feat: allow for reduced n srs for verification (#716)

.github/workflows/large-tests.yml

@@ -11,7 +11,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: nanoGPT Mock

.github/workflows/release.yml

@@ -45,7 +45,7 @@ jobs:
     steps:
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: Checkout repo

.github/workflows/rust.yml

@@ -26,7 +26,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: Build
@@ -38,7 +38,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: Docs
@@ -50,7 +50,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -73,7 +73,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -106,7 +106,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -139,7 +139,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -172,7 +172,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -189,7 +189,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: jetli/wasm-pack-action@v0.4.0
@@ -198,10 +198,8 @@ jobs:
         #   chromedriver-version: "115.0.5790.102"
       - name: Install wasm32-unknown-unknown
         run: rustup target add wasm32-unknown-unknown
-      - name: Install wasm runner
-        run: cargo install wasm-server-runner
       - name: Add rust-src
-        run: rustup component add rust-src --toolchain nightly-2023-08-24-x86_64-unknown-linux-gnu
+        run: rustup component add rust-src --toolchain nightly-2024-01-04-x86_64-unknown-linux-gnu
       - name: Run wasm verifier tests
         # on mac:
         #   AR=/opt/homebrew/opt/llvm/bin/llvm-ar CC=/opt/homebrew/opt/llvm/bin/clang wasm-pack test --firefox --headless -- -Z build-std="panic_abort,std" --features web
@@ -214,7 +212,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -231,7 +229,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -286,7 +284,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -312,7 +310,7 @@ jobs:
       - name: Install solc
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Install Anvil
-        run: cargo install --git https://github.com/foundry-rs/foundry --rev 95a93cd397f25f3f8d49d2851eb52bc2d52dd983 --profile local --locked anvil --force
+        run: cargo install --git https://github.com/foundry-rs/foundry --rev b320f350156a0fb15c2eb13dc380deb2367c4474 --profile local --locked anvil --force
       - name: KZG prove and verify tests (EVM + VK rendered seperately)
         run: cargo nextest run --release --verbose tests_evm::kzg_evm_prove_and_verify_render_seperately_ --test-threads 1
       - name: KZG prove and verify tests (EVM + kzg all)
@@ -345,18 +343,15 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: jetli/wasm-pack-action@v0.4.0
       - name: Add wasm32-unknown-unknown target
         run: rustup target add wasm32-unknown-unknown
 
-      - name: Install wasm-server-runner
-        run: cargo install wasm-server-runner
-
       - name: Add rust-src
-        run: rustup component add rust-src --toolchain nightly-2023-08-24-x86_64-unknown-linux-gnu
+        run: rustup component add rust-src --toolchain nightly-2024-01-04-x86_64-unknown-linux-gnu
       - uses: actions/checkout@v3
       - name: Use pnpm 8
         uses: pnpm/action-setup@v2
@@ -416,11 +411,11 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: Add rust-src
-        run: rustup component add rust-src --toolchain nightly-2023-08-24-x86_64-unknown-linux-gnu
+        run: rustup component add rust-src --toolchain nightly-2024-01-04-x86_64-unknown-linux-gnu
       - uses: actions/checkout@v3
       - uses: baptiste0928/cargo-install@v1
         with:
@@ -450,7 +445,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -460,7 +455,7 @@ jobs:
       - name: Install solc
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Install Anvil
-        run: cargo install --git https://github.com/foundry-rs/foundry --rev 95a93cd397f25f3f8d49d2851eb52bc2d52dd983 --profile local --locked anvil --force
+        run: cargo install --git https://github.com/foundry-rs/foundry --rev b320f350156a0fb15c2eb13dc380deb2367c4474 --profile local --locked anvil --force
       - name: fuzz tests (EVM)
         run: cargo nextest run --release --verbose tests_evm::kzg_evm_fuzz_ --test-threads 2
       # - name: fuzz tests
@@ -473,7 +468,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -491,7 +486,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -508,7 +503,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -525,7 +520,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -535,7 +530,7 @@ jobs:
       - name: Install solc
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Install Anvil
-        run: cargo install --git https://github.com/foundry-rs/foundry --rev 95a93cd397f25f3f8d49d2851eb52bc2d52dd983 --profile local --locked anvil --force
+        run: cargo install --git https://github.com/foundry-rs/foundry --rev b320f350156a0fb15c2eb13dc380deb2367c4474 --profile local --locked anvil --force
       - name: KZG prove and verify aggr tests
         run: cargo nextest run  --release --verbose tests_evm::kzg_evm_aggr_prove_and_verify_::t --test-threads 4 -- --include-ignored
 
@@ -546,7 +541,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -568,7 +563,7 @@ jobs:
           python-version: "3.7"
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - name: Install solc
@@ -576,7 +571,7 @@ jobs:
       - name: Setup Virtual Env and Install python dependencies
         run: python -m venv .env; source .env/bin/activate; pip install -r requirements.txt;
       - name: Install Anvil
-        run: cargo install --git https://github.com/foundry-rs/foundry --rev 95a93cd397f25f3f8d49d2851eb52bc2d52dd983 --profile local --locked anvil --force
+        run: cargo install --git https://github.com/foundry-rs/foundry --rev b320f350156a0fb15c2eb13dc380deb2367c4474 --profile local --locked anvil --force
       - name: Build python ezkl
         run: source .env/bin/activate; maturin develop --features python-bindings --release
       - name: Run pytest
@@ -592,7 +587,7 @@ jobs:
           python-version: "3.7"
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -623,7 +618,7 @@ jobs:
           python-version: "3.9"
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: baptiste0928/cargo-install@v1
@@ -633,7 +628,7 @@ jobs:
       - name: Install solc
         run: (hash svm 2>/dev/null || cargo install svm-rs) && svm install 0.8.20 && solc --version
       - name: Install Anvil
-        run: cargo install --git https://github.com/foundry-rs/foundry --rev 95a93cd397f25f3f8d49d2851eb52bc2d52dd983 --profile local --locked anvil --force
+        run: cargo install --git https://github.com/foundry-rs/foundry --rev b320f350156a0fb15c2eb13dc380deb2367c4474 --profile local --locked anvil --force
       - name: Setup Virtual Env and Install python dependencies
         run: python -m venv .env; source .env/bin/activate; pip install -r requirements.txt;
       - name: Build python ezkl

.github/workflows/wasm.yml

@@ -22,18 +22,15 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions-rs/toolchain@v1
         with:
-          toolchain: nightly-2023-08-24
+          toolchain: nightly-2024-01-04
           override: true
           components: rustfmt, clippy
       - uses: jetli/wasm-pack-action@v0.4.0
       - name: Add wasm32-unknown-unknown target
         run: rustup target add wasm32-unknown-unknown
 
-      - name: Install wasm-server-runner
-        run: cargo install wasm-server-runner
-
       - name: Add rust-src
-        run: rustup component add rust-src --toolchain nightly-2023-08-24-x86_64-unknown-linux-gnu
+        run: rustup component add rust-src --toolchain nightly-2024-01-04-x86_64-unknown-linux-gnu
       - name: Install binaryen
         run: |
             set -e

Cargo.toml

@@ -17,7 +17,7 @@ crate-type = ["cdylib", "rlib"]
 [dependencies]
 halo2_gadgets = { git = "https://github.com/zkonduit/halo2", branch= "main" }
 halo2_proofs = { git = "https://github.com/zkonduit/halo2", branch= "main"  }
-halo2curves = { version = "0.6.0", features = ["derive_serde"] }
+halo2curves = { git = "https://github.com/privacy-scaling-explorations/halo2curves", rev="9fff22c", features=["derive_serde"] }
 rand = { version = "0.8", default_features = false }
 itertools = { version = "0.10.3", default_features = false }
 clap = { version = "4.3.3", features = ["derive"]}
@@ -34,10 +34,11 @@ bincode = { version = "1.3.3", default_features = false }
 ark-std = { version = "^0.3.0", default-features = false }
 unzip-n = "0.1.2"
 num = "0.4.1"
+portable-atomic = "1.6.0"
 
 # evm related deps
 [target.'cfg(not(target_arch = "wasm32"))'.dependencies]
-ethers = { version = "2.0.7", default_features = false, features = ["ethers-solc"] }
+ethers = { version = "2.0.11", default_features = false, features = ["ethers-solc"] }
 indicatif = {version = "0.17.5", features = ["rayon"]}
 gag =  { version = "1.0.0", default_features = false}
 instant = { version = "0.1" }

examples/conv2d_mnist/main.rs

@@ -6,6 +6,7 @@ use ezkl::fieldutils;
 use ezkl::fieldutils::i32_to_felt;
 use ezkl::tensor::*;
 use halo2_proofs::dev::MockProver;
+use halo2_proofs::poly::commitment::Params;
 use halo2_proofs::poly::kzg::multiopen::{ProverSHPLONK, VerifierSHPLONK};
 use halo2_proofs::{
     circuit::{Layouter, SimpleFloorPlanner, Value},
@@ -489,6 +490,7 @@ pub fn runconv() {
         strategy,
         pi_for_real_prover,
         &mut transcript,
+        params.n(),
     );
     assert!(verify.is_ok());
 

examples/onnx/1l_tiny_div/gen.py

@@ -0,0 +1,39 @@
+from torch import nn
+import torch
+import json
+
+class Circuit(nn.Module):
+    def __init__(self, inplace=False):
+        super(Circuit, self).__init__()
+
+    def forward(self, x):
+        return x/ 10000
+
+
+circuit = Circuit()
+
+
+x = torch.empty(1, 8).random_(0, 2)
+
+out = circuit(x)
+
+print(out)
+
+torch.onnx.export(circuit, x, "network.onnx",
+                  export_params=True,        # store the trained parameter weights inside the model file
+                  opset_version=17,          # the ONNX version to export the model to
+                  do_constant_folding=True,  # whether to execute constant folding for optimization
+                  input_names=['input'],   # the model's input names
+                  output_names=['output'],  # the model's output names
+                  dynamic_axes={'input': {0: 'batch_size'},  # variable length axes
+                                'output': {0: 'batch_size'}})
+
+
+d1 = ((x).detach().numpy()).reshape([-1]).tolist()
+
+data = dict(
+    input_data=[d1],
+)
+
+# Serialize data into file:
+json.dump(data, open("input.json", 'w'))

examples/onnx/1l_tiny_div/input.json

@@ -0,0 +1 @@
+{"input_data": [[1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 0.0, 1.0]]}

src/circuit/modules/kzg.rs

@@ -41,7 +41,7 @@ pub struct KZGChip {
 }
 
 impl KZGChip {
-    /// Returns the number of inputs to the hash function
+    /// Commit to the message using the KZG commitment scheme
     pub fn commit(
         message: Vec<Fp>,
         degree: u32,

src/circuit/ops/base.rs

@@ -125,8 +125,8 @@ impl BaseOp {
             BaseOp::Sum => 1,
             BaseOp::SumInit => 1,
             BaseOp::Range { .. } => 1,
-            BaseOp::IsZero => 1,
-            BaseOp::IsBoolean => 1,
+            BaseOp::IsZero => 0,
+            BaseOp::IsBoolean => 0,
         }
     }
 

src/circuit/ops/chip.rs

@@ -276,9 +276,20 @@ impl<F: PrimeField + TensorType + PartialOrd> BaseConfig<F> {
 
                 let constraints = match base_op {
                     BaseOp::IsBoolean => {
-                        vec![(qis[1].clone()) * (qis[1].clone() - Expression::Constant(F::from(1)))]
+                        let expected_output: Tensor<Expression<F>> = output
+                            .query_rng(meta, *block_idx, *inner_col_idx, 0, 1)
+                            .expect("non accum: output query failed");
+
+                        let output = expected_output[base_op.constraint_idx()].clone();
+
+                        vec![(output.clone()) * (output.clone() - Expression::Constant(F::from(1)))]
+                    }
+                    BaseOp::IsZero => {
+                        let expected_output: Tensor<Expression<F>> = output
+                            .query_rng(meta, *block_idx, *inner_col_idx, 0, 1)
+                            .expect("non accum: output query failed");
+                        vec![expected_output[base_op.constraint_idx()].clone()]
                     }
-                    BaseOp::IsZero => vec![qis[1].clone()],
                     _ => {
                         let expected_output: Tensor<Expression<F>> = output
                             .query_rng(meta, *block_idx, *inner_col_idx, rotation_offset, rng)
@@ -512,10 +523,10 @@ impl<F: PrimeField + TensorType + PartialOrd> BaseConfig<F> {
 
         // we borrow mutably twice so we need to do this dance
 
-        let range_check = if !self.range_checks.contains_key(&range) {
+        let range_check = if let std::collections::btree_map::Entry::Vacant(e) = self.range_checks.entry(range) {
             // as all tables have the same input we see if there's another table who's input we can reuse
             let range_check = RangeCheck::<F>::configure(cs, range);
-            self.range_checks.insert(range, range_check.clone());
+            e.insert(range_check.clone());
             range_check
         } else {
             return Ok(());

src/circuit/ops/hybrid.rs

@@ -1,11 +1,11 @@
 use super::*;
 use crate::{
-    circuit::{self, layouts, utils, Tolerance},
+    circuit::{layouts, utils, Tolerance},
     fieldutils::{felt_to_i128, i128_to_felt},
+    graph::multiplier_to_scale,
     tensor::{self, Tensor, TensorError, TensorType, ValTensor},
 };
 use halo2curves::ff::PrimeField;
-use itertools::Itertools;
 use serde::{Deserialize, Serialize};
 // import run args from model
 
