: 'false'
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
-      - name: Setup Hyperstack dependencies
-        uses: ./.github/actions/gpu_setup
-        with:
-          cuda-version: ${{ matrix.cuda }}
-          gcc-version: ${{ matrix.gcc }}
-
       - name: Get benchmark details
         run: |
           {
@@ -167,8 +199,28 @@ jobs:
             echo "COMMIT_HASH=$(git describe --tags --dirty)";
           } >> "${GITHUB_ENV}"
 
+      # Re-export environment variables as dependencies setup perform this task in the previous job.
+      # Local env variables are cleaned at the end of each job.
+      - name: Export CUDA variables
+        shell: bash
+        run: |
+          CUDA_PATH=/usr/local/cuda-${{ matrix.cuda }}
+          echo "CUDA_PATH=$CUDA_PATH" >> "${GITHUB_ENV}"
+          echo "PATH=$PATH:$CUDA_PATH/bin" >> "${GITHUB_PATH}"
+          echo "LD_LIBRARY_PATH=$CUDA_PATH/lib64:$LD_LIBRARY_PATH" >> "${GITHUB_ENV}"
+          echo "CUDA_MODULE_LOADER=EAGER" >> "${GITHUB_ENV}"
+
+      - name: Export gcc and g++ variables
+        shell: bash
+        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: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -196,7 +248,7 @@ jobs:
           --bench-type ${{ matrix.bench_type }}
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}_${{ inputs.profile }}
           path: ${{ env.RESULTS_FILENAME }}
@@ -230,7 +282,7 @@ jobs:
 
   teardown-instance:
     name: Teardown instance (cuda-${{ inputs.profile }}-benchmarks)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-benchmarks, slack-notify ]
     runs-on: ubuntu-latest
     steps:

.github/workflows/benchmark_gpu_core_crypto.yml

@@ -1,5 +1,5 @@
 # Run core crypto benchmarks on an instance with CUDA and return parsed results to Slab CI bot.
-name: Core crypto GPU benchmarks
+name: Cuda - Core crypto benchmarks
 
 on:
   workflow_dispatch:
@@ -23,10 +23,16 @@ jobs:
     if: github.event_name != 'schedule' ||
       (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
-      runner-name: ${{ steps.start-instance.outputs.label }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
-      - name: Start instance
-        id: start-instance
+      - name: Start remote instance
+        id: start-remote-instance
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -36,6 +42,13 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
   cuda-core-crypto-benchmarks:
     name: Execute GPU core crypto benchmarks
     needs: setup-instance
@@ -57,6 +70,7 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -71,13 +85,15 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
       - name: Run benchmarks with AVX512
         run: |
+          make bench_ks_pbs_gpu
           make bench_pbs_gpu
+          make bench_pbs128_gpu
           make bench_ks_gpu
 
       - name: Parse results
@@ -94,7 +110,7 @@ jobs:
           --walk-subdirs
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_core_crypto
           path: ${{ env.RESULTS_FILENAME }}
@@ -128,7 +144,7 @@ jobs:
 
   teardown-instance:
     name: Teardown instance (cuda-integer-full-benchmarks)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-core-crypto-benchmarks, slack-notify ]
     runs-on: ubuntu-latest
     steps:

.github/workflows/benchmark_gpu_erc20_common.yml

@@ -50,10 +50,16 @@ jobs:
     if:  github.event_name == 'workflow_dispatch' ||
       (github.event_name == 'schedule' && github.repository == 'zama-ai/tfhe-rs')
     outputs:
-      runner-name: ${{ steps.start-instance.outputs.label }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
-      - name: Start instance
-        id: start-instance
+      - name: Start remote instance
+        id: start-remote-instance
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -63,6 +69,15 @@ jobs:
           backend: ${{ inputs.backend }}
           profile: ${{ inputs.profile }}
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' &&
+          steps.start-remote-instance.outcome == 'failure' &&
+          inputs.profile == 'single-h100'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
   cuda-erc20-benchmarks:
     name: Cuda ERC20 benchmarks (${{ inputs.profile }})
     needs: setup-instance
@@ -84,6 +99,7 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -98,7 +114,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -120,7 +136,7 @@ jobs:
           --name-suffix avx512
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_erc20_${{ inputs.profile }}
           path: ${{ env.RESULTS_FILENAME }}
@@ -154,7 +170,7 @@ jobs:
 
   teardown-instance:
     name: Teardown instance (cuda-erc20-${{ inputs.profile }}-benchmarks)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-erc20-benchmarks, slack-notify ]
     runs-on: ubuntu-latest
     steps:

.github/workflows/benchmark_gpu_weekly.yml

@@ -10,7 +10,7 @@ jobs:
   run-benchmarks-1-h100:
     name: Run benchmarks (1xH100)
     if: github.repository == 'zama-ai/tfhe-rs'
-    uses: ./.github/workflows/benchmark_gpu_integer_common.yml
+    uses: ./.github/workflows/benchmark_gpu_common.yml
     with:
       profile: single-h100
       hardware_name: n3-H100x1
@@ -23,7 +23,7 @@ jobs:
   run-benchmarks-2-h100:
     name: Run benchmarks (2xH100)
     if: github.repository == 'zama-ai/tfhe-rs'
-    uses: ./.github/workflows/benchmark_gpu_integer_common.yml
+    uses: ./.github/workflows/benchmark_gpu_common.yml
     with:
       profile: 2-h100
       hardware_name: n3-H100x2
@@ -36,7 +36,7 @@ jobs:
   run-benchmarks-8-h100:
     name: Run benchmarks (8xH100)
     if: github.repository == 'zama-ai/tfhe-rs'
-    uses: ./.github/workflows/benchmark_gpu_integer_common.yml
+    uses: ./.github/workflows/benchmark_gpu_common.yml
     with:
       profile: multi-h100
       hardware_name: n3-H100x8
@@ -49,7 +49,7 @@ jobs:
   run-benchmarks-l40:
     name: Run benchmarks (L40)
     if: github.repository == 'zama-ai/tfhe-rs'
-    uses: ./.github/workflows/benchmark_gpu_integer_common.yml
+    uses: ./.github/workflows/benchmark_gpu_common.yml
     with:
       profile: l40
       hardware_name: n3-L40x1

.github/workflows/benchmark_integer.yml

@@ -104,7 +104,7 @@ jobs:
     needs: [ prepare-matrix, setup-instance ]
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     timeout-minutes: 1440  # 24 hours
@@ -131,7 +131,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -172,7 +172,7 @@ jobs:
           --bench-type ${{ matrix.bench_type }}
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}_${{ matrix.bench_type }}
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_shortint.yml

@@ -70,7 +70,7 @@ jobs:
     needs: [ prepare-matrix, setup-instance ]
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     strategy:
@@ -94,7 +94,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -138,7 +138,7 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_shortint_${{ matrix.op_flavor }}
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_signed_integer.yml

@@ -104,7 +104,7 @@ jobs:
     needs: [ prepare-matrix, setup-instance ]
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     continue-on-error: true
     timeout-minutes: 1440  # 24 hours
@@ -131,7 +131,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -166,7 +166,7 @@ jobs:
           --bench-type ${{ matrix.bench_type }}
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_${{ matrix.command }}_${{ matrix.op_flavor }}_${{ matrix.bench_type }}
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_tfhe_fft.yml

@@ -45,7 +45,7 @@ jobs:
     name: Execute FFT benchmarks in EC2
     needs: setup-ec2
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-ec2.outputs.runner-name }}
     steps:
@@ -84,7 +84,7 @@ jobs:
           --name-suffix avx512
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_fft
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_tfhe_ntt.yml

@@ -45,7 +45,7 @@ jobs:
     name: Execute NTT benchmarks in EC2
     needs: setup-ec2
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-ec2.outputs.runner-name }}
     steps:
@@ -84,7 +84,7 @@ jobs:
           --name-suffix avx512
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_ntt
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_tfhe_zk_pok.yml

@@ -45,7 +45,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             zk_pok:
@@ -80,7 +80,7 @@ jobs:
     if: needs.setup-instance.result != 'skipped'
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.ref }}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
+      group: ${{ github.workflow_ref }}_${{github.event_name}}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -100,7 +100,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -132,7 +132,7 @@ jobs:
           --bench-type ${{ env.BENCH_TYPE }}
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_tfhe_zk_pok
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_wasm_client.yml

@@ -41,7 +41,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             wasm_bench:
@@ -100,7 +100,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -110,7 +110,7 @@ jobs:
 
       - name: Node cache restoration
         id: node-cache
-        uses: actions/cache/restore@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/restore@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         with:
           path: |
             ~/.nvm
@@ -123,7 +123,7 @@ jobs:
           make install_node
 
       - name: Node cache save
-        uses: actions/cache/save@1bd1e32a3bdc45362d1e726936510720a7c30a57 #v4.2.0
+        uses: actions/cache/save@5a3ec84eff668545956fd18022155c47e93e2684 #v4.2.3
         if: steps.node-cache.outputs.cache-hit != 'true'
         with:
           path: |
@@ -167,7 +167,7 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_wasm_${{ matrix.browser }}
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/benchmark_zk_pke.yml

@@ -48,7 +48,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             zk_pok:
@@ -118,7 +118,7 @@ jobs:
     if: needs.setup-instance.result != 'skipped'
     needs: [ prepare-matrix, setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.ref }}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
+      group: ${{ github.workflow_ref }}_${{github.event_name}}${{ github.ref == 'refs/heads/main' && github.sha || '' }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -142,7 +142,7 @@ jobs:
           } >> "${GITHUB_ENV}"
 
       - name: Install rust
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: nightly
 
@@ -179,7 +179,7 @@ jobs:
           --append-results
 
       - name: Upload parsed results artifact
-        uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
         with:
           name: ${{ github.sha }}_integer_zk
           path: ${{ env.RESULTS_FILENAME }}

.github/workflows/cargo_build.yml

@@ -28,7 +28,7 @@ jobs:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 

.github/workflows/cargo_test_fft.yml

@@ -3,16 +3,46 @@ name: Cargo Test tfhe-fft
 
 on:
   pull_request:
+  push:
+    branches:
+      - main
 
 env:
   CARGO_TERM_COLOR: always
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref }}
   cancel-in-progress: true
 
 jobs:
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: read
+    outputs:
+      fft_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.fft_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
+        with:
+          fetch-depth: 0
+          persist-credentials: 'false'
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
+        with:
+          files_yaml: |
+            fft:
+              - tfhe/Cargo.toml
+              - Makefile
+              - tfhe-fft/**
+              - '.github/workflows/cargo_test_fft.yml'
+
   cargo-tests-fft:
+    needs: should-run
+    if: needs.should-run.outputs.fft_test == 'true'
     runs-on: ${{ matrix.runner_type }}
     strategy:
       matrix:
@@ -39,6 +69,8 @@ jobs:
           make test_fft_no_std
 
   cargo-tests-fft-nightly:
+    needs: should-run
+    if: needs.should-run.outputs.fft_test == 'true'
     runs-on: ${{ matrix.runner_type }}
     strategy:
       matrix:
@@ -61,7 +93,9 @@ jobs:
           make test_fft_no_std_nightly
 
   cargo-tests-fft-node-js:
-    runs-on: "ubuntu-latest"
+    needs: should-run
+    if: needs.should-run.outputs.fft_test == 'true'
+    runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
 
@@ -69,3 +103,30 @@ jobs:
         run: |
           make install_node
           make test_fft_node_js_ci
+
+  cargo-tests-fft-successful:
+    needs: [ should-run, cargo-tests-fft, cargo-tests-fft-nightly, cargo-tests-fft-node-js ]
+    if: ${{ always() }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Tests do not need to run
+        if: needs.should-run.outputs.fft_test == 'false'
+        run: |
+          echo "tfhe-fft files haven't changed tests don't need to run"
+
+      - name: Check all tests passed
+        if: needs.should-run.outputs.fft_test == 'true' &&
+          needs.cargo-tests-fft.result == 'success' &&
+          needs.cargo-tests-fft-nightly.result == 'success' &&
+          needs.cargo-tests-fft-node-js.result == 'success'
+        run: |
+          echo "All tfhe-fft test passed"
+
+      - name: Check tests failure
+        if: needs.should-run.outputs.fft_test == 'true' &&
+          (needs.cargo-tests-fft.result != 'success' ||
+          needs.cargo-tests-fft-nightly.result != 'success' ||
+          needs.cargo-tests-fft-node-js.result != 'success')
+        run: |
+          echo "Some tfhe-fft tests failed"
+          exit 1

.github/workflows/cargo_test_ntt.yml

@@ -3,16 +3,46 @@ name: Cargo Test tfhe-ntt
 
 on:
   pull_request:
+  push:
+    branches:
+      - main
 
 env:
   CARGO_TERM_COLOR: always
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref }}
   cancel-in-progress: true
 
-jobs:
+jobs:  
+  should-run:
+    runs-on: ubuntu-latest
+    permissions:
+      pull-requests: read
+    outputs:
+      ntt_test: ${{ env.IS_PULL_REQUEST == 'false' || steps.changed-files.outputs.ntt_any_changed }}
+    steps:
+      - name: Checkout tfhe-rs
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
+        with:
+          fetch-depth: 0
+          persist-credentials: 'false'
+
+      - name: Check for file changes
+        id: changed-files
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
+        with:
+          files_yaml: |
+            ntt:
+              - tfhe/Cargo.toml
+              - Makefile
+              - tfhe-ntt/**
+              - '.github/workflows/cargo_test_ntt.yml'
+
   cargo-tests-ntt:
+    needs: should-run
+    if: needs.should-run.outputs.ntt_test == 'true'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -34,6 +64,8 @@ jobs:
         run: make test_ntt_no_std
 
   cargo-tests-ntt-nightly:
+    needs: should-run
+    if: needs.should-run.outputs.ntt_test == 'true'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -52,3 +84,28 @@ jobs:
 
       - name: Test no-std nightly
         run: make test_ntt_no_std_nightly
+
+  cargo-tests-ntt-successful:
+    needs: [ should-run, cargo-tests-ntt, cargo-tests-ntt-nightly ]
+    if: ${{ always() }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Tests do not need to run
+        if: needs.should-run.outputs.ntt_test == 'false'
+        run: |
+          echo "tfhe-ntt files haven't changed tests don't need to run"
+
+      - name: Check all tests success
+        if: needs.should-run.outputs.ntt_test == 'true' &&
+          needs.cargo-tests-ntt.result == 'success' &&
+          needs.cargo-tests-ntt-nightly.result == 'success'
+        run: |
+          echo "All tfhe-ntt tests passed"
+
+      - name: Check tests failure
+        if: needs.should-run.outputs.ntt_test == 'true' &&
+          (needs.cargo-tests-ntt.result != 'success' ||
+          needs.cargo-tests-ntt-nightly.result != 'success')
+        run: |
+          echo "Some tfhe-ntt tests failed"
+          exit 1

.github/workflows/code_coverage.yml

@@ -38,7 +38,7 @@ jobs:
     name: Code coverage tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.event_name }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}_${{ github.event_name }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     timeout-minutes: 5760 # 4 days
@@ -47,13 +47,13 @@ jobs:
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             tfhe:
@@ -83,7 +83,7 @@ jobs:
           make test_shortint_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@13ce06bfc6bbe3ecf90edbbf1bc32fe5978ca1d3
+        uses: codecov/codecov-action@0565863a31f2c772f9f0395002a31e3f06189574
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
@@ -97,7 +97,7 @@ jobs:
           make test_integer_cov
 
       - name: Upload tfhe coverage to Codecov
-        uses: codecov/codecov-action@13ce06bfc6bbe3ecf90edbbf1bc32fe5978ca1d3
+        uses: codecov/codecov-action@0565863a31f2c772f9f0395002a31e3f06189574
         if: steps.changed-files.outputs.tfhe_any_changed == 'true'
         with:
           token: ${{ secrets.CODECOV_TOKEN }}

.github/workflows/csprng_randomness_tests.yml

@@ -52,7 +52,7 @@ jobs:
     name: CSPRNG randomness tests
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     steps:
@@ -63,7 +63,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -77,7 +77,7 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "tfhe-csprng randomness check finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "tfhe-csprng randomness check finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
 
   teardown-instance:
     name: Teardown instance (csprng-randomness-tests)
@@ -102,4 +102,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (csprng-randomness-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (csprng-randomness-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/data_pr_close.yml

@@ -8,7 +8,7 @@ env:
   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 }}
+  PR_BRANCH: ${{ github.ref_name }}
   CLOSE_TYPE: ${{ github.event.pull_request.merged && 'merge' || 'close' }}
 
 # only trigger on pull request closed events

.github/workflows/gpu_4090_tests.yml

@@ -29,9 +29,10 @@ jobs:
       contains(github.event.label.name, '4090_test') ||
       (github.event_name == 'schedule' &&  github.repository == 'zama-ai/tfhe-rs')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ["self-hosted", "4090-desktop"]
+    timeout-minutes: 1440 # 24 hours
 
     steps:
       - name: Checkout tfhe-rs
@@ -41,7 +42,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -81,4 +82,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "CUDA RTX 4090 tests finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "CUDA RTX 4090 tests finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_fast_h100_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -40,7 +42,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -68,11 +70,17 @@ jobs:
       (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-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -82,6 +90,13 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -95,7 +110,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -114,6 +129,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -121,7 +137,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -150,16 +166,21 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Fast H100 tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Fast H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-h100-tests)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
@@ -180,4 +201,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_fast_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -39,7 +41,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -93,7 +95,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -119,7 +121,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -148,12 +150,17 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Base GPU tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Base GPU tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-tests)
@@ -178,4 +185,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_full_h100_tests.yml

@@ -20,10 +20,16 @@ jobs:
     name: Setup instance (cuda-h100-tests)
     runs-on: ubuntu-latest
     outputs:
-      runner-name: ${{ steps.start-instance.outputs.label }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
-      - name: Start instance
-        id: start-instance
+      - name: Start remote instance
+        id: start-remote-instance
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -33,11 +39,18 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
   cuda-tests-linux:
     name: CUDA H100 tests
     needs: [ setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -68,13 +81,14 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
           gcc-version: ${{ matrix.gcc }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -105,10 +119,11 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Full H100 tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Full 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.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
@@ -128,4 +143,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_full_multi_gpu_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -40,7 +42,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -95,7 +97,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -121,7 +123,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -153,12 +155,17 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Multi-GPU tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Multi-GPU tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-tests-multi-gpu)
@@ -183,4 +190,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-tests-multi-gpu) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-tests-multi-gpu) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_integer_long_run_tests.yml

@@ -1,4 +1,4 @@
-name: Long Run Tests on GPU
+name: Cuda - Long Run Tests on GPU
 
 env:
   CARGO_TERM_COLOR: always
@@ -15,8 +15,8 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   schedule:
-    # Weekly tests will be triggered each Friday at 9p.m.
-    - cron: "0 21 * * 5"
+    # Nightly tests will be triggered each evening 8p.m.
+    - cron: "0 20 * * *"
 
 jobs:
   setup-instance:
@@ -42,7 +42,7 @@ jobs:
     name: Long run GPU tests
     needs: [ setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}_${{github.event_name}}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -65,7 +65,7 @@ jobs:
           gcc-version: ${{ matrix.gcc }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 

.github/workflows/gpu_pcc.yml

@@ -11,6 +11,8 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -49,7 +51,7 @@ jobs:
     name: CUDA post-commit checks
     needs: setup-instance
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -81,7 +83,7 @@ jobs:
           sudo apt -y install "cuda-toolkit-${TOOLKIT_VERSION}" cmake-format
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -111,13 +113,18 @@ jobs:
         run: |
           make pcc_gpu
 
+      - name: Set pull-request URL
+        if: ${{ failure() && github.event_name == 'pull_request' }}
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Slack Notification
         if: ${{ failure() && env.SECRETS_AVAILABLE == 'true' }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "CUDA AWS post-commit checks finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "CUDA AWS post-commit checks finished with status: ${{ job.status }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-pcc)
@@ -142,4 +149,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-pcc) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-pcc) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_signed_integer_classic_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -40,7 +42,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -95,7 +97,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -121,7 +123,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -136,12 +138,17 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Integer GPU signed integer tests with classical PBS finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Integer GPU signed integer tests with classical PBS finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-signed-classic-tests)
@@ -166,4 +173,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-signed-classic-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-signed-classic-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_signed_integer_h100_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -41,7 +43,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -69,11 +71,17 @@ jobs:
       (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-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -83,6 +91,13 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -96,7 +111,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -115,6 +130,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -122,7 +138,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -137,16 +153,21 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Integer GPU H100 tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Integer GPU H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-h100-tests)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
@@ -167,4 +188,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_signed_integer_tests.yml

@@ -11,9 +11,11 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   FAST_TESTS: TRUE
   NIGHTLY_TESTS: FALSE
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -44,7 +46,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -63,6 +65,7 @@ jobs:
               - '.github/workflows/gpu_signed_integer_tests.yml'
               - scripts/integer-tests.sh
               - ci/slab.toml
+
   setup-instance:
     name: Setup instance (cuda-signed-integer-tests)
     runs-on: ubuntu-latest
@@ -98,7 +101,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -108,7 +111,7 @@ jobs:
         include:
           - os: ubuntu-22.04
             cuda: "12.2"
-            gcc: 11 
+            gcc: 11
     steps:
       - name: Checkout tfhe-rs
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683
@@ -124,7 +127,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -147,11 +150,17 @@ jobs:
     if: ${{ always() && needs.cuda-signed-integer-tests.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
+        if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-signed-integer-tests.result }}
-          SLACK_MESSAGE: "Base GPU tests finished with status: ${{ needs.cuda-signed-integer-tests.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Signed GPU tests finished with status: ${{ needs.cuda-signed-integer-tests.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-tests)
@@ -176,4 +185,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-signed-integer-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-signed-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_unsigned_integer_classic_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -41,7 +43,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -96,7 +98,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -122,7 +124,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -137,12 +139,17 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Unsigned integer GPU classic tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Unsigned integer GPU classic tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-unsigned-classic-tests)
@@ -167,4 +174,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-unsigned-classic-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-unsigned-classic-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_unsigned_integer_h100_tests.yml

@@ -11,7 +11,9 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -40,7 +42,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -68,11 +70,17 @@ jobs:
       (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-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      # Use permanent remote instance label first as on-demand remote instance label output is set before the end of start-remote-instance step.
+      # If the latter fails due to a failed GitHub action runner set up, we have to fallback on the permanent instance.
+      # Since the on-demand remote label is set before failure, we have to do the logical OR in this order,
+      # otherwise we'll try to run the next job on a non-existing on-demand instance.
+      runner-name: ${{ steps.use-permanent-instance.outputs.runner_group || steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
+      remote-instance-outcome: ${{ steps.start-remote-instance.outcome }}
     steps:
       - name: Start remote instance
         id: start-remote-instance
         if: env.SECRETS_AVAILABLE == 'true'
+        continue-on-error: true
         uses: zama-ai/slab-github-runner@79939325c3c429837c10d6041e4fd8589d328bac
         with:
           mode: start
@@ -82,6 +90,13 @@ jobs:
           backend: hyperstack
           profile: single-h100
 
+      # This will allow to fallback on permanent instances running on Hyperstack.
+      - name: Use permanent remote instance
+        id: use-permanent-instance
+        if: env.SECRETS_AVAILABLE == 'true' && steps.start-remote-instance.outcome == 'failure'
+        run: |
+          echo "runner_group=h100x1" >> "$GITHUB_OUTPUT"
+
       # This instance will be spawned especially for pull-request from forked repository
       - name: Start GitHub instance
         id: start-github-instance
@@ -95,7 +110,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -114,6 +129,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Setup Hyperstack dependencies
+        if: needs.setup-instance.outputs.remote-instance-outcome == 'success'
         uses: ./.github/actions/gpu_setup
         with:
           cuda-version: ${{ matrix.cuda }}
@@ -121,7 +137,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -136,16 +152,21 @@ jobs:
     if: ${{ always() && needs.cuda-tests-linux.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-tests-linux.result }}
-          SLACK_MESSAGE: "Unsigned integer GPU H100 tests finished with status: ${{ needs.cuda-tests-linux.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Unsigned integer GPU H100 tests finished with status: ${{ needs.cuda-tests-linux.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-h100-tests)
-    if: ${{ always() && needs.setup-instance.result == 'success' }}
+    if: ${{ always() && needs.setup-instance.outputs.remote-instance-outcome == 'success' }}
     needs: [ setup-instance, cuda-tests-linux ]
     runs-on: ubuntu-latest
     steps:
@@ -166,4 +187,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-h100-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/gpu_unsigned_integer_tests.yml

@@ -11,8 +11,11 @@ env:
   SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
   SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
   SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
+  SLACKIFY_MARKDOWN: true
   FAST_TESTS: TRUE
   NIGHTLY_TESTS: FALSE
+  IS_PULL_REQUEST: ${{ github.event_name == 'pull_request' }}
+  PULL_REQUEST_MD_LINK: ""
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
   # Secrets will be available only to zama-ai organization members
   SECRETS_AVAILABLE: ${{ secrets.JOB_SECRET != '' }}
@@ -22,7 +25,6 @@ on:
   # Allows you to run this workflow manually from the Actions tab as an alternative.
   workflow_dispatch:
   pull_request:
-    types: [ labeled ]
   schedule:
     # Nightly tests @ 1AM after each work day
     - cron: "0 1 * * MON-FRI"
@@ -44,7 +46,7 @@ jobs:
 
       - name: Check for file changes
         id: changed-files
-        uses: tj-actions/changed-files@dcc7a0cba800f454d79fff4b993e8c3555bcc0a8
+        uses: tj-actions/changed-files@26a38635fc1173cc5820336ce97be6188d0de9f5 # v46.0.2
         with:
           files_yaml: |
             gpu:
@@ -66,11 +68,11 @@ jobs:
 
   setup-instance:
     name: Setup instance (cuda-unsigned-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' ||
       needs.should-run.outputs.gpu_test == 'true'
-    runs-on: ubuntu-latest
     outputs:
       runner-name: ${{ steps.start-remote-instance.outputs.label || steps.start-github-instance.outputs.runner_group }}
     steps:
@@ -99,7 +101,7 @@ jobs:
     if: github.event_name != 'pull_request' ||
       (github.event_name == 'pull_request' && needs.setup-instance.result != 'skipped')
     concurrency:
-      group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+      group: ${{ github.workflow_ref }}
       cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     strategy:
@@ -125,7 +127,7 @@ jobs:
           github-instance: ${{ env.SECRETS_AVAILABLE == 'false' }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -148,12 +150,17 @@ jobs:
     if: ${{ always() && needs.cuda-unsigned-integer-tests.result != 'skipped' && failure() }}
     continue-on-error: true
     steps:
+      - name: Set pull-request URL
+        if: env.SECRETS_AVAILABLE == 'true' && github.event_name == 'pull_request'
+        run: |
+          echo "PULL_REQUEST_MD_LINK=[pull-request](${{ vars.PR_BASE_URL }}${{ github.event.pull_request.number }}), "  >> "${GITHUB_ENV}"
+
       - name: Send message
         if: env.SECRETS_AVAILABLE == 'true'
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ needs.cuda-unsigned-integer-tests.result }}
-          SLACK_MESSAGE: "Unsigned integer GPU tests finished with status: ${{ needs.cuda-unsigned-integer-tests.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Unsigned integer GPU tests finished with status: ${{ needs.cuda-unsigned-integer-tests.result }}. (${{ env.PULL_REQUEST_MD_LINK }}[action run](${{ env.ACTION_RUN_URL }}))"
 
   teardown-instance:
     name: Teardown instance (cuda-tests)
@@ -178,4 +185,4 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_MESSAGE: "Instance teardown (cuda-unsigned-integer-tests) finished with status: ${{ job.status }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "Instance teardown (cuda-unsigned-integer-tests) finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"

.github/workflows/integer_long_run_tests.yml

@@ -42,7 +42,7 @@ jobs:
     name: Long run CPU tests
     needs: [ setup-instance ]
     concurrency:
-      group: ${{ github.workflow }}_${{github.event_name}}_${{ github.ref }}
+      group: ${{ github.workflow_ref }}_${{github.event_name}}
       cancel-in-progress: true
     runs-on: ${{ needs.setup-instance.outputs.runner-name }}
     timeout-minutes: 4320 # 72 hours
@@ -54,7 +54,7 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 

.github/workflows/m1_tests.yml

@@ -24,7 +24,7 @@ env:
   CHECKOUT_TOKEN: ${{ secrets.REPO_CHECKOUT_TOKEN || secrets.GITHUB_TOKEN }}
 
 concurrency:
-  group: ${{ github.workflow }}_${{ github.head_ref || github.ref }}
+  group: ${{ github.workflow_ref }}
   cancel-in-progress: true
 
 jobs:
@@ -43,7 +43,7 @@ jobs:
           token: ${{ env.CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -195,7 +195,7 @@ jobs:
           SLACK_COLOR: ${{ needs.cargo-builds-m1.result }}
           SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
           SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
-          SLACK_MESSAGE: "M1 tests finished with status: ${{ needs.cargo-builds-m1.result }} on '${{ env.BRANCH }}'. (${{ env.ACTION_RUN_URL }})"
+          SLACK_MESSAGE: "M1 tests finished with status: ${{ needs.cargo-builds-m1.result }}. (${{ env.ACTION_RUN_URL }})"
           SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
           SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
           MSG_MINIMAL: event,action url,commit

.github/workflows/make_release.yml

@@ -51,7 +51,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate
           path: target/package/*.crate
@@ -62,7 +62,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -94,7 +94,7 @@ jobs:
         run: |
           echo "NPM_TAG=latest" >> "${GITHUB_ENV}"
       - name: Download artifact
-        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: crate
           path: target/package

.github/workflows/make_release_cuda.yml

@@ -65,7 +65,7 @@ jobs:
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
 
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 
@@ -99,7 +99,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -127,7 +127,7 @@ jobs:
       CUDA_PATH: /usr/local/cuda-${{ matrix.cuda }}
     steps:
       - name: Install latest stable
-        uses: dtolnay/rust-toolchain@a54c7afa936fefeb4456b2dd8068152669aa8203
+        uses: dtolnay/rust-toolchain@888c2e1ea69ab0d4330cbf0af1ecc7b68f368cc1
         with:
           toolchain: stable
 

.github/workflows/make_release_tfhe_csprng.yml

@@ -30,7 +30,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe-csprng
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate-tfhe-csprng
           path: target/package/*.crate
@@ -42,7 +42,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -66,7 +66,7 @@ jobs:
           fetch-depth: 0
           token: ${{ secrets.FHE_ACTIONS_TOKEN }}
       - name: Download artifact
-        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: crate-tfhe-csprng
           path: target/package

.github/workflows/make_release_tfhe_fft.yml

@@ -33,7 +33,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe-fft
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate
           path: target/package/*.crate
@@ -44,7 +44,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read

.github/workflows/make_release_tfhe_ntt.yml

@@ -33,7 +33,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe-ntt
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate
           path: target/package/*.crate
@@ -44,7 +44,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read

.github/workflows/make_release_tfhe_versionable.yml

@@ -2,14 +2,13 @@ name: Publish tfhe-versionable release
 
 on:
   workflow_dispatch:
-    inputs:
-      dry_run:
-        description: "Dry-run"
-        type: boolean
-        default: true
 
 env:
   ACTION_RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+  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:
   verify_tag:
@@ -19,6 +18,7 @@ jobs:
       READ_ORG_TOKEN: ${{ secrets.READ_ORG_TOKEN }}
 
   package-derive:
+    name: Package tfhe-versionable-derive Release
     runs-on: ubuntu-latest
     outputs:
       hash: ${{ steps.hash.outputs.hash }}
@@ -30,7 +30,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe-versionable-derive
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate-tfhe-versionable-derive
           path: target/package/*.crate
@@ -40,7 +40,7 @@ jobs:
 
   provenance-derive:
     needs: [package-derive]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -53,8 +53,8 @@ jobs:
       base64-subjects: ${{ needs.package-derive.outputs.hash }}
 
   publish_release-derive:
-    name: Publish tfhe-versionable Release
-    needs: [verify_tag, package-derive] # for comparing hashes
+    name: Publish tfhe-versionable-derive Release
+    needs: [ verify_tag, package-derive ] # for comparing hashes
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -64,7 +64,7 @@ jobs:
           persist-credentials: 'false'
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
       - name: Download artifact
-        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: crate-tfhe-versionable-derive
           path: target/package
@@ -72,7 +72,7 @@ jobs:
         env:
           CRATES_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
         run: |
-          cargo publish -p tfhe-versionable-derive --token ${{ env.CRATES_TOKEN }} ${{ env.DRY_RUN }}
+          cargo publish -p tfhe-versionable-derive --token ${{ env.CRATES_TOKEN }}
       - name: Generate hash
         id: published_hash
         run: cd target/package && echo "pub_hash=$(sha256sum ./*.crate | base64 -w0)" >> "${GITHUB_OUTPUT}"
@@ -82,24 +82,18 @@ jobs:
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990 # v2.3.2
         env:
           SLACK_COLOR: failure
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "SLSA tfhe-versionable-derive - hash comparison failure: (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990 # v2.3.2
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "tfhe-versionable-derive release finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
 
   package:
+    name: Package tfhe-versionable Release
+    needs: publish_release-derive
     runs-on: ubuntu-latest
     outputs:
       hash: ${{ steps.hash.outputs.hash }}
@@ -111,7 +105,7 @@ jobs:
       - name: Prepare package
         run: |
           cargo package -p tfhe-versionable
-      - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: crate-tfhe-versionable
           path: target/package/*.crate
@@ -120,8 +114,8 @@ jobs:
         run: cd target/package && echo "hash=$(sha256sum ./*.crate | base64 -w0)" >> "${GITHUB_OUTPUT}"
 
   provenance:
-    needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    needs: package
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -135,7 +129,7 @@ jobs:
 
   publish_release:
     name: Publish tfhe-versionable Release
-    needs: [package] # for comparing hashes
+    needs: package # for comparing hashes
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -143,7 +137,7 @@ jobs:
         with:
           fetch-depth: 0
       - name: Download artifact
-        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: crate-tfhe-versionable
           path: target/package
@@ -151,32 +145,21 @@ jobs:
         env:
           CRATES_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
         run: |
-          cargo publish -p tfhe-versionable --token ${{ env.CRATES_TOKEN }} ${{ env.DRY_RUN }}
-
+          cargo publish -p tfhe-versionable --token ${{ env.CRATES_TOKEN }}
       - name: Generate hash
         id: published_hash
         run: cd target/package && echo "pub_hash=$(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@c33737706dea87cd7784c687dadc9adf1be59990 # v2.3.2
         env:
           SLACK_COLOR: failure
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "SLSA tfhe-versionable - hash comparison failure: (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}
-
       - name: Slack Notification
         if: ${{ failure() }}
         continue-on-error: true
         uses: rtCamp/action-slack-notify@c33737706dea87cd7784c687dadc9adf1be59990 # v2.3.2
         env:
           SLACK_COLOR: ${{ job.status }}
-          SLACK_CHANNEL: ${{ secrets.SLACK_CHANNEL }}
-          SLACK_ICON: https://pbs.twimg.com/profile_images/1274014582265298945/OjBKP9kn_400x400.png
           SLACK_MESSAGE: "tfhe-versionable release finished with status: ${{ job.status }}. (${{ env.ACTION_RUN_URL }})"
-          SLACK_USERNAME: ${{ secrets.BOT_USERNAME }}
-          SLACK_WEBHOOK: ${{ secrets.SLACK_WEBHOOK }}

.github/workflows/make_release_zk_pok.yml

@@ -24,7 +24,7 @@ jobs:
         - name: Prepare package
           run: |
             cargo package -p tfhe-zk-pok
-        - uses: actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08 # v4.6.0
+        - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
           with:
             name: crate-zk-pok
             path: target/package/*.crate
@@ -34,7 +34,7 @@ jobs:
   provenance:
     if: ${{ !inputs.dry_run  }}
     needs: [package]
-    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.0.0
+    uses: slsa-framework/slsa-github-generator/.github/workflows/generator_generic_slsa3.yml@v2.1.0
     permissions:
       # Needed to detect the GitHub Actions environment
       actions: read
@@ -64,7 +64,7 @@ jobs:
           persist-credentials: 'false'
           token: ${{ secrets.REPO_CHECKOUT_TOKEN }}
       - name: Download artifact
-        uses: actions/download-artifact@fa0a91b85d4f404e444e00e005971372dc801d16 # v4.1.8
+        uses: actions/download-artifact@95815c38cf2ff2164869cbab79da8d1f422bc89e # v4.2.1
         with:
           name: crate-zk-pok
           path: target/package

.gitignore

@@ -37,3 +37,4 @@ package-lock.json
 
 # Dir used for backward compatibility test data
 tests/tfhe-backward-compat-data/
+ci/

CONTRIBUTING.md

@@ -0,0 +1,233 @@
+# Contributing to TFHE-rs
+
+This document provides guidance on how to contribute to **TFHE-rs**.
+
+There are two ways to contribute:
+
+- **Report issues:** Open issues on GitHub to report bugs, suggest improvements, or note typos.
+- **Submit codes**: To become an official contributor, you must sign our Contributor License Agreement (CLA). Our CLA-bot will guide you through this process when you open your first pull request.
+
+## 1. Setting up the project
+
+Start by [forking](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/fork-a-repo) the **TFHE-rs** repository.
+
+{% hint style="info" %}
+- **Rust version**:  Ensure that you use a Rust version >= 1.81 to compile **TFHE-rs**.
+- **Incompatibility**: AArch64-based machines are not yet supported for Windows as it's currently missing an entropy source to be able to seed the [CSPRNGs](https://en.wikipedia.org/wiki/Cryptographically_secure_pseudorandom_number_generator) used in **TFHE-rs**.
+- **Performance**: For optimal performance, it is highly recommended to run **TFHE-rs** code in release mode with cargo's `--release` flag.
+{% endhint %}
+
+To get more details about the library, please refer to the [documentation](https://docs.zama.ai/tfhe-rs).
+
+## 2. Creating a new branch
+
+When creating your branch, make sure to use the following format :
+
+```
+git checkout -b {feat|fix|docs|chore…}/short_description
+```
+
+For example:
+
+```
+git checkout -b feat/new_feature_X
+```
+
+## 3. Before committing
+
+### 3.1 Linting
+
+Each commit to **TFHE-rs** should conform to the standards of the project. In particular, every source code, docker or workflows files should be linted to prevent programmatic and stylistic errors.
+
+- Rust source code linters: `clippy`
+- Typescript/Javascript source code linters: `eslint`, `prettier`
+
+To apply automatic code formatting, run:
+
+```
+make fmt
+```
+
+You can perform linting of all Cargo targets with:
+
+```
+make clippy_all_targets
+```
+
+### 3.2 Testing
+
+Your contributions must include comprehensive documentation and tests without breaking existing tests. To run pre-commit checks, execute:
+
+```
+make pcc
+```
+
+This command ensure that all the targets in the library are building correctly.
+For a faster check, use:
+
+```
+make fpcc
+```
+
+If you're contributing to GPU code, run also:
+
+```
+make pcc_gpu
+```
+
+Unit testing suites are heavy and can require a lot of computing power and RAM availability.
+Whilst tests are run automatically in continuous integration pipeline, you can run tests locally.
+
+All unit tests have a command formatted as:
+
+```
+make test_*
+```
+
+Run `make help` to display a list of all the commands available.
+
+To quickly test your changes locally, follow these steps:
+ 1. Locate where the code has changed.
+ 2. Add (or modify) a Cargo test filter to the corresponding `make` target in Makefile.
+ 3. Run the target.
+
+{% hint style="success" %}
+`make test_<something>` will print the underlying cargo command in STDOUT. You can quickly test your changes by copy/pasting the command and then modify it to suit your needs.
+{% endhint %}
+
+For example, if you made changes in `tfhe/src/integer/*`, you can test them with the following steps:
+ 1. In `test_integer` target, replace the filter `-- integer::` by `-- my_new_test`.
+ 2. Run `make test_integer`.
+
+## 4. Committing
+
+**TFHE-rs** follows the conventional commit specification to maintain a consistent commit history, essential for Semantic Versioning ([semver.org](https://semver.org/)).
+Commit messages are automatically checked in CI and will be rejected if they do not comply, so make sure that you follow the commit conventions detailed on [this page]
+(https://www.conventionalcommits.org/en/v1.0.0/).
+
+## 5. Rebasing
+
+Before creating a pull request, rebase your branch on the repository's `main` branch. Merge commits are not permitted, thus rebasing ensures fewer conflicts and a smoother PR review process.
+
+## 6. Opening a Pull Request
+
+Once your changes are ready, open a pull request.
+
+For instructions on creating a PR from a fork, refer to GitHub's [official documentation](https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/proposing-changes-to-your-work-with-pull-requests/creating-a-pull-request-from-a-fork).
+
+## 7. Continuous integration
+
+Before a pull request can be merged, several test suites run automatically. Below is an overview of the CI process:
+
+```mermaid
+---
+title: Continuous Integration Process
+---
+sequenceDiagram
+    autonumber
+
+    participant Contributor
+    participant GitHub
+    participant Reviewer
+    participant CI-pipeline
+
+    Contributor ->> GitHub: Open pull-request
+    GitHub -->> Contributor: Ask for CLA signing (once)
+    loop
+        Reviewer ->> GitHub: Review code
+        Reviewer ->> CI-pipeline: Approve workflows (short-run)
+        CI-pipeline -->> GitHub: Send checks results
+        Contributor ->> GitHub: Make changes
+    end
+    Reviewer ->> GitHub: Pull-request approval
+    Reviewer ->> CI-pipeline: Approve workflows (long-run)
+    CI-pipeline -->> GitHub: Send checks results
+    Reviewer -->> GitHub: Merge if pipeline green
+```
+
+> [!Note]
+>Useful details:
+>* pipeline is triggered by humans
+>* review team is located in Paris timezone, pipeline launch will most likely happen during office hours
+>* direct changes to CI related files are not allowed for external contributors
+>* run `make pcc` to fix any build errors before pushing commits
+
+## 8. Data versioning
+
+Data serialized with TFHE-rs must remain backward compatible. This is done using the [tfhe-versionable](https://crates.io/crates/tfhe-versionable) crate.
+
+If you modify a type that derives `Versionize` in a backward-incompatible way, an upgrade implementation must be provided.
+
+For example, these changes are data breaking:
+ * Adding a field to a struct.
+ * Changing the order of the fields within a struct or the variants within an enum.
+ * Renaming a field of a struct or a variant of an enum.
+ * Changing the type of field in a struct or a variant in an enum.
+
+On the contrary, these changes are *not* data breaking:
+ * Renaming a type (unless it implements the `Named` trait).
+ * Adding a variant to the end of an enum.
+
+## Example: adding a field
+
+Suppose you want to add an i32 field to a type named `MyType`. The original type is defined as:
+```rust
+#[derive(Serialize, Deserialize, Versionize)]
+#[versionize(MyTypeVersions)]
+struct MyType {
+  val: u64,
+}
+```
+And you want to change it to:
+```rust
+#[derive(Serialize, Deserialize, Versionize)]
+#[versionize(MyTypeVersions)]
+struct MyType {
+  val: u64,
+  other_val: i32
+}
+```
+
+Follow these steps:
+
+ 1. Navigate to the definition of the dispatch enum of this type. This is the type inside the `#[versionize(MyTypeVersions)]` macro attribute. In general, this type has the same name as the base type with a `Versions` suffix. You should find something like
+
+```rust
+#[derive(VersionsDispatch)]
+enum MyTypeVersions {
+  V0(MyTypeV0),
+  V1(MyType)
+}
+```
+
+ 2. Add a new variant to the enum to preserve the previous version of the type. You can simply copy and paste the previous definition of the type and add a version suffix:
+
+```rust
+#[derive(Version)]
+struct MyTypeV1 {
+  val: u64,
+}
+
+#[derive(VersionsDispatch)]
+enum MyTypeVersions {
+  V0(MyTypeV0),
+  V1(MyTypeV1),
+  V2(MyType) // Here this points to your modified type
+}
+```
+
+ 3. Implement the `Upgrade` trait to define how we should go from the previous version to the current version:
+```rust
+impl Upgrade<MyType> for MyTypeV1 {
+  type Error = Infallible;
+
+   fn upgrade(self) -> Result<MyType, Self::Error> {
+       Ok(MyType {
+           val: self.val,
+           other_val: 0
+        })
+   }
+}
+```
+
+ 4. Fix the upgrade target of the previous version. In this example, `impl Upgrade<MyType> for MyTypeV0 {` should simply be changed to `impl Upgrade<MyTypeV1> for MyTypeV0 {`

Cargo.toml

@@ -6,7 +6,6 @@ members = [
     "tfhe-ntt",
     "tfhe-zk-pok",
     "tasks",
-    "apps/trivium",
     "tfhe-csprng",
     "backends/tfhe-cuda-backend",
     "utils/tfhe-versionable",
@@ -14,14 +13,18 @@ members = [
     "tests",
 ]
 
-exclude = ["tests/backward_compatibility_tests", "utils/tfhe-lints"]
+exclude = [
+    "tests/backward_compatibility_tests",
+    "utils/tfhe-lints",
+    "apps/trivium",
+]
 [workspace.dependencies]
 aligned-vec = { version = "0.6", default-features = false }
 bytemuck = "1.14.3"
 dyn-stack = { version = "0.11", default-features = false }
 itertools = "0.14"
 num-complex = "0.4"
-pulp = { version = "0.20", default-features = false }
+pulp = { version = "0.21", default-features = false }
 rand = "0.8"
 rayon = "1"
 serde = { version = "1.0", default-features = false }

Makefile

@@ -18,6 +18,8 @@ FAST_BENCH?=FALSE
 NIGHTLY_TESTS?=FALSE
 BENCH_OP_FLAVOR?=DEFAULT
 BENCH_TYPE?=latency
+BENCH_PARAM_TYPE?=classical
+BENCH_PARAMS_SET?=default
 NODE_VERSION=22.6
 BACKWARD_COMPAT_DATA_URL=https://github.com/zama-ai/tfhe-backward-compat-data.git
 BACKWARD_COMPAT_DATA_BRANCH?=$(shell ./scripts/backward_compat_data_version.py)
@@ -282,14 +284,14 @@ check_typos: install_typos_checker
 .PHONY: clippy_gpu # Run clippy lints on tfhe with "gpu" enabled
 clippy_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean,shortint,integer,internal-keycache,gpu \
+		--features=boolean,shortint,integer,internal-keycache,gpu,pbs-stats \
 		--all-targets \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: check_gpu # Run check on tfhe with "gpu" enabled
 check_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" check \
-		--features=boolean,shortint,integer,internal-keycache,gpu \
+		--features=boolean,shortint,integer,internal-keycache,gpu,pbs-stats \
 		--all-targets \
 		-p $(TFHE_SPEC)
 
@@ -348,6 +350,9 @@ clippy_integer: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
 		--features=integer,experimental \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
+	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
+		--features=integer,experimental,extended-types \
+		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: clippy # Run clippy lints enabling the boolean, shortint, integer
 clippy: install_rs_check_toolchain
@@ -380,16 +385,16 @@ clippy_rustdoc_gpu: install_rs_check_toolchain
 .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 \
-		--features=boolean-c-api,shortint-c-api,high-level-c-api \
+		--features=boolean-c-api,shortint-c-api,high-level-c-api,extended-types \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .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,high-level-client-js-wasm-api,zk-pok \
+		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-client-js-wasm-api,zk-pok,extended-types \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy \
-		--features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,high-level-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,extended-types \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: clippy_tasks # Run clippy lints on helper tasks crate.
@@ -399,16 +404,16 @@ clippy_tasks: install_rs_check_toolchain
 
 .PHONY: clippy_trivium # Run clippy lints on Trivium app
 clippy_trivium: install_rs_check_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
+	cd apps/trivium; RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
 		-p tfhe-trivium -- --no-deps -D warnings
 
 .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=boolean,shortint,integer,internal-keycache,zk-pok,strings \
+		--features=boolean,shortint,integer,internal-keycache,zk-pok,strings,pbs-stats,extended-types \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 	RUSTFLAGS="$(RUSTFLAGS)" cargo "$(CARGO_RS_CHECK_TOOLCHAIN)" clippy --all-targets \
-		--features=boolean,shortint,integer,internal-keycache,zk-pok,strings,experimental \
+		--features=boolean,shortint,integer,internal-keycache,zk-pok,strings,pbs-stats,extended-types,experimental \
 		-p $(TFHE_SPEC) -- --no-deps -D warnings
 
 .PHONY: clippy_tfhe_csprng # Run clippy lints on tfhe-csprng
@@ -431,6 +436,7 @@ clippy_versionable: install_rs_check_toolchain
 .PHONY: clippy_tfhe_lints # Run clippy lints on tfhe-lints
 clippy_tfhe_lints: install_cargo_dylint # the toolchain is selected with toolchain.toml
 	cd utils/tfhe-lints && \
+	rustup toolchain install && \
 	cargo clippy --all-targets -- --no-deps -D warnings
 
 .PHONY: clippy_all # Run all clippy targets
@@ -506,13 +512,13 @@ build_tfhe_coverage: install_rs_build_toolchain
 .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=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok \
+		--features=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,extended-types \
 		-p $(TFHE_SPEC)
 
 .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=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,gpu \
+		--features=boolean-c-api,shortint-c-api,high-level-c-api,zk-pok,extended-types,gpu \
 		-p $(TFHE_SPEC)
 
 .PHONY: build_c_api_experimental_deterministic_fft # Build the C API for boolean, shortint and integer with experimental deterministic FFT
@@ -526,7 +532,7 @@ 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
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok,extended-types
 
 .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
@@ -534,7 +540,7 @@ 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" rustup run $(RS_CHECK_TOOLCHAIN) \
 		wasm-pack build --release --target=web \
-		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,parallel-wasm-api,zk-pok \
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,parallel-wasm-api,zk-pok,extended-types \
 		-Z build-std=panic_abort,std && \
 	find pkg/snippets -type f -iname workerHelpers.js -exec sed -i "s|const pkg = await import('..\/..\/..');|const pkg = await import('..\/..\/..\/tfhe.js');|" {} \;
 	jq '.files += ["snippets"]' tfhe/pkg/package.json > tmp_pkg.json && mv -f tmp_pkg.json tfhe/pkg/package.json
@@ -544,7 +550,7 @@ build_node_js_api: install_rs_build_toolchain install_wasm_pack
 	cd tfhe && \
 	RUSTFLAGS="$(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
+		-- --features=boolean-client-js-wasm-api,shortint-client-js-wasm-api,integer-client-js-wasm-api,zk-pok,extended-types
 
 .PHONY: build_tfhe_csprng # Build tfhe_csprng
 build_tfhe_csprng: install_rs_build_toolchain
@@ -596,7 +602,7 @@ 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=integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key:: --test-threads=6
+		--features=integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key:: --test-threads=4
 	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --doc --profile $(CARGO_PROFILE) \
 		--features=integer,gpu -p $(TFHE_SPEC) -- integer::gpu::server_key::
 
@@ -868,12 +874,12 @@ test_examples: test_sha256_bool test_regex_engine
 
 .PHONY: test_trivium # Run tests for trivium
 test_trivium: install_rs_build_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
+	cd apps/trivium; RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
 		-p tfhe-trivium -- --test-threads=1 trivium::
 
 .PHONY: test_kreyvium # Run tests for kreyvium
 test_kreyvium: install_rs_build_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
+	cd apps/trivium; RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_BUILD_TOOLCHAIN) test --profile $(CARGO_PROFILE) \
 		-p tfhe-trivium -- --test-threads=1 kreyvium::
 
 .PHONY: test_tfhe_csprng # Run tfhe-csprng tests
@@ -907,6 +913,7 @@ test_versionable: install_rs_build_toolchain
 .PHONY: test_tfhe_lints # Run test on tfhe-lints
 test_tfhe_lints: install_cargo_dylint
 	cd utils/tfhe-lints && \
+	rustup toolchain install && \
 	cargo test
 
 # The backward compat data repo holds historical binary data but also rust code to generate and load them.
@@ -1071,35 +1078,42 @@ bench_integer: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-bench \
-	--features=integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_signed_integer # Run benchmarks for signed integer
 bench_signed_integer: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-signed-bench \
-	--features=integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_gpu # Run benchmarks for integer on GPU backend
 bench_integer_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-bench \
-	--features=integer,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,gpu,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
+
+.PHONY: bench_signed_integer_gpu # Run benchmarks for signed integer on GPU backend
+bench_signed_integer_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench integer-signed-bench \
+	--features=integer,gpu,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_compression # Run benchmarks for unsigned integer compression
 bench_integer_compression: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench	glwe_packing_compression-integer-bench \
-	--features=integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_compression_gpu
 bench_integer_compression_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench	glwe_packing_compression-integer-bench \
-	--features=integer,internal-keycache,gpu -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,gpu,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_multi_bit # Run benchmarks for unsigned integer using multi-bit parameters
 bench_integer_multi_bit: install_rs_check_toolchain
@@ -1107,7 +1121,7 @@ bench_integer_multi_bit: install_rs_check_toolchain
 	__TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-bench \
-	--features=integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_signed_integer_multi_bit # Run benchmarks for signed integer using multi-bit parameters
 bench_signed_integer_multi_bit: install_rs_check_toolchain
@@ -1115,7 +1129,7 @@ bench_signed_integer_multi_bit: install_rs_check_toolchain
 	__TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-signed-bench \
-	--features=integer,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_multi_bit_gpu # Run benchmarks for integer on GPU backend using multi-bit parameters
 bench_integer_multi_bit_gpu: install_rs_check_toolchain
@@ -1123,22 +1137,22 @@ bench_integer_multi_bit_gpu: install_rs_check_toolchain
 	__TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench integer-bench \
-	--features=integer,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) --
+	--features=integer,gpu,internal-keycache,nightly-avx512,pbs-stats -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
+.PHONY: bench_signed_integer_multi_bit_gpu # Run benchmarks for signed integer on GPU backend using multi-bit parameters
+bench_signed_integer_multi_bit_gpu: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAM_TYPE=MULTI_BIT \
 	__TFHE_RS_BENCH_OP_FLAVOR=$(BENCH_OP_FLAVOR) __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
-	--bench integer-bench \
-	--features=integer,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC) -- ::unsigned
+	--bench integer-signed-bench \
+	--features=integer,gpu,internal-keycache,nightly-avx512,pbs-stats -p $(TFHE_SPEC) --
 
 .PHONY: bench_integer_zk # Run benchmarks for integer encryption with ZK proofs
 bench_integer_zk: install_rs_check_toolchain
 	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_BENCH_TYPE=$(BENCH_TYPE) \
 	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench zk-pke-bench \
-	--features=integer,internal-keycache,zk-pok,nightly-avx512 \
+	--features=integer,internal-keycache,zk-pok,nightly-avx512,pbs-stats \
 	-p $(TFHE_SPEC) --
 
 .PHONY: bench_shortint # Run benchmarks for shortint
@@ -1171,31 +1185,52 @@ bench_boolean: install_rs_check_toolchain
 
 .PHONY: bench_pbs # Run benchmarks for PBS
 bench_pbs: install_rs_check_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench pbs-bench \
 	--features=boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
+.PHONY: bench_ks_pbs # Run benchmarks for KS-PBS
+bench_ks_pbs: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench ks-pbs-bench \
+	--features=boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
+
+.PHONY: bench_ks_pbs_gpu # Run benchmarks for KS-PBS on GPU backend
+bench_ks_pbs_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAM_TYPE=$(BENCH_PARAM_TYPE) __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench ks-pbs-bench \
+	--features=boolean,shortint,gpu,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=boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
+.PHONY: bench_pbs128_gpu # Run benchmarks for PBS using FFT 128 bits on GPU
+bench_pbs128_gpu: install_rs_check_toolchain
+	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	--bench pbs128-bench \
+	--features=boolean,shortint,gpu,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)" __TFHE_RS_FAST_BENCH=$(FAST_BENCH) cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_FAST_BENCH=$(FAST_BENCH) __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) \
+	cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench pbs-bench \
 	--features=boolean,shortint,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
 .PHONY: bench_ks # Run benchmarks for keyswitch
 bench_ks: install_rs_check_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench ks-bench \
 	--features=boolean,shortint,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 
 .PHONY: bench_ks_gpu # Run benchmarks for PBS on GPU backend
 bench_ks_gpu: install_rs_check_toolchain
-	RUSTFLAGS="$(RUSTFLAGS)" cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
+	RUSTFLAGS="$(RUSTFLAGS)" __TFHE_RS_PARAMS_SET=$(BENCH_PARAMS_SET) cargo $(CARGO_RS_CHECK_TOOLCHAIN) bench \
 	--bench ks-bench \
 	--features=boolean,shortint,gpu,internal-keycache,nightly-avx512 -p $(TFHE_SPEC)
 

README.md

@@ -10,7 +10,7 @@
 <hr/>
 
 <p align="center">
-  <a href="https://docs.zama.ai/tfhe-rs"> 📒 Documentation</a> | <a href="https://zama.ai/community"> 💛 Community support</a> | <a href="https://github.com/zama-ai/awesome-zama"> 📚 FHE resources by Zama</a>
+  <a href="https://github.com/zama-ai/tfhe-rs-handbook/blob/main/tfhe-rs-handbook.pdf"> 📃 Read Handbook</a> |<a href="https://docs.zama.ai/tfhe-rs"> 📒 Documentation</a> | <a href="https://zama.ai/community"> 💛 Community support</a> | <a href="https://github.com/zama-ai/awesome-zama"> 📚 FHE resources by Zama</a>
 </p>
 
 
@@ -67,6 +67,9 @@ production-ready library for all the advanced features of TFHE.
 
 ## Getting started
 
+> [!Important]
+> **TFHE-rs** released its first stable version v1.0.0 in February 2025, stabilizing the high-level API for the x86 CPU backend.
+
 ### 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`:
 
@@ -75,13 +78,13 @@ tfhe = { version = "*", features = ["boolean", "shortint", "integer"] }
 ```
 
 > [!Note]
-> Note: You need to use a Rust version >= 1.81 to compile TFHE-rs.
+> Note: You need to use Rust version >= 1.84 to compile TFHE-rs.
 
 > [!Note]
-> Note: aarch64-based machines are not yet supported for Windows as it's currently missing an entropy source to be able to seed the [CSPRNGs](https://en.wikipedia.org/wiki/Cryptographically_secure_pseudorandom_number_generator) used in TFHE-rs.
+> Note: AArch64-based machines are not supported for Windows as it's currently missing an entropy source to be able to seed the [CSPRNGs](https://en.wikipedia.org/wiki/Cryptographically_secure_pseudorandom_number_generator) used in TFHE-rs.
 
 <p align="right">
-  <a href="#about" > ↑ Back to top </a> 
+  <a href="#about" > ↑ Back to top </a>
 </p>
 
 ### A simple example
@@ -138,7 +141,7 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
 }
 ```
 
-To run this code, use the following command: 
+To run this code, use the following command:
 <p align="center"> <code> cargo run --release </code> </p>
 
 > [!Note]
@@ -148,12 +151,15 @@ to run in release mode with cargo's `--release` flag to have the best performanc
 *Find an example with more explanations in [this part of the documentation](https://docs.zama.ai/tfhe-rs/get-started/quick_start)*
 
 <p align="right">
-  <a href="#about" > ↑ Back to top </a> 
+  <a href="#about" > ↑ Back to top </a>
 </p>
 
 
 
-## Resources 
+## Resources
+
+### TFHE-rs Handbook
+A document containing scientific and technical details about algorithms implemented into the library is available here: [TFHE-rs: A (Practical) Handbook](https://github.com/zama-ai/tfhe-rs-handbook/blob/main/tfhe-rs-handbook.pdf).
 
 ### TFHE deep dive
 - [TFHE Deep Dive - Part I - Ciphertext types](https://www.zama.ai/post/tfhe-deep-dive-part-1)
@@ -176,7 +182,7 @@ to run in release mode with cargo's `--release` flag to have the best performanc
 
 Full, comprehensive documentation is available here: [https://docs.zama.ai/tfhe-rs](https://docs.zama.ai/tfhe-rs).
 <p align="right">
-  <a href="#about" > ↑ Back to top </a> 
+  <a href="#about" > ↑ Back to top </a>
 </p>
 
 
@@ -194,9 +200,13 @@ When a new update is published in the Lattice Estimator, we update parameters ac
 
 ### 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 = $2^{-64}$. 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]. 
+By default, the parameter sets used in the High-Level API with the x86 CPU backend have a failure probability $\le 2^{128}$ to securely work in the IND-CPA^D model using the algorithmic techniques provided in our code base [1].
+If you want to work within the IND-CPA security model, which is less strict than the IND-CPA-D model, the parameter sets can easily be changed and would have slightly better performance. More details can be found in the [TFHE-rs documentation](https://docs.zama.ai/tfhe-rs).
 
-[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
+The default parameters used in the High-Level API with the GPU backend are chosen considering the IND-CPA security model, and are selected with a bootstrapping failure probability fixed at $p_{error} \le 2^{-64}$. 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 [2].
+
+[1] Bernard, Olivier, et al. "Drifting Towards Better Error Probabilities in Fully Homomorphic Encryption Schemes". https://eprint.iacr.org/2024/1718.pdf
+[2] 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
 
@@ -245,7 +255,7 @@ This software is distributed under the **BSD-3-Clause-Clear** license. Read [thi
 >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> 
+  <a href="#about" > ↑ Back to top </a>
 </p>
 
 
@@ -259,8 +269,8 @@ This software is distributed under the **BSD-3-Clause-Clear** license. Read [thi
 </picture>
 </a>
 
-🌟 If you find this project helpful or interesting, please consider giving it a star on GitHub! Your support helps to grow the community and motivates further development. 
+🌟 If you find this project helpful or interesting, please consider giving it a star on GitHub! Your support helps to grow the community and motivates further development.
 
 <p align="right">
-  <a href="#about" > ↑ Back to top </a> 
+  <a href="#about" > ↑ Back to top </a>
 </p>

_typos.toml

@@ -13,3 +13,9 @@ extend-ignore-identifiers-re = [
     # Example in trivium
     "C9217BA0D762ACA1"
 ]
+
+[files]
+extend-exclude = [
+    "backends/tfhe-cuda-backend/cuda/src/fft128/twiddles.cu",
+    "backends/tfhe-cuda-backend/cuda/src/fft/twiddles.cu",
+]

apps/trivium/Cargo.toml

@@ -6,11 +6,17 @@ edition = "2021"
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
 [dependencies]
-rayon = { workspace = true }
-tfhe = { path = "../../tfhe", features = [ "boolean", "shortint", "integer" ] }
+rayon = "1"
+tfhe = { path = "../../tfhe", features = ["boolean", "shortint", "integer"] }
 
 [dev-dependencies]
-criterion = { version = "0.5.1", features = [ "html_reports" ]}
+criterion = { version = "0.5.1", features = ["html_reports"] }
+
+[profile.devo]
+inherits = "dev"
+opt-level = 3
+lto = "off"
+debug-assertions = false
 
 [[bench]]
 name = "trivium"

apps/trivium/README.md

@@ -129,7 +129,7 @@ Other sizes than 64 bit are expected to be available in the future.
 