@@ -13,6 +13,15 @@ use serde::{Deserialize, Serialize};
 /// An enum representing the operations that consist of both lookups and arithmetic operations.
 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub enum HybridOp {
+    Recip {
+        input_scale: utils::F32,
+        output_scale: utils::F32,
+        use_range_check_for_int: bool,
+    },
+    Div {
+        denom: utils::F32,
+        use_range_check_for_int: bool,
+    },
     ReduceMax {
         axes: Vec<usize>,
     },
@@ -59,14 +68,6 @@ pub enum HybridOp {
         dim: usize,
         num_classes: usize,
     },
-    GatherElements {
-        dim: usize,
-        constant_idx: Option<Tensor<usize>>,
-    },
-    ScatterElements {
-        dim: usize,
-        constant_idx: Option<Tensor<usize>>,
-    },
 }
 
 impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
@@ -74,7 +75,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
     fn requires_homogenous_input_scales(&self) -> Vec<usize> {
         match self {
             HybridOp::Greater | HybridOp::Less | HybridOp::Equals => vec![0, 1],
-            HybridOp::ScatterElements { .. } => vec![0, 2],
+            HybridOp::GreaterEqual | HybridOp::LessEqual => vec![0, 1],
             _ => vec![],
         }
     }
@@ -87,142 +88,42 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
     fn f(&self, inputs: &[Tensor<F>]) -> Result<ForwardResult<F>, TensorError> {
         let x = inputs[0].clone().map(|x| felt_to_i128(x));
 
-        let (res, intermediate_lookups) = match &self {
-            HybridOp::ReduceMax { axes, .. } => {
-                let res = tensor::ops::max_axes(&x, axes)?;
-                let max_minus_one =
-                    Tensor::from(vec![x.clone().into_iter().max().unwrap() - 1].into_iter());
-                let unit = Tensor::from(vec![1].into_iter());
-                // relu(x - max(x - 1)
-                let inter_1 = (x.clone() - max_minus_one)?;
-                // relu(1 - sum(relu(inter_1)))
-                let inter_2 = (unit
-                    - tensor::ops::sum(&tensor::ops::nonlinearities::leakyrelu(&inter_1, 0.0))?)?;
-
-                (res.clone(), vec![inter_1, inter_2])
-            }
-            HybridOp::ReduceMin { axes, .. } => {
-                let res = tensor::ops::min_axes(&x, axes)?;
-                let min_plus_one =
-                    Tensor::from(vec![x.clone().into_iter().min().unwrap() + 1].into_iter());
-                let unit = Tensor::from(vec![1].into_iter());
-                // relu(min(x + 1) - x)
-                let inter_1 = (min_plus_one - x.clone())?;
-                // relu(1 - sum(relu(inter_1)))
-                let inter_2 = (unit
-                    - tensor::ops::sum(&tensor::ops::nonlinearities::leakyrelu(&inter_1, 0.0))?)?;
-                (res.clone(), vec![inter_1, inter_2])
-            }
-            HybridOp::ReduceArgMax { dim } => {
-                let res = tensor::ops::argmax_axes(&x, *dim)?;
-                let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                let mut inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                let inter =
-                    Op::f(&HybridOp::ReduceMax { axes: vec![*dim] }, inputs)?.intermediate_lookups;
-                inter_equals.extend(inter);
-
-                (res.clone(), inter_equals)
-            }
-            HybridOp::ReduceArgMin { dim } => {
-                let res = tensor::ops::argmin_axes(&x, *dim)?;
-                let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                let mut inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                let inter =
-                    Op::f(&HybridOp::ReduceMin { axes: vec![*dim] }, inputs)?.intermediate_lookups;
-                inter_equals.extend(inter);
-
-                (res.clone(), inter_equals)
+        let res = match &self {
+            HybridOp::ReduceMax { axes, .. } => tensor::ops::max_axes(&x, axes)?,
+            HybridOp::ReduceMin { axes, .. } => tensor::ops::min_axes(&x, axes)?,
+            HybridOp::Div { denom, .. } => {
+                crate::tensor::ops::nonlinearities::const_div(&x, denom.0 as f64)
             }
+            HybridOp::Recip {
+                input_scale,
+                output_scale,
+                ..
+            } => crate::tensor::ops::nonlinearities::recip(
+                &x,
+                input_scale.0 as f64,
+                output_scale.0 as f64,
+            ),
+            HybridOp::ReduceArgMax { dim } => tensor::ops::argmax_axes(&x, *dim)?,
+            HybridOp::ReduceArgMin { dim } => tensor::ops::argmin_axes(&x, *dim)?,
             HybridOp::Gather { dim, constant_idx } => {
                 if let Some(idx) = constant_idx {
-                    log::debug!("idx: {}", idx.show());
-                    let res = tensor::ops::gather(&x, idx, *dim)?;
-                    (res.clone(), vec![])
+                    tensor::ops::gather(&x, idx, *dim)?
                 } else {
                     let y = inputs[1].clone().map(|x| felt_to_i128(x));
-                    let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                    let inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                    let res = tensor::ops::gather(&x, &y.map(|x| x as usize), *dim)?;
-                    (res.clone(), inter_equals)
+                    tensor::ops::gather(&x, &y.map(|x| x as usize), *dim)?
                 }
             }
             HybridOp::OneHot { dim, num_classes } => {
-                let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                let inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                let res = tensor::ops::one_hot(&x, *num_classes, *dim)?;
-                (res.clone(), inter_equals)
+                tensor::ops::one_hot(&x, *num_classes, *dim)?.clone()
             }
-            HybridOp::TopK { dim, k, largest } => {
-                let res = tensor::ops::topk_axes(&x, *k, *dim, *largest)?;
 
-                let mut inter_equals = x
-                    .clone()
-                    .into_iter()
-                    .flat_map(|elem| {
-                        tensor::ops::equals(&res, &vec![elem].into_iter().into())
-                            .unwrap()
-                            .1
-                    })
-                    .collect::<Vec<_>>();
-
-                // sort in descending order and take pairwise differences
-                inter_equals.push(
-                    x.into_iter()
-                        .sorted()
-                        .tuple_windows()
-                        .map(|(a, b)| b - a)
-                        .into(),
-                );
-
-                (res.clone(), inter_equals)
-            }
-            HybridOp::GatherElements { dim, constant_idx } => {
-                if let Some(idx) = constant_idx {
-                    log::debug!("idx: {}", idx.show());
-                    let res = tensor::ops::gather_elements(&x, idx, *dim)?;
-                    (res.clone(), vec![])
-                } else {
-                    let y = inputs[1].clone().map(|x| felt_to_i128(x));
-                    let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                    let inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                    let res = tensor::ops::gather_elements(&x, &y.map(|x| x as usize), *dim)?;
-                    (res.clone(), inter_equals)
-                }
-            }
-            HybridOp::ScatterElements { dim, constant_idx } => {
-                if let Some(idx) = constant_idx {
-                    log::debug!("idx: {}", idx.show());
-                    let src = inputs[1].clone().map(|x| felt_to_i128(x));
-                    let res = tensor::ops::scatter(&x, idx, &src, *dim)?;
-                    (res.clone(), vec![])
-                } else {
-                    let idx = inputs[1].clone().map(|x| felt_to_i128(x) as usize);
-                    let src = inputs[2].clone().map(|x| felt_to_i128(x));
-                    let indices = Tensor::from(0..x.dims()[*dim] as i128);
-                    let inter_equals: Vec<Tensor<i128>> = vec![indices.clone(), -indices];
-                    let res = tensor::ops::scatter(&x, &idx, &src, *dim)?;
-                    (res.clone(), inter_equals)
-                }
-            }
+            HybridOp::TopK { dim, k, largest } => tensor::ops::topk_axes(&x, *k, *dim, *largest)?,
             HybridOp::MaxPool2d {
                 padding,
                 stride,
                 pool_dims,
                 ..
-            } => {
-                let max_minus_one =
-                    Tensor::from(vec![x.clone().into_iter().max().unwrap() - 1].into_iter());
-                let unit = Tensor::from(vec![1].into_iter());
-                // relu(x - max(x - 1)
-                let inter_1 = (x.clone() - max_minus_one)?;
-                // relu(1 - sum(relu(inter_1)))
-                let inter_2 = (unit
-                    - tensor::ops::sum(&tensor::ops::nonlinearities::leakyrelu(&inter_1, 0.0))?)?;
-                (
-                    tensor::ops::max_pool2d(&x, padding, stride, pool_dims)?,
-                    vec![inter_1, inter_2],
-                )
-            }
+            } => tensor::ops::max_pool2d(&x, padding, stride, pool_dims)?,
             HybridOp::SumPool {
                 padding,
                 stride,
@@ -234,10 +135,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
             }
             HybridOp::RangeCheck(tol) => {
                 let y = inputs[1].clone().map(|x| felt_to_i128(x));
-                (
-                    tensor::ops::nonlinearities::range_check_percent(&[x, y], 128, 128, tol.val),
-                    vec![],
-                )
+                tensor::ops::nonlinearities::range_check_percent(&[x, y], 128, 128, tol.val)
             }
             HybridOp::Greater => {
                 let y = inputs[1].clone().map(|x| felt_to_i128(x));
@@ -264,14 +162,26 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
         // convert back to felt
         let output = res.map(|x| i128_to_felt(x));
 
-        Ok(ForwardResult {
-            output,
-            intermediate_lookups,
-        })
+        Ok(ForwardResult { output })
     }
 
     fn as_string(&self) -> String {
         match self {
+            HybridOp::Recip {
+                input_scale,
+                output_scale,
+                use_range_check_for_int,
+            } => format!(
+                "RECIP (input_scale={}, output_scale={}, use_range_check_for_int={})",
+                input_scale, output_scale, use_range_check_for_int
+            ),
+            HybridOp::Div {
+                denom,
+                use_range_check_for_int,
+            } => format!(
+                "DIV (denom={}, use_range_check_for_int={})",
+                denom, use_range_check_for_int
+            ),
             HybridOp::SumPool {
                 padding,
                 stride,
@@ -306,8 +216,6 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
             HybridOp::TopK { k, dim, largest } => {
                 format!("TOPK (k={}, dim={}, largest={})", k, dim, largest)
             }
-            HybridOp::GatherElements { dim, .. } => format!("GATHERELEMENTS (dim={})", dim),
-            HybridOp::ScatterElements { dim, .. } => format!("SCATTERELEMENTS (dim={})", dim),
             HybridOp::OneHot { dim, num_classes } => {
                 format!("ONEHOT (dim={}, num_classes={})", dim, num_classes)
             }
@@ -335,6 +243,55 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
                 *kernel_shape,
                 *normalized,
             )?,
+            HybridOp::Recip {
+                input_scale,
+                output_scale,
+                use_range_check_for_int,
+            } => {
+                if input_scale.0.fract() == 0.0
+                    && output_scale.0.fract() == 0.0
+                    && *use_range_check_for_int
+                {
+                    layouts::recip(
+                        config,
+                        region,
+                        values[..].try_into()?,
+                        i128_to_felt(input_scale.0 as i128),
+                        i128_to_felt(output_scale.0 as i128),
+                    )?
+                } else {
+                    layouts::nonlinearity(
+                        config,
+                        region,
+                        values.try_into()?,
+                        &LookupOp::Recip {
+                            input_scale: *input_scale,
+                            output_scale: *output_scale,
+                        },
+                    )?
+                }
+            }
+            HybridOp::Div {
+                denom,
+                use_range_check_for_int,
+                ..
+            } => {
+                if denom.0.fract() == 0.0 && *use_range_check_for_int {
+                    layouts::div(
+                        config,
+                        region,
+                        values[..].try_into()?,
+                        i128_to_felt(denom.0 as i128),
+                    )?
+                } else {
+                    layouts::nonlinearity(
+                        config,
+                        region,
+                        values.try_into()?,
+                        &LookupOp::Div { denom: *denom },
+                    )?
+                }
+            }
             HybridOp::Gather { dim, constant_idx } => {
                 if let Some(idx) = constant_idx {
                     tensor::ops::gather(values[0].get_inner_tensor()?, idx, *dim)?.into()
@@ -342,26 +299,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
                     layouts::gather(config, region, values[..].try_into()?, *dim)?
                 }
             }
-            HybridOp::GatherElements { dim, constant_idx } => {
-                if let Some(idx) = constant_idx {
-                    tensor::ops::gather_elements(values[0].get_inner_tensor()?, idx, *dim)?.into()
-                } else {
-                    layouts::gather_elements(config, region, values[..].try_into()?, *dim)?
-                }
-            }
-            HybridOp::ScatterElements { dim, constant_idx } => {
-                if let Some(idx) = constant_idx {
-                    tensor::ops::scatter(
-                        values[0].get_inner_tensor()?,
-                        idx,
-                        values[1].get_inner_tensor()?,
-                        *dim,
-                    )?
-                    .into()
-                } else {
-                    layouts::scatter_elements(config, region, values[..].try_into()?, *dim)?
-                }
-            }
+
             HybridOp::MaxPool2d {
                 padding,
                 stride,
@@ -422,86 +360,12 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for HybridOp {
             | HybridOp::OneHot { .. }
             | HybridOp::ReduceArgMin { .. } => 0,
             HybridOp::Softmax { .. } => 2 * in_scales[0],
+            HybridOp::Recip { output_scale, .. } => multiplier_to_scale(output_scale.0 as f64),
             _ => in_scales[0],
         };
         Ok(scale)
     }
 
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        match self {
-            HybridOp::ReduceMax { .. }
-            | HybridOp::ReduceMin { .. }
-            | HybridOp::MaxPool2d { .. } => Op::<F>::required_lookups(&LookupOp::ReLU),
-            HybridOp::Softmax { scale, .. } => {
-                vec![
-                    LookupOp::Exp { scale: *scale },
-                    LookupOp::Recip {
-                        scale: scale.0.powf(2.0).into(),
-                    },
-                ]
-            }
-            HybridOp::RangeCheck(tol) => {
-                let mut lookups = vec![];
-                if tol.val > 0.0 {
-                    let scale_squared = tol.scale.0.powf(2.0);
-                    lookups.extend([
-                        LookupOp::Recip {
-                            scale: scale_squared.into(),
-                        },
-                        LookupOp::GreaterThan {
-                            a: circuit::utils::F32((tol.val * scale_squared) / 100.0),
-                        },
-                    ]);
-                }
-                lookups
-            }
-            HybridOp::Greater { .. } | HybridOp::Less { .. } => {
-                vec![LookupOp::GreaterThan {
-                    a: circuit::utils::F32(0.),
-                }]
-            }
-            HybridOp::GreaterEqual { .. } | HybridOp::LessEqual { .. } => {
-                vec![LookupOp::GreaterThanEqual {
-                    a: circuit::utils::F32(0.),
-                }]
-            }
-            HybridOp::TopK { .. } => {
-                vec![
-                    LookupOp::GreaterThan {
-                        a: circuit::utils::F32(0.),
-                    },
-                    LookupOp::KroneckerDelta,
-                ]
-            }
-            HybridOp::Gather {
-                constant_idx: None, ..
-            }
-            | HybridOp::OneHot { .. }
-            | HybridOp::GatherElements {
-                constant_idx: None, ..
-            }
-            | HybridOp::ScatterElements {
-                constant_idx: None, ..
-            }
-            | HybridOp::Equals { .. } => {
-                vec![LookupOp::KroneckerDelta]
-            }
-            HybridOp::ReduceArgMax { .. } | HybridOp::ReduceArgMin { .. } => {
-                vec![LookupOp::ReLU, LookupOp::KroneckerDelta]
-            }
-            HybridOp::SumPool {
-                kernel_shape,
-                normalized: true,
-                ..
-            } => {
-                vec![LookupOp::Div {
-                    denom: utils::F32((kernel_shape.0 * kernel_shape.1) as f32),
-                }]
-            }
-            _ => vec![],
-        }
-    }
-
     fn clone_dyn(&self) -> Box<dyn Op<F>> {
         Box::new(self.clone()) // Forward to the derive(Clone) impl
     }

src/circuit/ops/layouts.rs

@@ -18,7 +18,10 @@ use super::{
     region::RegionCtx,
 };
 use crate::{
-    circuit::{ops::base::BaseOp, utils},
+    circuit::{
+        ops::base::BaseOp,
+        utils::{self},
+    },
     fieldutils::{felt_to_i128, i128_to_felt},
     tensor::{
         get_broadcasted_shape,
@@ -61,7 +64,7 @@ pub fn div<F: PrimeField + TensorType + PartialOrd>(
     let input = value[0].clone();
     let input_dims = input.dims();
 
-    let range_check_bracket = felt_to_i128(div) - 1;
+    let range_check_bracket = felt_to_i128(div) / 2;
 
     let mut divisor = Tensor::from(vec![ValType::Constant(div)].into_iter());
     divisor.set_visibility(&crate::graph::Visibility::Fixed);
@@ -72,8 +75,7 @@ pub fn div<F: PrimeField + TensorType + PartialOrd>(
 
     let mut claimed_output: ValTensor<F> = if is_assigned {
         let input_evals = input.get_int_evals()?;
-        let divisor_evals = divisor.get_int_evals()?;
-        tensor::ops::div(&[input_evals.clone(), divisor_evals.clone()])?
+        tensor::ops::nonlinearities::const_div(&input_evals.clone(), felt_to_i128(div) as f64)
             .iter()
             .map(|x| Ok(Value::known(i128_to_felt(*x))))
             .collect::<Result<Tensor<Value<F>>, Box<dyn Error>>>()?
@@ -94,6 +96,8 @@ pub fn div<F: PrimeField + TensorType + PartialOrd>(
         BaseOp::Mult,
     )?;
 
+    log::debug!("product: {:?}", product.get_int_evals()?);
+
     let diff_with_input = pairwise(
         config,
         region,
@@ -111,6 +115,64 @@ pub fn div<F: PrimeField + TensorType + PartialOrd>(
     Ok(claimed_output)
 }
 
+/// recip accumulated layout
+pub fn recip<F: PrimeField + TensorType + PartialOrd>(
+    config: &BaseConfig<F>,
+    region: &mut RegionCtx<F>,
+    value: &[ValTensor<F>; 1],
+    input_scale: F,
+    output_scale: F,
+) -> Result<ValTensor<F>, Box<dyn Error>> {
+    let input = value[0].clone();
+    let input_dims = input.dims();
+
+    let range_check_bracket = felt_to_i128(output_scale * input_scale) / 2;
+
+    let is_assigned = !input.any_unknowns()?;
+
+    let mut claimed_output: ValTensor<F> = if is_assigned {
+        let input_evals = input.get_int_evals()?;
+        tensor::ops::nonlinearities::recip(
+            &input_evals,
+            felt_to_i128(input_scale) as f64,
+            felt_to_i128(output_scale) as f64,
+        )
+        .iter()
+        .map(|x| Ok(Value::known(i128_to_felt(*x))))
+        .collect::<Result<Tensor<Value<F>>, Box<dyn Error>>>()?
+        .into()
+    } else {
+        Tensor::new(
+            Some(&vec![Value::<F>::unknown(); input.len()]),
+            &[input.len()],
+        )?
+        .into()
+    };
+    claimed_output.reshape(input_dims)?;
+
+    // this is now of scale 2 * scale
+    let product = pairwise(
+        config,
+        region,
+        &[claimed_output.clone(), input.clone()],
+        BaseOp::Mult,
+    )?;
+
+    log::debug!("product: {:?}", product.get_int_evals()?);
+
+    log::debug!("range_check_bracket: {:?}", range_check_bracket);
+
+    // at most the error should be in the original unit scale's range
+    range_check(
+        config,
+        region,
+        &[product],
+        &(range_check_bracket, 3 * range_check_bracket),
+    )?;
+
+    Ok(claimed_output)
+}
+
 /// Dot product accumulated layout
 pub fn dot<F: PrimeField + TensorType + PartialOrd>(
     config: &BaseConfig<F>,
@@ -152,7 +214,7 @@ pub fn dot<F: PrimeField + TensorType + PartialOrd>(
 
     let mut assigned_len = 0;
     for (i, input) in values.iter_mut().enumerate() {
-        input.pad_to_zero_rem(block_width)?;
+        input.pad_to_zero_rem(block_width, ValType::Constant(F::ZERO))?;
         let inp = {
             let (res, len) = region.assign_with_duplication(
                 &config.inputs[i],
@@ -708,14 +770,7 @@ fn one_hot<F: PrimeField + TensorType + PartialOrd>(
     let assigned_input = region.assign(&config.inputs[0], &input)?;
 
     // now assert all elems are 0 or 1
-    let assigned_output = region.assign(&config.inputs[1], &output)?;
-    if !region.is_dummy() {
-        for i in 0..assigned_output.len() {
-            let (x, y, z) = config.output.cartesian_coord(region.linear_coord() + i);
-            let selector = config.selectors.get(&(BaseOp::IsBoolean, x, y));
-            region.enable(selector, z)?;
-        }
-    }
+    let assigned_output = boolean_identity(config, region, &[output.clone()], true)?;
     region.increment(std::cmp::max(assigned_output.len(), assigned_input.len()));
 
     let sum = sum(config, region, &[assigned_output.clone()])?;
@@ -1087,7 +1142,7 @@ pub fn sum<F: PrimeField + TensorType + PartialOrd>(
     let assigned_len: usize;
     let input = {
         let mut input = values[0].clone();
-        input.pad_to_zero_rem(block_width)?;
+        input.pad_to_zero_rem(block_width, ValType::Constant(F::ZERO))?;
         let (res, len) =
             region.assign_with_duplication(&config.inputs[1], &input, &config.check_mode, false)?;
         assigned_len = len;
@@ -1156,7 +1211,7 @@ pub fn prod<F: PrimeField + TensorType + PartialOrd>(
     let assigned_len: usize;
     let input = {
         let mut input = values[0].clone();
-        input.pad_to_zero_rem(block_width)?;
+        input.pad_to_zero_rem(block_width, ValType::Constant(F::ONE))?;
         let (res, len) =
             region.assign_with_duplication(&config.inputs[1], &input, &config.check_mode, false)?;
         assigned_len = len;
@@ -1620,10 +1675,28 @@ pub fn equals<F: PrimeField + TensorType + PartialOrd>(
     values: &[ValTensor<F>; 2],
 ) -> Result<ValTensor<F>, Box<dyn Error>> {
     let diff = pairwise(config, region, values, BaseOp::Sub)?;
+    let diff_inverse = diff.inverse()?;
+    let product_diff_and_invert =
+        pairwise(config, region, &[diff.clone(), diff_inverse], BaseOp::Mult)?;
 
-    let res = nonlinearity(config, region, &[diff], &LookupOp::KroneckerDelta)?;
+    // constant of 1
+    let mut ones = Tensor::from(vec![ValType::Constant(F::from(1))].into_iter());
+    ones.set_visibility(&crate::graph::Visibility::Fixed);
 
-    Ok(res)
+    // subtract
+    let output = pairwise(
+        config,
+        region,
+        &[ones.into(), product_diff_and_invert],
+        BaseOp::Sub,
+    )?;
+
+    // take the product of diff and output
+    let prod_check = pairwise(config, region, &[diff, output.clone()], BaseOp::Mult)?;
+
+    is_zero_identity(config, region, &[prod_check], false)?;
+
+    Ok(output)
 }
 
 /// Xor boolean operation
@@ -1687,21 +1760,7 @@ pub fn iff<F: PrimeField + TensorType + PartialOrd>(
             .into();
 
     // make sure mask is boolean
-    let assigned_mask = region.assign(&config.inputs[1], mask)?;
-
-    // Enable the selectors
-    if !region.is_dummy() {
-        (0..assigned_mask.len())
-            .map(|i| {
-                let (x, y, z) = config.inputs[1].cartesian_coord(region.linear_coord() + i);
-                let selector = config.selectors.get(&(BaseOp::IsBoolean, x, y));
-                region.enable(selector, z)?;
-                Ok(())
-            })
-            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
-    }
-
-    region.increment(assigned_mask.len());
+    let assigned_mask = boolean_identity(config, region, &[mask.clone()], true)?;
 
     let one_minus_mask = pairwise(config, region, &[unit, assigned_mask.clone()], BaseOp::Sub)?;
 
@@ -1799,13 +1858,11 @@ pub fn sumpool<F: PrimeField + TensorType + PartialOrd>(
     last_elem.reshape(&[&[batch_size, image_channels], shape].concat())?;
 
     if normalized {
-        last_elem = nonlinearity(
+        last_elem = div(
             config,
             region,
             &[last_elem],
-            &LookupOp::Div {
-                denom: utils::F32((kernel_shape.0 * kernel_shape.1) as f32),
-            },
+            F::from((kernel_shape.0 * kernel_shape.1) as u64),
         )?;
     }
     Ok(last_elem)
@@ -2302,18 +2359,60 @@ pub fn identity<F: PrimeField + TensorType + PartialOrd>(
     Ok(output)
 }
 
+/// is zero identity constraint. Usually used to constrain an instance column to an advice so the returned cells / values can be operated upon.
+pub fn is_zero_identity<F: PrimeField + TensorType + PartialOrd>(
+    config: &BaseConfig<F>,
+    region: &mut RegionCtx<F>,
+    values: &[ValTensor<F>; 1],
+    assign: bool,
+) -> Result<ValTensor<F>, Box<dyn Error>> {
+    let output = if assign || !values[0].get_const_indices()?.is_empty() {
+        let output = region.assign(&config.output, &values[0])?;
+        region.increment(output.len());
+        output
+    } else {
+        values[0].clone()
+    };
+    // Enable the selectors
+    if !region.is_dummy() {
+        (0..output.len())
+            .map(|j| {
+                let index = region.linear_coord() - j - 1;
+
+                let (x, y, z) = config.output.cartesian_coord(index);
+                let selector = config.selectors.get(&(BaseOp::IsZero, x, y));
+
+                region.enable(selector, z)?;
+                Ok(())
+            })
+            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
+    }
+
+    Ok(output)
+}
+
 /// Boolean identity constraint. Usually used to constrain an instance column to an advice so the returned cells / values can be operated upon.
 pub fn boolean_identity<F: PrimeField + TensorType + PartialOrd>(
     config: &BaseConfig<F>,
     region: &mut RegionCtx<F>,
     values: &[ValTensor<F>; 1],
+    assign: bool,
 ) -> Result<ValTensor<F>, Box<dyn Error>> {
-    let output = region.assign(&config.inputs[1], &values[0])?;
+    let output = if assign || !values[0].get_const_indices()?.is_empty() {
+        // get zero constants indices
+        let output = region.assign(&config.output, &values[0])?;
+        region.increment(output.len());
+        output
+    } else {
+        values[0].clone()
+    };
     // Enable the selectors
     if !region.is_dummy() {
         (0..output.len())
             .map(|j| {
-                let (x, y, z) = config.inputs[1].cartesian_coord(region.linear_coord() + j);
+                let index = region.linear_coord() - j - 1;
+
+                let (x, y, z) = config.output.cartesian_coord(index);
                 let selector = config.selectors.get(&(BaseOp::IsBoolean, x, y));
 
                 region.enable(selector, z)?;
@@ -2321,7 +2420,6 @@ pub fn boolean_identity<F: PrimeField + TensorType + PartialOrd>(
             })
             .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
     }
-    region.increment(output.len());
 