 # FHE shortint Trivium implementation
 
-The same implementation is also available for generic Ciphertexts representing bits (meant to be used with parameters `V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128`).
+The same implementation is also available for generic Ciphertexts representing bits (meant to be used with parameters `V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128`).
 It uses a lower level API of tfhe-rs, so the syntax is a little bit different. It also implements the `TransCiphering` trait. For optimization purposes, it does not internally run
 on the same cryptographic parameters as the high level API of tfhe-rs. As such, it requires the usage of a casting key, to switch from one parameter space to another, which makes
 its setup a little more intricate.
@@ -137,10 +137,10 @@ its setup a little more intricate.
 Example code:
 ```rust
 use tfhe::shortint::prelude::*;
-use tfhe::shortint::parameters::v1_0::{
-    V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+use tfhe::shortint::parameters::v1_1::{
+    V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
 };
 use tfhe::{ConfigBuilder, generate_keys, FheUint64};
 use tfhe::prelude::*;
@@ -148,17 +148,17 @@ use tfhe_trivium::TriviumStreamShortint;
 
 fn test_shortint() {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
-    let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+    let (client_key, server_key): (ClientKey, ServerKey) = gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB".to_string();

apps/trivium/benches/kreyvium_shortint.rs

@@ -1,9 +1,9 @@
 use criterion::Criterion;
 use tfhe::prelude::*;
-use tfhe::shortint::parameters::v1_0::{
-    V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
+use tfhe::shortint::parameters::v1_1::{
+    V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
 };
 use tfhe::shortint::prelude::*;
 use tfhe::{generate_keys, ConfigBuilder, FheUint64};
@@ -11,19 +11,19 @@ use tfhe_trivium::{KreyviumStreamShortint, TransCiphering};
 
 pub fn kreyvium_shortint_warmup(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();
@@ -64,19 +64,19 @@ pub fn kreyvium_shortint_warmup(c: &mut Criterion) {
 
 pub fn kreyvium_shortint_gen(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();
@@ -112,19 +112,19 @@ pub fn kreyvium_shortint_gen(c: &mut Criterion) {
 
 pub fn kreyvium_shortint_trans(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();

apps/trivium/benches/trivium_shortint.rs

@@ -1,9 +1,9 @@
 use criterion::Criterion;
 use tfhe::prelude::*;
-use tfhe::shortint::parameters::v1_0::{
-    V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
+use tfhe::shortint::parameters::v1_1::{
+    V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
 };
 use tfhe::shortint::prelude::*;
 use tfhe::{generate_keys, ConfigBuilder, FheUint64};
@@ -11,19 +11,19 @@ use tfhe_trivium::{TransCiphering, TriviumStreamShortint};
 
 pub fn trivium_shortint_warmup(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB".to_string();
@@ -64,19 +64,19 @@ pub fn trivium_shortint_warmup(c: &mut Criterion) {
 
 pub fn trivium_shortint_gen(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB".to_string();
@@ -112,19 +112,19 @@ pub fn trivium_shortint_gen(c: &mut Criterion) {
 
 pub fn trivium_shortint_trans(c: &mut Criterion) {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB".to_string();

apps/trivium/src/kreyvium/test.rs

@@ -1,9 +1,9 @@
 use crate::{KreyviumStream, KreyviumStreamByte, KreyviumStreamShortint, TransCiphering};
 use tfhe::prelude::*;
-use tfhe::shortint::parameters::v1_0::{
-    V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
+use tfhe::shortint::parameters::v1_1::{
+    V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
 };
 use tfhe::{generate_keys, ConfigBuilder, FheBool, FheUint64, FheUint8};
 // Values for these tests come from the github repo renaud1239/Kreyvium,
@@ -221,19 +221,19 @@ use tfhe::shortint::prelude::*;
 #[test]
 fn kreyvium_test_shortint_long() {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB000000000000".to_string();

apps/trivium/src/trans_ciphering/mod.rs

@@ -48,6 +48,8 @@ fn transcipher_from_1_1_stream(
 ) -> FheUint64 {
     assert_eq!(stream.len(), 64);
 
+    let id_lut = internal_server_key.generate_lookup_table(|x| x);
+
     let pairs = (0..32)
         .into_par_iter()
         .map(|i| {
@@ -57,10 +59,11 @@ fn transcipher_from_1_1_stream(
             let b0 = &stream[8 * byte_idx + 2 * pair_idx];
             let b1 = &stream[8 * byte_idx + 2 * pair_idx + 1];
 
-            casting_key.cast(
-                &internal_server_key
-                    .unchecked_add(b0, &internal_server_key.unchecked_scalar_mul(b1, 2)),
-            )
+            let mut combined = internal_server_key
+                .unchecked_add(b0, &internal_server_key.unchecked_scalar_mul(b1, 2));
+            internal_server_key.apply_lookup_table_assign(&mut combined, &id_lut);
+
+            casting_key.cast(&combined)
         })
         .collect::<Vec<_>>();
 

apps/trivium/src/trivium/test.rs

@@ -1,9 +1,9 @@
 use crate::{TransCiphering, TriviumStream, TriviumStreamByte, TriviumStreamShortint};
 use tfhe::prelude::*;
-use tfhe::shortint::parameters::v1_0::{
-    V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
-    V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
+use tfhe::shortint::parameters::v1_1::{
+    V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128,
+    V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128,
 };
 use tfhe::{generate_keys, ConfigBuilder, FheBool, FheUint64, FheUint8};
 // Values for these tests come from the github repo cantora/avr-crypto-lib, commit 2a5b018,
@@ -357,19 +357,19 @@ use tfhe::shortint::prelude::*;
 #[test]
 fn trivium_test_shortint_long() {
     let config = ConfigBuilder::default()
-        .use_custom_parameters(V1_0_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
+        .use_custom_parameters(V1_1_PARAM_MESSAGE_2_CARRY_2_KS_PBS_GAUSSIAN_2M128)
         .build();
     let (hl_client_key, hl_server_key) = generate_keys(config);
     let underlying_ck: tfhe::shortint::ClientKey = (*hl_client_key.as_ref()).clone().into();
     let underlying_sk: tfhe::shortint::ServerKey = (*hl_server_key.as_ref()).clone().into();
 
     let (client_key, server_key): (ClientKey, ServerKey) =
-        gen_keys(V1_0_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
+        gen_keys(V1_1_PARAM_MESSAGE_1_CARRY_1_KS_PBS_GAUSSIAN_2M128);
 
     let ksk = KeySwitchingKey::new(
         (&client_key, Some(&server_key)),
         (&underlying_ck, &underlying_sk),
-        V1_0_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
+        V1_1_PARAM_KEYSWITCH_1_1_KS_PBS_TO_2_2_KS_PBS_GAUSSIAN_2M128,
     );
 
     let key_string = "0053A6F94C9FF24598EB".to_string();

backends/tfhe-cuda-backend/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tfhe-cuda-backend"
-version = "0.8.0"
+version = "0.9.0-alpha.0"
 edition = "2021"
 authors = ["Zama team"]
 license = "BSD-3-Clause-Clear"

backends/tfhe-cuda-backend/README.md

@@ -13,24 +13,26 @@ and forth between the CPU and the GPU, to create and destroy Cuda streams, etc.:
 - `cuda_get_number_of_gpus`
 - `cuda_synchronize_device`
 The cryptographic operations it provides are:
-- an amortized implementation of the TFHE programmable bootstrap: `cuda_bootstrap_amortized_lwe_ciphertext_vector_32` and `cuda_bootstrap_amortized_lwe_ciphertext_vector_64`
-- a low latency implementation of the TFHE programmable bootstrap: `cuda_bootstrap_low latency_lwe_ciphertext_vector_32` and `cuda_bootstrap_low_latency_lwe_ciphertext_vector_64`
-- the keyswitch: `cuda_keyswitch_lwe_ciphertext_vector_32` and `cuda_keyswitch_lwe_ciphertext_vector_64`
-- the larger precision programmable bootstrap (wop PBS, which supports up to 16 bits of message while the classical PBS only supports up to 8 bits of message) and its sub-components: `cuda_wop_pbs_64`, `cuda_extract_bits_64`, `cuda_circuit_bootstrap_64`, `cuda_cmux_tree_64`, `cuda_blind_rotation_sample_extraction_64`
-- acceleration for leveled operations: `cuda_negate_lwe_ciphertext_vector_64`, `cuda_add_lwe_ciphertext_vector_64`, `cuda_add_lwe_ciphertext_vector_plaintext_vector_64`, `cuda_mult_lwe_ciphertext_vector_cleartext_vector`.
+- an implementation of the classical TFHE programmable bootstrap,
+- an implementation of the multi-bit TFHE programmable bootstrap,
+- the keyswitch,
+- acceleration for leveled operations,
+- acceleration for arithmetics over encrypted integers of arbitrary size, 
+- acceleration for integer compression/decompression.
 
 ## Dependencies
 
 **Disclaimer**: Compilation on Windows/Mac is not supported yet. Only Nvidia GPUs are supported. 
 
-- nvidia driver - for example, if you're running Ubuntu 20.04 check this [page](https://linuxconfig.org/how-to-install-the-nvidia-drivers-on-ubuntu-20-04-focal-fossa-linux) for installation
+- nvidia driver - for example, if you're running Ubuntu 20.04 check this [page](https://linuxconfig.org/how-to-install-the-nvidia-drivers-on-ubuntu-20-04-focal-fossa-linux) for installation. You need an Nvidia GPU with Compute Capability >= 3.0
 - [nvcc](https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html) >= 10.0
 - [gcc](https://gcc.gnu.org/) >= 8.0 - check this [page](https://gist.github.com/ax3l/9489132) for more details about nvcc/gcc compatible versions
 - [cmake](https://cmake.org/) >= 3.24
+- libclang, to match Rust bingen [requirements](https://rust-lang.github.io/rust-bindgen/requirements.html) >= 9.0
 
 ## Build
 
-The Cuda project held in `tfhe-cuda-backend` can be compiled independently from TFHE-rs in the following way:
+The Cuda project held in `tfhe-cuda-backend` can be compiled independently of TFHE-rs in the following way:
 ```
 git clone git@github.com:zama-ai/tfhe-rs
 cd backends/tfhe-cuda-backend/cuda

backends/tfhe-cuda-backend/build.rs

@@ -62,6 +62,7 @@ fn main() {
             "cuda/include/integer/integer.h",
             "cuda/include/keyswitch.h",
             "cuda/include/linear_algebra.h",
+            "cuda/include/fft/fft128.h",
             "cuda/include/pbs/programmable_bootstrap.h",
             "cuda/include/pbs/programmable_bootstrap_multibit.h",
         ];

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

@@ -18,7 +18,7 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu_64(void *stream,
 void cuda_glwe_sample_extract_64(void *stream, uint32_t gpu_index,
                                  void *lwe_array_out, void const *glwe_array_in,
                                  uint32_t const *nth_array, uint32_t num_nths,
-                                 uint32_t glwe_dimension,
+                                 uint32_t lwe_per_glwe, uint32_t glwe_dimension,
                                  uint32_t polynomial_size);
 }
 #endif

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

@@ -52,7 +52,7 @@ void *cuda_malloc_async(uint64_t size, cudaStream_t stream, uint32_t gpu_index);
 
 void cuda_check_valid_malloc(uint64_t size, uint32_t gpu_index);
 
-void cuda_memcpy_async_to_gpu(void *dest, void *src, uint64_t size,
+void cuda_memcpy_async_to_gpu(void *dest, const void *src, uint64_t size,
                               cudaStream_t stream, uint32_t gpu_index);
 
 void cuda_memcpy_async_gpu_to_gpu(void *dest, void const *src, uint64_t size,
@@ -76,7 +76,7 @@ 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);
+uint32_t cuda_get_max_shared_memory(uint32_t gpu_index);
 
 bool cuda_check_support_cooperative_groups();
 

backends/tfhe-cuda-backend/cuda/include/fft/fft128.h

@@ -0,0 +1,17 @@
+#include <stdint.h>
+extern "C" {
+void cuda_fourier_transform_forward_as_torus_f128_async(
+    void *stream, uint32_t gpu_index, void *re0, void *re1, void *im0,
+    void *im1, void const *standard, uint32_t const N,
+    const uint32_t number_of_samples);
+
+void cuda_fourier_transform_forward_as_integer_f128_async(
+    void *stream, uint32_t gpu_index, void *re0, void *re1, void *im0,
+    void *im1, void const *standard, uint32_t const N,
+    const uint32_t number_of_samples);
+
+void cuda_fourier_transform_backward_as_torus_f128_async(
+    void *stream, uint32_t gpu_index, void *standard, void const *re0,
+    void const *re1, void const *im0, void const *im1, uint32_t const N,
+    const uint32_t number_of_samples);
+}

backends/tfhe-cuda-backend/cuda/include/integer/compression/compression_utilities.h

@@ -102,9 +102,7 @@ template <typename Torus> struct int_decompression {
       // Example: in the 2_2 case we are mapping a 2 bits message onto a 4 bits
       // space, we want to keep the original 2 bits value in the 4 bits space,
       // so we apply the identity and the encoding will rescale it for us.
-      auto decompression_rescale_f = [encryption_params](Torus x) -> Torus {
-        return x;
-      };
+      auto decompression_rescale_f = [](Torus x) -> Torus { return x; };
 
       auto effective_compression_message_modulus =
           encryption_params.carry_modulus;

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

@@ -44,6 +44,7 @@ typedef struct {
   uint64_t *degrees;
   uint64_t *noise_levels;
   uint32_t num_radix_blocks;
+  uint32_t max_num_radix_blocks;
   uint32_t lwe_dimension;
 } CudaRadixCiphertextFFI;
 
@@ -113,7 +114,8 @@ void scratch_cuda_full_propagation_64(
 
 void cuda_full_propagation_64_inplace(void *const *streams,
                                       uint32_t const *gpu_indexes,
-                                      uint32_t gpu_count, void *input_blocks,
+                                      uint32_t gpu_count,
+                                      CudaRadixCiphertextFFI *input_blocks,
                                       int8_t *mem_ptr, void *const *ksks,
                                       void *const *bsks, uint32_t num_blocks);
 
@@ -132,10 +134,11 @@ void scratch_cuda_integer_mult_radix_ciphertext_kb_64(
 
 void cuda_integer_mult_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *radix_lwe_out, void const *radix_lwe_left, bool const is_bool_left,
-    void const *radix_lwe_right, bool const is_bool_right, void *const *bsks,
-    void *const *ksks, int8_t *mem_ptr, uint32_t polynomial_size,
-    uint32_t num_blocks);
+    CudaRadixCiphertextFFI *radix_lwe_out,
+    CudaRadixCiphertextFFI const *radix_lwe_left, bool const is_bool_left,
+    CudaRadixCiphertextFFI const *radix_lwe_right, bool const is_bool_right,
+    void *const *bsks, void *const *ksks, int8_t *mem_ptr,
+    uint32_t polynomial_size, uint32_t num_blocks);
 
 void cleanup_cuda_integer_mult(void *const *streams,
                                uint32_t const *gpu_indexes, uint32_t gpu_count,
@@ -145,12 +148,13 @@ void cuda_negate_integer_radix_ciphertext_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
     CudaRadixCiphertextFFI *lwe_array_out,
     CudaRadixCiphertextFFI const *lwe_array_in, uint32_t message_modulus,
-    uint32_t carry_modulus);
+    uint32_t carry_modulus, uint32_t num_radix_blocks);
 
 void cuda_scalar_addition_integer_radix_ciphertext_64_inplace(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
     CudaRadixCiphertextFFI *lwe_array, void const *scalar_input,
-    uint32_t num_scalars, uint32_t message_modulus, uint32_t carry_modulus);
+    void const *h_scalar_input, uint32_t num_scalars, uint32_t message_modulus,
+    uint32_t carry_modulus);
 
 void scratch_cuda_integer_radix_logical_scalar_shift_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
@@ -218,15 +222,17 @@ void scratch_cuda_integer_radix_comparison_kb_64(
 
 void cuda_comparison_integer_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_1, void const *lwe_array_2,
-    int8_t *mem_ptr, void *const *bsks, void *const *ksks,
-    uint32_t lwe_ciphertext_count);
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_1,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks);
 
 void cuda_scalar_comparison_integer_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_in, void const *scalar_blocks,
-    int8_t *mem_ptr, void *const *bsks, void *const *ksks,
-    uint32_t lwe_ciphertext_count, uint32_t num_scalar_blocks);
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in, void const *scalar_blocks,
+    void const *h_scalar_blocks, int8_t *mem_ptr, void *const *bsks,
+    void *const *ksks, uint32_t num_scalar_blocks);
 
 void cleanup_cuda_integer_comparison(void *const *streams,
                                      uint32_t const *gpu_indexes,
@@ -252,8 +258,8 @@ void cuda_scalar_bitop_integer_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
     CudaRadixCiphertextFFI *lwe_array_out,
     CudaRadixCiphertextFFI const *lwe_array_input, void const *clear_blocks,
-    uint32_t num_clear_blocks, int8_t *mem_ptr, void *const *bsks,
-    void *const *ksks);
+    void const *h_clear_blocks, uint32_t num_clear_blocks, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks);
 
 void cleanup_cuda_integer_bitop(void *const *streams,
                                 uint32_t const *gpu_indexes, uint32_t gpu_count,
@@ -373,9 +379,9 @@ void scratch_cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
 
 void cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
     void *const *streams, uint32_t const *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 *const *bsks, void *const *ksks,
-    uint32_t num_blocks_in_radix);
+    CudaRadixCiphertextFFI *radix_lwe_out,
+    CudaRadixCiphertextFFI *radix_lwe_vec, int8_t *mem_ptr, void *const *bsks,
+    void *const *ksks);
 
 void cleanup_cuda_integer_radix_partial_sum_ciphertexts_vec(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
@@ -391,10 +397,10 @@ void scratch_cuda_integer_scalar_mul_kb_64(
 
 void cuda_scalar_multiplication_integer_radix_ciphertext_64_inplace(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array, uint64_t const *decomposed_scalar,
+    CudaRadixCiphertextFFI *lwe_array, uint64_t const *decomposed_scalar,
     uint64_t const *has_at_least_one_set, int8_t *mem_ptr, void *const *bsks,
-    void *const *ksks, uint32_t lwe_dimension, uint32_t polynomial_size,
-    uint32_t message_modulus, uint32_t num_blocks, uint32_t num_scalars);
+    void *const *ksks, uint32_t polynomial_size, uint32_t message_modulus,
+    uint32_t num_scalars);
 
 void cleanup_cuda_integer_radix_scalar_mul(void *const *streams,
                                            uint32_t const *gpu_indexes,
@@ -412,9 +418,10 @@ void scratch_cuda_integer_div_rem_radix_ciphertext_kb_64(
 
 void cuda_integer_div_rem_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *quotient, void *remainder, void const *numerator, void const *divisor,
-    bool is_signed, int8_t *mem_ptr, void *const *bsks, void *const *ksks,
-    uint32_t num_blocks_in_radix);
+    CudaRadixCiphertextFFI *quotient, CudaRadixCiphertextFFI *remainder,
+    CudaRadixCiphertextFFI const *numerator,
+    CudaRadixCiphertextFFI const *divisor, bool is_signed, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks);
 
 void cleanup_cuda_integer_div_rem(void *const *streams,
                                   uint32_t const *gpu_indexes,
@@ -474,7 +481,8 @@ void scratch_cuda_integer_are_all_comparisons_block_true_kb_64(
 
 void cuda_integer_are_all_comparisons_block_true_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_in, int8_t *mem_ptr,
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks);
 
 void cleanup_cuda_integer_are_all_comparisons_block_true(
@@ -492,7 +500,8 @@ void scratch_cuda_integer_is_at_least_one_comparisons_block_true_kb_64(
 
 void cuda_integer_is_at_least_one_comparisons_block_true_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_in, int8_t *mem_ptr,
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks);
 
 void cleanup_cuda_integer_is_at_least_one_comparisons_block_true(

backends/tfhe-cuda-backend/cuda/include/integer/radix_ciphertext.h

@@ -1,8 +1,11 @@
 #ifndef CUDA_RADIX_CIPHERTEXT_H
 #define CUDA_RADIX_CIPHERTEXT_H
 
-void release_radix_ciphertext(cudaStream_t const stream,
-                              uint32_t const gpu_index,
-                              CudaRadixCiphertextFFI *data);
+void release_radix_ciphertext_async(cudaStream_t const stream,
+                                    uint32_t const gpu_index,
+                                    CudaRadixCiphertextFFI *data);
+
+void reset_radix_ciphertext_blocks(CudaRadixCiphertextFFI *data,
+                                   uint32_t new_num_blocks);
 
 #endif

backends/tfhe-cuda-backend/cuda/include/pbs/pbs_multibit_utilities.h

@@ -5,12 +5,12 @@
 
 template <typename Torus>
 bool supports_distributed_shared_memory_on_multibit_programmable_bootstrap(
-    uint32_t polynomial_size, int max_shared_memory);
+    uint32_t polynomial_size, uint32_t max_shared_memory);
 
 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, int max_shared_memory);
+    uint32_t level_count, uint32_t max_shared_memory);
 
 #if CUDA_ARCH >= 900
 template <typename Torus>

backends/tfhe-cuda-backend/cuda/include/pbs/pbs_utilities.h

@@ -9,20 +9,26 @@
 template <typename Torus>
 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
+  size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
+  return sizeof(Torus) * polynomial_size + // accumulator_rotated
+         sizeof(double) * 2 * double_count * polynomial_size /
+             2; // accumulator fft
 }
 template <typename Torus>
 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
+  size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
+  return sizeof(Torus) * polynomial_size + // accumulator
+         sizeof(double) * 2 * double_count * polynomial_size /
+             2; // accumulator fft
 }
 
 template <typename Torus>
 uint64_t
 get_buffer_size_partial_sm_programmable_bootstrap(uint32_t polynomial_size) {
-  return sizeof(double2) * polynomial_size / 2; // accumulator fft
+  size_t double_count = (sizeof(Torus) == 16) ? 2 : 1;
+  return sizeof(double) * 2 * double_count * polynomial_size /
+         2; // accumulator fft
 }
 
 template <typename Torus>
@@ -61,7 +67,7 @@ get_buffer_size_partial_sm_programmable_bootstrap_cg(uint32_t polynomial_size) {
 
 template <typename Torus>
 bool supports_distributed_shared_memory_on_classic_programmable_bootstrap(
-    uint32_t polynomial_size, int max_shared_memory);
+    uint32_t polynomial_size, uint32_t max_shared_memory);
 
 template <typename Torus, PBS_TYPE pbs_type> struct pbs_buffer;
 
@@ -215,6 +221,155 @@ template <typename Torus> struct pbs_buffer<Torus, PBS_TYPE::CLASSICAL> {
   }
 };
 
+template <PBS_TYPE pbs_type> struct pbs_buffer_128;
+
+template <> struct pbs_buffer_128<PBS_TYPE::CLASSICAL> {
+  int8_t *d_mem;
+
+  __uint128_t *global_accumulator;
+  double *global_join_buffer;
+
+  PBS_VARIANT pbs_variant;
+
+  pbs_buffer_128(cudaStream_t stream, uint32_t gpu_index,
+                 uint32_t glwe_dimension, uint32_t polynomial_size,
+                 uint32_t level_count, uint32_t input_lwe_ciphertext_count,
+                 PBS_VARIANT pbs_variant, bool allocate_gpu_memory) {
+    cuda_set_device(gpu_index);
+    this->pbs_variant = pbs_variant;
+
+    auto max_shared_memory = cuda_get_max_shared_memory(gpu_index);
+    size_t global_join_buffer_size = (glwe_dimension + 1) * level_count *
+                                     input_lwe_ciphertext_count *
+                                     polynomial_size / 2 * sizeof(double) * 4;
+
+    if (allocate_gpu_memory) {
+      switch (pbs_variant) {
+      case PBS_VARIANT::DEFAULT: {
+        uint64_t full_sm_step_one =
+            get_buffer_size_full_sm_programmable_bootstrap_step_one<
+                __uint128_t>(polynomial_size);
+        uint64_t full_sm_step_two =
+            get_buffer_size_full_sm_programmable_bootstrap_step_two<
+                __uint128_t>(polynomial_size);
+        uint64_t partial_sm =
+            get_buffer_size_partial_sm_programmable_bootstrap<__uint128_t>(
+                polynomial_size);
+
+        uint64_t partial_dm_step_one = full_sm_step_one - partial_sm;
+        uint64_t partial_dm_step_two = full_sm_step_two - partial_sm;
+        uint64_t full_dm = full_sm_step_one;
+
+        uint64_t device_mem = 0;
+        if (max_shared_memory < partial_sm) {
+          device_mem = full_dm * input_lwe_ciphertext_count * level_count *
+                       (glwe_dimension + 1);
+        } else if (max_shared_memory < full_sm_step_two) {
+          device_mem =
+              (partial_dm_step_two + partial_dm_step_one * level_count) *
+              input_lwe_ciphertext_count * (glwe_dimension + 1);
+        } else if (max_shared_memory < full_sm_step_one) {
+          device_mem = partial_dm_step_one * input_lwe_ciphertext_count *
+                       level_count * (glwe_dimension + 1);
+        }
+        // Otherwise, both kernels run all in shared memory
+        d_mem = (int8_t *)cuda_malloc_async(device_mem, stream, gpu_index);
+
+        global_join_buffer = (double *)cuda_malloc_async(
+            global_join_buffer_size, stream, gpu_index);
+
+        global_accumulator = (__uint128_t *)cuda_malloc_async(
+            (glwe_dimension + 1) * input_lwe_ciphertext_count *
+                polynomial_size * sizeof(__uint128_t),
+            stream, gpu_index);
+      } break;
+      case PBS_VARIANT::CG: {
+        uint64_t full_sm =
+            get_buffer_size_full_sm_programmable_bootstrap_cg<__uint128_t>(
+                polynomial_size);
+        uint64_t partial_sm =
+            get_buffer_size_partial_sm_programmable_bootstrap_cg<__uint128_t>(
+                polynomial_size);
+
+        uint64_t partial_dm = full_sm - partial_sm;
+        uint64_t full_dm = full_sm;
+        uint64_t device_mem = 0;
+
+        if (max_shared_memory < partial_sm) {
+          device_mem = full_dm * input_lwe_ciphertext_count * level_count *
+                       (glwe_dimension + 1);
+        } else if (max_shared_memory < full_sm) {
+          device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
+                       (glwe_dimension + 1);
+        }
+
+        // Otherwise, both kernels run all in shared memory
+        d_mem = (int8_t *)cuda_malloc_async(device_mem, stream, gpu_index);
+
+        global_join_buffer = (double *)cuda_malloc_async(
+            global_join_buffer_size, stream, gpu_index);
+      } break;
+#if CUDA_ARCH >= 900
+      case PBS_VARIANT::TBC: {
+
+        bool supports_dsm =
+            supports_distributed_shared_memory_on_classic_programmable_bootstrap<
+                __uint128_t>(polynomial_size, max_shared_memory);
+
+        uint64_t full_sm =
+            get_buffer_size_full_sm_programmable_bootstrap_tbc<__uint128_t>(
+                polynomial_size);
+        uint64_t partial_sm =
+            get_buffer_size_partial_sm_programmable_bootstrap_tbc<__uint128_t>(
+                polynomial_size);
+        uint64_t minimum_sm_tbc = 0;
+        if (supports_dsm)
+          minimum_sm_tbc =
+              get_buffer_size_sm_dsm_plus_tbc_classic_programmable_bootstrap<
+                  __uint128_t>(polynomial_size);
+
+        uint64_t partial_dm = full_sm - partial_sm;
+        uint64_t full_dm = full_sm;
+        uint64_t device_mem = 0;
+
+        // There is a minimum amount of memory we need to run the TBC PBS, which
+        // is minimum_sm_tbc. We know that minimum_sm_tbc bytes are available
+        // because otherwise the previous check would have redirected
+        // computation to some other variant. If over that we don't have more
+        // partial_sm bytes, TBC PBS will run on NOSM. If we have partial_sm but
+        // not full_sm bytes, it will run on PARTIALSM. Otherwise, FULLSM.
+        //
+        // NOSM mode actually requires minimum_sm_tbc shared memory bytes.
+        if (max_shared_memory < partial_sm + minimum_sm_tbc) {
+          device_mem = full_dm * input_lwe_ciphertext_count * level_count *
+                       (glwe_dimension + 1);
+        } else if (max_shared_memory < full_sm + minimum_sm_tbc) {
+          device_mem = partial_dm * input_lwe_ciphertext_count * level_count *
+                       (glwe_dimension + 1);
+        }
+
+        // Otherwise, both kernels run all in shared memory
+        d_mem = (int8_t *)cuda_malloc_async(device_mem, stream, gpu_index);
+
+        global_join_buffer = (double *)cuda_malloc_async(
+            global_join_buffer_size, stream, gpu_index);
+      } break;
+#endif
+      default:
+        PANIC("Cuda error (PBS): unsupported implementation variant.")
+      }
+    }
+  }
+
+  void release(cudaStream_t stream, uint32_t gpu_index) {
+    cuda_drop_async(d_mem, stream, gpu_index);
+    cuda_drop_async(global_join_buffer, stream, gpu_index);
+
+    if (pbs_variant == DEFAULT)
+      cuda_drop_async(global_accumulator, stream, gpu_index);
+  }
+};
+
 template <typename Torus>
 uint64_t get_buffer_size_programmable_bootstrap_cg(
     uint32_t glwe_dimension, uint32_t polynomial_size, uint32_t level_count,
@@ -245,7 +400,7 @@ bool has_support_to_cuda_programmable_bootstrap_cg(uint32_t glwe_dimension,
                                                    uint32_t polynomial_size,
                                                    uint32_t level_count,
                                                    uint32_t num_samples,
-                                                   int max_shared_memory);
+                                                   uint32_t max_shared_memory);
 
 template <typename Torus>
 void cuda_programmable_bootstrap_cg_lwe_ciphertext_vector(

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

@@ -20,6 +20,11 @@ void cuda_convert_lwe_programmable_bootstrap_key_64(
     uint32_t input_lwe_dim, uint32_t glwe_dim, uint32_t level_count,
     uint32_t polynomial_size);
 
+void cuda_convert_lwe_programmable_bootstrap_key_128(
+    void *stream, uint32_t gpu_index, void *dest, void const *src,
+    uint32_t input_lwe_dim, uint32_t glwe_dim, uint32_t level_count,
+    uint32_t polynomial_size);
+
 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,
@@ -62,6 +67,11 @@ void scratch_cuda_programmable_bootstrap_64(
     uint32_t polynomial_size, uint32_t level_count,
     uint32_t input_lwe_ciphertext_count, bool allocate_gpu_memory);
 
+void scratch_cuda_programmable_bootstrap_128(
+    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, bool allocate_gpu_memory);
+
 void cuda_programmable_bootstrap_lwe_ciphertext_vector_32(
     void *stream, uint32_t gpu_index, void *lwe_array_out,
     void const *lwe_output_indexes, void const *lut_vector,
@@ -80,7 +90,17 @@ void cuda_programmable_bootstrap_lwe_ciphertext_vector_64(
     uint32_t polynomial_size, uint32_t base_log, uint32_t level_count,
     uint32_t num_samples, uint32_t num_many_lut, uint32_t lut_stride);
 
+void cuda_programmable_bootstrap_lwe_ciphertext_vector_128(
+    void *stream, uint32_t gpu_index, void *lwe_array_out,
+    void const *lut_vector, void const *lwe_array_in,
+    void const *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);
+
 void cleanup_cuda_programmable_bootstrap(void *stream, uint32_t gpu_index,
                                          int8_t **pbs_buffer);
+
+void cleanup_cuda_programmable_bootstrap_128(void *stream, uint32_t gpu_index,
+                                             int8_t **pbs_buffer);
 }
 #endif // CUDA_BOOTSTRAP_H

backends/tfhe-cuda-backend/cuda/include/pbs/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, int max_shared_memory);
+    uint32_t num_samples, uint32_t max_shared_memory);
 
 void cuda_convert_lwe_multi_bit_programmable_bootstrap_key_64(
     void *stream, uint32_t gpu_index, void *dest, void const *src,

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

@@ -24,7 +24,7 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu_64(void *stream,
 void cuda_glwe_sample_extract_64(void *stream, uint32_t gpu_index,
                                  void *lwe_array_out, void const *glwe_array_in,
                                  uint32_t const *nth_array, uint32_t num_nths,
-                                 uint32_t glwe_dimension,
+                                 uint32_t lwe_per_glwe, uint32_t glwe_dimension,
                                  uint32_t polynomial_size) {
 
   switch (polynomial_size) {
@@ -32,43 +32,43 @@ void cuda_glwe_sample_extract_64(void *stream, uint32_t gpu_index,
     host_sample_extract<uint64_t, AmortizedDegree<256>>(
         static_cast<cudaStream_t>(stream), gpu_index, (uint64_t *)lwe_array_out,
         (uint64_t const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, 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 const *)glwe_array_in, (uint32_t const *)nth_array, num_nths,
-        glwe_dimension);
+        lwe_per_glwe, glwe_dimension);
     break;
   default:
     PANIC("Cuda error: unsupported polynomial size. Supported "

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

@@ -28,7 +28,7 @@ void cuda_convert_lwe_ciphertext_vector_to_cpu(cudaStream_t stream,
 
 template <typename Torus, class params>
 __global__ void sample_extract(Torus *lwe_array_out, Torus const *glwe_array_in,
-                               uint32_t const *nth_array,
+                               uint32_t const *nth_array, uint32_t lwe_per_glwe,
                                uint32_t glwe_dimension) {
 
   const int input_id = blockIdx.x;
@@ -39,28 +39,28 @@ __global__ void sample_extract(Torus *lwe_array_out, Torus const *glwe_array_in,
   auto lwe_out = lwe_array_out + input_id * lwe_output_size;
 
   // We assume each GLWE will store the first polynomial_size inputs
-  uint32_t lwe_per_glwe = params::degree;
   auto glwe_in = glwe_array_in + (input_id / lwe_per_glwe) * glwe_input_size;
 
-  // nth is ensured to be in [0, lwe_per_glwe)
-  auto nth = nth_array[input_id] % lwe_per_glwe;
+  // nth is ensured to be in [0, params::degree)
+  auto nth = nth_array[input_id] % params::degree;
 
   sample_extract_mask<Torus, params>(lwe_out, glwe_in, glwe_dimension, nth);
   sample_extract_body<Torus, params>(lwe_out, glwe_in, glwe_dimension, nth);
 }
 
+// lwe_per_glwe LWEs will be extracted per GLWE ciphertext, thus we need to have
+// enough indexes
 template <typename Torus, class params>
-__host__ void host_sample_extract(cudaStream_t stream, uint32_t gpu_index,
-                                  Torus *lwe_array_out,
-                                  Torus const *glwe_array_in,
-                                  uint32_t const *nth_array, uint32_t num_nths,
-                                  uint32_t glwe_dimension) {
+__host__ void
+host_sample_extract(cudaStream_t stream, uint32_t gpu_index,
+                    Torus *lwe_array_out, Torus const *glwe_array_in,
+                    uint32_t const *nth_array, uint32_t num_nths,
+                    uint32_t lwe_per_glwe, uint32_t glwe_dimension) {
   cuda_set_device(gpu_index);
-
   dim3 grid(num_nths);
   dim3 thds(params::degree / params::opt);
   sample_extract<Torus, params><<<grid, thds, 0, stream>>>(
-      lwe_array_out, glwe_array_in, nth_array, glwe_dimension);
+      lwe_array_out, glwe_array_in, nth_array, lwe_per_glwe, glwe_dimension);
   check_cuda_error(cudaGetLastError());
 }
 

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

@@ -2,7 +2,6 @@
 #define CNCRT_FAST_KS_CUH
 
 #undef NDEBUG
-#include <assert.h>
 
 #include "device.h"
 #include "gadget.cuh"

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

@@ -3,6 +3,7 @@
 
 #include "crypto/torus.cuh"
 #include "device.h"
+#include "fft128/f128.cuh"
 #include <cstdint>
 
 /**
@@ -42,6 +43,13 @@ public:
     }
   }
 
+  __device__ void decompose_and_compress_next_128(double *result) {
+    for (int j = 0; j < num_poly; j++) {
+      auto result_slice = result + j * params::degree / 2 * 4;
+      decompose_and_compress_next_polynomial_128(result_slice, j);
+    }
+  }
+
   // Decomposes a single polynomial
   __device__ void decompose_and_compress_next_polynomial(double2 *result,
                                                          int j) {
@@ -75,10 +83,58 @@ public:
     synchronize_threads_in_block();
   }
 
+  // Decomposes a single polynomial
+  __device__ void decompose_and_compress_next_polynomial_128(double *result,
+                                                             int j) {
+    uint32_t tid = threadIdx.x;
+    auto state_slice = &state[j * params::degree];
+    for (int i = 0; i < params::opt / 2; i++) {
+      auto input1 = &state_slice[tid];
+      auto input2 = &state_slice[tid + params::degree / 2];
+      T res_re = *input1 & mask_mod_b;
+      T res_im = *input2 & mask_mod_b;
+
+      *input1 >>= base_log; // Update state
+      *input2 >>= base_log; // Update state
+
+      T carry_re = ((res_re - 1ll) | *input1) & res_re;
+      T carry_im = ((res_im - 1ll) | *input2) & res_im;
+      carry_re >>= (base_log - 1);
+      carry_im >>= (base_log - 1);
+
+      *input1 += carry_re; // Update state
+      *input2 += carry_im; // Update state
+
+      res_re -= carry_re << base_log;
+      res_im -= carry_im << base_log;
+
+      auto out_re = u128_to_signed_to_f128(res_re);
+      auto out_im = u128_to_signed_to_f128(res_im);
+
+      auto out_re_hi = result + 0 * params::degree / 2;
+      auto out_re_lo = result + 1 * params::degree / 2;
+      auto out_im_hi = result + 2 * params::degree / 2;
+      auto out_im_lo = result + 3 * params::degree / 2;
+
+      out_re_hi[tid] = out_re.hi;
+      out_re_lo[tid] = out_re.lo;
+      out_im_hi[tid] = out_im.hi;
+      out_im_lo[tid] = out_im.lo;
+
+      tid += params::degree / params::opt;
+    }
+    synchronize_threads_in_block();
+  }
+
   __device__ void decompose_and_compress_level(double2 *result, int level) {
     for (int i = 0; i < level_count - level; i++)
       decompose_and_compress_next(result);
   }
+
+  __device__ void decompose_and_compress_level_128(double *result, int level) {
+    for (int i = 0; i < level_count - level; i++)
+      decompose_and_compress_next_128(result);
+  }
 };
 
 template <typename Torus>

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

@@ -135,7 +135,7 @@ bool cuda_check_support_thread_block_clusters() {
 }
 
 /// Copy memory to the GPU asynchronously
-void cuda_memcpy_async_to_gpu(void *dest, void *src, uint64_t size,
+void cuda_memcpy_async_to_gpu(void *dest, const void *src, uint64_t size,
                               cudaStream_t stream, uint32_t gpu_index) {
   if (size == 0)
     return;
@@ -304,8 +304,8 @@ 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) {
-  int max_shared_memory = 0;
+uint32_t cuda_get_max_shared_memory(uint32_t gpu_index) {
+  auto max_shared_memory = 0;
 #if CUDA_ARCH == 900
   max_shared_memory = 226000;
 #elif CUDA_ARCH == 890
@@ -321,5 +321,5 @@ int cuda_get_max_shared_memory(uint32_t gpu_index) {
                          gpu_index);
   check_cuda_error(cudaGetLastError());
 #endif
-  return max_shared_memory;
+  return (uint32_t)(max_shared_memory);
 }

backends/tfhe-cuda-backend/cuda/src/fft128/f128.cuh

@@ -0,0 +1,410 @@
+
+#ifndef CUDA_FFT128_F128_CUH
+#define CUDA_FFT128_F128_CUH
+
+#include <cstdint>
+
+struct alignas(16) f128 {
+  double hi;
+  double lo;
+
+  // Default and parameterized constructors
+  __host__ __device__ f128() : hi(0.0), lo(0.0) {}
+  __host__ __device__ f128(double high, double low) : hi(high), lo(low) {}
+
+  // Quick two-sum
+  __host__ __device__ __forceinline__ static f128 quick_two_sum(double a,
+                                                                double b) {
+#ifdef __CUDA_ARCH__
+    double s = __dadd_rn(a, b);
+    return f128(s, __dsub_rn(b, __dsub_rn(s, a)));
+#else
+    double s = a + b;
+    return f128(s, b - (s - a));
+#endif
+  }
+
+  // Two-sum
+  __host__ __device__ __forceinline__ static f128 two_sum(double a, double b) {
+#ifdef __CUDA_ARCH__
+    double s = __dadd_rn(a, b);
+    double bb = __dsub_rn(s, a);
+    return f128(s, __dadd_rn(__dsub_rn(a, __dsub_rn(s, bb)), __dsub_rn(b, bb)));
+#else
+    double s = a + b;
+    double bb = s - a;
+    return f128(s, (a - (s - bb)) + (b - bb));
+#endif
+  }
+
+  // Two-product
+  __host__ __device__ __forceinline__ static f128 two_prod(double a, double b) {
+
+#ifdef __CUDA_ARCH__
+    double p = __dmul_rn(a, b);
+    double p2 = __fma_rn(a, b, -p);
+#else
+    double p = a * b;
+    double p2 = fma(a, b, -p);
+#endif
+    return f128(p, p2);
+  }
+
+  __host__ __device__ __forceinline__ static f128 two_diff(double a, double b) {
+#ifdef __CUDA_ARCH__
+    double s = __dsub_rn(a, b);
+    double bb = __dsub_rn(s, a);
+    return f128(s, __dsub_rn(__dsub_rn(a, __dsub_rn(s, bb)), __dadd_rn(b, bb)));
+#else
+    double s = a - b;
+    double bb = s - a;
+    return f128(s, (a - (s - bb)) - (b + bb));
+#endif
+  }
+
+  // Addition
+  __host__ __device__ static f128 add(const f128 &a, const f128 &b) {
+    auto s = two_sum(a.hi, b.hi);
+    auto t = two_sum(a.lo, b.lo);
+
+    double hi = s.hi;
+    double lo = s.lo + t.hi;
+    hi = hi + lo;
+    lo = lo - (hi - s.hi);
+
+    return f128(hi, lo + t.lo);
+  }
+
+  // Addition with estimate
+  __host__ __device__ static f128 add_estimate(const f128 &a, const f128 &b) {
+    auto se = two_sum(a.hi, b.hi);
+#ifdef __CUDA_ARCH__
+    se.lo = __dadd_rn(se.lo, __dadd_rn(a.lo, b.lo));
+#else
+    se.lo += (a.lo + b.lo);
+#endif
+
+    return quick_two_sum(se.hi, se.lo);
+  }
+
+  // Subtraction with estimate
+  __host__ __device__ static f128 sub_estimate(const f128 &a, const f128 &b) {
+    f128 se = two_diff(a.hi, b.hi);
+#ifdef __CUDA_ARCH__
+    se.lo = __dadd_rn(se.lo, a.lo);
+    se.lo = __dsub_rn(se.lo, b.lo);
+#else
+    se.lo += a.lo;
+    se.lo -= b.lo;
+#endif
+    return quick_two_sum(se.hi, se.lo);
+  }
+
+  // Subtraction
+  __host__ __device__ static f128 sub(const f128 &a, const f128 &b) {
+    auto s = two_diff(a.hi, b.hi);
+    auto t = two_diff(a.lo, b.lo);
+    s = quick_two_sum(s.hi, s.lo + t.hi);
+    return quick_two_sum(s.hi, s.lo + t.lo);
+  }
+
+  // Multiplication
+  __host__ __device__ static f128 mul(const f128 &a, const f128 &b) {
+    auto p = two_prod(a.hi, b.hi);
+#ifdef __CUDA_ARCH__
+    double a_0_x_b_1 = __dmul_rn(a.hi, b.lo);
+    double a_1_x_b_0 = __dmul_rn(a.lo, b.hi);
+    p.lo = __dadd_rn(p.lo, __dadd_rn(a_0_x_b_1, a_1_x_b_0));
+#else
+    p.lo += (a.hi * b.lo + a.lo * b.hi);
+#endif
+    p = quick_two_sum(p.hi, p.lo);
+    return p;
+  }
+
+  __host__ __device__ static f128 add_f64_f64(const double a, const double b) {
+    return two_sum(a, b);
+  }
+
+  __host__ __device__ static f128 f128_floor(const f128 &x) {
+    double x0_floor = floor(x.hi);
+    if (x0_floor == x.hi) {
+      return add_f64_f64(x0_floor, floor(x.lo));
+    }
+
+    return f128(x0_floor, 0.0);
+  }
+
+  __host__ __device__ static void
+  cplx_f128_mul_assign(f128 &c_re, f128 &c_im, const f128 &a_re,
+                       const f128 &a_im, const f128 &b_re, const f128 &b_im) {
+    auto a_re_x_b_re = mul(a_re, b_re);
+    auto a_re_x_b_im = mul(a_re, b_im);
+    auto a_im_x_b_re = mul(a_im, b_re);
+    auto a_im_x_b_im = mul(a_im, b_im);
+
+    c_re = sub_estimate(a_re_x_b_re, a_im_x_b_im);
+    c_im = add_estimate(a_im_x_b_re, a_re_x_b_im);
+  }
+
+  __host__ __device__ static void
+  cplx_f128_sub_assign(f128 &c_re, f128 &c_im, const f128 &a_re,
+                       const f128 &a_im, const f128 &b_re, const f128 &b_im) {
+    c_re = sub_estimate(a_re, b_re);
+    c_im = sub_estimate(a_im, b_im);
+  }
+  __host__ __device__ static void
+  cplx_f128_add_assign(f128 &c_re, f128 &c_im, const f128 &a_re,
+                       const f128 &a_im, const f128 &b_re, const f128 &b_im) {
+    c_re = add_estimate(a_re, b_re);
+    c_im = add_estimate(a_im, b_im);
+  }
+};
+
+struct f128x2 {
+  f128 re;
+  f128 im;
+
+  __host__ __device__ f128x2() : re(), im() {}
+
+  __host__ __device__ f128x2(const f128 &real, const f128 &imag)
+      : re(real), im(imag) {}
+
+  __host__ __device__ f128x2(double real, double imag)
+      : re(real, 0.0), im(imag, 0.0) {}
+
+  __host__ __device__ explicit f128x2(double real)
+      : re(real, 0.0), im(0.0, 0.0) {}
+
+  __host__ __device__ f128x2(const f128x2 &other)
+      : re(other.re), im(other.im) {}
+
+  __host__ __device__ f128x2(f128x2 &&other) noexcept
+      : re(std::move(other.re)), im(std::move(other.im)) {}
+
+  __host__ __device__ f128x2 &operator=(const f128x2 &other) {
+    if (this != &other) {
+      re = other.re;
+      im = other.im;
+    }
+    return *this;
+  }
+
+  __host__ __device__ f128x2 &operator=(f128x2 &&other) noexcept {
+    if (this != &other) {
+      re = std::move(other.re);
+      im = std::move(other.im);
+    }
+    return *this;
+  }
+
+  __host__ __device__ f128x2 conjugate() const {
+    return f128x2(re, f128(-im.hi, -im.lo));
+  }
+
+  __host__ __device__ f128 norm_squared() const {
+    return f128::add(f128::mul(re, re), f128::mul(im, im));
+  }
+
+  __host__ __device__ void zero() {
+    re = f128(0.0, 0.0);
+    im = f128(0.0, 0.0);
+  }
+
+  // Addition
+  __host__ __device__ friend f128x2 operator+(const f128x2 &a,
+                                              const f128x2 &b) {
+    return f128x2(f128::add(a.re, b.re), f128::add(a.im, b.im));
+  }
+
+  // Subtraction
+  __host__ __device__ friend f128x2 operator-(const f128x2 &a,
+                                              const f128x2 &b) {
+    return f128x2(f128::add(a.re, f128(-b.re.hi, -b.re.lo)),
+                  f128::add(a.im, f128(-b.im.hi, -b.im.lo)));
+  }
+
+  // Multiplication (complex multiplication)
+  __host__ __device__ friend f128x2 operator*(const f128x2 &a,
+                                              const f128x2 &b) {
+    f128 real_part =
+        f128::add(f128::mul(a.re, b.re),
+                  f128(-f128::mul(a.im, b.im).hi, -f128::mul(a.im, b.im).lo));
+    f128 imag_part = f128::add(f128::mul(a.re, b.im), f128::mul(a.im, b.re));
+    return f128x2(real_part, imag_part);
+  }
+
+  // Addition-assignment operator
+  __host__ __device__ f128x2 &operator+=(const f128x2 &other) {
+    re = f128::add(re, other.re);
+    im = f128::add(im, other.im);
+    return *this;
+  }
+
+  // Subtraction-assignment operator
+  __host__ __device__ f128x2 &operator-=(const f128x2 &other) {
+    re = f128::add(re, f128(-other.re.hi, -other.re.lo));
+    im = f128::add(im, f128(-other.im.hi, -other.im.lo));
+    return *this;
+  }
+
+  // Multiplication-assignment operator
+  __host__ __device__ f128x2 &operator*=(const f128x2 &other) {
+    f128 new_re =
+        f128::add(f128::mul(re, other.re), f128(-f128::mul(im, other.im).hi,
+                                                -f128::mul(im, other.im).lo));
+    f128 new_im = f128::add(f128::mul(re, other.im), f128::mul(im, other.re));
+    re = new_re;
+    im = new_im;
+    return *this;
+  }
+};
+
+__host__ __device__ inline uint64_t double_to_bits(double d) {
+  uint64_t bits = *reinterpret_cast<uint64_t *>(&d);
+  return bits;
+}
+
+__host__ __device__ inline double bits_to_double(uint64_t bits) {
+  double d = *reinterpret_cast<double *>(&bits);
+  return d;
+}
+
+__host__ __device__ inline double u128_to_f64(__uint128_t x) {
+  const __uint128_t ONE = 1;
+  const double A = ONE << 52;
+  const double B = ONE << 104;
+  const double C = ONE << 76;
+  const double D = 340282366920938500000000000000000000000.;
+
+  const __uint128_t threshold = (ONE << 104);
+
+  if (x < threshold) {
+    uint64_t A_bits = double_to_bits(A);
+
+    __uint128_t shifted = (x << 12);
+    uint64_t lower64 = static_cast<uint64_t>(shifted);
+    lower64 >>= 12;
+
+    uint64_t bits_l = A_bits | lower64;
+    double l_temp = bits_to_double(bits_l);
+    double l = l_temp - A;
+
+    uint64_t B_bits = double_to_bits(B);
+    uint64_t top64 = static_cast<uint64_t>(x >> 52);
+    uint64_t bits_h = B_bits | top64;
+    double h_temp = bits_to_double(bits_h);
+    double h = h_temp - B;
+
+    return (l + h);
+
+  } else {
+    uint64_t C_bits = double_to_bits(C);
+
+    __uint128_t shifted = (x >> 12);
+    uint64_t lower64 = static_cast<uint64_t>(shifted);
+    lower64 >>= 12;
+
+    uint64_t x_lo = static_cast<uint64_t>(x);
+    uint64_t mask_part = (x_lo & 0xFFFFFFULL);
+
+    uint64_t bits_l = C_bits | lower64 | mask_part;
+    double l_temp = bits_to_double(bits_l);
+    double l = l_temp - C;
+
+    uint64_t D_bits = double_to_bits(D);
+    uint64_t top64 = static_cast<uint64_t>(x >> 76);
+    uint64_t bits_h = D_bits | top64;
+    double h_temp = bits_to_double(bits_h);
+    double h = h_temp - D;
+
+    return (l + h);
+  }
+}
+
+__host__ __device__ inline __uint128_t f64_to_u128(const double f) {
+  const __uint128_t ONE = 1;
+  const uint64_t f_bits = double_to_bits(f);
+  if (f_bits < 1023ull << 52) {
+    return 0;
+  } else {
+    const __uint128_t m = ONE << 127 | (__uint128_t)f_bits << 75;
+    const uint64_t s = 1150 - (f_bits >> 52);
+    if (s >= 128) {
+      return 0;
+    } else {
+      return m >> s;
+    }
+  }
+}
+
+__host__ __device__ inline __uint128_t f64_to_i128(const double f) {
+  // Get raw bits of the double
+  const uint64_t f_bits = double_to_bits(f);
+
+  // Remove sign bit (equivalent to Rust's !0 >> 1 mask)
+  const uint64_t a = f_bits & 0x7FFFFFFFFFFFFFFFull;
+
+  // Check if value is in [0, 1) range
+  if (a < (1023ull << 52)) {
+    return 0;
+  }
+
+  // Reconstruct mantissa with implicit leading 1
+  const __uint128_t m =
+      (__uint128_t{1} << 127) | (static_cast<__uint128_t>(a) << 75);
+
+  // Calculate shift amount based on exponent
+  const uint64_t exponent = a >> 52;
+  const uint64_t s = 1150 - exponent;
+
+  // Perform unsigned right shift
+  const __uint128_t u = m >> s;
+
+  // Apply sign (check original sign bit)
+  const __int128_t result = static_cast<__int128_t>(u);
+  return (f_bits >> 63) ? -result : result;
+}
+
+__host__ __device__ inline double i128_to_f64(__int128_t const x) {
+  uint64_t sign = static_cast<uint64_t>(x >> 64) & (1ULL << 63);
+  __uint128_t abs =
+      (x < 0) ? static_cast<__uint128_t>(-x) : static_cast<__uint128_t>(x);
+
+  return bits_to_double(double_to_bits(u128_to_f64(abs)) | sign);
+}
+__host__ __device__ inline f128 u128_to_signed_to_f128(__uint128_t x) {
+  const double first_approx = i128_to_f64(x);
+  const uint64_t sign_bit = double_to_bits(first_approx) & (1ull << 63);
+  const __uint128_t first_approx_roundtrip =
+      f64_to_u128((first_approx < 0) ? -first_approx : first_approx);
+  const __uint128_t first_approx_roundtrip_signed =
+      (sign_bit == (1ull << 63)) ? -first_approx_roundtrip
+                                 : first_approx_roundtrip;
+
+  double correction = i128_to_f64(x - first_approx_roundtrip_signed);
+
+  return f128(first_approx, correction);
+}
+
+__host__ __device__ inline __uint128_t u128_from_torus_f128(const f128 &a) {
+  auto x = f128::sub_estimate(a, f128::f128_floor(a));
+  const double normalization = 340282366920938500000000000000000000000.;
+#ifdef __CUDA_ARCH__
+  x.hi = __dmul_rn(x.hi, normalization);
+  x.lo = __dmul_rn(x.lo, normalization);
+#else
+  x.hi *= normalization;
+  x.lo *= normalization;
+#endif
+
+  // TODO has to be round
+  x = f128::f128_floor(x);
+
+  __uint128_t x0 = f64_to_u128(x.hi);
+  __int128_t x1 = f64_to_i128(x.lo);
+
+  return x0 + x1;
+}
+
+#endif

backends/tfhe-cuda-backend/cuda/src/fft128/fft128.cu

@@ -0,0 +1,163 @@
+#include "fft128.cuh"
+
+void cuda_fourier_transform_forward_as_integer_f128_async(
+    void *stream, uint32_t gpu_index, void *re0, void *re1, void *im0,
+    void *im1, void const *standard, const uint32_t N,
+    const uint32_t number_of_samples) {
+  switch (N) {
+  case 64:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<64>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 128:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<128>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 256:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<256>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 512:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<512>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 1024:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<1024>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 2048:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 4096:
+    host_fourier_transform_forward_as_integer_f128<AmortizedDegree<4096>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  default:
+    PANIC("Cuda error (f128 fft): unsupported polynomial size. Supported "
+          "N's are powers of two"
+          " in the interval [64..4096].")
+  }
+}
+
+void cuda_fourier_transform_forward_as_torus_f128_async(
+    void *stream, uint32_t gpu_index, void *re0, void *re1, void *im0,
+    void *im1, void const *standard, const uint32_t N,
+    const uint32_t number_of_samples) {
+  switch (N) {
+  case 64:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<64>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 128:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<128>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 256:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<256>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 512:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<512>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 1024:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<1024>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 2048:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  case 4096:
+    host_fourier_transform_forward_as_torus_f128<AmortizedDegree<4096>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (double *)re0,
+        (double *)re1, (double *)im0, (double *)im1,
+        (__uint128_t const *)standard, N, number_of_samples);
+    break;
+  default:
+    PANIC("Cuda error (f128 fft): unsupported polynomial size. Supported "
+          "N's are powers of two"
+          " in the interval [64..4096].")
+  }
+}
+
+void cuda_fourier_transform_backward_as_torus_f128_async(
+    void *stream, uint32_t gpu_index, void *standard, void const *re0,
+    void const *re1, void const *im0, void const *im1, const uint32_t N,
+    const uint32_t number_of_samples) {
+  switch (N) {
+  case 64:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<64>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 128:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<128>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 256:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<256>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 512:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<512>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 1024:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<1024>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 2048:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<2048>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  case 4096:
+    host_fourier_transform_backward_as_torus_f128<AmortizedDegree<4096>>(
+        static_cast<cudaStream_t>(stream), gpu_index, (__uint128_t *)standard,
+        (double const *)re0, (double const *)re1, (double const *)im0,
+        (double const *)im1, N, number_of_samples);
+    break;
+  default:
+    PANIC("Cuda error (f128 ifft): unsupported polynomial size. Supported "
+          "N's are powers of two"
+          " in the interval [64..4096].")
+  }
+}