     Ok(output)
 }
@@ -2371,6 +2469,8 @@ pub fn range_check<F: PrimeField + TensorType + PartialOrd>(
     values: &[ValTensor<F>; 1],
     range: &crate::circuit::table::Range,
 ) -> Result<ValTensor<F>, Box<dyn Error>> {
+    region.add_used_range_check(*range)?;
+
     // time the entire operation
     let timer = instant::Instant::now();
 
@@ -2388,7 +2488,7 @@ pub fn range_check<F: PrimeField + TensorType + PartialOrd>(
                 let (x, y, z) = config
                     .lookup_input
                     .cartesian_coord(region.linear_coord() + i);
-                let selector = config.range_check_selectors.get(&(range.clone(), x, y));
+                let selector = config.range_check_selectors.get(&(*range, x, y));
                 region.enable(selector, z)?;
                 Ok(())
             })
@@ -2415,6 +2515,8 @@ pub fn nonlinearity<F: PrimeField + TensorType + PartialOrd>(
     values: &[ValTensor<F>; 1],
     nl: &LookupOp,
 ) -> Result<ValTensor<F>, Box<dyn Error>> {
+    region.add_used_lookup(nl.clone(), values)?;
+
     // time the entire operation
     let timer = instant::Instant::now();
 
@@ -2496,22 +2598,6 @@ pub fn nonlinearity<F: PrimeField + TensorType + PartialOrd>(
     Ok(output)
 }
 
-/// mean function layout
-pub fn mean<F: PrimeField + TensorType + PartialOrd>(
-    config: &BaseConfig<F>,
-    region: &mut RegionCtx<F>,
-    values: &[ValTensor<F>; 1],
-    scale: usize,
-) -> Result<ValTensor<F>, Box<dyn Error>> {
-    let x = &values[0];
-
-    let sum_x = sum(config, region, &[x.clone()])?;
-    let nl = LookupOp::Div {
-        denom: utils::F32((scale * x.len()) as f32),
-    };
-    nonlinearity(config, region, &[sum_x], &nl)
-}
-
 /// Argmax
 pub fn argmax<F: PrimeField + TensorType + PartialOrd>(
     config: &BaseConfig<F>,
@@ -2624,24 +2710,8 @@ pub fn max<F: PrimeField + TensorType + PartialOrd>(
     )?;
     // relu(x - max(x - 1))
     let relu = nonlinearity(config, region, &[diff], &LookupOp::ReLU)?;
-
-    let len = relu.dims().iter().product();
-
-    // y_i*(1 - y_i) =0 // assert the values are either 0 or 1
-    region.assign(&config.inputs[1], &relu)?;
-
-    if !region.is_dummy() {
-        (0..len)
-            .map(|i| {
-                let (x, y, z) = config.inputs[1].cartesian_coord(region.linear_coord() + i);
-                let selector = config.selectors.get(&(BaseOp::IsBoolean, x, y));
-                region.enable(selector, z)?;
-                Ok(())
-            })
-            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
-    }
-
-    region.increment(len);
+    // constraining relu(x - max(x - 1)) = 0/1
+    boolean_identity(config, region, &[relu.clone()], false)?;
 
     // sum(relu(x - max(x - 1)))
     let sum_relu = sum(config, region, &[relu])?;
@@ -2652,13 +2722,7 @@ pub fn max<F: PrimeField + TensorType + PartialOrd>(
         nonlinearity(config, region, &[one_minus_sum_relu], &LookupOp::ReLU)?;
 
     // constraining 1 - sum(relu(x - max(x - 1))) = 0
-    region.assign(&config.inputs[1], &relu_one_minus_sum_relu)?;
-
-    let (x, y, z) = config.output.cartesian_coord(region.linear_coord());
-    let selector = config.selectors.get(&(BaseOp::IsZero, x, y));
-    region.enable(selector, z)?;
-
-    region.increment(relu_one_minus_sum_relu.len());
+    is_zero_identity(config, region, &[relu_one_minus_sum_relu], false)?;
 
     Ok(assigned_max_val)
 }
@@ -2703,23 +2767,8 @@ pub fn min<F: PrimeField + TensorType + PartialOrd>(
 
     // relu(min(x + 1)  - x)
     let relu = nonlinearity(config, region, &[diff], &LookupOp::ReLU)?;
-
-    let len = relu.dims().iter().product();
-
-    region.assign(&config.inputs[1], &relu)?;
-    // y_i*(1 - y_i) =0 // assert the values are either 0 or 1
-    if !region.is_dummy() {
-        (0..len)
-            .map(|i| {
-                let (x, y, z) = config.inputs[1].cartesian_coord(region.linear_coord() + i);
-                let selector = config.selectors.get(&(BaseOp::IsBoolean, x, y));
-                region.enable(selector, z)?;
-                Ok(())
-            })
-            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
-    }
-
-    region.increment(len);
+    // constraining relu(min(x + 1) - x) = 0/1
+    boolean_identity(config, region, &[relu.clone()], false)?;
 
     // sum(relu(min(x + 1) - x))
     let sum_relu = sum(config, region, &[relu])?;
@@ -2730,14 +2779,8 @@ pub fn min<F: PrimeField + TensorType + PartialOrd>(
     let relu_one_minus_sum_relu =
         nonlinearity(config, region, &[one_minus_sum_relu], &LookupOp::ReLU)?;
 
-    region.assign(&config.inputs[1], &relu_one_minus_sum_relu)?;
-
     // constraining product to 0
-    let (x, y, z) = config.output.cartesian_coord(region.linear_coord());
-    let selector = config.selectors.get(&(BaseOp::IsZero, x, y));
-    region.enable(selector, z)?;
-
-    region.increment(relu_one_minus_sum_relu.len());
+    is_zero_identity(config, region, &[relu_one_minus_sum_relu], false)?;
 
     Ok(assigned_min_val)
 }
@@ -2884,7 +2927,8 @@ pub fn softmax<F: PrimeField + TensorType + PartialOrd>(
         &[denom],
         // we set to input scale + output_scale so the output scale is output)scale
         &LookupOp::Recip {
-            scale: scale.0.powf(2.0).into(),
+            input_scale: scale,
+            output_scale: scale,
         },
     )?;
 
@@ -2912,19 +2956,22 @@ pub fn range_check_percent<F: PrimeField + TensorType + PartialOrd>(
     // Calculate the difference between the expected output and actual output
     let diff = pairwise(config, region, values, BaseOp::Sub)?;
 
-    let scale_squared = scale.0.powf(2.0);
     // Calculate the reciprocal of the expected output tensor, scaling by double the scaling factor
     let recip = nonlinearity(
         config,
         region,
         &[values[0].clone()],
         &LookupOp::Recip {
-            scale: scale_squared.into(),
+            input_scale: scale,
+            output_scale: scale,
         },
     )?;
+
     // Multiply the difference by the recip
     let product = pairwise(config, region, &[diff, recip], BaseOp::Mult)?;
 
+    let scale_squared = scale.0 * scale.0;
+
     // Use the greater than look up table to check if the percent error is within the tolerance for upper bound
     let tol = tol / 100.0;
     let upper_bound = nonlinearity(
@@ -2952,15 +2999,8 @@ pub fn range_check_percent<F: PrimeField + TensorType + PartialOrd>(
     // Add the lower_bound and upper_bound
     let sum = pairwise(config, region, &[lower_bound, upper_bound], BaseOp::Add)?;
 
-    // Assign the sum tensor to the inputs
-    region.assign(&config.inputs[1], &sum)?;
-
     // Constrain the sum to be all zeros
-    let (x, y, z) = config.output.cartesian_coord(region.linear_coord());
-    let selector = config.selectors.get(&(BaseOp::IsZero, x, y));
-    region.enable(selector, z)?;
-
-    region.increment(sum.len());
+    is_zero_identity(config, region, &[sum.clone()], false)?;
 
     Ok(sum)
 }

src/circuit/ops/lookup.rs

@@ -5,7 +5,7 @@ use std::error::Error;
 use crate::{
     circuit::{layouts, table::Range, utils},
     fieldutils::{felt_to_i128, i128_to_felt},
-    graph::{multiplier_to_scale, scale_to_multiplier},
+    graph::multiplier_to_scale,
     tensor::{self, Tensor, TensorError, TensorType},
 };
 
@@ -17,42 +17,112 @@ use halo2curves::ff::PrimeField;
 #[derive(Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord, Deserialize, Serialize)]
 pub enum LookupOp {
     Abs,
-    Div { denom: utils::F32 },
-    Cast { scale: utils::F32 },
+    Div {
+        denom: utils::F32,
+    },
+    Cast {
+        scale: utils::F32,
+    },
     ReLU,
-    Max { scale: utils::F32, a: utils::F32 },
-    Min { scale: utils::F32, a: utils::F32 },
-    Ceil { scale: utils::F32 },
-    Floor { scale: utils::F32 },
-    Round { scale: utils::F32 },
-    RoundHalfToEven { scale: utils::F32 },
-    Sqrt { scale: utils::F32 },
-    Rsqrt { scale: utils::F32 },
-    Recip { scale: utils::F32 },
-    LeakyReLU { slope: utils::F32 },
-    Sigmoid { scale: utils::F32 },
-    Ln { scale: utils::F32 },
-    Exp { scale: utils::F32 },
-    Cos { scale: utils::F32 },
-    ACos { scale: utils::F32 },
-    Cosh { scale: utils::F32 },
-    ACosh { scale: utils::F32 },
-    Sin { scale: utils::F32 },
-    ASin { scale: utils::F32 },
-    Sinh { scale: utils::F32 },
-    ASinh { scale: utils::F32 },
-    Tan { scale: utils::F32 },
-    ATan { scale: utils::F32 },
-    Tanh { scale: utils::F32 },
-    ATanh { scale: utils::F32 },
-    Erf { scale: utils::F32 },
-    GreaterThan { a: utils::F32 },
-    LessThan { a: utils::F32 },
-    GreaterThanEqual { a: utils::F32 },
-    LessThanEqual { a: utils::F32 },
+    Max {
+        scale: utils::F32,
+        a: utils::F32,
+    },
+    Min {
+        scale: utils::F32,
+        a: utils::F32,
+    },
+    Ceil {
+        scale: utils::F32,
+    },
+    Floor {
+        scale: utils::F32,
+    },
+    Round {
+        scale: utils::F32,
+    },
+    RoundHalfToEven {
+        scale: utils::F32,
+    },
+    Sqrt {
+        scale: utils::F32,
+    },
+    Rsqrt {
+        scale: utils::F32,
+    },
+    Recip {
+        input_scale: utils::F32,
+        output_scale: utils::F32,
+    },
+    LeakyReLU {
+        slope: utils::F32,
+    },
+    Sigmoid {
+        scale: utils::F32,
+    },
+    Ln {
+        scale: utils::F32,
+    },
+    Exp {
+        scale: utils::F32,
+    },
+    Cos {
+        scale: utils::F32,
+    },
+    ACos {
+        scale: utils::F32,
+    },
+    Cosh {
+        scale: utils::F32,
+    },
+    ACosh {
+        scale: utils::F32,
+    },
+    Sin {
+        scale: utils::F32,
+    },
+    ASin {
+        scale: utils::F32,
+    },
+    Sinh {
+        scale: utils::F32,
+    },
+    ASinh {
+        scale: utils::F32,
+    },
+    Tan {
+        scale: utils::F32,
+    },
+    ATan {
+        scale: utils::F32,
+    },
+    Tanh {
+        scale: utils::F32,
+    },
+    ATanh {
+        scale: utils::F32,
+    },
+    Erf {
+        scale: utils::F32,
+    },
+    GreaterThan {
+        a: utils::F32,
+    },
+    LessThan {
+        a: utils::F32,
+    },
+    GreaterThanEqual {
+        a: utils::F32,
+    },
+    LessThanEqual {
+        a: utils::F32,
+    },
     Sign,
     KroneckerDelta,
-    Pow { scale: utils::F32, a: utils::F32 },
+    Pow {
+        scale: utils::F32,
+        a: utils::F32,
+    },
 }
 
 impl LookupOp {
@@ -120,7 +190,14 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for LookupOp {
                 &x,
                 f32::from(*scale).into(),
             )),
-            LookupOp::Recip { scale } => Ok(tensor::ops::nonlinearities::recip(&x, scale.into())),
+            LookupOp::Recip {
+                input_scale,
+                output_scale,
+            } => Ok(tensor::ops::nonlinearities::recip(
+                &x,
+                input_scale.into(),
+                output_scale.into(),
+            )),
             LookupOp::ReLU => Ok(tensor::ops::nonlinearities::leakyrelu(&x, 0_f64)),
 
             LookupOp::LeakyReLU { slope: a } => {
@@ -150,10 +227,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for LookupOp {
 
         let output = res.map(|x| i128_to_felt(x));
 
-        Ok(ForwardResult {
-            output,
-            intermediate_lookups: vec![],
-        })
+        Ok(ForwardResult { output })
     }
 