backends/tfhe-cuda-backend/cuda/src/fft128/fft128.cuh

@@ -0,0 +1,669 @@
+#ifndef CUDA_FFT128_CUH
+#define CUDA_FFT128_CUH
+
+#include "f128.cuh"
+#include "fft/fft128.h"
+#include "polynomial/functions.cuh"
+#include "polynomial/parameters.cuh"
+#include "twiddles.cuh"
+#include "types/complex/operations.cuh"
+#include <iostream>
+
+using Index = unsigned;
+
+#define NEG_TWID(i)                                                            \
+  f128x2(f128(neg_twiddles_re_hi[(i)], neg_twiddles_re_lo[(i)]),               \
+         f128(neg_twiddles_im_hi[(i)], neg_twiddles_im_lo[(i)]))
+
+#define F64x4_TO_F128x2(f128x2_reg, ind)                                       \
+  f128x2_reg.re.hi = dt_re_hi[ind];                                            \
+  f128x2_reg.re.lo = dt_re_lo[ind];                                            \
+  f128x2_reg.im.hi = dt_im_hi[ind];                                            \
+  f128x2_reg.im.lo = dt_im_lo[ind]
+
+#define F128x2_TO_F64x4(f128x2_reg, ind)                                       \
+  dt_re_hi[ind] = f128x2_reg.re.hi;                                            \
+  dt_re_lo[ind] = f128x2_reg.re.lo;                                            \
+  dt_im_hi[ind] = f128x2_reg.im.hi;                                            \
+  dt_im_lo[ind] = f128x2_reg.im.lo
+
+template <class params>
+__device__ void negacyclic_forward_fft_f128(double *dt_re_hi, double *dt_re_lo,
+                                            double *dt_im_hi,
+                                            double *dt_im_lo) {
+
+  __syncthreads();
+  constexpr Index BUTTERFLY_DEPTH = params::opt >> 1;
+  constexpr Index LOG2_DEGREE = params::log2_degree;
+  constexpr Index HALF_DEGREE = params::degree >> 1;
+  constexpr Index STRIDE = params::degree / params::opt;
+
+  f128x2 u[BUTTERFLY_DEPTH], v[BUTTERFLY_DEPTH], w;
+
+  Index tid = threadIdx.x;
+
+  // load into registers
+#pragma unroll
+  for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+    F64x4_TO_F128x2(u[i], tid);
+    F64x4_TO_F128x2(v[i], tid + HALF_DEGREE);
+    tid += STRIDE;
+  }
+
+  // level 1
+  // we don't make actual complex multiplication on level1 since we have only
+  // one twiddle, it's real and image parts are equal, so we can multiply
+  // it with simpler operations
+#pragma unroll
+  for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+    auto ww = NEG_TWID(1);
+    f128::cplx_f128_mul_assign(w.re, w.im, v[i].re, v[i].im, NEG_TWID(1).re,
+                               NEG_TWID(1).im);
+    f128::cplx_f128_sub_assign(v[i].re, v[i].im, u[i].re, u[i].im, w.re, w.im);
+    f128::cplx_f128_add_assign(u[i].re, u[i].im, u[i].re, u[i].im, w.re, w.im);
+  }
+
+  Index twiddle_shift = 1;
+  for (Index l = LOG2_DEGREE - 1; l >= 1; --l) {
+    Index lane_mask = 1 << (l - 1);
+    Index thread_mask = (1 << l) - 1;
+    twiddle_shift <<= 1;
+
+    tid = threadIdx.x;
+    __syncthreads();
+#pragma unroll
+    for (Index i = 0; i < BUTTERFLY_DEPTH; i++) {
+      Index rank = tid & thread_mask;
+      bool u_stays_in_register = rank < lane_mask;
+      F128x2_TO_F64x4(((u_stays_in_register) ? v[i] : u[i]), tid);
+      tid = tid + STRIDE;
+    }
+    __syncthreads();
+
+    tid = threadIdx.x;
+#pragma unroll
+    for (Index i = 0; i < BUTTERFLY_DEPTH; i++) {
+      Index rank = tid & thread_mask;
+      bool u_stays_in_register = rank < lane_mask;
+      F64x4_TO_F128x2(w, tid ^ lane_mask);
+      u[i] = (u_stays_in_register) ? u[i] : w;
+      v[i] = (u_stays_in_register) ? w : v[i];
+      w = NEG_TWID(tid / lane_mask + twiddle_shift);
+      f128::cplx_f128_mul_assign(w.re, w.im, v[i].re, v[i].im, w.re, w.im);
+      f128::cplx_f128_sub_assign(v[i].re, v[i].im, u[i].re, u[i].im, w.re,
+                                 w.im);
+      f128::cplx_f128_add_assign(u[i].re, u[i].im, u[i].re, u[i].im, w.re,
+                                 w.im);
+      tid = tid + STRIDE;
+    }
+  }
+  __syncthreads();
+
+  //   store registers in SM
+  tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < BUTTERFLY_DEPTH; i++) {
+    F128x2_TO_F64x4(u[i], tid * 2);
+    F128x2_TO_F64x4(v[i], (tid * 2 + 1));
+    tid = tid + STRIDE;
+  }
+  __syncthreads();
+}
+
+template <class params>
+__device__ void negacyclic_backward_fft_f128(double *dt_re_hi, double *dt_re_lo,
+                                             double *dt_im_hi,
+                                             double *dt_im_lo) {
+  __syncthreads();
+  constexpr Index BUTTERFLY_DEPTH = params::opt >> 1;
+  constexpr Index LOG2_DEGREE = params::log2_degree;
+  constexpr Index DEGREE = params::degree;
+  constexpr Index HALF_DEGREE = params::degree >> 1;
+  constexpr Index STRIDE = params::degree / params::opt;
+
+  size_t tid = threadIdx.x;
+  f128x2 u[BUTTERFLY_DEPTH], v[BUTTERFLY_DEPTH], w;
+
+  // load into registers and divide by compressed polynomial size
+#pragma unroll
+  for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+    F64x4_TO_F128x2(u[i], 2 * tid);
+    F64x4_TO_F128x2(v[i], 2 * tid + 1);
+    tid += STRIDE;
+  }
+
+  Index twiddle_shift = DEGREE;
+  for (Index l = 1; l <= LOG2_DEGREE - 1; ++l) {
+    Index lane_mask = 1 << (l - 1);
+    Index thread_mask = (1 << l) - 1;
+    tid = threadIdx.x;
+    twiddle_shift >>= 1;
+
+    // at this point registers are ready for the  butterfly
+    tid = threadIdx.x;
+    __syncthreads();
+#pragma unroll
+    for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+      w = (u[i] - v[i]);
+      u[i] += v[i];
+      v[i] = w * NEG_TWID(tid / lane_mask + twiddle_shift).conjugate();
+
+      // keep one of the register for next iteration and store another one in sm
+      Index rank = tid & thread_mask;
+      bool u_stays_in_register = rank < lane_mask;
+      F128x2_TO_F64x4(((u_stays_in_register) ? v[i] : u[i]), tid);
+
+      tid = tid + STRIDE;
+    }
+    __syncthreads();
+
+    // prepare registers for next butterfly iteration
+    tid = threadIdx.x;
+#pragma unroll
+    for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+      Index rank = tid & thread_mask;
+      bool u_stays_in_register = rank < lane_mask;
+      F64x4_TO_F128x2(w, tid ^ lane_mask);
+
+      u[i] = (u_stays_in_register) ? u[i] : w;
+      v[i] = (u_stays_in_register) ? w : v[i];
+
+      tid = tid + STRIDE;
+    }
+  }
+
+  // last iteration
+  for (Index i = 0; i < BUTTERFLY_DEPTH; ++i) {
+    w = (u[i] - v[i]);
+    u[i] = u[i] + v[i];
+    v[i] = w * NEG_TWID(1).conjugate();
+  }
+  __syncthreads();
+  // store registers in SM
+  tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < BUTTERFLY_DEPTH; i++) {
+    F128x2_TO_F64x4(u[i], tid);
+    F128x2_TO_F64x4(v[i], tid + HALF_DEGREE);
+
+    tid = tid + STRIDE;
+  }
+  __syncthreads();
+}
+
+// params is expected to be full degree not half degree
+template <class params>
+__device__ void convert_u128_to_f128_as_integer(
+    double *out_re_hi, double *out_re_lo, double *out_im_hi, double *out_im_lo,
+    const __uint128_t *in_re, const __uint128_t *in_im) {
+
+  Index tid = threadIdx.x;
+  // #pragma unroll
+  for (Index i = 0; i < params::opt / 2; i++) {
+    auto out_re = u128_to_signed_to_f128(in_re[tid]);
+    auto out_im = u128_to_signed_to_f128(in_im[tid]);
+
+    out_re_hi[tid] = out_re.hi;
+    out_re_lo[tid] = out_re.lo;
+    out_im_hi[tid] = out_im.hi;
+    out_im_lo[tid] = out_im.lo;
+
+    tid += params::degree / params::opt;
+  }
+}
+
+// params is expected to be full degree not half degree
+template <class params>
+__device__ void convert_u128_to_f128_as_torus(
+    double *out_re_hi, double *out_re_lo, double *out_im_hi, double *out_im_lo,
+    const __uint128_t *in_re, const __uint128_t *in_im) {
+
+  const double normalization = pow(2., -128.);
+  Index tid = threadIdx.x;
+  // #pragma unroll
+  for (Index i = 0; i < params::opt / 2; i++) {
+    auto out_re = u128_to_signed_to_f128(in_re[tid]);
+    auto out_im = u128_to_signed_to_f128(in_im[tid]);
+
+    out_re_hi[tid] = out_re.hi * normalization;
+    out_re_lo[tid] = out_re.lo * normalization;
+    out_im_hi[tid] = out_im.hi * normalization;
+    out_im_lo[tid] = out_im.lo * normalization;
+
+    tid += params::degree / params::opt;
+  }
+}
+
+template <class params>
+__device__ void
+convert_f128_to_u128_as_torus(__uint128_t *out_re, __uint128_t *out_im,
+                              const double *in_re_hi, const double *in_re_lo,
+                              const double *in_im_hi, const double *in_im_lo) {
+
+  const double normalization = 1. / (params::degree / 2);
+  Index tid = threadIdx.x;
+  // #pragma unroll
+  for (Index i = 0; i < params::opt / 2; i++) {
+
+    f128 in_re(in_re_hi[tid] * normalization, in_re_lo[tid] * normalization);
+    f128 in_im(in_im_hi[tid] * normalization, in_im_lo[tid] * normalization);
+
+    out_re[tid] = u128_from_torus_f128(in_re);
+    out_im[tid] = u128_from_torus_f128(in_im);
+
+    tid += params::degree / params::opt;
+  }
+}
+
+// params is expected to be full degree not half degree
+template <class params>
+__global__ void
+batch_convert_u128_to_f128_as_integer(double *out_re_hi, double *out_re_lo,
+                                      double *out_im_hi, double *out_im_lo,
+                                      const __uint128_t *in) {
+
+  convert_u128_to_f128_as_integer<params>(
+      &out_re_hi[blockIdx.x * params::degree / 2],
+      &out_re_lo[blockIdx.x * params::degree / 2],
+      &out_im_hi[blockIdx.x * params::degree / 2],
+      &out_im_lo[blockIdx.x * params::degree / 2],
+      &in[blockIdx.x * params::degree],
+      &in[blockIdx.x * params::degree + params::degree / 2]);
+}
+
+// params is expected to be full degree not half degree
+// converts standqard input into complex<128> represented by 4 double
+// with following pattern: [re_hi_0, re_hi_1, ... re_hi_n, re_lo_0, re_lo_1,
+// ... re_lo_n, im_hi_0, im_hi_1, ..., im_hi_n,  im_lo_0, im_lo_1, ..., im_lo_n]
+template <class params>
+__global__ void
+batch_convert_u128_to_f128_as_torus(double *out_re_hi, double *out_re_lo,
+                                    double *out_im_hi, double *out_im_lo,
+                                    const __uint128_t *in) {
+
+  convert_u128_to_f128_as_torus<params>(
+      &out_re_hi[blockIdx.x * params::degree / 2],
+      &out_re_lo[blockIdx.x * params::degree / 2],
+      &out_im_hi[blockIdx.x * params::degree / 2],
+      &out_im_lo[blockIdx.x * params::degree / 2],
+      &in[blockIdx.x * params::degree],
+      &in[blockIdx.x * params::degree + params::degree / 2]);
+}
+
+// params is expected to be full degree not half degree
+// converts standqard input into complex<128> represented by 4 double
+// with following pattern: [re_hi_0, re_lo_0, im_hi_0, im_lo_0, re_hi_1,
+// re_lo_1, im_hi_1, im_lo_1,
+// ...,re_hi_n, re_lo_n, im_hi_n, im_lo_n, ]
+template <class params>
+__global__ void
+batch_convert_u128_to_f128_strided_as_torus(double *d_out,
+                                            const __uint128_t *d_in) {
+
+  constexpr size_t chunk_size = params::degree / 2 * 4;
+  double *chunk = &d_out[blockIdx.x * chunk_size];
+  double *out_re_hi = &chunk[0 * params::degree / 2];
+  double *out_re_lo = &chunk[1 * params::degree / 2];
+  double *out_im_hi = &chunk[2 * params::degree / 2];
+  double *out_im_lo = &chunk[3 * params::degree / 2];
+
+  convert_u128_to_f128_as_torus<params>(
+      out_re_hi, out_re_lo, out_im_hi, out_im_lo,
+      &d_in[blockIdx.x * params::degree],
+      &d_in[blockIdx.x * params::degree + params::degree / 2]);
+}
+
+// params is expected to be full degree not half degree
+template <class params>
+__global__ void batch_convert_f128_to_u128_as_torus(__uint128_t *out,
+                                                    const double *in_re_hi,
+                                                    const double *in_re_lo,
+                                                    const double *in_im_hi,
+                                                    const double *in_im_lo) {
+
+  convert_f128_to_u128_as_torus<params>(
+      &out[blockIdx.x * params::degree],
+      &out[blockIdx.x * params::degree + params::degree / 2],
+      &in_re_hi[blockIdx.x * params::degree / 2],
+      &in_re_lo[blockIdx.x * params::degree / 2],
+      &in_im_hi[blockIdx.x * params::degree / 2],
+      &in_im_lo[blockIdx.x * params::degree / 2]);
+}
+
+template <class params, sharedMemDegree SMD>
+__global__ void
+batch_NSMFFT_128(double *in_re_hi, double *in_re_lo, double *in_im_hi,
+                 double *in_im_lo, double *out_re_hi, double *out_re_lo,
+                 double *out_im_hi, double *out_im_lo, double *buffer) {
+  extern __shared__ double sharedMemoryFFT128[];
+  double *re_hi, *re_lo, *im_hi, *im_lo;
+
+  if (SMD == NOSM) {
+    re_hi =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 0];
+    re_lo =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 1];
+    im_hi =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 2];
+    im_lo =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 3];
+  } else {
+    re_hi = &sharedMemoryFFT128[params::degree / 2 * 0];
+    re_lo = &sharedMemoryFFT128[params::degree / 2 * 1];
+    im_hi = &sharedMemoryFFT128[params::degree / 2 * 2];
+    im_lo = &sharedMemoryFFT128[params::degree / 2 * 3];
+  }
+
+  Index tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < params::opt / 2; ++i) {
+    re_hi[tid] = in_re_hi[blockIdx.x * (params::degree / 2) + tid];
+    re_lo[tid] = in_re_lo[blockIdx.x * (params::degree / 2) + tid];
+    im_hi[tid] = in_im_hi[blockIdx.x * (params::degree / 2) + tid];
+    im_lo[tid] = in_im_lo[blockIdx.x * (params::degree / 2) + tid];
+    tid += params::degree / params::opt;
+  }
+  __syncthreads();
+  if constexpr (params::fft_direction == 1) {
+    negacyclic_backward_fft_f128<HalfDegree<params>>(re_hi, re_lo, im_hi,
+                                                     im_lo);
+  } else {
+    negacyclic_forward_fft_f128<HalfDegree<params>>(re_hi, re_lo, im_hi, im_lo);
+  }
+  __syncthreads();
+  tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < params::opt / 2; ++i) {
+    out_re_hi[blockIdx.x * (params::degree / 2) + tid] = re_hi[tid];
+    out_re_lo[blockIdx.x * (params::degree / 2) + tid] = re_lo[tid];
+    out_im_hi[blockIdx.x * (params::degree / 2) + tid] = im_hi[tid];
+    out_im_lo[blockIdx.x * (params::degree / 2) + tid] = im_lo[tid];
+    tid += params::degree / params::opt;
+  }
+}
+
+template <class params, sharedMemDegree SMD>
+__global__ void batch_NSMFFT_strided_128(double *d_in, double *d_out,
+                                         double *buffer) {
+  extern __shared__ double sharedMemoryFFT128[];
+  double *re_hi, *re_lo, *im_hi, *im_lo;
+
+  if (SMD == NOSM) {
+    re_hi =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 0];
+    re_lo =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 1];
+    im_hi =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 2];
+    im_lo =
+        &buffer[blockIdx.x * params::degree / 2 * 4 + params::degree / 2 * 3];
+  } else {
+    re_hi = &sharedMemoryFFT128[params::degree / 2 * 0];
+    re_lo = &sharedMemoryFFT128[params::degree / 2 * 1];
+    im_hi = &sharedMemoryFFT128[params::degree / 2 * 2];
+    im_lo = &sharedMemoryFFT128[params::degree / 2 * 3];
+  }
+
+  constexpr size_t chunk_size = params::degree / 2 * 4;
+  double *chunk = &d_in[blockIdx.x * chunk_size];
+  double *tmp_re_hi = &chunk[0 * params::degree / 2];
+  double *tmp_re_lo = &chunk[1 * params::degree / 2];
+  double *tmp_im_hi = &chunk[2 * params::degree / 2];
+  double *tmp_im_lo = &chunk[3 * params::degree / 2];
+
+  Index tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < params::opt / 2; ++i) {
+    re_hi[tid] = tmp_re_hi[tid];
+    re_lo[tid] = tmp_re_lo[tid];
+    im_hi[tid] = tmp_im_hi[tid];
+    im_lo[tid] = tmp_im_lo[tid];
+    tid += params::degree / params::opt;
+  }
+  __syncthreads();
+  if constexpr (params::fft_direction == 1) {
+    negacyclic_backward_fft_f128<HalfDegree<params>>(re_hi, re_lo, im_hi,
+                                                     im_lo);
+  } else {
+    negacyclic_forward_fft_f128<HalfDegree<params>>(re_hi, re_lo, im_hi, im_lo);
+  }
+  __syncthreads();
+
+  chunk = &d_out[blockIdx.x * chunk_size];
+  tmp_re_hi = &chunk[0 * params::degree / 2];
+  tmp_re_lo = &chunk[1 * params::degree / 2];
+  tmp_im_hi = &chunk[2 * params::degree / 2];
+  tmp_im_lo = &chunk[3 * params::degree / 2];
+
+  tid = threadIdx.x;
+#pragma unroll
+  for (Index i = 0; i < params::opt / 2; ++i) {
+    tmp_re_hi[tid] = re_hi[tid];
+    tmp_re_lo[tid] = re_lo[tid];
+    tmp_im_hi[tid] = im_hi[tid];
+    tmp_im_lo[tid] = im_lo[tid];
+    tid += params::degree / params::opt;
+  }
+}
+
+template <class params>
+__host__ void host_fourier_transform_forward_as_integer_f128(
+    cudaStream_t stream, uint32_t gpu_index, double *re0, double *re1,
+    double *im0, double *im1, const __uint128_t *standard, const uint32_t N,
+    const uint32_t number_of_samples) {
+
+  // allocate device buffers
+  double *d_re0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_re1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  __uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
+      N * sizeof(__uint128_t), stream, gpu_index);
+
+  // copy input into device
+  cuda_memcpy_async_to_gpu(d_standard, standard, N * sizeof(__uint128_t),
+                           stream, gpu_index);
+
+  // setup launch parameters
+  size_t required_shared_memory_size = sizeof(double) * N / 2 * 4;
+  int grid_size = number_of_samples;
+  int block_size = params::degree / params::opt;
+  bool full_sm =
+      (required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
+  size_t buffer_size = full_sm ? 0 : (size_t)number_of_samples * N / 2 * 4;
+  size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
+  double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
+
+  // configure shared memory for batch fft kernel
+  if (full_sm) {
+    check_cuda_error(cudaFuncSetAttribute(
+        batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>,
+        cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
+    check_cuda_error(cudaFuncSetCacheConfig(
+        batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>,
+        cudaFuncCachePreferShared));
+  }
+
+  // convert u128 into 4 x double
+  batch_convert_u128_to_f128_as_integer<params>
+      <<<grid_size, block_size, 0, stream>>>(d_re0, d_re1, d_im0, d_im1,
+                                             d_standard);
+
+  // call negacyclic 128 bit forward fft.
+  if (full_sm) {
+    batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  } else {
+    batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, NOSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  }
+
+  cuda_memcpy_async_to_cpu(re0, d_re0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(re1, d_re1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(im0, d_im0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(im1, d_im1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+
+  cuda_drop_async(d_standard, stream, gpu_index);
+  cuda_drop_async(d_re0, stream, gpu_index);
+  cuda_drop_async(d_re1, stream, gpu_index);
+  cuda_drop_async(d_im0, stream, gpu_index);
+  cuda_drop_async(d_im1, stream, gpu_index);
+}
+
+template <class params>
+__host__ void host_fourier_transform_forward_as_torus_f128(
+    cudaStream_t stream, uint32_t gpu_index, double *re0, double *re1,
+    double *im0, double *im1, const __uint128_t *standard, const uint32_t N,
+    const uint32_t number_of_samples) {
+
+  // allocate device buffers
+  double *d_re0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_re1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  __uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
+      N * sizeof(__uint128_t), stream, gpu_index);
+
+  // copy input into device
+  cuda_memcpy_async_to_gpu(d_standard, standard, N * sizeof(__uint128_t),
+                           stream, gpu_index);
+
+  // setup launch parameters
+  size_t required_shared_memory_size = sizeof(double) * N / 2 * 4;
+  int grid_size = number_of_samples;
+  int block_size = params::degree / params::opt;
+  bool full_sm =
+      (required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
+  size_t buffer_size = full_sm ? 0 : (size_t)number_of_samples * N / 2 * 4;
+  size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
+  double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
+
+  // configure shared memory for batch fft kernel
+  if (full_sm) {
+    check_cuda_error(cudaFuncSetAttribute(
+        batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>,
+        cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
+    check_cuda_error(cudaFuncSetCacheConfig(
+        batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>,
+        cudaFuncCachePreferShared));
+  }
+
+  // convert u128 into 4 x double
+  batch_convert_u128_to_f128_as_torus<params>
+      <<<grid_size, block_size, 0, stream>>>(d_re0, d_re1, d_im0, d_im1,
+                                             d_standard);
+
+  // call negacyclic 128 bit forward fft.
+  if (full_sm) {
+    batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, FULLSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  } else {
+    batch_NSMFFT_128<FFTDegree<params, ForwardFFT>, NOSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  }
+
+  cuda_memcpy_async_to_cpu(re0, d_re0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(re1, d_re1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(im0, d_im0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_cpu(im1, d_im1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+
+  cuda_drop_async(d_standard, stream, gpu_index);
+  cuda_drop_async(d_re0, stream, gpu_index);
+  cuda_drop_async(d_re1, stream, gpu_index);
+  cuda_drop_async(d_im0, stream, gpu_index);
+  cuda_drop_async(d_im1, stream, gpu_index);
+}
+
+template <class params>
+__host__ void host_fourier_transform_backward_as_torus_f128(
+    cudaStream_t stream, uint32_t gpu_index, __uint128_t *standard,
+    double const *re0, double const *re1, double const *im0, double const *im1,
+    const uint32_t N, const uint32_t number_of_samples) {
+
+  // allocate device buffers
+  double *d_re0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_re1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im0 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  double *d_im1 =
+      (double *)cuda_malloc_async(N / 2 * sizeof(double), stream, gpu_index);
+  __uint128_t *d_standard = (__uint128_t *)cuda_malloc_async(
+      N * sizeof(__uint128_t), stream, gpu_index);
+
+  //  // copy input into device
+  cuda_memcpy_async_to_gpu(d_re0, re0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_gpu(d_re1, re1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_gpu(d_im0, im0, N / 2 * sizeof(double), stream,
+                           gpu_index);
+  cuda_memcpy_async_to_gpu(d_im1, im1, N / 2 * sizeof(double), stream,
+                           gpu_index);
+
+  // setup launch parameters
+  size_t required_shared_memory_size = sizeof(double) * N / 2 * 4;
+  int grid_size = number_of_samples;
+  int block_size = params::degree / params::opt;
+  bool full_sm =
+      (required_shared_memory_size <= cuda_get_max_shared_memory(gpu_index));
+  size_t buffer_size = full_sm ? 0 : (size_t)number_of_samples * N / 2 * 4;
+  size_t shared_memory_size = full_sm ? required_shared_memory_size : 0;
+  double *buffer = (double *)cuda_malloc_async(buffer_size, stream, gpu_index);
+
+  // configure shared memory for batch fft kernel
+  if (full_sm) {
+    check_cuda_error(cudaFuncSetAttribute(
+        batch_NSMFFT_128<FFTDegree<params, BackwardFFT>, FULLSM>,
+        cudaFuncAttributeMaxDynamicSharedMemorySize, shared_memory_size));
+    check_cuda_error(cudaFuncSetCacheConfig(
+        batch_NSMFFT_128<FFTDegree<params, BackwardFFT>, FULLSM>,
+        cudaFuncCachePreferShared));
+    batch_NSMFFT_128<FFTDegree<params, BackwardFFT>, FULLSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  } else {
+    batch_NSMFFT_128<FFTDegree<params, BackwardFFT>, NOSM>
+        <<<grid_size, block_size, shared_memory_size, stream>>>(
+            d_re0, d_re1, d_im0, d_im1, d_re0, d_re1, d_im0, d_im1, buffer);
+  }
+
+  batch_convert_f128_to_u128_as_torus<params>
+      <<<grid_size, block_size, 0, stream>>>(d_standard, d_re0, d_re1, d_im0,
+                                             d_im1);
+
+  cuda_memcpy_async_to_cpu(standard, d_standard, N * sizeof(__uint128_t),
+                           stream, gpu_index);
+  cuda_drop_async(d_standard, stream, gpu_index);
+  cuda_drop_async(d_re0, stream, gpu_index);
+  cuda_drop_async(d_re1, stream, gpu_index);
+  cuda_drop_async(d_im0, stream, gpu_index);
+  cuda_drop_async(d_im1, stream, gpu_index);
+}
+
+#undef NEG_TWID
+#undef F64x4_TO_F128x2
+#undef F128x2_TO_F64x4
+
+#endif // TFHE_RS_BACKENDS_TFHE_CUDA_BACKEND_CUDA_SRC_FFT128_FFT128_CUH_

backends/tfhe-cuda-backend/cuda/src/fft128/twiddles.cuh

@@ -0,0 +1,11 @@
+#ifndef CUDA_FFT128_TWIDDLES_CUH
+#define CUDA_FFT128_TWIDDLES_CUH
+
+/*
+ * 'negtwiddles' are stored in device memory to profit caching
+ */
+extern __device__ double neg_twiddles_re_hi[4096];
+extern __device__ double neg_twiddles_re_lo[4096];
+extern __device__ double neg_twiddles_im_hi[4096];
+extern __device__ double neg_twiddles_im_lo[4096];
+#endif

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

@@ -28,44 +28,6 @@ __host__ void scratch_cuda_integer_abs_kb(
                                   num_blocks, allocate_gpu_memory);
 }
 
-template <typename Torus>
-__host__ void legacy_host_integer_abs_kb_async(
-    cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *ct, void *const *bsks, uint64_t *const *ksks,
-    int_abs_buffer<uint64_t> *mem_ptr, bool is_signed, uint32_t num_blocks) {
-  if (!is_signed)
-    return;
-
-  auto radix_params = mem_ptr->params;
-  auto mask = (Torus *)(mem_ptr->mask->ptr);
-
-  auto big_lwe_dimension = radix_params.big_lwe_dimension;
-  auto big_lwe_size = big_lwe_dimension + 1;
-  auto big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
-  uint32_t num_bits_in_ciphertext =
-      (31 - __builtin_clz(radix_params.message_modulus)) * num_blocks;
-
-  cuda_memcpy_async_gpu_to_gpu(mask, ct, num_blocks * big_lwe_size_bytes,
-                               streams[0], gpu_indexes[0]);
-
-  legacy_host_integer_radix_arithmetic_scalar_shift_kb_inplace<Torus>(
-      streams, gpu_indexes, gpu_count, mask, num_bits_in_ciphertext - 1,
-      mem_ptr->arithmetic_scalar_shift_mem, bsks, ksks, num_blocks);
-  legacy_host_addition<Torus>(streams[0], gpu_indexes[0], ct, mask, ct,
-                              radix_params.big_lwe_dimension, num_blocks);
-
-  uint32_t requested_flag = outputFlag::FLAG_NONE;
-  uint32_t uses_carry = 0;
-  legacy_host_propagate_single_carry<Torus>(
-      streams, gpu_indexes, gpu_count, ct, nullptr, nullptr, mem_ptr->scp_mem,
-      bsks, ksks, num_blocks, requested_flag, uses_carry);
-
-  // legacy bitop
-  legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, ct, mask, ct, bsks, ksks, num_blocks,
-      mem_ptr->bitxor_mem->lut, mem_ptr->bitxor_mem->params.message_modulus);
-}
-
 template <typename Torus>
 __host__ void
 host_integer_abs_kb(cudaStream_t const *streams, uint32_t const *gpu_indexes,

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

@@ -19,6 +19,12 @@ __host__ void host_integer_radix_bitop_kb(
     CudaRadixCiphertextFFI const *lwe_array_2, int_bitop_buffer<Torus> *mem_ptr,
     void *const *bsks, Torus *const *ksks) {
 
+  if (lwe_array_out->num_radix_blocks != lwe_array_1->num_radix_blocks ||
+      lwe_array_out->num_radix_blocks != lwe_array_2->num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks must be equal")
+  if (lwe_array_out->lwe_dimension != lwe_array_1->lwe_dimension ||
+      lwe_array_out->lwe_dimension != lwe_array_2->lwe_dimension)
+    PANIC("Cuda error: input and output lwe dimension must be equal")
   auto lut = mem_ptr->lut;
   uint64_t degrees[lwe_array_1->num_radix_blocks];
   if (mem_ptr->op == BITOP_TYPE::BITAND) {

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

@@ -7,11 +7,20 @@
 template <typename Torus>
 __host__ void zero_out_if(cudaStream_t const *streams,
                           uint32_t const *gpu_indexes, uint32_t gpu_count,
-                          Torus *lwe_array_out, Torus const *lwe_array_input,
-                          Torus const *lwe_condition,
+                          CudaRadixCiphertextFFI *lwe_array_out,
+                          CudaRadixCiphertextFFI const *lwe_array_input,
+                          CudaRadixCiphertextFFI const *lwe_condition,
                           int_zero_out_if_buffer<Torus> *mem_ptr,
                           int_radix_lut<Torus> *predicate, void *const *bsks,
                           Torus *const *ksks, uint32_t num_radix_blocks) {
+  if (lwe_array_out->num_radix_blocks < num_radix_blocks ||
+      lwe_array_input->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: input or output radix ciphertexts does not have enough "
+          "blocks")
+  if (lwe_array_out->lwe_dimension != lwe_array_input->lwe_dimension ||
+      lwe_array_input->lwe_dimension != lwe_condition->lwe_dimension)
+    PANIC("Cuda error: input and output radix ciphertexts must have the same "
+          "lwe dimension")
   cuda_set_device(gpu_indexes[0]);
   auto params = mem_ptr->params;
 
@@ -21,56 +30,11 @@ __host__ void zero_out_if(cudaStream_t const *streams,
   host_pack_bivariate_blocks_with_single_block<Torus>(
       streams, gpu_indexes, gpu_count, tmp_lwe_array_input,
       predicate->lwe_indexes_in, lwe_array_input, lwe_condition,
-      predicate->lwe_indexes_in, params.big_lwe_dimension,
-      params.message_modulus, num_radix_blocks);
+      predicate->lwe_indexes_in, params.message_modulus, num_radix_blocks);
 
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
       streams, gpu_indexes, gpu_count, lwe_array_out, tmp_lwe_array_input, bsks,
-      ksks, num_radix_blocks, predicate);
-}
-
-template <typename Torus>
-__host__ void legacy_host_integer_radix_cmux_kb(
-    cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_condition,
-    Torus const *lwe_array_true, Torus const *lwe_array_false,
-    int_cmux_buffer<Torus> *mem_ptr, void *const *bsks, Torus *const *ksks,
-    uint32_t num_radix_blocks) {
-
-  auto params = mem_ptr->params;
-  Torus lwe_size = params.big_lwe_dimension + 1;
-  Torus radix_lwe_size = lwe_size * num_radix_blocks;
-  cuda_memcpy_async_gpu_to_gpu(mem_ptr->buffer_in->ptr, lwe_array_true,
-                               radix_lwe_size * sizeof(Torus), streams[0],
-                               gpu_indexes[0]);
-  cuda_memcpy_async_gpu_to_gpu(
-      (Torus *)(mem_ptr->buffer_in->ptr) + radix_lwe_size, lwe_array_false,
-      radix_lwe_size * sizeof(Torus), streams[0], gpu_indexes[0]);
-  for (uint i = 0; i < 2 * num_radix_blocks; i++) {
-    cuda_memcpy_async_gpu_to_gpu(
-        (Torus *)(mem_ptr->condition_array->ptr) + i * lwe_size, lwe_condition,
-        lwe_size * sizeof(Torus), streams[0], gpu_indexes[0]);
-  }
-  legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, (Torus *)(mem_ptr->buffer_out->ptr),
-      (Torus *)(mem_ptr->buffer_in->ptr),
-      (Torus *)(mem_ptr->condition_array->ptr), bsks, ksks,
-      2 * num_radix_blocks, mem_ptr->predicate_lut, params.message_modulus);
-
-  // 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
-  // have kept its value
-  auto mem_true = (Torus *)(mem_ptr->buffer_out->ptr);
-  auto ptr = (Torus *)mem_ptr->buffer_out->ptr;
-  auto mem_false = &ptr[radix_lwe_size];
-  auto added_cts = mem_true;
-  legacy_host_addition<Torus>(streams[0], gpu_indexes[0], added_cts, mem_true,
-                              mem_false, params.big_lwe_dimension,
-                              num_radix_blocks);
-
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, added_cts, bsks, ksks,
-      num_radix_blocks, mem_ptr->message_extract_lut);
+      ksks, predicate, num_radix_blocks);
 }
 
 template <typename Torus>

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

@@ -38,21 +38,26 @@ void scratch_cuda_integer_radix_comparison_kb_64(
 
 void cuda_comparison_integer_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_1, void const *lwe_array_2,
-    int8_t *mem_ptr, void *const *bsks, void *const *ksks,
-    uint32_t num_radix_blocks) {
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_1,
+    CudaRadixCiphertextFFI const *lwe_array_2, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks) {
 
+  if (lwe_array_1->num_radix_blocks != lwe_array_1->num_radix_blocks)
+    PANIC("Cuda error: input num radix blocks must be the same")
+  // The output ciphertext might be a boolean block or a radix ciphertext
+  // depending on the case (eq/gt vs max/min) so the amount of blocks to
+  // consider for calculation is the one of the input
+  auto num_radix_blocks = lwe_array_1->num_radix_blocks;
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;
   switch (buffer->op) {
   case EQ:
   case NE:
     host_integer_radix_equality_check_kb<uint64_t>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(lwe_array_out),
-        static_cast<const uint64_t *>(lwe_array_1),
-        static_cast<const uint64_t *>(lwe_array_2), buffer, bsks,
-        (uint64_t **)(ksks), num_radix_blocks);
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, lwe_array_out,
+        lwe_array_1, lwe_array_2, buffer, bsks, (uint64_t **)(ksks),
+        num_radix_blocks);
     break;
   case GT:
   case GE:
@@ -62,23 +67,18 @@ void cuda_comparison_integer_radix_ciphertext_kb_64(
       PANIC("Cuda error (comparisons): the number of radix blocks has to be "
             "even.")
     host_integer_radix_difference_check_kb<uint64_t>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(lwe_array_out),
-        static_cast<const uint64_t *>(lwe_array_1),
-        static_cast<const uint64_t *>(lwe_array_2), buffer,
-        buffer->diff_buffer->operator_f, bsks, (uint64_t **)(ksks),
-        num_radix_blocks);
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, lwe_array_out,
+        lwe_array_1, lwe_array_2, buffer, buffer->diff_buffer->operator_f, bsks,
+        (uint64_t **)(ksks), num_radix_blocks);
     break;
   case MAX:
   case MIN:
     if (num_radix_blocks % 2 != 0)
       PANIC("Cuda error (max/min): the number of radix blocks has to be even.")
     host_integer_radix_maxmin_kb<uint64_t>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(lwe_array_out),
-        static_cast<const uint64_t *>(lwe_array_1),
-        static_cast<const uint64_t *>(lwe_array_2), buffer, bsks,
-        (uint64_t **)(ksks), num_radix_blocks);
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, lwe_array_out,
+        lwe_array_1, lwe_array_2, buffer, bsks, (uint64_t **)(ksks),
+        num_radix_blocks);
     break;
   default:
     PANIC("Cuda error: integer operation not supported")
@@ -117,17 +117,16 @@ void scratch_cuda_integer_are_all_comparisons_block_true_kb_64(
 
 void cuda_integer_are_all_comparisons_block_true_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_in, int8_t *mem_ptr,
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks) {
 
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;
 
   host_integer_are_all_comparisons_block_true_kb<uint64_t>(
-      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-      static_cast<uint64_t *>(lwe_array_out),
-      static_cast<const uint64_t *>(lwe_array_in), buffer, bsks,
-      (uint64_t **)(ksks), num_radix_blocks);
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count, lwe_array_out,
+      lwe_array_in, buffer, bsks, (uint64_t **)(ksks), num_radix_blocks);
 }
 
 void cleanup_cuda_integer_are_all_comparisons_block_true(
@@ -161,17 +160,16 @@ void scratch_cuda_integer_is_at_least_one_comparisons_block_true_kb_64(
 
 void cuda_integer_is_at_least_one_comparisons_block_true_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *lwe_array_out, void const *lwe_array_in, int8_t *mem_ptr,
+    CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in, int8_t *mem_ptr,
     void *const *bsks, void *const *ksks, uint32_t num_radix_blocks) {
 
   int_comparison_buffer<uint64_t> *buffer =
       (int_comparison_buffer<uint64_t> *)mem_ptr;
 
   host_integer_is_at_least_one_comparisons_block_true_kb<uint64_t>(
-      (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-      static_cast<uint64_t *>(lwe_array_out),
-      static_cast<const uint64_t *>(lwe_array_in), buffer, bsks,
-      (uint64_t **)(ksks), num_radix_blocks);
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count, lwe_array_out,
+      lwe_array_in, buffer, bsks, (uint64_t **)(ksks), num_radix_blocks);
 }
 
 void cleanup_cuda_integer_is_at_least_one_comparisons_block_true(

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

@@ -8,6 +8,7 @@
 #include "integer/integer_utilities.h"
 #include "integer/negation.cuh"
 #include "integer/scalar_addition.cuh"
+#include "integer/subtraction.cuh"
 #include "pbs/programmable_bootstrap_classic.cuh"
 #include "pbs/programmable_bootstrap_multibit.cuh"
 #include "types/complex/operations.cuh"
@@ -58,10 +59,17 @@ __host__ void accumulate_all_blocks(cudaStream_t stream, uint32_t gpu_index,
 template <typename Torus>
 __host__ void are_all_comparisons_block_true(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_in,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in,
     int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
     Torus *const *ksks, uint32_t num_radix_blocks) {
 