     /// Returns the name of the operation
@@ -173,7 +247,13 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for LookupOp {
             LookupOp::GreaterThanEqual { .. } => "GREATER_THAN_EQUAL".into(),
             LookupOp::LessThan { .. } => "LESS_THAN".into(),
             LookupOp::LessThanEqual { .. } => "LESS_THAN_EQUAL".into(),
-            LookupOp::Recip { scale, .. } => format!("RECIP(scale={})", scale),
+            LookupOp::Recip {
+                input_scale,
+                output_scale,
+            } => format!(
+                "RECIP(input_scale={}, output_scale={})",
+                input_scale, output_scale
+            ),
             LookupOp::Div { denom, .. } => format!("DIV(denom={})", denom),
             LookupOp::Cast { scale } => format!("CAST(scale={})", scale),
             LookupOp::Ln { scale } => format!("LN(scale={})", scale),
@@ -220,12 +300,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for LookupOp {
                 let in_scale = inputs_scale[0];
                 in_scale + multiplier_to_scale(1. / scale.0 as f64)
             }
-            LookupOp::Recip { scale } => {
-                let mut out_scale = inputs_scale[0];
-                out_scale +=
-                    multiplier_to_scale(scale.0 as f64 / scale_to_multiplier(out_scale).powf(2.0));
-                out_scale
-            }
+            LookupOp::Recip { output_scale, .. } => multiplier_to_scale(output_scale.into()),
             LookupOp::Sign
             | LookupOp::GreaterThan { .. }
             | LookupOp::LessThan { .. }
@@ -237,10 +312,6 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for LookupOp {
         Ok(scale)
     }
 
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        vec![self.clone()]
-    }
-
     fn clone_dyn(&self) -> Box<dyn Op<F>> {
         Box::new(self.clone()) // Forward to the derive(Clone) impl
     }

src/circuit/ops/mod.rs

@@ -10,8 +10,6 @@ use halo2curves::ff::PrimeField;
 
 use self::{lookup::LookupOp, region::RegionCtx};
 
-use super::table::Range;
-
 ///
 pub mod base;
 ///
@@ -31,7 +29,6 @@ pub mod region;
 #[derive(Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)]
 pub struct ForwardResult<F: PrimeField + TensorType + PartialOrd> {
     pub(crate) output: Tensor<F>,
-    pub(crate) intermediate_lookups: Vec<Tensor<i128>>,
 }
 
 /// A trait representing operations that can be represented as constraints in a circuit.
@@ -57,16 +54,6 @@ pub trait Op<F: PrimeField + TensorType + PartialOrd>: std::fmt::Debug + Send +
         vec![]
     }
 
-    /// Returns the lookups required by the operation.
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        vec![]
-    }
-
-    /// Returns the range checks required by the operation.
-    fn required_range_checks(&self) -> Vec<Range> {
-        vec![]
-    }
-
     /// Returns true if the operation is an input.
     fn is_input(&self) -> bool {
         false
@@ -190,7 +177,6 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for Input {
     fn f(&self, x: &[Tensor<F>]) -> Result<ForwardResult<F>, TensorError> {
         Ok(ForwardResult {
             output: x[0].clone(),
-            intermediate_lookups: vec![],
         })
     }
 
@@ -213,6 +199,7 @@ impl<F: PrimeField + TensorType + PartialOrd> Op<F> for Input {
                         config,
                         region,
                         values[..].try_into()?,
+                        true,
                     )?))
                 }
                 _ => Ok(Some(super::layouts::identity(
@@ -315,10 +302,7 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
     fn f(&self, _: &[Tensor<F>]) -> Result<ForwardResult<F>, TensorError> {
         let output = self.quantized_values.clone();
 
-        Ok(ForwardResult {
-            output,
-            intermediate_lookups: vec![],
-        })
+        Ok(ForwardResult { output })
     }
 
     fn as_string(&self) -> String {

src/circuit/ops/poly.rs

@@ -1,5 +1,6 @@
 use crate::{
     circuit::layouts,
+    fieldutils::felt_to_i128,
     tensor::{self, Tensor, TensorError},
 };
 
@@ -9,6 +10,14 @@ use super::{base::BaseOp, *};
 /// An enum representing the operations that can be expressed as arithmetic (non lookup) operations.
 #[derive(Clone, Debug, Serialize, Deserialize)]
 pub enum PolyOp {
+    GatherElements {
+        dim: usize,
+        constant_idx: Option<Tensor<usize>>,
+    },
+    ScatterElements {
+        dim: usize,
+        constant_idx: Option<Tensor<usize>>,
+    },
     MultiBroadcastTo {
         shape: Vec<usize>,
     },
@@ -33,7 +42,9 @@ pub enum PolyOp {
     Sub,
     Neg,
     Mult,
-    Identity,
+    Identity {
+        out_scale: Option<crate::Scale>,
+    },
     Reshape(Vec<usize>),
     MoveAxis {
         source: usize,
@@ -79,13 +90,17 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
 
     fn as_string(&self) -> String {
         match &self {
+            PolyOp::GatherElements { dim, .. } => format!("GATHERELEMENTS (dim={})", dim),
+            PolyOp::ScatterElements { dim, .. } => format!("SCATTERELEMENTS (dim={})", dim),
             PolyOp::MultiBroadcastTo { shape } => format!("MULTIBROADCASTTO (shape={:?})", shape),
             PolyOp::MoveAxis { .. } => "MOVEAXIS".into(),
             PolyOp::Downsample { .. } => "DOWNSAMPLE".into(),
             PolyOp::Resize { .. } => "RESIZE".into(),
             PolyOp::Iff => "IFF".into(),
             PolyOp::Einsum { equation, .. } => format!("EINSUM {}", equation),
-            PolyOp::Identity => "IDENTITY".into(),
+            PolyOp::Identity { out_scale } => {
+                format!("IDENTITY (out_scale={:?})", out_scale)
+            }
             PolyOp::Reshape(shape) => format!("RESHAPE (shape={:?})", shape),
             PolyOp::Flatten(_) => "FLATTEN".into(),
             PolyOp::Pad(_) => "PAD".into(),
@@ -135,7 +150,7 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
             PolyOp::Resize { scale_factor } => tensor::ops::resize(&inputs[0], scale_factor),
             PolyOp::Iff => tensor::ops::iff(&inputs[0], &inputs[1], &inputs[2]),
             PolyOp::Einsum { equation } => tensor::ops::einsum(equation, &inputs),
-            PolyOp::Identity => Ok(inputs[0].clone()),
+            PolyOp::Identity { .. } => Ok(inputs[0].clone()),
             PolyOp::Reshape(new_dims) => {
                 let mut t = inputs[0].clone();
                 t.reshape(new_dims)?;
@@ -199,14 +214,36 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
                 if 1 != inputs.len() {
                     return Err(TensorError::DimMismatch("slice inputs".to_string()));
                 }
-                Ok(tensor::ops::slice(&inputs[0], axis, start, end)?)
+                tensor::ops::slice(&inputs[0], axis, start, end)
+            }
+            PolyOp::GatherElements { dim, constant_idx } => {
+                let x = inputs[0].clone();
+                let y = if let Some(idx) = constant_idx {
+                    idx.clone()
+                } else {
+                    inputs[1].clone().map(|x| felt_to_i128(x) as usize)
+                };
+                tensor::ops::gather_elements(&x, &y, *dim)
+            }
+            PolyOp::ScatterElements { dim, constant_idx } => {
+                let x = inputs[0].clone();
+
+                let idx = if let Some(idx) = constant_idx {
+                    idx.clone()
+                } else {
+                    inputs[1].clone().map(|x| felt_to_i128(x) as usize)
+                };
+
+                let src = if let Some(_) = constant_idx {
+                    inputs[1].clone()
+                } else {
+                    inputs[2].clone()
+                };
+                tensor::ops::scatter(&x, &idx, &src, *dim)
             }
         }?;
 
-        Ok(ForwardResult {
-            output: res,
-            intermediate_lookups: vec![],
-        })
+        Ok(ForwardResult { output: res })
     }
 
     fn layout(
@@ -247,6 +284,26 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
             PolyOp::Conv { padding, stride } => {
                 layouts::conv(config, region, values[..].try_into()?, *padding, *stride)?
             }
+            PolyOp::GatherElements { dim, constant_idx } => {
+                if let Some(idx) = constant_idx {
+                    tensor::ops::gather_elements(values[0].get_inner_tensor()?, idx, *dim)?.into()
+                } else {
+                    layouts::gather_elements(config, region, values[..].try_into()?, *dim)?
+                }
+            }
+            PolyOp::ScatterElements { dim, constant_idx } => {
+                if let Some(idx) = constant_idx {
+                    tensor::ops::scatter(
+                        values[0].get_inner_tensor()?,
+                        idx,
+                        values[1].get_inner_tensor()?,
+                        *dim,
+                    )?
+                    .into()
+                } else {
+                    layouts::scatter_elements(config, region, values[..].try_into()?, *dim)?
+                }
+            }
             PolyOp::DeConv {
                 padding,
                 output_padding,
@@ -264,7 +321,7 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
             PolyOp::Mult => {
                 layouts::pairwise(config, region, values[..].try_into()?, BaseOp::Mult)?
             }
-            PolyOp::Identity => layouts::identity(config, region, values[..].try_into()?)?,
+            PolyOp::Identity { .. } => layouts::identity(config, region, values[..].try_into()?)?,
             PolyOp::Reshape(d) | PolyOp::Flatten(d) => layouts::reshape(values[..].try_into()?, d)?,
             PolyOp::Pad(p) => {
                 if values.len() != 1 {
@@ -322,9 +379,8 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
                 output_scale
             }
             PolyOp::Add => {
-                let mut scale_a = 0;
-                let scale_b = in_scales[0];
-                scale_a += in_scales[1];
+                let scale_a = in_scales[0];
+                let scale_b = in_scales[1];
                 assert_eq!(scale_a, scale_b);
                 scale_a
             }
@@ -336,19 +392,21 @@ impl<F: PrimeField + TensorType + PartialOrd + Serialize + for<'de> Deserialize<
             }
             PolyOp::Reshape(_) | PolyOp::Flatten(_) => in_scales[0],
             PolyOp::Pow(pow) => in_scales[0] * (*pow as crate::Scale),
+            PolyOp::Identity { out_scale } => out_scale.unwrap_or(in_scales[0]),
             _ => in_scales[0],
         };
         Ok(scale)
     }
 
     fn requires_homogenous_input_scales(&self) -> Vec<usize> {
-        if matches!(
-            self,
-            PolyOp::Add { .. } | PolyOp::Sub | PolyOp::Concat { .. }
-        ) {
+        if matches!(self, PolyOp::Add { .. } | PolyOp::Sub) {
             vec![0, 1]
         } else if matches!(self, PolyOp::Iff) {
             vec![1, 2]
+        } else if matches!(self, PolyOp::Concat { .. }) {
+            (0..100).collect()
+        } else if matches!(self, PolyOp::ScatterElements { .. }) {
+            vec![0, 2]
         } else {
             vec![]
         }

src/circuit/ops/region.rs

@@ -1,4 +1,7 @@
-use crate::tensor::{Tensor, TensorError, TensorType, ValTensor, ValType, VarTensor};
+use crate::{
+    circuit::table::Range,
+    tensor::{Tensor, TensorError, TensorType, ValTensor, ValType, VarTensor},
+};
 use halo2_proofs::{
     circuit::Region,
     plonk::{Error, Selector},
@@ -7,9 +10,16 @@ use halo2curves::ff::PrimeField;
 use std::{
     cell::RefCell,
     collections::HashSet,
-    sync::atomic::{AtomicUsize, Ordering},
+    sync::{
+        atomic::{AtomicUsize, Ordering},
+        Arc, Mutex,
+    },
 };
 
+use portable_atomic::AtomicI128 as AtomicInt;
+
+use super::lookup::LookupOp;
+
 /// Region error
 #[derive(Debug, thiserror::Error)]
 pub enum RegionError {
@@ -56,6 +66,10 @@ pub struct RegionCtx<'a, F: PrimeField + TensorType + PartialOrd> {
     linear_coord: usize,
     num_inner_cols: usize,
     total_constants: usize,
+    used_lookups: HashSet<LookupOp>,
+    used_range_checks: HashSet<Range>,
+    max_lookup_inputs: i128,
+    min_lookup_inputs: i128,
 }
 
 impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
@@ -75,6 +89,10 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
             row,
             linear_coord,
             total_constants: 0,
+            used_lookups: HashSet::new(),
+            used_range_checks: HashSet::new(),
+            max_lookup_inputs: 0,
+            min_lookup_inputs: 0,
         }
     }
     /// Create a new region context from a wrapped region
@@ -90,6 +108,10 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
             linear_coord,
             row,
             total_constants: 0,
+            used_lookups: HashSet::new(),
+            used_range_checks: HashSet::new(),
+            max_lookup_inputs: 0,
+            min_lookup_inputs: 0,
         }
     }
 
@@ -104,6 +126,10 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
             linear_coord,
             row,
             total_constants: 0,
+            used_lookups: HashSet::new(),
+            used_range_checks: HashSet::new(),
+            max_lookup_inputs: 0,
+            min_lookup_inputs: 0,
         }
     }
 
@@ -111,8 +137,10 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
     pub fn new_dummy_with_constants(
         row: usize,
         linear_coord: usize,
-        constants: usize,
+        total_constants: usize,
         num_inner_cols: usize,
+        used_lookups: HashSet<LookupOp>,
+        used_range_checks: HashSet<Range>,
     ) -> RegionCtx<'a, F> {
         let region = None;
         RegionCtx {
@@ -120,7 +148,11 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
             num_inner_cols,
             linear_coord,
             row,
-            total_constants: constants,
+            total_constants,
+            used_lookups,
+            used_range_checks,
+            max_lookup_inputs: 0,
+            min_lookup_inputs: 0,
         }
     }
 