+  if (lwe_array_out->lwe_dimension != lwe_array_in->lwe_dimension)
+    PANIC("Cuda error: input and output lwe dimensions must be the same")
+  if (lwe_array_in->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: input num radix blocks should not be lower "
+          "than the number of blocks to operate on")
+
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto glwe_dimension = params.glwe_dimension;
@@ -76,10 +84,9 @@ __host__ void are_all_comparisons_block_true(
   uint32_t total_modulus = message_modulus * carry_modulus;
   uint32_t max_value = (total_modulus - 1) / (message_modulus - 1);
 
-  cuda_memcpy_async_gpu_to_gpu(tmp_out, lwe_array_in,
-                               num_radix_blocks * (big_lwe_dimension + 1) *
-                                   sizeof(Torus),
-                               streams[0], gpu_indexes[0]);
+  copy_radix_ciphertext_slice_async<Torus>(streams[0], gpu_indexes[0], tmp_out,
+                                           0, num_radix_blocks, lwe_array_in, 0,
+                                           num_radix_blocks);
 
   uint32_t remaining_blocks = num_radix_blocks;
 
@@ -89,8 +96,9 @@ __host__ void are_all_comparisons_block_true(
 
     // Since all blocks encrypt either 0 or 1, we can sum max_value of them
     // 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 input_blocks = (Torus *)tmp_out->ptr;
+    auto accumulator_ptr =
+        (Torus *)are_all_block_true_buffer->tmp_block_accumulated->ptr;
     auto is_max_value_lut = are_all_block_true_buffer->is_max_value;
     uint32_t chunk_lengths[num_chunks];
     auto begin_remaining_blocks = remaining_blocks;
@@ -98,15 +106,15 @@ __host__ void are_all_comparisons_block_true(
       uint32_t chunk_length =
           std::min(max_value, begin_remaining_blocks - i * max_value);
       chunk_lengths[i] = chunk_length;
-      accumulate_all_blocks<Torus>(streams[0], gpu_indexes[0], accumulator,
+      accumulate_all_blocks<Torus>(streams[0], gpu_indexes[0], accumulator_ptr,
                                    input_blocks, big_lwe_dimension,
                                    chunk_length);
 
-      accumulator += (big_lwe_dimension + 1);
+      accumulator_ptr += (big_lwe_dimension + 1);
       remaining_blocks -= (chunk_length - 1);
       input_blocks += (big_lwe_dimension + 1) * chunk_length;
     }
-    accumulator = are_all_block_true_buffer->tmp_block_accumulated;
+    auto accumulator = are_all_block_true_buffer->tmp_block_accumulated;
 
     // Selects a LUT
     int_radix_lut<Torus> *lut;
@@ -127,7 +135,7 @@ __host__ void are_all_comparisons_block_true(
             polynomial_size, message_modulus, carry_modulus,
             is_equal_to_num_blocks_lut_f);
 
-        Torus *h_lut_indexes = (Torus *)malloc(num_chunks * sizeof(Torus));
+        Torus *h_lut_indexes = is_max_value_lut->h_lut_indexes;
         for (int index = 0; index < num_chunks; index++) {
           if (index == num_chunks - 1) {
             h_lut_indexes[index] = 1;
@@ -139,8 +147,6 @@ __host__ void are_all_comparisons_block_true(
                                  h_lut_indexes, num_chunks * sizeof(Torus),
                                  streams[0], gpu_indexes[0]);
         is_max_value_lut->broadcast_lut(streams, gpu_indexes, 0);
-        cuda_synchronize_stream(streams[0], gpu_indexes[0]);
-        free(h_lut_indexes);
       }
       lut = is_max_value_lut;
     }
@@ -148,14 +154,24 @@ __host__ void are_all_comparisons_block_true(
     // Applies the LUT
     if (remaining_blocks == 1) {
       // In the last iteration we copy the output to the final address
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
           streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsks,
-          ksks, 1, lut);
+          ksks, lut, 1);
+      // Reset max_value_lut_indexes before returning, otherwise if the lut is
+      // reused the lut indexes will be wrong
+      memset(is_max_value_lut->h_lut_indexes, 0,
+             is_max_value_lut->num_blocks * sizeof(Torus));
+      cuda_memcpy_async_to_gpu(is_max_value_lut->get_lut_indexes(0, 0),
+                               is_max_value_lut->h_lut_indexes,
+                               is_max_value_lut->num_blocks * sizeof(Torus),
+                               streams[0], gpu_indexes[0]);
+      is_max_value_lut->broadcast_lut(streams, gpu_indexes, 0);
+      reset_radix_ciphertext_blocks(lwe_array_out, 1);
       return;
     } else {
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
           streams, gpu_indexes, gpu_count, tmp_out, accumulator, bsks, ksks,
-          num_chunks, lut);
+          lut, num_chunks);
     }
   }
 }
@@ -169,10 +185,17 @@ __host__ void are_all_comparisons_block_true(
 template <typename Torus>
 __host__ void is_at_least_one_comparisons_block_true(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_in,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in,
     int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
     Torus *const *ksks, uint32_t num_radix_blocks) {
 
+  if (lwe_array_out->lwe_dimension != lwe_array_in->lwe_dimension)
+    PANIC("Cuda error: input lwe dimensions must be the same")
+
+  if (lwe_array_in->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: input num radix blocks should not be lower "
+          "than the number of blocks to operate on")
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -183,10 +206,9 @@ __host__ void is_at_least_one_comparisons_block_true(
   uint32_t total_modulus = message_modulus * carry_modulus;
   uint32_t max_value = (total_modulus - 1) / (message_modulus - 1);
 
-  cuda_memcpy_async_gpu_to_gpu(mem_ptr->tmp_lwe_array_out, lwe_array_in,
-                               num_radix_blocks * (big_lwe_dimension + 1) *
-                                   sizeof(Torus),
-                               streams[0], gpu_indexes[0]);
+  copy_radix_ciphertext_slice_async<Torus>(
+      streams[0], gpu_indexes[0], mem_ptr->tmp_lwe_array_out, 0,
+      num_radix_blocks, lwe_array_in, 0, num_radix_blocks);
 
   uint32_t remaining_blocks = num_radix_blocks;
   while (remaining_blocks > 0) {
@@ -195,8 +217,8 @@ __host__ void is_at_least_one_comparisons_block_true(
 
     // Since all blocks encrypt either 0 or 1, we can sum max_value of them
     // as in the worst case we will be adding `max_value` ones
-    auto input_blocks = mem_ptr->tmp_lwe_array_out;
-    auto accumulator = buffer->tmp_block_accumulated;
+    auto input_blocks = (Torus *)mem_ptr->tmp_lwe_array_out->ptr;
+    auto accumulator = (Torus *)buffer->tmp_block_accumulated->ptr;
     uint32_t chunk_lengths[num_chunks];
     auto begin_remaining_blocks = remaining_blocks;
     for (int i = 0; i < num_chunks; i++) {
@@ -211,7 +233,6 @@ __host__ void is_at_least_one_comparisons_block_true(
       remaining_blocks -= (chunk_length - 1);
       input_blocks += (big_lwe_dimension + 1) * chunk_length;
     }
-    accumulator = buffer->tmp_block_accumulated;
 
     // Selects a LUT
     int_radix_lut<Torus> *lut = mem_ptr->eq_buffer->is_non_zero_lut;
@@ -219,45 +240,35 @@ __host__ void is_at_least_one_comparisons_block_true(
     // Applies the LUT
     if (remaining_blocks == 1) {
       // In the last iteration we copy the output to the final address
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, lwe_array_out, accumulator, bsks,
-          ksks, 1, lut);
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, lwe_array_out,
+          buffer->tmp_block_accumulated, bsks, ksks, lut, 1);
       return;
     } else {
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
           streams, gpu_indexes, gpu_count, mem_ptr->tmp_lwe_array_out,
-          accumulator, bsks, ksks, num_chunks, lut);
+          buffer->tmp_block_accumulated, bsks, ksks, lut, num_chunks);
     }
   }
 }
 
-// This takes an input slice of blocks.
-//
-// Each block can encrypt any value as long as its < message_modulus.
-//
-// It will compare blocks with 0, for either equality or difference.
-//
-// This returns a Vec of block, where each block encrypts 1 or 0
-// depending of if all blocks matched with the comparison type with 0.
-//
-// E.g. For ZeroComparisonType::Equality, if all input blocks are zero
-// than all returned block will encrypt 1
-//
-// The returned Vec will have less block than the number of input blocks.
-// The returned blocks potentially needs to be 'reduced' to one block
-// with eg are_all_comparisons_block_true.
-//
-// This function exists because sometimes it is faster to concatenate
-// multiple vec of 'boolean' shortint block before reducing them with
-// are_all_comparisons_block_true
 template <typename Torus>
-__host__ void host_compare_with_zero_equality(
+__host__ void host_compare_blocks_with_zero(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_in,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in,
     int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
     Torus *const *ksks, int32_t num_radix_blocks,
     int_radix_lut<Torus> *zero_comparison) {
 
+  if (num_radix_blocks == 0)
+    return;
+  if (lwe_array_out->lwe_dimension != lwe_array_in->lwe_dimension)
+    PANIC("Cuda error: input lwe dimensions must be the same")
+  if (lwe_array_in->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: input num radix blocks should not be lower "
+          "than the number of blocks to operate on")
+
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -274,21 +285,19 @@ __host__ void host_compare_with_zero_equality(
   uint32_t num_elements_to_fill_carry = (total_modulus - 1) / message_max;
 
   size_t big_lwe_size = big_lwe_dimension + 1;
-  size_t big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
-
   int num_sum_blocks = 0;
   // Accumulator
   auto sum = lwe_array_out;
 
   if (num_radix_blocks == 1) {
     // Just copy
-    cuda_memcpy_async_gpu_to_gpu(sum, lwe_array_in, big_lwe_size_bytes,
-                                 streams[0], gpu_indexes[0]);
+    copy_radix_ciphertext_slice_async<Torus>(streams[0], gpu_indexes[0], sum, 0,
+                                             1, lwe_array_in, 0, 1);
     num_sum_blocks = 1;
   } else {
     uint32_t remainder_blocks = num_radix_blocks;
-    auto sum_i = sum;
-    auto chunk = lwe_array_in;
+    auto sum_i = (Torus *)sum->ptr;
+    auto chunk = (Torus *)lwe_array_in->ptr;
     while (remainder_blocks > 1) {
       uint32_t chunk_size =
           std::min(remainder_blocks, num_elements_to_fill_carry);
@@ -305,28 +314,31 @@ __host__ void host_compare_with_zero_equality(
     }
   }
 
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, sum, sum, bsks, ksks, num_sum_blocks,
-      zero_comparison);
-  are_all_comparisons_block_true<Torus>(streams, gpu_indexes, gpu_count,
-                                        lwe_array_out, sum, mem_ptr, bsks, ksks,
-                                        num_sum_blocks);
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      streams, gpu_indexes, gpu_count, lwe_array_out, sum, bsks, ksks,
+      zero_comparison, num_sum_blocks);
+  reset_radix_ciphertext_blocks(lwe_array_out, num_sum_blocks);
 }
 
 template <typename Torus>
 __host__ void host_integer_radix_equality_check_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_1,
-    Torus const *lwe_array_2, int_comparison_buffer<Torus> *mem_ptr,
-    void *const *bsks, Torus *const *ksks, uint32_t num_radix_blocks) {
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_1,
+    CudaRadixCiphertextFFI const *lwe_array_2,
+    int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
+    Torus *const *ksks, uint32_t num_radix_blocks) {
 
+  if (lwe_array_out->lwe_dimension != lwe_array_1->lwe_dimension ||
+      lwe_array_out->lwe_dimension != lwe_array_2->lwe_dimension)
+    PANIC("Cuda error: input lwe dimensions must be the same")
   auto eq_buffer = mem_ptr->eq_buffer;
 
   // Applies the LUT for the comparison operation
   auto comparisons = mem_ptr->tmp_block_comparisons;
-  legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+  integer_radix_apply_bivariate_lookup_table_kb<Torus>(
       streams, gpu_indexes, gpu_count, comparisons, lwe_array_1, lwe_array_2,
-      bsks, ksks, num_radix_blocks, eq_buffer->operator_lut,
+      bsks, ksks, eq_buffer->operator_lut, num_radix_blocks,
       eq_buffer->operator_lut->params.message_modulus);
 
   // This takes a Vec of blocks, where each block is either 0 or 1.
@@ -341,10 +353,16 @@ __host__ void host_integer_radix_equality_check_kb(
 template <typename Torus>
 __host__ void compare_radix_blocks_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_left,
-    Torus const *lwe_array_right, int_comparison_buffer<Torus> *mem_ptr,
-    void *const *bsks, Torus *const *ksks, uint32_t num_radix_blocks) {
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_left,
+    CudaRadixCiphertextFFI const *lwe_array_right,
+    int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
+    Torus *const *ksks, uint32_t num_radix_blocks) {
 
+  if (lwe_array_out->lwe_dimension != lwe_array_left->lwe_dimension ||
+      lwe_array_out->lwe_dimension != lwe_array_right->lwe_dimension)
+    PANIC("Cuda error: input and output radix ciphertexts should have the same "
+          "lwe dimension")
   auto params = mem_ptr->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
   auto message_modulus = params.message_modulus;
@@ -364,35 +382,43 @@ __host__ void compare_radix_blocks_kb(
   // space, so (-1) % (4 * 4) = 15 = 1|1111 We then add one and get 0 = 0|0000
 
   // Subtract
-  // Here we need the true lwe sub, not the one that comes from shortint.
-  host_subtraction<Torus>(streams[0], gpu_indexes[0], lwe_array_out,
-                          lwe_array_left, lwe_array_right, big_lwe_dimension,
-                          num_radix_blocks);
+  host_subtraction<Torus>(
+      streams[0], gpu_indexes[0], (Torus *)lwe_array_out->ptr,
+      (Torus *)lwe_array_left->ptr, (Torus *)lwe_array_right->ptr,
+      big_lwe_dimension, num_radix_blocks);
 
   // Apply LUT to compare to 0
   auto is_non_zero_lut = mem_ptr->eq_buffer->is_non_zero_lut;
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
       streams, gpu_indexes, gpu_count, lwe_array_out, lwe_array_out, bsks, ksks,
-      num_radix_blocks, is_non_zero_lut);
+      is_non_zero_lut, num_radix_blocks);
 
   // Add one
   // Here Lhs can have the following values: (-1) % (message modulus * carry
   // modulus), 0, 1 So the output values after the addition will be: 0, 1, 2
-  legacy_host_integer_radix_add_scalar_one_inplace<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, big_lwe_dimension,
-      num_radix_blocks, message_modulus, carry_modulus);
+  host_integer_radix_add_scalar_one_inplace<Torus>(
+      streams, gpu_indexes, gpu_count, lwe_array_out, message_modulus,
+      carry_modulus);
 }
 
 // Reduces a vec containing shortint blocks that encrypts a sign
 // (inferior, equal, superior) to one single shortint block containing the
 // final sign
 template <typename Torus>
-__host__ void tree_sign_reduction(
-    cudaStream_t const *streams, uint32_t const *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 *const *bsks,
-    Torus *const *ksks, uint32_t num_radix_blocks) {
+__host__ void
+tree_sign_reduction(cudaStream_t const *streams, uint32_t const *gpu_indexes,
+                    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+                    CudaRadixCiphertextFFI *lwe_block_comparisons,
+                    int_tree_sign_reduction_buffer<Torus> *tree_buffer,
+                    std::function<Torus(Torus)> sign_handler_f,
+                    void *const *bsks, Torus *const *ksks,
+                    uint32_t num_radix_blocks) {
+
+  if (lwe_array_out->lwe_dimension != lwe_block_comparisons->lwe_dimension)
+    PANIC("Cuda error: input lwe dimensions must be the same")
+  if (lwe_block_comparisons->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: block comparisons num radix blocks should not be lower "
+          "than the number of blocks to operate on")
 
   auto params = tree_buffer->params;
   auto big_lwe_dimension = params.big_lwe_dimension;
@@ -405,37 +431,31 @@ __host__ void tree_sign_reduction(
   // Reduces a vec containing shortint blocks that encrypts a sign
   // (inferior, equal, superior) to one single shortint block containing the
   // final sign
-  size_t big_lwe_size = big_lwe_dimension + 1;
-  size_t big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
-
   auto x = tree_buffer->tmp_x;
   auto y = tree_buffer->tmp_y;
   if (x != lwe_block_comparisons)
-    cuda_memcpy_async_gpu_to_gpu(x, lwe_block_comparisons,
-                                 big_lwe_size_bytes * num_radix_blocks,
-                                 streams[0], gpu_indexes[0]);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], x, 0, num_radix_blocks,
+        lwe_block_comparisons, 0, num_radix_blocks);
 
   uint32_t partial_block_count = num_radix_blocks;
 
   auto inner_tree_leaf = tree_buffer->tree_inner_leaf_lut;
   while (partial_block_count > 2) {
-    pack_blocks<Torus>(streams[0], gpu_indexes[0], y, x, big_lwe_dimension,
-                       partial_block_count, 4);
+    pack_blocks<Torus>(streams[0], gpu_indexes[0], y, x, partial_block_count,
+                       4);
 
-    legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, x, y, bsks, ksks,
-        partial_block_count >> 1, inner_tree_leaf);
+    integer_radix_apply_univariate_lookup_table_kb<Torus>(
+        streams, gpu_indexes, gpu_count, x, y, bsks, ksks, inner_tree_leaf,
+        partial_block_count >> 1);
 
     if ((partial_block_count % 2) != 0) {
       partial_block_count >>= 1;
       partial_block_count++;
 
-      auto last_y_block = y + (partial_block_count - 1) * big_lwe_size;
-      auto last_x_block = x + (partial_block_count - 1) * big_lwe_size;
-
-      cuda_memcpy_async_gpu_to_gpu(last_x_block, last_y_block,
-                                   big_lwe_size_bytes, streams[0],
-                                   gpu_indexes[0]);
+      copy_radix_ciphertext_slice_async<Torus>(
+          streams[0], gpu_indexes[0], x, partial_block_count - 1,
+          partial_block_count, y, partial_block_count - 1, partial_block_count);
     } else {
       partial_block_count >>= 1;
     }
@@ -446,8 +466,8 @@ __host__ void tree_sign_reduction(
   std::function<Torus(Torus)> f;
 
   if (partial_block_count == 2) {
-    pack_blocks<Torus>(streams[0], gpu_indexes[0], y, x, big_lwe_dimension,
-                       partial_block_count, 4);
+    pack_blocks<Torus>(streams[0], gpu_indexes[0], y, x, partial_block_count,
+                       4);
 
     f = [block_selector_f, sign_handler_f](Torus x) -> Torus {
       int msb = (x >> 2) & 3;
@@ -468,58 +488,64 @@ __host__ void tree_sign_reduction(
   last_lut->broadcast_lut(streams, gpu_indexes, 0);
 
   // Last leaf
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out, y, bsks, ksks, 1,
-      last_lut);
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      streams, gpu_indexes, gpu_count, lwe_array_out, y, bsks, ksks, last_lut,
+      1);
 }
 
 template <typename Torus>
 __host__ void host_integer_radix_difference_check_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_left,
-    Torus const *lwe_array_right, int_comparison_buffer<Torus> *mem_ptr,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_left,
+    CudaRadixCiphertextFFI const *lwe_array_right,
+    int_comparison_buffer<Torus> *mem_ptr,
     std::function<Torus(Torus)> reduction_lut_f, void *const *bsks,
     Torus *const *ksks, uint32_t num_radix_blocks) {
 
+  if (lwe_array_out->lwe_dimension != lwe_array_left->lwe_dimension ||
+      lwe_array_out->lwe_dimension != lwe_array_right->lwe_dimension)
+    PANIC("Cuda error: input lwe dimensions must be the same")
+
   auto diff_buffer = mem_ptr->diff_buffer;
 
   auto params = mem_ptr->params;
-  auto big_lwe_dimension = params.big_lwe_dimension;
-  auto big_lwe_size = big_lwe_dimension + 1;
   auto message_modulus = params.message_modulus;
   auto carry_modulus = params.carry_modulus;
 
   uint32_t packed_num_radix_blocks = num_radix_blocks;
-  Torus *lhs = (Torus *)lwe_array_left;
-  Torus *rhs = (Torus *)lwe_array_right;
+  CudaRadixCiphertextFFI lhs;
+  as_radix_ciphertext_slice<Torus>(&lhs, diff_buffer->tmp_packed, 0,
+                                   num_radix_blocks / 2);
+  CudaRadixCiphertextFFI rhs;
+  as_radix_ciphertext_slice<Torus>(&rhs, diff_buffer->tmp_packed,
+                                   num_radix_blocks / 2, num_radix_blocks);
   if (carry_modulus >= message_modulus) {
     // Packing is possible
     // Pack inputs
-    Torus *packed_left = diff_buffer->tmp_packed;
-    Torus *packed_right =
-        diff_buffer->tmp_packed + num_radix_blocks / 2 * big_lwe_size;
     // In case the ciphertext is signed, the sign block and the one before it
     // are handled separately
     if (mem_ptr->is_signed) {
       packed_num_radix_blocks -= 2;
     }
-    pack_blocks<Torus>(streams[0], gpu_indexes[0], packed_left, lwe_array_left,
-                       big_lwe_dimension, packed_num_radix_blocks,
-                       message_modulus);
-    pack_blocks<Torus>(streams[0], gpu_indexes[0], packed_right,
-                       lwe_array_right, big_lwe_dimension,
+    pack_blocks<Torus>(streams[0], gpu_indexes[0], &lhs, lwe_array_left,
+                       packed_num_radix_blocks, message_modulus);
+    pack_blocks<Torus>(streams[0], gpu_indexes[0], &rhs, lwe_array_right,
                        packed_num_radix_blocks, message_modulus);
     // From this point we have half number of blocks
     packed_num_radix_blocks /= 2;
 
     // Clean noise
     auto identity_lut = mem_ptr->identity_lut;
-    legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-        streams, gpu_indexes, gpu_count, packed_left, packed_left, bsks, ksks,
-        2 * packed_num_radix_blocks, identity_lut);
-
-    lhs = packed_left;
-    rhs = packed_right;
+    integer_radix_apply_univariate_lookup_table_kb<Torus>(
+        streams, gpu_indexes, gpu_count, diff_buffer->tmp_packed,
+        diff_buffer->tmp_packed, bsks, ksks, identity_lut,
+        2 * packed_num_radix_blocks);
+  } else {
+    as_radix_ciphertext_slice<Torus>(&lhs, lwe_array_left, 0,
+                                     lwe_array_left->num_radix_blocks);
+    as_radix_ciphertext_slice<Torus>(&rhs, lwe_array_right, 0,
+                                     lwe_array_right->num_radix_blocks);
   }
 
   // comparisons will be assigned
@@ -532,7 +558,7 @@ __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<Torus>(streams, gpu_indexes, gpu_count, comparisons,
-                                   lhs, rhs, mem_ptr, bsks, ksks,
+                                   &lhs, &rhs, mem_ptr, bsks, ksks,
                                    packed_num_radix_blocks);
     num_comparisons = packed_num_radix_blocks;
   } else {
@@ -540,38 +566,59 @@ __host__ void host_integer_radix_difference_check_kb(
     if (carry_modulus >= message_modulus) {
       // Compare (num_radix_blocks - 2) / 2 packed blocks
       compare_radix_blocks_kb<Torus>(streams, gpu_indexes, gpu_count,
-                                     comparisons, lhs, rhs, mem_ptr, bsks, ksks,
-                                     packed_num_radix_blocks);
+                                     comparisons, &lhs, &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;
-      Torus *packed_left = diff_buffer->tmp_packed;
-      Torus *packed_right =
-          diff_buffer->tmp_packed + num_radix_blocks / 2 * big_lwe_size;
-      Torus *last_left_block_before_sign_block =
-          packed_left + packed_num_radix_blocks * big_lwe_size;
-      Torus *last_right_block_before_sign_block =
-          packed_right + packed_num_radix_blocks * big_lwe_size;
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, last_left_block_before_sign_block,
-          lwe_array_left + (num_radix_blocks - 2) * big_lwe_size, bsks, ksks, 1,
-          identity_lut);
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, last_right_block_before_sign_block,
-          lwe_array_right + (num_radix_blocks - 2) * big_lwe_size, bsks, ksks,
-          1, identity_lut);
+      CudaRadixCiphertextFFI last_left_block_before_sign_block;
+      as_radix_ciphertext_slice<Torus>(
+          &last_left_block_before_sign_block, diff_buffer->tmp_packed,
+          packed_num_radix_blocks, packed_num_radix_blocks + 1);
+      CudaRadixCiphertextFFI shifted_lwe_array_left;
+      as_radix_ciphertext_slice<Torus>(&shifted_lwe_array_left, lwe_array_left,
+                                       num_radix_blocks - 2,
+                                       num_radix_blocks - 1);
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, &last_left_block_before_sign_block,
+          &shifted_lwe_array_left, bsks, ksks, identity_lut, 1);
+
+      CudaRadixCiphertextFFI last_right_block_before_sign_block;
+      as_radix_ciphertext_slice<Torus>(
+          &last_right_block_before_sign_block, diff_buffer->tmp_packed,
+          num_radix_blocks / 2 + packed_num_radix_blocks,
+          num_radix_blocks / 2 + packed_num_radix_blocks + 1);
+      CudaRadixCiphertextFFI shifted_lwe_array_right;
+      as_radix_ciphertext_slice<Torus>(&shifted_lwe_array_right,
+                                       lwe_array_right, num_radix_blocks - 2,
+                                       num_radix_blocks - 1);
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, &last_right_block_before_sign_block,
+          &shifted_lwe_array_right, bsks, ksks, identity_lut, 1);
+
+      CudaRadixCiphertextFFI shifted_comparisons;
+      as_radix_ciphertext_slice<Torus>(&shifted_comparisons, comparisons,
+                                       packed_num_radix_blocks,
+                                       packed_num_radix_blocks + 1);
       compare_radix_blocks_kb<Torus>(
-          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, bsks, ksks, 1);
+          streams, gpu_indexes, gpu_count, &shifted_comparisons,
+          &last_left_block_before_sign_block,
+          &last_right_block_before_sign_block, mem_ptr, bsks, ksks, 1);
+
       // Compare the sign block separately
-      legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-          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, bsks, ksks,
-          1, mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
+      as_radix_ciphertext_slice<Torus>(&shifted_comparisons, comparisons,
+                                       packed_num_radix_blocks + 1,
+                                       packed_num_radix_blocks + 2);
+      CudaRadixCiphertextFFI last_left_block;
+      as_radix_ciphertext_slice<Torus>(&last_left_block, lwe_array_left,
+                                       num_radix_blocks - 1, num_radix_blocks);
+      CudaRadixCiphertextFFI last_right_block;
+      as_radix_ciphertext_slice<Torus>(&last_right_block, lwe_array_right,
+                                       num_radix_blocks - 1, num_radix_blocks);
+      integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, &shifted_comparisons,
+          &last_left_block, &last_right_block, bsks, ksks, mem_ptr->signed_lut,
+          1, mem_ptr->signed_lut->params.message_modulus);
       num_comparisons = packed_num_radix_blocks + 2;
 
     } else {
@@ -579,12 +626,19 @@ __host__ void host_integer_radix_difference_check_kb(
           streams, gpu_indexes, gpu_count, comparisons, lwe_array_left,
           lwe_array_right, mem_ptr, bsks, ksks, num_radix_blocks - 1);
       // Compare the sign block separately
-      legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-          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, bsks, ksks,
-          1, mem_ptr->signed_lut, mem_ptr->signed_lut->params.message_modulus);
+      CudaRadixCiphertextFFI shifted_comparisons;
+      as_radix_ciphertext_slice<Torus>(&shifted_comparisons, comparisons,
+                                       num_radix_blocks - 1, num_radix_blocks);
+      CudaRadixCiphertextFFI last_left_block;
+      as_radix_ciphertext_slice<Torus>(&last_left_block, lwe_array_left,
+                                       num_radix_blocks - 1, num_radix_blocks);
+      CudaRadixCiphertextFFI last_right_block;
+      as_radix_ciphertext_slice<Torus>(&last_right_block, lwe_array_right,
+                                       num_radix_blocks - 1, num_radix_blocks);
+      integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, &shifted_comparisons,
+          &last_left_block, &last_right_block, bsks, ksks, mem_ptr->signed_lut,
+          1, mem_ptr->signed_lut->params.message_modulus);
       num_comparisons = num_radix_blocks;
     }
   }
@@ -612,32 +666,42 @@ __host__ void scratch_cuda_integer_radix_comparison_check_kb(
 template <typename Torus>
 __host__ void host_integer_radix_maxmin_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_left,
-    Torus const *lwe_array_right, int_comparison_buffer<Torus> *mem_ptr,
-    void *const *bsks, Torus *const *ksks, uint32_t total_num_radix_blocks) {
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_left,
+    CudaRadixCiphertextFFI const *lwe_array_right,
+    int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
+    Torus *const *ksks, uint32_t num_radix_blocks) {
+
+  if (lwe_array_out->lwe_dimension != lwe_array_left->lwe_dimension ||
+      lwe_array_out->lwe_dimension != lwe_array_right->lwe_dimension)
+    PANIC("Cuda error: input and output lwe dimensions must be the same")
+  if (lwe_array_out->num_radix_blocks < num_radix_blocks ||
+      lwe_array_left->num_radix_blocks < num_radix_blocks ||
+      lwe_array_right->num_radix_blocks < num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks should not be lower "
+          "than the number of blocks to operate on")
 
   // Compute the sign
   host_integer_radix_difference_check_kb<Torus>(
       streams, gpu_indexes, gpu_count, mem_ptr->tmp_lwe_array_out,
       lwe_array_left, lwe_array_right, mem_ptr, mem_ptr->identity_lut_f, bsks,
-      ksks, total_num_radix_blocks);
+      ksks, num_radix_blocks);
 
   // Selector
-  legacy_host_integer_radix_cmux_kb<Torus>(
-      streams, gpu_indexes, gpu_count, lwe_array_out,
-      mem_ptr->tmp_lwe_array_out, lwe_array_left, lwe_array_right,
-      mem_ptr->cmux_buffer, bsks, ksks, total_num_radix_blocks);
+  host_integer_radix_cmux_kb<Torus>(streams, gpu_indexes, gpu_count,
+                                    lwe_array_out, mem_ptr->tmp_lwe_array_out,
+                                    lwe_array_left, lwe_array_right,
+                                    mem_ptr->cmux_buffer, bsks, ksks);
 }
 
 template <typename Torus>
 __host__ void host_integer_are_all_comparisons_block_true_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_in,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in,
     int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
     Torus *const *ksks, uint32_t num_radix_blocks) {
 
-  auto eq_buffer = mem_ptr->eq_buffer;
-
   // It returns a block encrypting 1 if all input blocks are 1
   // otherwise the block encrypts 0
   are_all_comparisons_block_true<Torus>(streams, gpu_indexes, gpu_count,
@@ -648,12 +712,11 @@ __host__ void host_integer_are_all_comparisons_block_true_kb(
 template <typename Torus>
 __host__ void host_integer_is_at_least_one_comparisons_block_true_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *lwe_array_out, Torus const *lwe_array_in,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *lwe_array_out,
+    CudaRadixCiphertextFFI const *lwe_array_in,
     int_comparison_buffer<Torus> *mem_ptr, void *const *bsks,
     Torus *const *ksks, uint32_t num_radix_blocks) {
 