@@ -160,6 +192,7 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
 
     /// Create a new region context per loop iteration
     /// hacky but it works
+
     pub fn dummy_loop<T: TensorType + Send + Sync>(
         &mut self,
         output: &mut Tensor<T>,
@@ -170,6 +203,12 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
         let row = AtomicUsize::new(self.row());
         let linear_coord = AtomicUsize::new(self.linear_coord());
         let constants = AtomicUsize::new(self.total_constants());
+        let max_lookup_inputs =
+            AtomicInt::new(self.max_lookup_inputs().try_into().unwrap_or_default());
+        let min_lookup_inputs =
+            AtomicInt::new(self.min_lookup_inputs().try_into().unwrap_or_default());
+        let lookups = Arc::new(Mutex::new(self.used_lookups.clone()));
+        let range_checks = Arc::new(Mutex::new(self.used_range_checks.clone()));
 
         *output = output
             .par_enum_map(|idx, _| {
@@ -177,12 +216,16 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
                 let starting_offset = row.load(Ordering::SeqCst);
                 let starting_linear_coord = linear_coord.load(Ordering::SeqCst);
                 let starting_constants = constants.load(Ordering::SeqCst);
+                // get inner value of the locked lookups
+
                 // we need to make sure that the region is not shared between threads
                 let mut local_reg = Self::new_dummy_with_constants(
                     starting_offset,
                     starting_linear_coord,
                     starting_constants,
                     self.num_inner_cols,
+                    HashSet::new(),
+                    HashSet::new(),
                 );
                 let res = inner_loop_function(idx, &mut local_reg);
                 // we update the offset and constants
@@ -195,6 +238,19 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
                     local_reg.total_constants() - starting_constants,
                     Ordering::SeqCst,
                 );
+
+                let local_max_lookup_inputs =
+                    local_reg.max_lookup_inputs().try_into().unwrap_or_default();
+                let local_min_lookup_inputs =
+                    local_reg.min_lookup_inputs().try_into().unwrap_or_default();
+
+                max_lookup_inputs.fetch_max(local_max_lookup_inputs, Ordering::SeqCst);
+                min_lookup_inputs.fetch_min(local_min_lookup_inputs, Ordering::SeqCst);
+                // update the lookups
+                let mut lookups = lookups.lock().unwrap();
+                lookups.extend(local_reg.used_lookups());
+                let mut range_checks = range_checks.lock().unwrap();
+                range_checks.extend(local_reg.used_range_checks());
                 res
             })
             .map_err(|e| {
@@ -203,7 +259,56 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
             })?;
         self.total_constants = constants.into_inner();
         self.linear_coord = linear_coord.into_inner();
+        #[allow(trivial_numeric_casts)]
+        {
+            self.max_lookup_inputs = max_lookup_inputs.into_inner() as i128;
+            self.min_lookup_inputs = min_lookup_inputs.into_inner() as i128;
+        }
         self.row = row.into_inner();
+        self.used_lookups = Arc::try_unwrap(lookups)
+            .map_err(|e| RegionError::from(format!("dummy_loop: failed to get lookups: {:?}", e)))?
+            .into_inner()
+            .map_err(|e| {
+                RegionError::from(format!("dummy_loop: failed to get lookups: {:?}", e))
+            })?;
+        self.used_range_checks = Arc::try_unwrap(range_checks)
+            .map_err(|e| {
+                RegionError::from(format!("dummy_loop: failed to get range checks: {:?}", e))
+            })?
+            .into_inner()
+            .map_err(|e| {
+                RegionError::from(format!("dummy_loop: failed to get range checks: {:?}", e))
+            })?;
+
+        Ok(())
+    }
+
+    /// Update the max and min from inputs
+    pub fn update_max_min_lookup_inputs(
+        &mut self,
+        inputs: &[ValTensor<F>],
+    ) -> Result<(), Box<dyn std::error::Error>> {
+        let (mut min, mut max) = (0, 0);
+        for i in inputs {
+            max = max.max(i.get_int_evals()?.into_iter().max().unwrap_or_default());
+            min = min.min(i.get_int_evals()?.into_iter().min().unwrap_or_default());
+        }
+        self.max_lookup_inputs = self.max_lookup_inputs.max(max);
+        self.min_lookup_inputs = self.min_lookup_inputs.min(min);
+        Ok(())
+    }
+
+    /// Update the max and min from inputs
+    pub fn update_max_min_lookup_range(
+        &mut self,
+        range: Range,
+    ) -> Result<(), Box<dyn std::error::Error>> {
+        if range.0 > range.1 {
+            return Err("update_max_min_lookup_range: invalid range".into());
+        }
+
+        self.max_lookup_inputs = self.max_lookup_inputs.max(range.1);
+        self.min_lookup_inputs = self.min_lookup_inputs.min(range.0);
         Ok(())
     }
 
@@ -212,15 +317,20 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
         self.region.is_none()
     }
 
-    /// duplicate_dummy
-    pub fn duplicate_dummy(&self) -> Self {
-        Self {
-            region: None,
-            linear_coord: self.linear_coord,
-            num_inner_cols: self.num_inner_cols,
-            row: self.row,
-            total_constants: self.total_constants,
-        }
+    /// add used lookup
+    pub fn add_used_lookup(
+        &mut self,
+        lookup: LookupOp,
+        inputs: &[ValTensor<F>],
+    ) -> Result<(), Box<dyn std::error::Error>> {
+        self.used_lookups.insert(lookup);
+        self.update_max_min_lookup_inputs(inputs)
+    }
+
+    /// add used range check
+    pub fn add_used_range_check(&mut self, range: Range) -> Result<(), Box<dyn std::error::Error>> {
+        self.used_range_checks.insert(range);
+        self.update_max_min_lookup_range(range)
     }
 
     /// Get the offset
@@ -238,6 +348,26 @@ impl<'a, F: PrimeField + TensorType + PartialOrd> RegionCtx<'a, F> {
         self.total_constants
     }
 
+    /// get used lookups
+    pub fn used_lookups(&self) -> HashSet<LookupOp> {
+        self.used_lookups.clone()
+    }
+
+    /// get used range checks
+    pub fn used_range_checks(&self) -> HashSet<Range> {
+        self.used_range_checks.clone()
+    }
+
+    /// max lookup inputs
+    pub fn max_lookup_inputs(&self) -> i128 {
+        self.max_lookup_inputs
+    }
+
+    /// min lookup inputs
+    pub fn min_lookup_inputs(&self) -> i128 {
+        self.min_lookup_inputs
+    }
+
     /// Assign a constant value
     pub fn assign_constant(&mut self, var: &VarTensor, value: F) -> Result<ValType<F>, Error> {
         self.total_constants += 1;

src/circuit/tests.rs

@@ -246,7 +246,7 @@ mod matmul_col_overflow {
 #[cfg(test)]
 #[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
 mod matmul_col_ultra_overflow_double_col {
-    use halo2_proofs::poly::commitment::ParamsProver;
+    use halo2_proofs::poly::commitment::{Params, ParamsProver};
 
     use super::*;
 
@@ -349,8 +349,13 @@ mod matmul_col_ultra_overflow_double_col {
         let strategy =
             halo2_proofs::poly::kzg::strategy::SingleStrategy::new(params.verifier_params());
         let vk = pk.get_vk();
-        let result =
-            crate::pfsys::verify_proof_circuit_kzg(params.verifier_params(), proof, vk, strategy);
+        let result = crate::pfsys::verify_proof_circuit_kzg(
+            params.verifier_params(),
+            proof,
+            vk,
+            strategy,
+            params.n(),
+        );
 
         assert!(result.is_ok());
 
@@ -361,7 +366,7 @@ mod matmul_col_ultra_overflow_double_col {
 #[cfg(test)]
 #[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
 mod matmul_col_ultra_overflow {
-    use halo2_proofs::poly::commitment::ParamsProver;
+    use halo2_proofs::poly::commitment::{Params, ParamsProver};
 
     use super::*;
 
@@ -463,8 +468,13 @@ mod matmul_col_ultra_overflow {
         let strategy =
             halo2_proofs::poly::kzg::strategy::SingleStrategy::new(params.verifier_params());
         let vk = pk.get_vk();
-        let result =
-            crate::pfsys::verify_proof_circuit_kzg(params.verifier_params(), proof, vk, strategy);
+        let result = crate::pfsys::verify_proof_circuit_kzg(
+            params.verifier_params(),
+            proof,
+            vk,
+            strategy,
+            params.n(),
+        );
 
         assert!(result.is_ok());
 
@@ -1140,7 +1150,7 @@ mod conv {
 #[cfg(test)]
 #[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
 mod conv_col_ultra_overflow {
-    use halo2_proofs::poly::commitment::ParamsProver;
+    use halo2_proofs::poly::commitment::{Params, ParamsProver};
 
     use super::*;
 
@@ -1262,8 +1272,13 @@ mod conv_col_ultra_overflow {
         let strategy =
             halo2_proofs::poly::kzg::strategy::SingleStrategy::new(params.verifier_params());
         let vk = pk.get_vk();
-        let result =
-            crate::pfsys::verify_proof_circuit_kzg(params.verifier_params(), proof, vk, strategy);
+        let result = crate::pfsys::verify_proof_circuit_kzg(
+            params.verifier_params(),
+            proof,
+            vk,
+            strategy,
+            params.n(),
+        );
 
         assert!(result.is_ok());
 
@@ -1275,7 +1290,7 @@ mod conv_col_ultra_overflow {
 // not wasm 32 unknown
 #[cfg(not(all(target_arch = "wasm32", target_os = "unknown")))]
 mod conv_relu_col_ultra_overflow {
-    use halo2_proofs::poly::commitment::ParamsProver;
+    use halo2_proofs::poly::commitment::{Params, ParamsProver};
 
     use super::*;
 
@@ -1412,8 +1427,13 @@ mod conv_relu_col_ultra_overflow {
         let strategy =
             halo2_proofs::poly::kzg::strategy::SingleStrategy::new(params.verifier_params());
         let vk = pk.get_vk();
-        let result =
-            crate::pfsys::verify_proof_circuit_kzg(params.verifier_params(), proof, vk, strategy);
+        let result = crate::pfsys::verify_proof_circuit_kzg(
+            params.verifier_params(),
+            proof,
+            vk,
+            strategy,
+            params.n(),
+        );
 
         assert!(result.is_ok());
 
@@ -2154,7 +2174,7 @@ mod rangecheckpercent {
         }
 
         fn configure(cs: &mut ConstraintSystem<F>) -> Self::Config {
-            let scale = utils::F32(SCALE.pow(2) as f32);
+            let scale = utils::F32(SCALE as f32);
             let a = VarTensor::new_advice(cs, K, 1, LEN);
             let b = VarTensor::new_advice(cs, K, 1, LEN);
             let output = VarTensor::new_advice(cs, K, 1, LEN);
@@ -2162,11 +2182,12 @@ mod rangecheckpercent {
                 Self::Config::configure(cs, &[a.clone(), b.clone()], &output, CheckMode::SAFE);
             // set up a new GreaterThan and Recip tables
             let nl = &LookupOp::GreaterThan {
-                a: circuit::utils::F32((RANGE * scale.0) / 100.0),
+                a: circuit::utils::F32((RANGE * SCALE.pow(2) as f32) / 100.0),
             };
             config
                 .configure_lookup(cs, &b, &output, &a, (-32768, 32768), K, nl)
                 .unwrap();
+
             config
                 .configure_lookup(
                     cs,
@@ -2175,7 +2196,10 @@ mod rangecheckpercent {
                     &a,
                     (-32768, 32768),
                     K,
-                    &LookupOp::Recip { scale },
+                    &LookupOp::Recip {
+                        input_scale: scale,
+                        output_scale: scale,
+                    },
                 )
                 .unwrap();
             config
@@ -2339,7 +2363,7 @@ mod lookup_ultra_overflow {
     use halo2_proofs::{
         circuit::{Layouter, SimpleFloorPlanner, Value},
         plonk::{Circuit, ConstraintSystem, Error},
-        poly::commitment::ParamsProver,
+        poly::commitment::{Params, ParamsProver},
     };
 
     #[derive(Clone)]
@@ -2443,8 +2467,13 @@ mod lookup_ultra_overflow {
         let strategy =
             halo2_proofs::poly::kzg::strategy::SingleStrategy::new(params.verifier_params());
         let vk = pk.get_vk();
-        let result =
-            crate::pfsys::verify_proof_circuit_kzg(params.verifier_params(), proof, vk, strategy);
+        let result = crate::pfsys::verify_proof_circuit_kzg(
+            params.verifier_params(),
+            proof,
+            vk,
+            strategy,
+            params.n(),
+        );
 
         assert!(result.is_ok());
 
@@ -2511,7 +2540,8 @@ mod softmax {
                     (-32768, 32768),
                     K,
                     &LookupOp::Recip {
-                        scale: SCALE.powf(2.0).into(),
+                        input_scale: SCALE.into(),
+                        output_scale: SCALE.into(),
                     },
                 )
                 .unwrap();

src/commands.rs

@@ -85,6 +85,8 @@ pub const DEFAULT_VK_SOL: &str = "vk.sol";
 pub const DEFAULT_VK_ABI: &str = "vk.abi";
 /// Default scale rebase multipliers for calibration
 pub const DEFAULT_SCALE_REBASE_MULTIPLIERS: &str = "1,2,10";
+/// Default use reduced srs for verification
+pub const DEFAULT_USE_REDUCED_SRS_FOR_VERIFICATION: &str = "false";
 
 impl std::fmt::Display for TranscriptType {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
@@ -695,6 +697,9 @@ pub enum Commands {
         /// The path to SRS, if None will use $EZKL_REPO_PATH/srs/kzg{logrows}.srs
         #[arg(long)]
         srs_path: Option<PathBuf>,
+        /// Reduce SRS logrows to the number of instances rather than the number of logrows used for proofs (only works if the srs were generated in the same ceremony)
+        #[arg(long, default_value = DEFAULT_USE_REDUCED_SRS_FOR_VERIFICATION)]
+        reduced_srs: Option<bool>,
     },
     /// Verifies an aggregate proof, returning accept or reject
     VerifyAggr {
@@ -795,12 +800,4 @@ pub enum Commands {
         #[arg(long)]
         addr_vk: Option<H160>,
     },
-
-    /// Print the proof in hexadecimal
-    #[command(name = "print-proof-hex")]
-    PrintProofHex {
-        /// The path to the proof file
-        #[arg(long, default_value = DEFAULT_PROOF)]
-        proof_path: PathBuf,
-    },
 }

src/execute.rs

@@ -366,7 +366,8 @@ pub async fn run(command: Commands) -> Result<String, Box<dyn Error>> {
             settings_path,
             vk_path,
             srs_path,
-        } => verify(proof_path, settings_path, vk_path, srs_path)
+            reduced_srs,
+        } => verify(proof_path, settings_path, vk_path, srs_path, reduced_srs)
             .map(|e| serde_json::to_string(&e).unwrap()),
         Commands::VerifyAggr {
             proof_path,
@@ -440,7 +441,6 @@ pub async fn run(command: Commands) -> Result<String, Box<dyn Error>> {
             addr_da,
             addr_vk,
         } => verify_evm(proof_path, addr_verifier, rpc_url, addr_da, addr_vk).await,
-        Commands::PrintProofHex { proof_path } => print_proof_hex(proof_path),
     }
 }
 
@@ -630,6 +630,10 @@ pub(crate) async fn gen_witness(
     if let Some(output_path) = output {
         serde_json::to_writer(&File::create(output_path)?, &witness)?;
     }
+
+    // print the witness in debug
+    debug!("witness: \n {}", witness.as_json()?.to_colored_json_auto()?);
+
     Ok(witness)
 }
 
@@ -722,11 +726,11 @@ impl AccuracyResults {
             let percentage_error = error.enum_map(|i, x| Ok::<_, TensorError>(x / original[i]))?;
             let abs_percentage_error = percentage_error.map(|x| x.abs());
 
-            errors.extend(error.into_iter());
-            abs_errors.extend(abs_error.into_iter());
-            squared_errors.extend(squared_error.into_iter());
-            percentage_errors.extend(percentage_error.into_iter());
-            abs_percentage_errors.extend(abs_percentage_error.into_iter());
+            errors.extend(error);
+            abs_errors.extend(abs_error);
+            squared_errors.extend(squared_error);
+            percentage_errors.extend(percentage_error);
+            abs_percentage_errors.extend(abs_percentage_error);
         }
 
         let mean_percent_error =
@@ -737,22 +741,22 @@ impl AccuracyResults {
         let median_error = errors[errors.len() / 2];
         let max_error = *errors
             .iter()
-            .max_by(|a, b| a.partial_cmp(b).unwrap())
+            .max_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
             .unwrap();
         let min_error = *errors
             .iter()
-            .min_by(|a, b| a.partial_cmp(b).unwrap())
+            .min_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
             .unwrap();
 
         let mean_abs_error = abs_errors.iter().sum::<f32>() / abs_errors.len() as f32;
         let median_abs_error = abs_errors[abs_errors.len() / 2];
         let max_abs_error = *abs_errors
             .iter()
-            .max_by(|a, b| a.partial_cmp(b).unwrap())
+            .max_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
             .unwrap();
         let min_abs_error = *abs_errors
             .iter()
-            .min_by(|a, b| a.partial_cmp(b).unwrap())
+            .min_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal))
             .unwrap();
 
         let mean_squared_error = squared_errors.iter().sum::<f32>() / squared_errors.len() as f32;
@@ -822,10 +826,7 @@ pub(crate) fn calibrate(
     let range = if let Some(scales) = scales {
         scales
     } else {
-        match target {
-            CalibrationTarget::Resources { .. } => (8..10).collect::<Vec<crate::Scale>>(),
-            CalibrationTarget::Accuracy => (10..14).collect::<Vec<crate::Scale>>(),
-        }
+        (10..14).collect::<Vec<crate::Scale>>()
     };
 
     let div_rebasing = if let Some(div_rebasing) = div_rebasing {
@@ -1168,16 +1169,6 @@ pub(crate) fn mock(
     Ok(String::new())
 }
 
-pub(crate) fn print_proof_hex(proof_path: PathBuf) -> Result<String, Box<dyn Error>> {
-    let proof = Snark::load::<KZGCommitmentScheme<Bn256>>(&proof_path)?;
-    for instance in proof.instances {
-        println!("{:?}", instance);
-    }
-    let hex_str = hex::encode(proof.proof);
-    info!("0x{}", hex_str);
-    Ok(format!("0x{}", hex_str))
-}
-
 #[cfg(feature = "render")]
 pub(crate) fn render(
     model: PathBuf,
@@ -1724,6 +1715,7 @@ pub(crate) fn fuzz(
             bad_proof,
             pk.get_vk(),
             strategy.clone(),
+            params.n(),
         )
         .map_err(|_| ())
     };
@@ -1754,6 +1746,7 @@ pub(crate) fn fuzz(
             bad_proof,
             pk.get_vk(),
             strategy.clone(),
+            params.n(),
         )
         .map_err(|_| ())
     };
@@ -1790,6 +1783,7 @@ pub(crate) fn fuzz(
             proof.clone(),
             bad_vk,
             strategy.clone(),
+            params.n(),
         )
         .map_err(|_| ())
     };
@@ -1821,6 +1815,7 @@ pub(crate) fn fuzz(
             bad_proof,
             pk.get_vk(),
             strategy.clone(),
+            params.n(),
         )
         .map_err(|_| ())
     };
@@ -1856,6 +1851,7 @@ pub(crate) fn fuzz(
             bad_proof,
             pk.get_vk(),
             strategy.clone(),
+            params.n(),
         )
         .map_err(|_| ())
     };
@@ -2041,15 +2037,33 @@ pub(crate) fn verify(
     settings_path: PathBuf,
     vk_path: PathBuf,
     srs_path: Option<PathBuf>,
+    reduced_srs: Option<bool>,
 ) -> Result<bool, Box<dyn Error>> {
     let circuit_settings = GraphSettings::load(&settings_path)?;
-    let params = load_params_cmd(srs_path, circuit_settings.run_args.logrows)?;
+
+    let params = if let Some(reduced_srs) = reduced_srs {
+        if reduced_srs {
+            load_params_cmd(srs_path, circuit_settings.log2_total_instances())?
+        } else {
+            load_params_cmd(srs_path, circuit_settings.run_args.logrows)?
+        }
+    } else {
+        load_params_cmd(srs_path, circuit_settings.run_args.logrows)?
+    };
+
     let proof = Snark::load::<KZGCommitmentScheme<Bn256>>(&proof_path)?;
 
     let strategy = KZGSingleStrategy::new(params.verifier_params());
-    let vk = load_vk::<KZGCommitmentScheme<Bn256>, Fr, GraphCircuit>(vk_path, circuit_settings)?;
+    let vk =
+        load_vk::<KZGCommitmentScheme<Bn256>, Fr, GraphCircuit>(vk_path, circuit_settings.clone())?;
     let now = Instant::now();
-    let result = verify_proof_circuit_kzg(params.verifier_params(), proof, &vk, strategy);
+    let result = verify_proof_circuit_kzg(
+        params.verifier_params(),
+        proof,
+        &vk,
+        strategy,
+        1 << circuit_settings.run_args.logrows,
+    );
     let elapsed = now.elapsed();
     info!(
         "verify took {}.{}",
@@ -2073,7 +2087,7 @@ pub(crate) fn verify_aggr(
     let strategy = AccumulatorStrategy::new(params.verifier_params());
     let vk = load_vk::<KZGCommitmentScheme<Bn256>, Fr, AggregationCircuit>(vk_path, ())?;
     let now = Instant::now();
-    let result = verify_proof_circuit_kzg(&params, proof, &vk, strategy);
+    let result = verify_proof_circuit_kzg(&params, proof, &vk, strategy, 1 << logrows);
 
     let elapsed = now.elapsed();
     info!(

src/graph/mod.rs

@@ -35,6 +35,8 @@ use halo2_proofs::{
 };
 use halo2curves::bn256::{self, Bn256, Fr as Fp, G1Affine};
 use halo2curves::ff::PrimeField;
+#[cfg(not(target_arch = "wasm32"))]
+use lazy_static::lazy_static;
 use log::{debug, error, info, trace, warn};
 use maybe_rayon::prelude::{IntoParallelRefIterator, ParallelIterator};
 pub use model::*;
@@ -61,6 +63,20 @@ pub const RANGE_MULTIPLIER: i128 = 2;
 /// Max representation of a lookup table input
 pub const MAX_LOOKUP_ABS: i128 = 8 * 2_i128.pow(MAX_PUBLIC_SRS);
 
+#[cfg(not(target_arch = "wasm32"))]
+lazy_static! {
+    /// Max circuit area
+    pub static ref EZKL_MAX_CIRCUIT_AREA: Option<usize> =
+        if let Ok(max_circuit_area) = std::env::var("EZKL_MAX_CIRCUIT_AREA") {
+            Some(max_circuit_area.parse().unwrap_or(0))
+        } else {
+            None
+        };
+}
+
+#[cfg(target_arch = "wasm32")]
+const EZKL_MAX_CIRCUIT_AREA: Option<usize> = None;
+
 /// circuit related errors.
 #[derive(Debug, Error)]
 pub enum GraphError {
@@ -456,6 +472,14 @@ impl GraphSettings {
         instances
     }
 
+    /// calculate the log2 of the total number of instances
+    pub fn log2_total_instances(&self) -> u32 {
+        let sum = self.total_instances().iter().sum::<usize>();
+
+        // max between 1 and the log2 of the sums
+        std::cmp::max((sum as f64).log2().ceil() as u32, 1)
+    }
+
     /// save params to file
     pub fn save(&self, path: &std::path::PathBuf) -> Result<(), std::io::Error> {
         let encoded = serde_json::to_string(&self)?;
@@ -530,6 +554,7 @@ impl GraphSettings {
 pub struct GraphConfig {
     model_config: ModelConfig,
     module_configs: ModuleConfigs,
+    circuit_size: CircuitSize,
 }
 
 /// Defines the circuit for a computational graph / model loaded from a `.onnx` file.
@@ -809,7 +834,7 @@ impl GraphCircuit {
         let shapes = self.model().graph.input_shapes()?;
         let scales = self.model().graph.get_input_scales();
         let input_types = self.model().graph.get_input_types()?;
-        info!("input scales: {:?}", scales);
+        debug!("input scales: {:?}", scales);
 
         match &data.input_data {
             DataSource::File(file_data) => {
@@ -828,7 +853,7 @@ impl GraphCircuit {
         let shapes = self.model().graph.input_shapes()?;
         let scales = self.model().graph.get_input_scales();
         let input_types = self.model().graph.get_input_types()?;
-        info!("input scales: {:?}", scales);
+        debug!("input scales: {:?}", scales);
 
         self.process_data_source(&data.input_data, shapes, scales, input_types)
             .await
@@ -968,8 +993,8 @@ impl GraphCircuit {
             lookup_safety_margin * max_lookup_inputs,
         );
         if lookup_safety_margin == 1 {
-            margin.0 += 1;
-            margin.1 += 1;
+            margin.0 += 4;
+            margin.1 += 4;
         }
 
         margin
@@ -1028,7 +1053,7 @@ impl GraphCircuit {
                 "extended k is too large to accommodate the quotient polynomial with logrows {}",
                 min_logrows
             );
-            error!("{}", err_string);
+            debug!("{}", err_string);
             return Err(err_string.into());
         }
 
@@ -1048,7 +1073,7 @@ impl GraphCircuit {
                 "extended k is too large to accommodate the quotient polynomial with logrows {}",
                 max_logrows
             );
-            error!("{}", err_string);
+            debug!("{}", err_string);
             return Err(err_string.into());
         }
 
@@ -1114,7 +1139,7 @@ impl GraphCircuit {
 
         settings_mut.run_args.logrows = std::cmp::min(max_logrows, settings_mut.run_args.logrows);
 
-        info!(
+        debug!(
             "setting lookup_range to: {:?}, setting logrows to: {}",
             self.settings().run_args.lookup_range,
             self.settings().run_args.logrows
@@ -1366,7 +1391,6 @@ impl GraphCircuit {
     }
 }
 
-#[cfg(not(target_arch = "wasm32"))]
 #[derive(Clone, Debug, Default, Serialize, Deserialize)]
 struct CircuitSize {
     num_instances: usize,
@@ -1374,20 +1398,22 @@ struct CircuitSize {
     num_fixed: usize,
     num_challenges: usize,
     num_selectors: usize,
+    logrows: u32,
 }
 