-  auto eq_buffer = mem_ptr->eq_buffer;
-
   // It returns a block encrypting 1 if all input blocks are 1
   // otherwise the block encrypts 0
   is_at_least_one_comparisons_block_true<Torus>(

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

@@ -43,6 +43,9 @@ __global__ void pack(Torus *array_out, Torus *array_in, uint32_t log_modulus,
   }
 }
 
+/// Packs `num_lwes` LWE-ciphertext contained in `num_glwes` GLWE-ciphertext in
+/// a compressed array This function follows the naming used in the CPU
+/// implementation
 template <typename Torus>
 __host__ void host_pack(cudaStream_t stream, uint32_t gpu_index,
                         Torus *array_out, Torus *array_in, uint32_t num_glwes,
@@ -55,26 +58,23 @@ __host__ void host_pack(cudaStream_t stream, uint32_t gpu_index,
 
   auto log_modulus = mem_ptr->storage_log_modulus;
   // [0..num_glwes-1) GLWEs
-  auto in_len = (compression_params.glwe_dimension + 1) *
-                compression_params.polynomial_size;
+  auto in_len = num_glwes * compression_params.glwe_dimension *
+                    compression_params.polynomial_size +
+                num_lwes;
+
   auto number_bits_to_pack = in_len * log_modulus;
-  auto nbits = sizeof(Torus) * 8;
+
   // number_bits_to_pack.div_ceil(Scalar::BITS)
+  auto nbits = sizeof(Torus) * 8;
   auto out_len = (number_bits_to_pack + nbits - 1) / nbits;
 
-  // Last GLWE
-  number_bits_to_pack = in_len * log_modulus;
-  auto last_out_len = (number_bits_to_pack + nbits - 1) / nbits;
-
-  auto num_coeffs = (num_glwes - 1) * out_len + last_out_len;
-
   int num_blocks = 0, num_threads = 0;
-  getNumBlocksAndThreads(num_coeffs, 1024, num_blocks, num_threads);
+  getNumBlocksAndThreads(out_len, 1024, num_blocks, num_threads);
 
   dim3 grid(num_blocks);
   dim3 threads(num_threads);
   pack<Torus><<<grid, threads, 0, stream>>>(array_out, array_in, log_modulus,
-                                            num_coeffs, in_len, out_len);
+                                            out_len, in_len, out_len);
   check_cuda_error(cudaGetLastError());
 }
 
@@ -99,14 +99,13 @@ host_integer_compress(cudaStream_t const *streams, uint32_t const *gpu_indexes,
   uint32_t lwe_in_size = input_lwe_dimension + 1;
   uint32_t glwe_out_size = (compression_params.glwe_dimension + 1) *
                            compression_params.polynomial_size;
-  uint32_t num_glwes_for_compression =
-      num_radix_blocks / mem_ptr->lwe_per_glwe + 1;
+  uint32_t num_glwes =
+      (num_radix_blocks + mem_ptr->lwe_per_glwe - 1) / mem_ptr->lwe_per_glwe;
 
   // Keyswitch LWEs to GLWE
   auto tmp_glwe_array_out = mem_ptr->tmp_glwe_array_out;
   cuda_memset_async(tmp_glwe_array_out, 0,
-                    num_glwes_for_compression *
-                        (compression_params.glwe_dimension + 1) *
+                    num_glwes * (compression_params.glwe_dimension + 1) *
                         compression_params.polynomial_size * sizeof(Torus),
                     streams[0], gpu_indexes[0]);
   auto fp_ks_buffer = mem_ptr->fp_ks_buffer;
@@ -131,23 +130,21 @@ host_integer_compress(cudaStream_t const *streams, uint32_t const *gpu_indexes,
   // Modulus switch
   host_modulus_switch_inplace<Torus>(
       streams[0], gpu_indexes[0], tmp_glwe_array_out,
-      num_glwes_for_compression * (compression_params.glwe_dimension + 1) *
-          compression_params.polynomial_size,
+      num_glwes * compression_params.glwe_dimension *
+              compression_params.polynomial_size +
+          num_radix_blocks,
       mem_ptr->storage_log_modulus);
 
   host_pack<Torus>(streams[0], gpu_indexes[0], glwe_array_out,
-                   tmp_glwe_array_out, num_glwes_for_compression,
-                   num_radix_blocks, mem_ptr);
+                   tmp_glwe_array_out, num_glwes, num_radix_blocks, mem_ptr);
 }
 
 template <typename Torus>
 __global__ void extract(Torus *glwe_array_out, Torus const *array_in,
-                        uint32_t index, uint32_t log_modulus,
-                        uint32_t input_len, uint32_t initial_out_len) {
+                        uint32_t log_modulus, uint32_t initial_out_len) {
   auto nbits = sizeof(Torus) * 8;
 
   auto i = threadIdx.x + blockIdx.x * blockDim.x;
-  auto chunk_array_in = array_in + index * input_len;
   if (i < initial_out_len) {
     // Unpack
     Torus mask = ((Torus)1 << log_modulus) - 1;
@@ -161,12 +158,11 @@ __global__ void extract(Torus *glwe_array_out, Torus const *array_in,
 
     Torus unpacked_i;
     if (start_block == end_block_inclusive) {
-      auto single_part = chunk_array_in[start_block] >> start_remainder;
+      auto single_part = array_in[start_block] >> start_remainder;
       unpacked_i = single_part & mask;
     } else {
-      auto first_part = chunk_array_in[start_block] >> start_remainder;
-      auto second_part = chunk_array_in[start_block + 1]
-                         << (nbits - start_remainder);
+      auto first_part = array_in[start_block] >> start_remainder;
+      auto second_part = array_in[start_block + 1] << (nbits - start_remainder);
 
       unpacked_i = (first_part | second_part) & mask;
     }
@@ -177,6 +173,7 @@ __global__ void extract(Torus *glwe_array_out, Torus const *array_in,
 }
 
 /// Extracts the glwe_index-nth GLWE ciphertext
+/// This function follows the naming used in the CPU implementation
 template <typename Torus>
 __host__ void host_extract(cudaStream_t stream, uint32_t gpu_index,
                            Torus *glwe_array_out, Torus const *array_in,
@@ -188,36 +185,51 @@ __host__ void host_extract(cudaStream_t stream, uint32_t gpu_index,
   cuda_set_device(gpu_index);
 
   auto compression_params = mem_ptr->compression_params;
-
   auto log_modulus = mem_ptr->storage_log_modulus;
+  auto glwe_ciphertext_size = (compression_params.glwe_dimension + 1) *
+                              compression_params.polynomial_size;
+
+  uint32_t body_count = mem_ptr->body_count;
+  auto num_glwes = (body_count + compression_params.polynomial_size - 1) /
+                   compression_params.polynomial_size;
+
+  // Compressed length of the compressed GLWE we want to extract
+  if (mem_ptr->body_count % compression_params.polynomial_size == 0)
+    body_count = compression_params.polynomial_size;
+  else if (glwe_index == num_glwes - 1)
+    body_count = mem_ptr->body_count % compression_params.polynomial_size;
+  else
+    body_count = compression_params.polynomial_size;
 
-  uint32_t body_count =
-      std::min(mem_ptr->body_count, compression_params.polynomial_size);
   auto initial_out_len =
       compression_params.glwe_dimension * compression_params.polynomial_size +
       body_count;
 
-  auto compressed_glwe_accumulator_size =
-      (compression_params.glwe_dimension + 1) *
-      compression_params.polynomial_size;
-  auto number_bits_to_unpack = compressed_glwe_accumulator_size * log_modulus;
+  // Calculates how many bits this particular GLWE shall use
+  auto number_bits_to_unpack = initial_out_len * log_modulus;
   auto nbits = sizeof(Torus) * 8;
-  // number_bits_to_unpack.div_ceil(Scalar::BITS)
   auto input_len = (number_bits_to_unpack + nbits - 1) / nbits;
 
-  // We assure the tail of the glwe is zeroed
-  auto zeroed_slice = glwe_array_out + initial_out_len;
-  cuda_memset_async(zeroed_slice, 0,
-                    (compression_params.polynomial_size - body_count) *
-                        sizeof(Torus),
-                    stream, gpu_index);
+  // Calculates how many bits a full-packed GLWE shall use
+  number_bits_to_unpack = glwe_ciphertext_size * log_modulus;
+  auto len = (number_bits_to_unpack + nbits - 1) / nbits;
+  // Uses that length to set the input pointer
+  auto chunk_array_in = array_in + glwe_index * len;
+
+  // Ensure the tail of the GLWE is zeroed
+  if (initial_out_len < glwe_ciphertext_size) {
+    auto zeroed_slice = glwe_array_out + initial_out_len;
+    cuda_memset_async(glwe_array_out, 0,
+                      (glwe_ciphertext_size - initial_out_len) * sizeof(Torus),
+                      stream, gpu_index);
+  }
+
   int num_blocks = 0, num_threads = 0;
   getNumBlocksAndThreads(initial_out_len, 128, num_blocks, num_threads);
   dim3 grid(num_blocks);
   dim3 threads(num_threads);
-  extract<Torus><<<grid, threads, 0, stream>>>(glwe_array_out, array_in,
-                                               glwe_index, log_modulus,
-                                               input_len, initial_out_len);
+  extract<Torus><<<grid, threads, 0, stream>>>(glwe_array_out, chunk_array_in,
+                                               log_modulus, initial_out_len);
   check_cuda_error(cudaGetLastError());
 }
 
@@ -235,18 +247,13 @@ __host__ void host_integer_decompress(
 
   auto compression_params = h_mem_ptr->compression_params;
   auto lwe_per_glwe = compression_params.polynomial_size;
-  if (indexes_array_size > lwe_per_glwe)
-    PANIC("Cuda error: too many LWEs to decompress. The number of LWEs should "
-          "be smaller than "
-          "polynomial_size.")
 
   auto num_radix_blocks = h_mem_ptr->num_radix_blocks;
   if (num_radix_blocks != indexes_array_size)
     PANIC("Cuda error: wrong number of LWEs in decompress: the number of LWEs "
           "should be the same as indexes_array_size.")
 
-  // the first element is the last index in h_indexes_array that lies in the
-  // related GLWE
+  // the first element is the number of LWEs that lies in the related GLWE
   std::vector<std::pair<int, Torus *>> glwe_vec;
 
   // Extract all GLWEs
@@ -257,7 +264,7 @@ __host__ void host_integer_decompress(
   auto extracted_glwe = h_mem_ptr->tmp_extracted_glwe;
   host_extract<Torus>(streams[0], gpu_indexes[0], extracted_glwe,
                       d_packed_glwe_in, current_glwe_index, h_mem_ptr);
-  glwe_vec.push_back(std::make_pair(0, extracted_glwe));
+  glwe_vec.push_back(std::make_pair(1, extracted_glwe));
   for (int i = 1; i < indexes_array_size; i++) {
     auto glwe_index = h_indexes_array[i] / lwe_per_glwe;
     if (glwe_index != current_glwe_index) {
@@ -266,10 +273,10 @@ __host__ void host_integer_decompress(
       // Extracts a new GLWE
       host_extract<Torus>(streams[0], gpu_indexes[0], extracted_glwe,
                           d_packed_glwe_in, glwe_index, h_mem_ptr);
-      glwe_vec.push_back(std::make_pair(i, extracted_glwe));
+      glwe_vec.push_back(std::make_pair(1, extracted_glwe));
     } else {
-      // Updates the index
-      glwe_vec.back().first++;
+      // Updates the quantity
+      ++glwe_vec.back().first;
     }
   }
   // Sample extract all LWEs
@@ -279,17 +286,16 @@ __host__ void host_integer_decompress(
   uint32_t current_idx = 0;
   auto d_indexes_array_chunk = d_indexes_array;
   for (const auto &max_idx_and_glwe : glwe_vec) {
-    uint32_t last_idx = max_idx_and_glwe.first;
+    const auto num_lwes = max_idx_and_glwe.first;
     extracted_glwe = max_idx_and_glwe.second;
 
-    auto num_lwes = last_idx + 1 - current_idx;
-    cuda_glwe_sample_extract_64(streams[0], gpu_indexes[0], extracted_lwe,
-                                extracted_glwe, d_indexes_array_chunk, num_lwes,
-                                compression_params.glwe_dimension,
-                                compression_params.polynomial_size);
+    cuda_glwe_sample_extract_64(
+        streams[0], gpu_indexes[0], extracted_lwe, extracted_glwe,
+        d_indexes_array_chunk, num_lwes, compression_params.polynomial_size,
+        compression_params.glwe_dimension, compression_params.polynomial_size);
     d_indexes_array_chunk += num_lwes;
     extracted_lwe += num_lwes * lwe_accumulator_size;
-    current_idx = last_idx;
+    current_idx += num_lwes;
   }
 
   // Reset

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

@@ -22,18 +22,16 @@ void scratch_cuda_integer_div_rem_radix_ciphertext_kb_64(
 
 void cuda_integer_div_rem_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *quotient, void *remainder, void const *numerator, void const *divisor,
-    bool is_signed, int8_t *mem_ptr, void *const *bsks, void *const *ksks,
-    uint32_t num_blocks) {
+    CudaRadixCiphertextFFI *quotient, CudaRadixCiphertextFFI *remainder,
+    CudaRadixCiphertextFFI const *numerator,
+    CudaRadixCiphertextFFI const *divisor, bool is_signed, int8_t *mem_ptr,
+    void *const *bsks, void *const *ksks) {
 
   auto mem = (int_div_rem_memory<uint64_t> *)mem_ptr;
 
   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<const uint64_t *>(numerator),
-      static_cast<const uint64_t *>(divisor), is_signed, bsks,
-      (uint64_t **)(ksks), mem, num_blocks);
+      (cudaStream_t *)(streams), gpu_indexes, gpu_count, quotient, remainder,
+      numerator, divisor, is_signed, bsks, (uint64_t **)(ksks), mem);
 }
 
 void cleanup_cuda_integer_div_rem(void *const *streams,

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

@@ -19,146 +19,6 @@
 #include <string>
 #include <vector>
 
-int ceil_div(int a, int b) { return (a + b - 1) / b; }
-
-// struct makes it easier to use list of ciphertexts and move data between them
-// struct does not allocate or drop any memory,
-// keeps track on number of ciphertexts inside list.
-template <typename Torus> struct lwe_ciphertext_list {
-  Torus *data;
-  size_t max_blocks;
-  size_t len;
-  int_radix_params params;
-
-  size_t big_lwe_size;
-  size_t big_lwe_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);
-    big_lwe_dimension = params.big_lwe_dimension;
-    len = max_blocks;
-  }
-
-  // copies ciphertexts from Torus*, starting from `starting_block` including
-  // `finish_block`, does not change the value of self len
-  void copy_from(Torus *src, size_t start_block, size_t finish_block,
-                 cudaStream_t stream, uint32_t gpu_index) {
-    size_t tmp_len = finish_block - start_block + 1;
-    cuda_memcpy_async_gpu_to_gpu(data, &src[start_block * big_lwe_size],
-                                 tmp_len * big_lwe_size_bytes, stream,
-                                 gpu_index);
-  }
-
-  // copies ciphertexts from lwe_ciphertext_list, starting from `starting_block`
-  // including `finish_block`, does not change the value of self len
-  void copy_from(const lwe_ciphertext_list &src, size_t start_block,
-                 size_t finish_block, cudaStream_t stream, uint32_t gpu_index) {
-    copy_from(src.data, start_block, finish_block, stream, gpu_index);
-  }
-
-  // copies ciphertexts from Torus*, starting from `starting_block`
-  // including `finish_block`, updating the value of self len
-  void clone_from(Torus *src, size_t start_block, size_t finish_block,
-                  cudaStream_t stream, uint32_t gpu_index) {
-    len = finish_block - start_block + 1;
-
-    cuda_memcpy_async_gpu_to_gpu(data, &src[start_block * big_lwe_size],
-                                 len * big_lwe_size_bytes, stream, gpu_index);
-  }
-
-  // copies ciphertexts from ciphertexts_list, starting from `starting_block`
-  // including `finish_block`, updating the value of self len
-  void clone_from(const lwe_ciphertext_list &src, size_t start_block,
-                  size_t finish_block, cudaStream_t stream,
-                  uint32_t gpu_index) {
-    clone_from(src.data, start_block, finish_block, stream, gpu_index);
-  }
-
-  // assign zero to blocks starting from `start_block` including `finish_block`
-  void assign_zero(size_t start_block, size_t finish_block, cudaStream_t stream,
-                   uint32_t gpu_index) {
-    auto size = finish_block - start_block + 1;
-    cuda_memset_async(&data[start_block * big_lwe_size], 0,
-                      size * big_lwe_size_bytes, stream, gpu_index);
-  }
-
-  // return pointer to last block
-  Torus *last_block() { return &data[(len - 1) * big_lwe_size]; }
-
-  // return pointer to first_block
-  Torus *first_block() { return data; }
-
-  // return block with `index`
-  Torus *get_block(size_t index) {
-    assert(index < len);
-    return &data[index * big_lwe_size];
-  }
-
-  bool is_empty() { return len == 0; }
-
-  // does not dop actual memory from `data`, only reduces value of `len` by one
-  void pop() {
-    if (len > 0)
-      len--;
-    else
-      assert(len > 0);
-  }
-
-  // insert ciphertext at index `ind`
-  void insert(size_t ind, Torus *ciphertext_block, cudaStream_t stream,
-              uint32_t gpu_index) {
-    assert(ind <= len);
-    assert(len < max_blocks);
-
-    size_t insert_offset = ind * big_lwe_size;
-
-    for (size_t i = len; i > ind; i--) {
-      Torus *src = &data[(i - 1) * big_lwe_size];
-      Torus *dst = &data[i * big_lwe_size];
-      cuda_memcpy_async_gpu_to_gpu(dst, src, big_lwe_size_bytes, stream,
-                                   gpu_index);
-    }
-
-    cuda_memcpy_async_gpu_to_gpu(&data[insert_offset], ciphertext_block,
-                                 big_lwe_size_bytes, stream, gpu_index);
-    len++;
-  }
-
-  // push ciphertext at the end of `data`
-  void push(Torus *ciphertext_block, cudaStream_t stream, uint32_t gpu_index) {
-    assert(len < max_blocks);
-
-    size_t offset = len * big_lwe_size;
-    cuda_memcpy_async_gpu_to_gpu(&data[offset], ciphertext_block,
-                                 big_lwe_size_bytes, stream, gpu_index);
-    len++;
-  }
-
-  // duplicate ciphertext into `number_of_blocks` ciphertexts
-  void fill_with_same_ciphertext(Torus *ciphertext, size_t number_of_blocks,
-                                 cudaStream_t stream, uint32_t gpu_index) {
-    assert(number_of_blocks <= max_blocks);
-
-    for (size_t i = 0; i < number_of_blocks; i++) {
-      Torus *dest = &data[i * big_lwe_size];
-      cuda_memcpy_async_gpu_to_gpu(dest, ciphertext, big_lwe_size_bytes, stream,
-                                   gpu_index);
-    }
-
-    len = number_of_blocks;
-  }
-
-  // used for debugging, prints body of each ciphertext.
-  void print_blocks_body(const char *name) {
-    for (int i = 0; i < len; i++) {
-      print_debug(name, &data[i * big_lwe_size + big_lwe_dimension], 1);
-    }
-  }
-};
-
 template <typename Torus>
 __host__ void scratch_cuda_integer_div_rem_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
@@ -173,67 +33,50 @@ __host__ void scratch_cuda_integer_div_rem_kb(
 template <typename Torus>
 __host__ void host_unsigned_integer_div_rem_kb(
     cudaStream_t const *streams, uint32_t const *gpu_indexes,
-    uint32_t gpu_count, Torus *quotient, Torus *remainder,
-    Torus const *numerator, Torus const *divisor, void *const *bsks,
-    uint64_t *const *ksks, unsigned_int_div_rem_memory<uint64_t> *mem_ptr,
-    uint32_t num_blocks) {
+    uint32_t gpu_count, CudaRadixCiphertextFFI *quotient,
+    CudaRadixCiphertextFFI *remainder, CudaRadixCiphertextFFI const *numerator,
+    CudaRadixCiphertextFFI const *divisor, void *const *bsks,
+    uint64_t *const *ksks, unsigned_int_div_rem_memory<uint64_t> *mem_ptr) {
 
+  if (remainder->num_radix_blocks != numerator->num_radix_blocks ||
+      remainder->num_radix_blocks != divisor->num_radix_blocks ||
+      remainder->num_radix_blocks != quotient->num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks must be equal")
+  if (remainder->lwe_dimension != numerator->lwe_dimension ||
+      remainder->lwe_dimension != divisor->lwe_dimension ||
+      remainder->lwe_dimension != quotient->lwe_dimension)
+    PANIC("Cuda error: input and output lwe dimension must be equal")
   auto radix_params = mem_ptr->params;
-
-  auto big_lwe_dimension = radix_params.big_lwe_dimension;
-  auto big_lwe_size = big_lwe_dimension + 1;
-  auto big_lwe_size_bytes = big_lwe_size * sizeof(Torus);
+  auto num_blocks = quotient->num_radix_blocks;
 
   uint32_t message_modulus = radix_params.message_modulus;
-  uint32_t carry_modulus = radix_params.carry_modulus;
   uint32_t num_bits_in_message = 31 - __builtin_clz(message_modulus);
+
   uint32_t total_bits = num_bits_in_message * num_blocks;
 
   // put temporary buffers in lwe_ciphertext_list for easy use
-  lwe_ciphertext_list<Torus> remainder1(mem_ptr->remainder1, radix_params,
-                                        num_blocks);
-  lwe_ciphertext_list<Torus> remainder2(mem_ptr->remainder2, radix_params,
-                                        num_blocks);
-  lwe_ciphertext_list<Torus> numerator_block_stack(
-      mem_ptr->numerator_block_stack, radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> numerator_block_1(mem_ptr->numerator_block_1,
-                                               radix_params, 1);
-  lwe_ciphertext_list<Torus> tmp_radix(mem_ptr->tmp_radix, radix_params,
-                                       num_blocks + 1);
-  lwe_ciphertext_list<Torus> interesting_remainder1(
-      mem_ptr->interesting_remainder1, radix_params, num_blocks + 1);
-  lwe_ciphertext_list<Torus> interesting_remainder2(
-      mem_ptr->interesting_remainder2, radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> interesting_divisor(mem_ptr->interesting_divisor,
-                                                 radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> divisor_ms_blocks(mem_ptr->divisor_ms_blocks,
-                                               radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> new_remainder(mem_ptr->new_remainder, radix_params,
-                                           num_blocks);
-  lwe_ciphertext_list<Torus> subtraction_overflowed(
-      mem_ptr->subtraction_overflowed, radix_params, 1);
-  lwe_ciphertext_list<Torus> did_not_overflow(mem_ptr->did_not_overflow,
-                                              radix_params, 1);
-  lwe_ciphertext_list<Torus> overflow_sum(mem_ptr->overflow_sum, radix_params,
-                                          1);
-  lwe_ciphertext_list<Torus> overflow_sum_radix(mem_ptr->overflow_sum_radix,
-                                                radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> tmp_1(mem_ptr->tmp_1, radix_params, num_blocks);
-  lwe_ciphertext_list<Torus> at_least_one_upper_block_is_non_zero(
-      mem_ptr->at_least_one_upper_block_is_non_zero, radix_params, 1);
-  lwe_ciphertext_list<Torus> cleaned_merged_interesting_remainder(
-      mem_ptr->cleaned_merged_interesting_remainder, radix_params, num_blocks);
+  auto remainder1 = mem_ptr->remainder1;
+  auto remainder2 = mem_ptr->remainder2;
+  auto numerator_block_stack = mem_ptr->numerator_block_stack;
+  auto interesting_remainder1 = mem_ptr->interesting_remainder1;
+  auto interesting_remainder2 = mem_ptr->interesting_remainder2;
+  auto interesting_divisor = mem_ptr->interesting_divisor;
+  auto divisor_ms_blocks = mem_ptr->divisor_ms_blocks;
+  auto new_remainder = mem_ptr->new_remainder;
+  auto subtraction_overflowed = mem_ptr->subtraction_overflowed;
+  auto overflow_sum = mem_ptr->overflow_sum;
+  auto overflow_sum_radix = mem_ptr->overflow_sum_radix;
+  auto at_least_one_upper_block_is_non_zero =
+      mem_ptr->at_least_one_upper_block_is_non_zero;
+  auto cleaned_merged_interesting_remainder =
+      mem_ptr->cleaned_merged_interesting_remainder;
 
-  numerator_block_stack.clone_from((Torus *)numerator, 0, num_blocks - 1,
-                                   streams[0], gpu_indexes[0]);
-  remainder1.assign_zero(0, num_blocks - 1, streams[0], gpu_indexes[0]);
-  remainder2.assign_zero(0, num_blocks - 1, streams[0], gpu_indexes[0]);
-
-  cuda_memset_async(quotient, 0, big_lwe_size_bytes * num_blocks, streams[0],
-                    gpu_indexes[0]);
+  copy_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                     numerator_block_stack, numerator);
+  set_zero_radix_ciphertext_slice_async<Torus>(streams[0], gpu_indexes[0],
+                                               quotient, 0, num_blocks);
 
   for (int i = total_bits - 1; i >= 0; i--) {
-    uint32_t block_of_bit = i / num_bits_in_message;
     uint32_t pos_in_block = i % num_bits_in_message;
     uint32_t msb_bit_set = total_bits - 1 - i;
     uint32_t last_non_trivial_block = msb_bit_set / num_bits_in_message;
@@ -242,16 +85,27 @@ __host__ void host_unsigned_integer_div_rem_kb(
     // and all blocks after it are also trivial zeros
     // This number is in range 1..=num_bocks -1
     uint32_t first_trivial_block = last_non_trivial_block + 1;
+    reset_radix_ciphertext_blocks(interesting_remainder1, first_trivial_block);
+    reset_radix_ciphertext_blocks(interesting_remainder2, first_trivial_block);
+    reset_radix_ciphertext_blocks(interesting_divisor, first_trivial_block);
+    reset_radix_ciphertext_blocks(divisor_ms_blocks,
+                                  num_blocks -
+                                      (msb_bit_set + 1) / num_bits_in_message);
 
-    interesting_remainder1.clone_from(remainder1, 0, last_non_trivial_block,
-                                      streams[0], gpu_indexes[0]);
-    interesting_remainder2.clone_from(remainder2, 0, last_non_trivial_block,
-                                      streams[0], gpu_indexes[0]);
-    interesting_divisor.clone_from((Torus *)divisor, 0, last_non_trivial_block,
-                                   streams[0], gpu_indexes[0]);
-    divisor_ms_blocks.clone_from((Torus *)divisor,
-                                 (msb_bit_set + 1) / num_bits_in_message,
-                                 num_blocks - 1, streams[0], gpu_indexes[0]);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], interesting_remainder1, 0,
+        first_trivial_block, remainder1, 0, first_trivial_block);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], interesting_remainder2, 0,
+        first_trivial_block, remainder2, 0, first_trivial_block);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], interesting_divisor, 0, first_trivial_block,
+        divisor, 0, first_trivial_block);
+    if ((msb_bit_set + 1) / num_bits_in_message < num_blocks)
+      copy_radix_ciphertext_slice_async<Torus>(
+          streams[0], gpu_indexes[0], divisor_ms_blocks, 0,
+          num_blocks - (msb_bit_set + 1) / num_bits_in_message, divisor,
+          (msb_bit_set + 1) / num_bits_in_message, num_blocks);
 
     // We split the divisor at a block position, when in reality the split
     // should be at a bit position meaning that potentially (depending on
@@ -285,16 +139,21 @@ __host__ void host_unsigned_integer_div_rem_kb(
       // Shift the mask so that we will only keep bits we should
       uint32_t shifted_mask = full_message_mask >> shift_amount;
 
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, interesting_divisor.last_block(),
-          interesting_divisor.last_block(), bsks, ksks, 1,
-          mem_ptr->masking_luts_1[shifted_mask]);
+      CudaRadixCiphertextFFI last_interesting_divisor_block;
+      as_radix_ciphertext_slice<Torus>(
+          &last_interesting_divisor_block, interesting_divisor,
+          interesting_divisor->num_radix_blocks - 1,
+          interesting_divisor->num_radix_blocks);
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, &last_interesting_divisor_block,
+          &last_interesting_divisor_block, bsks, ksks,
+          mem_ptr->masking_luts_1[shifted_mask], 1);
     }; // trim_last_interesting_divisor_bits
 
     auto trim_first_divisor_ms_bits = [&](cudaStream_t const *streams,
                                           uint32_t const *gpu_indexes,
                                           uint32_t gpu_count) {
-      if (divisor_ms_blocks.is_empty() ||
+      if (divisor_ms_blocks->num_radix_blocks == 0 ||
           ((msb_bit_set + 1) % num_bits_in_message) == 0) {
         return;
       }
@@ -314,10 +173,9 @@ __host__ void host_unsigned_integer_div_rem_kb(
       // the estimated degree of the output is < msg_modulus
       shifted_mask = shifted_mask & full_message_mask;
 
-      legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
-          streams, gpu_indexes, gpu_count, divisor_ms_blocks.first_block(),
-          divisor_ms_blocks.first_block(), bsks, ksks, 1,
-          mem_ptr->masking_luts_2[shifted_mask]);
+      integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, divisor_ms_blocks, divisor_ms_blocks,
+          bsks, ksks, mem_ptr->masking_luts_2[shifted_mask], 1);
     }; // trim_first_divisor_ms_bits
 
     // This does
@@ -332,46 +190,49 @@ __host__ void host_unsigned_integer_div_rem_kb(
     auto left_shift_interesting_remainder1 = [&](cudaStream_t const *streams,
                                                  uint32_t const *gpu_indexes,
                                                  uint32_t gpu_count) {
-      numerator_block_1.clone_from(
-          numerator_block_stack, numerator_block_stack.len - 1,
-          numerator_block_stack.len - 1, streams[0], gpu_indexes[0]);
-      numerator_block_stack.pop();
-      interesting_remainder1.insert(0, numerator_block_1.first_block(),
-                                    streams[0], gpu_indexes[0]);
+      pop_radix_ciphertext_block_async<Torus>(streams[0], gpu_indexes[0],
+                                              mem_ptr->numerator_block_1,
+                                              numerator_block_stack);
+      insert_block_in_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                                    mem_ptr->numerator_block_1,
+                                                    interesting_remainder1, 0);
 
-      legacy_host_integer_radix_logical_scalar_shift_kb_inplace<Torus>(
-          streams, gpu_indexes, gpu_count, interesting_remainder1.data, 1,
-          mem_ptr->shift_mem_1, bsks, ksks, interesting_remainder1.len);
+      host_integer_radix_logical_scalar_shift_kb_inplace<Torus>(
+          streams, gpu_indexes, gpu_count, interesting_remainder1, 1,
+          mem_ptr->shift_mem_1, bsks, ksks,
+          interesting_remainder1->num_radix_blocks);
 
-      tmp_radix.clone_from(interesting_remainder1, 0,
-                           interesting_remainder1.len - 1, streams[0],
-                           gpu_indexes[0]);
+      reset_radix_ciphertext_blocks(mem_ptr->tmp_radix,
+                                    interesting_remainder1->num_radix_blocks);
+      copy_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                         mem_ptr->tmp_radix,
+                                         interesting_remainder1);
 
-      legacy_host_radix_blocks_rotate_left<Torus>(
-          streams, gpu_indexes, gpu_count, interesting_remainder1.data,
-          tmp_radix.data, 1, interesting_remainder1.len, big_lwe_size);
+      host_radix_blocks_rotate_left<Torus>(
+          streams, gpu_indexes, gpu_count, interesting_remainder1,
+          mem_ptr->tmp_radix, 1, interesting_remainder1->num_radix_blocks);
 
-      numerator_block_1.clone_from(
-          interesting_remainder1, interesting_remainder1.len - 1,
-          interesting_remainder1.len - 1, streams[0], gpu_indexes[0]);
-
-      interesting_remainder1.pop();
+      pop_radix_ciphertext_block_async<Torus>(streams[0], gpu_indexes[0],
+                                              mem_ptr->numerator_block_1,
+                                              interesting_remainder1);
 
       if (pos_in_block != 0) {
         // We have not yet extracted all the bits from this numerator
         // so, we put it back on the front so that it gets taken next
         // iteration
-        numerator_block_stack.push(numerator_block_1.first_block(), streams[0],
-                                   gpu_indexes[0]);
+        push_block_to_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                                    mem_ptr->numerator_block_1,
+                                                    numerator_block_stack);
       }
     }; // left_shift_interesting_remainder1
 
     auto left_shift_interesting_remainder2 = [&](cudaStream_t const *streams,
                                                  uint32_t const *gpu_indexes,
                                                  uint32_t gpu_count) {
-      legacy_host_integer_radix_logical_scalar_shift_kb_inplace<Torus>(
-          streams, gpu_indexes, gpu_count, interesting_remainder2.data, 1,
-          mem_ptr->shift_mem_2, bsks, ksks, interesting_remainder2.len);
+      host_integer_radix_logical_scalar_shift_kb_inplace<Torus>(
+          streams, gpu_indexes, gpu_count, interesting_remainder2, 1,
+          mem_ptr->shift_mem_2, bsks, ksks,
+          interesting_remainder2->num_radix_blocks);
     }; // left_shift_interesting_remainder2
 
     for (uint j = 0; j < gpu_count; j++) {
@@ -400,24 +261,31 @@ __host__ void host_unsigned_integer_div_rem_kb(
     // if interesting_remainder1 == 0 -> interesting_remainder2 != 0
     // In practice interesting_remainder1 contains the numerator bit,
     // but in that position, interesting_remainder2 always has a 0
-    auto &merged_interesting_remainder = interesting_remainder1;
+    auto merged_interesting_remainder = interesting_remainder1;
 
-    legacy_host_addition<Torus>(
-        streams[0], gpu_indexes[0], merged_interesting_remainder.data,
-        merged_interesting_remainder.data, interesting_remainder2.data,
-        radix_params.big_lwe_dimension, merged_interesting_remainder.len);
+    host_addition<Torus>(streams[0], gpu_indexes[0],
+                         merged_interesting_remainder,
+                         merged_interesting_remainder, interesting_remainder2,
+                         merged_interesting_remainder->num_radix_blocks);
 
     // after create_clean_version_of_merged_remainder
     // `merged_interesting_remainder` will be reused as
     // `cleaned_merged_interesting_remainder`
-    cleaned_merged_interesting_remainder.clone_from(
-        merged_interesting_remainder, 0, merged_interesting_remainder.len - 1,
-        streams[0], gpu_indexes[0]);
+    reset_radix_ciphertext_blocks(
+        cleaned_merged_interesting_remainder,
+        merged_interesting_remainder->num_radix_blocks);
+    copy_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                       cleaned_merged_interesting_remainder,
+                                       merged_interesting_remainder);
 
-    assert(merged_interesting_remainder.len == interesting_divisor.len);
+    if (merged_interesting_remainder->num_radix_blocks !=
+        interesting_divisor->num_radix_blocks)
+      PANIC("Cuda error: merged interesting remainder and interesting divisor "
+            "should have the same number of blocks")
 
     // `new_remainder` is not initialized yet, so need to set length
-    new_remainder.len = merged_interesting_remainder.len;
+    reset_radix_ciphertext_blocks(
+        new_remainder, merged_interesting_remainder->num_radix_blocks);
 
     // fills:
     //  `new_remainder` - radix ciphertext
@@ -427,22 +295,24 @@ __host__ void host_unsigned_integer_div_rem_kb(
                                   uint32_t gpu_count) {
       uint32_t compute_borrow = 1;
       uint32_t uses_input_borrow = 0;
-      auto first_indexes = mem_ptr->first_indexes_for_overflow_sub
-                               [merged_interesting_remainder.len - 1];
-      auto second_indexes = mem_ptr->second_indexes_for_overflow_sub
-                                [merged_interesting_remainder.len - 1];
+      auto first_indexes =
+          mem_ptr->first_indexes_for_overflow_sub
+              [merged_interesting_remainder->num_radix_blocks - 1];
+      auto second_indexes =
+          mem_ptr->second_indexes_for_overflow_sub
+              [merged_interesting_remainder->num_radix_blocks - 1];
       auto scalar_indexes =
-          mem_ptr
-              ->scalars_for_overflow_sub[merged_interesting_remainder.len - 1];
+          mem_ptr->scalars_for_overflow_sub
+              [merged_interesting_remainder->num_radix_blocks - 1];
       mem_ptr->overflow_sub_mem->update_lut_indexes(
           streams, gpu_indexes, first_indexes, second_indexes, scalar_indexes,
-          merged_interesting_remainder.len);
-      legacy_host_integer_overflowing_sub<uint64_t>(
-          streams, gpu_indexes, gpu_count, new_remainder.data,
-          (uint64_t *)merged_interesting_remainder.data,
-          interesting_divisor.data, subtraction_overflowed.data,
-          (const Torus *)nullptr, mem_ptr->overflow_sub_mem, bsks, ksks,
-          merged_interesting_remainder.len, compute_borrow, uses_input_borrow);
+          merged_interesting_remainder->num_radix_blocks);
+      host_integer_overflowing_sub<uint64_t>(
+          streams, gpu_indexes, gpu_count, new_remainder,
+          merged_interesting_remainder, interesting_divisor,
+          subtraction_overflowed, (const CudaRadixCiphertextFFI *)nullptr,
+          mem_ptr->overflow_sub_mem, bsks, ksks, compute_borrow,
+          uses_input_borrow);
     };
 
     // fills:
@@ -450,27 +320,27 @@ __host__ void host_unsigned_integer_div_rem_kb(
     auto check_divisor_upper_blocks = [&](cudaStream_t const *streams,
                                           uint32_t const *gpu_indexes,
                                           uint32_t gpu_count) {
-      auto &trivial_blocks = divisor_ms_blocks;
-      if (trivial_blocks.is_empty()) {
-        cuda_memset_async(at_least_one_upper_block_is_non_zero.first_block(), 0,
-                          big_lwe_size_bytes, streams[0], gpu_indexes[0]);
+      auto trivial_blocks = divisor_ms_blocks;
+      if (trivial_blocks->num_radix_blocks == 0) {
+        set_zero_radix_ciphertext_slice_async<Torus>(
+            streams[0], gpu_indexes[0], at_least_one_upper_block_is_non_zero, 0,
+            1);
       } else {
 
         // We could call unchecked_scalar_ne
         // But we are in the special case where scalar == 0
         // So we can skip some stuff
-        host_compare_with_zero_equality<Torus>(
-            streams, gpu_indexes, gpu_count, tmp_1.data, trivial_blocks.data,
-            mem_ptr->comparison_buffer, bsks, ksks, trivial_blocks.len,
+        host_compare_blocks_with_zero<Torus>(
+            streams, gpu_indexes, gpu_count, mem_ptr->tmp_1, trivial_blocks,
+            mem_ptr->comparison_buffer, bsks, ksks,
+            trivial_blocks->num_radix_blocks,
             mem_ptr->comparison_buffer->eq_buffer->is_non_zero_lut);
 
-        tmp_1.len =
-            ceil_div(trivial_blocks.len, message_modulus * carry_modulus - 1);
-
         is_at_least_one_comparisons_block_true<Torus>(
             streams, gpu_indexes, gpu_count,
-            at_least_one_upper_block_is_non_zero.data, tmp_1.data,
-            mem_ptr->comparison_buffer, bsks, ksks, tmp_1.len);
+            at_least_one_upper_block_is_non_zero, mem_ptr->tmp_1,
+            mem_ptr->comparison_buffer, bsks, ksks,
+            mem_ptr->tmp_1->num_radix_blocks);
       }
     };
 
@@ -481,12 +351,12 @@ __host__ void host_unsigned_integer_div_rem_kb(
     auto create_clean_version_of_merged_remainder =
         [&](cudaStream_t const *streams, uint32_t const *gpu_indexes,
             uint32_t gpu_count) {
-          legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+          integer_radix_apply_univariate_lookup_table_kb<Torus>(
               streams, gpu_indexes, gpu_count,
-              cleaned_merged_interesting_remainder.data,
-              cleaned_merged_interesting_remainder.data, bsks, ksks,
-              cleaned_merged_interesting_remainder.len,
-              mem_ptr->message_extract_lut_1);
+              cleaned_merged_interesting_remainder,
+              cleaned_merged_interesting_remainder, bsks, ksks,
+              mem_ptr->message_extract_lut_1,
+              cleaned_merged_interesting_remainder->num_radix_blocks);
         };
 
     // phase 2
@@ -507,55 +377,56 @@ __host__ void host_unsigned_integer_div_rem_kb(
       cuda_synchronize_stream(mem_ptr->sub_streams_3[j], gpu_indexes[j]);
     }
 
-    legacy_host_addition<Torus>(streams[0], gpu_indexes[0], overflow_sum.data,
-                                subtraction_overflowed.data,
-                                at_least_one_upper_block_is_non_zero.data,
-                                radix_params.big_lwe_dimension, 1);
+    host_addition<Torus>(streams[0], gpu_indexes[0], overflow_sum,
+                         subtraction_overflowed,
+                         at_least_one_upper_block_is_non_zero, 1);
 
     int factor = (i) ? 3 : 2;
     int factor_lut_id = factor - 2;
-    overflow_sum_radix.fill_with_same_ciphertext(
-        overflow_sum.first_block(), cleaned_merged_interesting_remainder.len,
-        streams[0], gpu_indexes[0]);
+    for (size_t i = 0;
+         i < cleaned_merged_interesting_remainder->num_radix_blocks; i++) {
+      copy_radix_ciphertext_slice_async<Torus>(streams[0], gpu_indexes[0],
+                                               overflow_sum_radix, i, i + 1,
+                                               overflow_sum, 0, 1);
+    }
 
     auto conditionally_zero_out_merged_interesting_remainder =
         [&](cudaStream_t const *streams, uint32_t const *gpu_indexes,
             uint32_t gpu_count) {
-          legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+          integer_radix_apply_bivariate_lookup_table_kb<Torus>(
               streams, gpu_indexes, gpu_count,
-              cleaned_merged_interesting_remainder.data,
-              cleaned_merged_interesting_remainder.data,
-              overflow_sum_radix.data, bsks, ksks,
-              cleaned_merged_interesting_remainder.len,
-              mem_ptr->zero_out_if_overflow_did_not_happen[factor_lut_id],
-              factor);
+              cleaned_merged_interesting_remainder,
+              cleaned_merged_interesting_remainder, overflow_sum_radix, bsks,
+              ksks, mem_ptr->zero_out_if_overflow_did_not_happen[factor_lut_id],
+              cleaned_merged_interesting_remainder->num_radix_blocks, factor);
         };
 
     auto conditionally_zero_out_merged_new_remainder =
         [&](cudaStream_t const *streams, uint32_t const *gpu_indexes,
             uint32_t gpu_count) {
-          legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
-              streams, gpu_indexes, gpu_count, new_remainder.data,
-              new_remainder.data, overflow_sum_radix.data, bsks, ksks,
-              new_remainder.len,
-              mem_ptr->zero_out_if_overflow_happened[factor_lut_id], factor);
+          integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+              streams, gpu_indexes, gpu_count, new_remainder, new_remainder,
+              overflow_sum_radix, bsks, ksks,
+              mem_ptr->zero_out_if_overflow_happened[factor_lut_id],
+              new_remainder->num_radix_blocks, factor);
         };
 
     auto set_quotient_bit = [&](cudaStream_t const *streams,
                                 uint32_t const *gpu_indexes,
                                 uint32_t gpu_count) {
-      legacy_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, bsks, ksks, 1,
-          mem_ptr->merge_overflow_flags_luts[pos_in_block],
+      uint32_t block_of_bit = i / num_bits_in_message;
+      integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+          streams, gpu_indexes, gpu_count, mem_ptr->did_not_overflow,
+          subtraction_overflowed, at_least_one_upper_block_is_non_zero, bsks,
+          ksks, mem_ptr->merge_overflow_flags_luts[pos_in_block], 1,
           mem_ptr->merge_overflow_flags_luts[pos_in_block]
               ->params.message_modulus);
 
-      legacy_host_addition<Torus>(
-          streams[0], gpu_indexes[0], &quotient[block_of_bit * big_lwe_size],
-          &quotient[block_of_bit * big_lwe_size], did_not_overflow.data,
-          radix_params.big_lwe_dimension, 1);
+      CudaRadixCiphertextFFI quotient_block;
+      as_radix_ciphertext_slice<Torus>(&quotient_block, quotient, block_of_bit,
+                                       block_of_bit + 1);
+      host_addition<Torus>(streams[0], gpu_indexes[0], &quotient_block,
+                           &quotient_block, mem_ptr->did_not_overflow, 1);
     };
 
     for (uint j = 0; j < gpu_count; j++) {
@@ -575,32 +446,40 @@ __host__ void host_unsigned_integer_div_rem_kb(
       cuda_synchronize_stream(mem_ptr->sub_streams_3[j], gpu_indexes[j]);
     }
 
-    assert(first_trivial_block - 1 == cleaned_merged_interesting_remainder.len);
-    assert(first_trivial_block - 1 == new_remainder.len);
+    if (first_trivial_block !=
+        cleaned_merged_interesting_remainder->num_radix_blocks)
+      PANIC("Cuda error: first_trivial_block should be equal to "
+            "clean_merged_interesting_remainder num blocks")
+    if (first_trivial_block != new_remainder->num_radix_blocks)
+      PANIC("Cuda error: first_trivial_block should be equal to new_remainder "
+            "num blocks")
 
-    remainder1.copy_from(cleaned_merged_interesting_remainder, 0,
-                         first_trivial_block - 1, streams[0], gpu_indexes[0]);
-    remainder2.copy_from(new_remainder, 0, first_trivial_block - 1, streams[0],
-                         gpu_indexes[0]);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], remainder1, 0, first_trivial_block,
+        cleaned_merged_interesting_remainder, 0, first_trivial_block);
+    copy_radix_ciphertext_slice_async<Torus>(
+        streams[0], gpu_indexes[0], remainder2, 0, first_trivial_block,
+        new_remainder, 0, first_trivial_block);
   }
 
-  assert(remainder1.len == remainder2.len);
+  if (remainder1->num_radix_blocks != remainder2->num_radix_blocks)
+    PANIC("Cuda error: remainder1 and remainder2 should have the same number "
+          "of blocks")
 
   // Clean the quotient and remainder
   // as even though they have no carries, they are not at nominal noise level
-  legacy_host_addition<Torus>(streams[0], gpu_indexes[0], remainder,
-                              remainder1.data, remainder2.data,
-                              radix_params.big_lwe_dimension, remainder1.len);
+  host_addition<Torus>(streams[0], gpu_indexes[0], remainder, remainder1,
+                       remainder2, remainder1->num_radix_blocks);
 
   for (uint j = 0; j < gpu_count; j++) {
     cuda_synchronize_stream(streams[j], gpu_indexes[j]);
   }
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
       mem_ptr->sub_streams_1, gpu_indexes, gpu_count, remainder, remainder,
-      bsks, ksks, num_blocks, mem_ptr->message_extract_lut_1);
-  legacy_integer_radix_apply_univariate_lookup_table_kb<Torus>(
+      bsks, ksks, mem_ptr->message_extract_lut_1, num_blocks);
+  integer_radix_apply_univariate_lookup_table_kb<Torus>(
       mem_ptr->sub_streams_2, gpu_indexes, gpu_count, quotient, quotient, bsks,
-      ksks, num_blocks, mem_ptr->message_extract_lut_2);
+      ksks, mem_ptr->message_extract_lut_2, num_blocks);
   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]);
@@ -608,42 +487,43 @@ __host__ void host_unsigned_integer_div_rem_kb(
 }
 
 template <typename Torus>
-__host__ void host_integer_div_rem_kb(cudaStream_t const *streams,
-                                      uint32_t const *gpu_indexes,
-                                      uint32_t gpu_count, Torus *quotient,
-                                      Torus *remainder, Torus const *numerator,
-                                      Torus const *divisor, bool is_signed,
-                                      void *const *bsks, uint64_t *const *ksks,
-                                      int_div_rem_memory<uint64_t> *int_mem_ptr,
-                                      uint32_t num_blocks) {
+__host__ void host_integer_div_rem_kb(
+    cudaStream_t const *streams, uint32_t const *gpu_indexes,
+    uint32_t gpu_count, CudaRadixCiphertextFFI *quotient,
+    CudaRadixCiphertextFFI *remainder, CudaRadixCiphertextFFI const *numerator,
+    CudaRadixCiphertextFFI const *divisor, bool is_signed, void *const *bsks,
+    uint64_t *const *ksks, int_div_rem_memory<uint64_t> *int_mem_ptr) {
+  if (remainder->num_radix_blocks != numerator->num_radix_blocks ||
+      remainder->num_radix_blocks != divisor->num_radix_blocks ||
+      remainder->num_radix_blocks != quotient->num_radix_blocks)
+    PANIC("Cuda error: input and output num radix blocks must be equal")
+  if (remainder->lwe_dimension != numerator->lwe_dimension ||
+      remainder->lwe_dimension != divisor->lwe_dimension ||
+      remainder->lwe_dimension != quotient->lwe_dimension)
+    PANIC("Cuda error: input and output lwe dimension must be equal")
 
+  auto num_blocks = quotient->num_radix_blocks;
   if (is_signed) {
     auto radix_params = int_mem_ptr->params;
-    uint32_t big_lwe_size = radix_params.big_lwe_dimension + 1;
 
     // temporary memory
-    lwe_ciphertext_list<Torus> positive_numerator(
-        int_mem_ptr->positive_numerator, radix_params, num_blocks);
-    lwe_ciphertext_list<Torus> positive_divisor(int_mem_ptr->positive_divisor,
-                                                radix_params, num_blocks);
-
-    positive_numerator.clone_from((Torus *)numerator, 0, num_blocks - 1,
-                                  streams[0], gpu_indexes[0]);
-    positive_divisor.clone_from((Torus *)divisor, 0, num_blocks - 1, streams[0],
-                                gpu_indexes[0]);
+    auto positive_numerator = int_mem_ptr->positive_numerator;
+    auto positive_divisor = int_mem_ptr->positive_divisor;
+    copy_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                       positive_numerator, numerator);
+    copy_radix_ciphertext_async<Torus>(streams[0], gpu_indexes[0],
+                                       positive_divisor, divisor);
 
     for (uint j = 0; j < gpu_count; j++) {
       cuda_synchronize_stream(streams[j], gpu_indexes[j]);
     }
 
-    legacy_host_integer_abs_kb_async<Torus>(
-        int_mem_ptr->sub_streams_1, gpu_indexes, gpu_count,
-        positive_numerator.data, bsks, ksks, int_mem_ptr->abs_mem_1, true,
-        num_blocks);
-    legacy_host_integer_abs_kb_async<Torus>(
-        int_mem_ptr->sub_streams_2, gpu_indexes, gpu_count,
-        positive_divisor.data, bsks, ksks, int_mem_ptr->abs_mem_2, true,
-        num_blocks);
+    host_integer_abs_kb<Torus>(int_mem_ptr->sub_streams_1, gpu_indexes,
+                               gpu_count, positive_numerator, bsks, ksks,
+                               int_mem_ptr->abs_mem_1, true);
+    host_integer_abs_kb<Torus>(int_mem_ptr->sub_streams_2, gpu_indexes,
+                               gpu_count, positive_divisor, bsks, ksks,
+                               int_mem_ptr->abs_mem_2, true);
     for (uint j = 0; j < int_mem_ptr->active_gpu_count; j++) {
       cuda_synchronize_stream(int_mem_ptr->sub_streams_1[j], gpu_indexes[j]);
       cuda_synchronize_stream(int_mem_ptr->sub_streams_2[j], gpu_indexes[j]);
@@ -651,15 +531,19 @@ __host__ void host_integer_div_rem_kb(cudaStream_t const *streams,
 
     host_unsigned_integer_div_rem_kb<Torus>(
         int_mem_ptr->sub_streams_1, gpu_indexes, gpu_count, quotient, remainder,
-        positive_numerator.data, positive_divisor.data, bsks, ksks,
-        int_mem_ptr->unsigned_mem, num_blocks);
+        positive_numerator, positive_divisor, bsks, ksks,
+        int_mem_ptr->unsigned_mem);
 
-    legacy_integer_radix_apply_bivariate_lookup_table_kb<Torus>(
+    CudaRadixCiphertextFFI numerator_sign;
+    as_radix_ciphertext_slice<Torus>(&numerator_sign, numerator, num_blocks - 1,
+                                     num_blocks);
+    CudaRadixCiphertextFFI divisor_sign;
+    as_radix_ciphertext_slice<Torus>(&divisor_sign, divisor, num_blocks - 1,
+                                     num_blocks);
+    integer_radix_apply_bivariate_lookup_table_kb<Torus>(
         int_mem_ptr->sub_streams_2, gpu_indexes, gpu_count,
-        int_mem_ptr->sign_bits_are_different,
-        &numerator[big_lwe_size * (num_blocks - 1)],
-        &divisor[big_lwe_size * (num_blocks - 1)], bsks, ksks, 1,
-        int_mem_ptr->compare_signed_bits_lut,
+        int_mem_ptr->sign_bits_are_different, &numerator_sign, &divisor_sign,
+        bsks, ksks, int_mem_ptr->compare_signed_bits_lut, 1,
         int_mem_ptr->compare_signed_bits_lut->params.message_modulus);
 
     for (uint j = 0; j < int_mem_ptr->active_gpu_count; j++) {
@@ -667,40 +551,37 @@ __host__ void host_integer_div_rem_kb(cudaStream_t const *streams,
       cuda_synchronize_stream(int_mem_ptr->sub_streams_2[j], gpu_indexes[j]);
     }
 
-    legacy_host_integer_radix_negation(
-        int_mem_ptr->sub_streams_1, gpu_indexes, gpu_count,
-        int_mem_ptr->negated_quotient, quotient, radix_params.big_lwe_dimension,
-        num_blocks, radix_params.message_modulus, radix_params.carry_modulus);
+    host_integer_radix_negation<Torus>(int_mem_ptr->sub_streams_1, gpu_indexes,
+                                       gpu_count, int_mem_ptr->negated_quotient,
+                                       quotient, radix_params.message_modulus,
+                                       radix_params.carry_modulus, num_blocks);
 
     uint32_t requested_flag = outputFlag::FLAG_NONE;
     uint32_t uses_carry = 0;
-    legacy_host_propagate_single_carry<Torus>(
-        int_mem_ptr->sub_streams_1, gpu_indexes, gpu_count,
-        int_mem_ptr->negated_quotient, nullptr, nullptr, int_mem_ptr->scp_mem_1,
-        bsks, ksks, num_blocks, requested_flag, uses_carry);
+    host_propagate_single_carry<Torus>(int_mem_ptr->sub_streams_1, gpu_indexes,
+                                       gpu_count, int_mem_ptr->negated_quotient,
+                                       nullptr, nullptr, int_mem_ptr->scp_mem_1,
+                                       bsks, ksks, requested_flag, uses_carry);
 
-    legacy_host_integer_radix_negation(
+    host_integer_radix_negation<Torus>(
         int_mem_ptr->sub_streams_2, gpu_indexes, gpu_count,
-        int_mem_ptr->negated_remainder, remainder,
-        radix_params.big_lwe_dimension, num_blocks,
-        radix_params.message_modulus, radix_params.carry_modulus);
+        int_mem_ptr->negated_remainder, remainder, radix_params.message_modulus,
+        radix_params.carry_modulus, num_blocks);
 
-    legacy_host_propagate_single_carry<Torus>(
+    host_propagate_single_carry<Torus>(
         int_mem_ptr->sub_streams_2, gpu_indexes, gpu_count,
         int_mem_ptr->negated_remainder, nullptr, nullptr,
-        int_mem_ptr->scp_mem_2, bsks, ksks, num_blocks, requested_flag,
-        uses_carry);
+        int_mem_ptr->scp_mem_2, bsks, ksks, requested_flag, uses_carry);
 
-    legacy_host_integer_radix_cmux_kb<Torus>(
+    host_integer_radix_cmux_kb<Torus>(
         int_mem_ptr->sub_streams_1, gpu_indexes, gpu_count, quotient,
         int_mem_ptr->sign_bits_are_different, int_mem_ptr->negated_quotient,
-        quotient, int_mem_ptr->cmux_quotient_mem, bsks, ksks, num_blocks);
+        quotient, int_mem_ptr->cmux_quotient_mem, bsks, ksks);
 
-    legacy_host_integer_radix_cmux_kb<Torus>(
+    host_integer_radix_cmux_kb<Torus>(
         int_mem_ptr->sub_streams_2, gpu_indexes, gpu_count, remainder,
-        &numerator[big_lwe_size * (num_blocks - 1)],
-        int_mem_ptr->negated_remainder, remainder,
-        int_mem_ptr->cmux_remainder_mem, bsks, ksks, num_blocks);
+        &numerator_sign, int_mem_ptr->negated_remainder, remainder,
+        int_mem_ptr->cmux_remainder_mem, bsks, ksks);
 
     for (uint j = 0; j < int_mem_ptr->active_gpu_count; j++) {
       cuda_synchronize_stream(int_mem_ptr->sub_streams_1[j], gpu_indexes[j]);
@@ -709,7 +590,7 @@ __host__ void host_integer_div_rem_kb(cudaStream_t const *streams,
   } else {
     host_unsigned_integer_div_rem_kb<Torus>(
         streams, gpu_indexes, gpu_count, quotient, remainder, numerator,
-        divisor, bsks, ksks, int_mem_ptr->unsigned_mem, num_blocks);
+        divisor, bsks, ksks, int_mem_ptr->unsigned_mem);
   }
 }
 

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

@@ -4,17 +4,17 @@
 
 void cuda_full_propagation_64_inplace(void *const *streams,
                                       uint32_t const *gpu_indexes,
-                                      uint32_t gpu_count, void *input_blocks,
+                                      uint32_t gpu_count,
+                                      CudaRadixCiphertextFFI *input_blocks,
                                       int8_t *mem_ptr, void *const *ksks,
                                       void *const *bsks, uint32_t num_blocks) {
 
   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);
+  host_full_propagate_inplace<uint64_t>((cudaStream_t *)(streams), gpu_indexes,
+                                        gpu_count, input_blocks, buffer,
+                                        (uint64_t **)(ksks), bsks, num_blocks);
 }
 
 void scratch_cuda_full_propagation_64(
@@ -315,11 +315,9 @@ void cuda_integer_compute_prefix_sum_hillis_steele_64(
     CudaRadixCiphertextFFI *generates_or_propagates, int8_t *mem_ptr,
     void *const *ksks, void *const *bsks, uint32_t num_radix_blocks) {
 
-  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, output_radix_lwe,
-      generates_or_propagates, params, (int_radix_lut<uint64_t> *)mem_ptr, bsks,
+      generates_or_propagates, (int_radix_lut<uint64_t> *)mem_ptr, bsks,
       (uint64_t **)(ksks), num_radix_blocks);
 }
 
@@ -338,3 +336,20 @@ void cuda_integer_reverse_blocks_64_inplace(void *const *streams,
   host_radix_blocks_reverse_inplace<uint64_t>((cudaStream_t *)(streams),
                                               gpu_indexes, lwe_array);
 }
+
+void reverseArray(uint64_t arr[], size_t n) {
+  size_t start = 0;
+  size_t end = n - 1;
+
+  // Swap elements from the start with elements from the end
+  while (start < end) {
+    // Swap arr[start] and arr[end]
+    uint64_t temp = arr[start];
+    arr[start] = arr[end];
+    arr[end] = temp;
+
+    // Move towards the middle
+    start++;
+    end--;
+  }
+}

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

@@ -13,7 +13,7 @@
  * as output ids, -1 value as an input id means that zero ciphertext will be
  * copied on output index.
  */
-void generate_ids_update_degrees(int *terms_degree, size_t *h_lwe_idx_in,
+void generate_ids_update_degrees(uint64_t *terms_degree, size_t *h_lwe_idx_in,
                                  size_t *h_lwe_idx_out,
                                  int32_t *h_smart_copy_in,
                                  int32_t *h_smart_copy_out, size_t ch_amount,
@@ -127,66 +127,53 @@ void scratch_cuda_integer_mult_radix_ciphertext_kb_64(
  */
 void cuda_integer_mult_radix_ciphertext_kb_64(
     void *const *streams, uint32_t const *gpu_indexes, uint32_t gpu_count,
-    void *radix_lwe_out, void const *radix_lwe_left, bool const is_bool_left,
-    void const *radix_lwe_right, bool const is_bool_right, void *const *bsks,
-    void *const *ksks, int8_t *mem_ptr, uint32_t polynomial_size,
-    uint32_t num_blocks) {
+    CudaRadixCiphertextFFI *radix_lwe_out,
+    CudaRadixCiphertextFFI const *radix_lwe_left, bool const is_bool_left,
+    CudaRadixCiphertextFFI const *radix_lwe_right, bool const is_bool_right,
+    void *const *bsks, void *const *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, AmortizedDegree<256>>(
-        (cudaStream_t *)(streams), gpu_indexes, gpu_count,
-        static_cast<uint64_t *>(radix_lwe_out),
-        static_cast<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   case 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<const uint64_t *>(radix_lwe_left), is_bool_left,
-        static_cast<const uint64_t *>(radix_lwe_right), is_bool_right, bsks,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_left, is_bool_left, radix_lwe_right, is_bool_right, bsks,
         (uint64_t **)(ksks), (int_mul_memory<uint64_t> *)mem_ptr, num_blocks);
     break;
   default:
@@ -226,79 +213,77 @@ void scratch_cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
 
 void cuda_integer_radix_partial_sum_ciphertexts_vec_kb_64(
     void *const *streams, uint32_t const *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 *const *bsks, void *const *ksks,
-    uint32_t num_blocks_in_radix) {
+    CudaRadixCiphertextFFI *radix_lwe_out,
+    CudaRadixCiphertextFFI *radix_lwe_vec, int8_t *mem_ptr, void *const *bsks,
+    void *const *ksks) {
 
   auto mem = (int_sum_ciphertexts_vec_memory<uint64_t> *)mem_ptr;
-
-  int *terms_degree =
-      (int *)malloc(num_blocks_in_radix * num_radix_in_vec * sizeof(int));
-
-  for (int i = 0; i < num_radix_in_vec * num_blocks_in_radix; i++) {
-    terms_degree[i] = mem->params.message_modulus - 1;
+  if (radix_lwe_vec->num_radix_blocks % radix_lwe_out->num_radix_blocks != 0)
+    PANIC("Cuda error: input vector length should be a multiple of the "
+          "output's number of radix blocks")
+  // FIXME: this should not be necessary, we should make sure sum_ctxt works in
+  // the general case
+  for (int i = 0; i < radix_lwe_vec->num_radix_blocks; i++) {
+    radix_lwe_vec->degrees[i] = mem->params.message_modulus - 1;
   }
-
   switch (mem->params.polynomial_size) {
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   case 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, bsks,
-        (uint64_t **)(ksks), mem, num_blocks_in_radix, num_radix_in_vec,
+        (cudaStream_t *)(streams), gpu_indexes, gpu_count, radix_lwe_out,
+        radix_lwe_vec, bsks, (uint64_t **)(ksks), mem,
+        radix_lwe_out->num_radix_blocks,
+        radix_lwe_vec->num_radix_blocks / radix_lwe_out->num_radix_blocks,
         nullptr);
     break;
   default:
     PANIC("Cuda error (integer multiplication): unsupported polynomial size. "
           "Supported N's are powers of two in the interval [256..16384].")
   }
-
-  free(terms_degree);
 }
 
 void cleanup_cuda_integer_radix_partial_sum_ciphertexts_vec(