-#[cfg(not(target_arch = "wasm32"))]
 impl CircuitSize {
-    pub fn from_cs(cs: &ConstraintSystem<Fp>) -> Self {
+    pub fn from_cs(cs: &ConstraintSystem<Fp>, logrows: u32) -> Self {
         CircuitSize {
             num_instances: cs.num_instance_columns(),
             num_advice_columns: cs.num_advice_columns(),
             num_fixed: cs.num_fixed_columns(),
             num_challenges: cs.num_challenges(),
             num_selectors: cs.num_selectors(),
+            logrows,
         }
     }
 
+    #[cfg(not(target_arch = "wasm32"))]
     /// Export the ezkl configuration as json
     pub fn as_json(&self) -> Result<String, Box<dyn std::error::Error>> {
         let serialized = match serde_json::to_string(&self) {
@@ -1398,6 +1424,25 @@ impl CircuitSize {
         };
         Ok(serialized)
     }
+
+    /// number of columns
+    pub fn num_columns(&self) -> usize {
+        self.num_instances + self.num_advice_columns + self.num_fixed
+    }
+
+    /// area of the circuit
+    pub fn area(&self) -> usize {
+        self.num_columns() * (1 << self.logrows)
+    }
+
+    /// area less than max
+    pub fn area_less_than_max(&self) -> bool {
+        if EZKL_MAX_CIRCUIT_AREA.is_some() {
+            self.area() < EZKL_MAX_CIRCUIT_AREA.unwrap()
+        } else {
+            true
+        }
+    }
 }
 
 impl Circuit<Fp> for GraphCircuit {
@@ -1472,10 +1517,12 @@ impl Circuit<Fp> for GraphCircuit {
             (cs.degree() as f32).log2().ceil()
         );
 
+        let circuit_size = CircuitSize::from_cs(cs, params.run_args.logrows);
+
         #[cfg(not(target_arch = "wasm32"))]
         info!(
             "circuit size: \n {}",
-            CircuitSize::from_cs(cs)
+            circuit_size
                 .as_json()
                 .unwrap()
                 .to_colored_json_auto()
@@ -1485,6 +1532,7 @@ impl Circuit<Fp> for GraphCircuit {
         GraphConfig {
             model_config,
             module_configs,
+            circuit_size,
         }
     }
 
@@ -1497,6 +1545,16 @@ impl Circuit<Fp> for GraphCircuit {
         config: Self::Config,
         mut layouter: impl Layouter<Fp>,
     ) -> Result<(), PlonkError> {
+        // check if the circuit area is less than the max
+        if !config.circuit_size.area_less_than_max() {
+            error!(
+                "circuit area {} is larger than the max allowed area {}",
+                config.circuit_size.area(),
+                EZKL_MAX_CIRCUIT_AREA.unwrap()
+            );
+            return Err(PlonkError::Synthesis);
+        }
+
         trace!("Setting input in synthesize");
         let input_vis = &self.settings().run_args.input_visibility;
         let output_vis = &self.settings().run_args.output_visibility;

src/graph/model.rs

@@ -10,7 +10,6 @@ use crate::circuit::table::Range;
 use crate::circuit::Input;
 use crate::circuit::InputType;
 use crate::circuit::Unknown;
-use crate::fieldutils::felt_to_i128;
 use crate::tensor::ValType;
 use crate::{
     circuit::{lookup::LookupOp, BaseConfig as PolyConfig, CheckMode, Op},
@@ -68,6 +67,16 @@ pub struct ForwardResult {
     pub min_lookup_inputs: i128,
 }
 
+impl From<DummyPassRes> for ForwardResult {
+    fn from(res: DummyPassRes) -> Self {
+        Self {
+            outputs: res.outputs,
+            max_lookup_inputs: res.max_lookup_inputs,
+            min_lookup_inputs: res.min_lookup_inputs,
+        }
+    }
+}
+
 /// A circuit configuration for the entirety of a model loaded from an Onnx file.
 #[derive(Clone, Debug)]
 pub struct ModelConfig {
@@ -80,6 +89,27 @@ pub struct ModelConfig {
 /// Representation of execution graph
 pub type NodeGraph = BTreeMap<usize, NodeType>;
 
+/// A struct for loading from an Onnx file and converting a computational graph to a circuit.
+#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
+pub struct DummyPassRes {
+    /// number of rows use
+    pub num_rows: usize,
+    /// linear coordinate
+    pub linear_coord: usize,
+    /// total const size
+    pub total_const_size: usize,
+    /// lookup ops
+    pub lookup_ops: HashSet<LookupOp>,
+    /// range checks
+    pub range_checks: HashSet<Range>,
+    /// max lookup inputs
+    pub max_lookup_inputs: i128,
+    /// min lookup inputs
+    pub min_lookup_inputs: i128,
+    /// outputs
+    pub outputs: Vec<Tensor<Fp>>,
+}
+
 /// A struct for loading from an Onnx file and converting a computational graph to a circuit.
 #[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
 pub struct Model {
@@ -234,20 +264,7 @@ impl NodeType {
             NodeType::SubGraph { out_dims, .. } => out_dims.clone(),
         }
     }
-    /// Returns the lookups required by a graph
-    pub fn required_lookups(&self) -> Vec<LookupOp> {
-        match self {
-            NodeType::Node(n) => n.opkind.required_lookups(),
-            NodeType::SubGraph { model, .. } => model.required_lookups(),
-        }
-    }
-    /// Returns the lookups required by a graph
-    pub fn required_range_checks(&self) -> Vec<Range> {
-        match self {
-            NodeType::Node(n) => n.opkind.required_range_checks(),
-            NodeType::SubGraph { model, .. } => model.required_range_checks(),
-        }
-    }
+
     /// Returns the scales of the node's output.
     pub fn out_scales(&self) -> Vec<crate::Scale> {
         match self {
@@ -432,23 +449,6 @@ impl ParsedNodes {
 }
 
 impl Model {
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        self.graph
-            .nodes
-            .values()
-            .flat_map(|n| n.required_lookups())
-            .collect_vec()
-    }
-
-    ///
-    fn required_range_checks(&self) -> Vec<Range> {
-        self.graph
-            .nodes
-            .values()
-            .flat_map(|n| n.required_range_checks())
-            .collect_vec()
-    }
-
     /// Creates a `Model` from a specified path to an Onnx file.
     /// # Arguments
     /// * `reader` - A reader for an Onnx file.
@@ -493,7 +493,7 @@ impl Model {
     ) -> Result<GraphSettings, Box<dyn Error>> {
         let instance_shapes = self.instance_shapes()?;
         #[cfg(not(target_arch = "wasm32"))]
-        info!(
+        debug!(
             "{} {} {}",
             "model has".blue(),
             instance_shapes.len().to_string().blue(),
@@ -501,42 +501,39 @@ impl Model {
         );
         // this is the total number of variables we will need to allocate
         // for the circuit
-        let (num_rows, linear_coord, total_const_size) =
-            self.dummy_layout(run_args, &self.graph.input_shapes()?)?;
+        let default_value = if !self.visibility.input.is_fixed() {
+            ValType::Value(Value::<Fp>::unknown())
+        } else {
+            ValType::Constant(Fp::ONE)
+        };
 
-        // extract the requisite lookup ops from the model
-        let mut lookup_ops: Vec<LookupOp> = self.required_lookups();
-        // extract the requisite lookup ops from the model
-        let mut range_checks: Vec<Range> = self.required_range_checks();
+        let inputs: Vec<ValTensor<Fp>> = self
+            .graph
+            .input_shapes()?
+            .iter()
+            .map(|shape| {
+                let mut t: ValTensor<Fp> =
+                    vec![default_value.clone(); shape.iter().product()].into();
+                t.reshape(shape)?;
+                Ok(t)
+            })
+            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
+
+        let res = self.dummy_layout(run_args, &inputs)?;
 
         // if we're using percentage tolerance, we need to add the necessary range check ops for it.
 
-        if run_args.tolerance.val > 0.0 {
-            for scale in self.graph.get_output_scales()? {
-                let mut tolerance = run_args.tolerance;
-                tolerance.scale = scale_to_multiplier(scale).into();
-                let opkind: Box<dyn Op<Fp>> = Box::new(HybridOp::RangeCheck(tolerance));
-                lookup_ops.extend(opkind.required_lookups());
-            }
-        }
-
-        let set: HashSet<_> = lookup_ops.drain(..).collect(); // dedup
-        lookup_ops.extend(set.into_iter().sorted());
-
-        let set: HashSet<_> = range_checks.drain(..).collect(); // dedup
-        range_checks.extend(set.into_iter().sorted());
-
         Ok(GraphSettings {
             run_args: run_args.clone(),
             model_instance_shapes: instance_shapes,
             module_sizes: crate::graph::modules::ModuleSizes::default(),
-            num_rows,
-            total_assignments: linear_coord,
-            required_lookups: lookup_ops,
-            required_range_checks: range_checks,
+            num_rows: res.num_rows,
+            total_assignments: res.linear_coord,
+            required_lookups: res.lookup_ops.into_iter().collect(),
+            required_range_checks: res.range_checks.into_iter().collect(),
             model_output_scales: self.graph.get_output_scales()?,
             model_input_scales: self.graph.get_input_scales(),
-            total_const_size,
+            total_const_size: res.total_const_size,
             check_mode,
             version: env!("CARGO_PKG_VERSION").to_string(),
             num_blinding_factors: None,
@@ -558,204 +555,12 @@ impl Model {
     /// * `model_inputs` - A vector of [Tensor]s to use as inputs to the model.
     /// * `run_args` - [RunArgs]
     pub fn forward(&self, model_inputs: &[Tensor<Fp>]) -> Result<ForwardResult, Box<dyn Error>> {
-        let mut results: BTreeMap<&usize, Vec<Tensor<Fp>>> = BTreeMap::new();
-        let mut max_lookup_inputs = 0;
-        let mut min_lookup_inputs = 0;
-
-        let input_shapes = self.graph.input_shapes()?;
-
-        for (i, input_idx) in self.graph.inputs.iter().enumerate() {
-            let mut input = model_inputs[i].clone();
-            input.reshape(&input_shapes[i])?;
-            results.insert(input_idx, vec![input]);
-        }
-
-        for (idx, n) in self.graph.nodes.iter() {
-            let mut inputs = vec![];
-            if n.is_input() {
-                let t = results.get(idx).ok_or(GraphError::MissingResults)?[0].clone();
-                inputs.push(t);
-            } else {
-                for (idx, outlet) in n.inputs().iter() {
-                    match results.get(&idx) {
-                        Some(value) => inputs.push(value[*outlet].clone()),
-                        None => return Err(Box::new(GraphError::MissingNode(*idx))),
-                    }
-                }
-            };
-
-            debug!("executing {}: {}", idx, n.as_str());
-            debug!("dims: {:?}", n.out_dims());
-            debug!(
-                "input_dims: {:?}",
-                inputs.iter().map(|x| x.dims()).collect::<Vec<_>>()
-            );
-
-            if n.is_lookup() {
-                let (mut min, mut max) = (0, 0);
-                for i in &inputs {
-                    max = max.max(
-                        i.iter()
-                            .map(|x| felt_to_i128(*x))
-                            .max()
-                            .ok_or("missing max")?,
-                    );
-                    min = min.min(
-                        i.iter()
-                            .map(|x| felt_to_i128(*x))
-                            .min()
-                            .ok_or("missing min")?,
-                    );
-                }
-                max_lookup_inputs = max_lookup_inputs.max(max);
-                min_lookup_inputs = min_lookup_inputs.min(min);
-                debug!("max lookup inputs: {}", max);
-                debug!("min lookup inputs: {}", min);
-            }
-
-            match n {
-                NodeType::Node(n) => {
-                    // execute the op
-                    let start = instant::Instant::now();
-                    let mut res = Op::<Fp>::f(&n.opkind, &inputs)?;
-                    res.output.reshape(&n.out_dims)?;
-                    let elapsed = start.elapsed();
-                    trace!("op took: {:?}", elapsed);
-                    // see if any of the intermediate lookup calcs are the max
-                    if !res.intermediate_lookups.is_empty() {
-                        let (mut min, mut max) = (0, 0);
-                        for i in &res.intermediate_lookups {
-                            max = max.max(i.clone().into_iter().max().ok_or("missing max")?);
-                            min = min.min(i.clone().into_iter().min().ok_or("missing min")?);
-                        }
-                        max_lookup_inputs = max_lookup_inputs.max(max);
-                        min_lookup_inputs = min_lookup_inputs.min(min);
-                        debug!("intermediate max lookup inputs: {}", max);
-                        debug!("intermediate min lookup inputs: {}", min);
-                    }
-                    debug!(
-                        "------------ output node int {}: {} \n ------------ float: {} \n ------------ max: {} \n ------------ min: {} \n ------------ scale: {}",
-                        idx,
-                        res.output.map(crate::fieldutils::felt_to_i32).show(),
-                        res.output
-                            .map(|x| crate::fieldutils::felt_to_f64(x)
-                                / scale_to_multiplier(n.out_scale))
-                            .show(),
-                        res.output.clone().into_iter().map(crate::fieldutils::felt_to_i128).max().unwrap_or(0),
-                        res.output.clone().into_iter().map(crate::fieldutils::felt_to_i128).min().unwrap_or(0),
-                        n.out_scale
-                    );
-                    results.insert(idx, vec![res.output]);
-                }
-                NodeType::SubGraph {
-                    model,
-                    output_mappings,
-                    input_mappings,
-                    inputs: input_tuple,
-                    ..
-                } => {
-                    let orig_inputs = inputs.clone();
-                    let input_mappings = input_mappings.clone();
-
-                    let input_dims = inputs.iter().map(|inp| inp.dims());
-                    let num_iter = number_of_iterations(&input_mappings, input_dims.collect());
-
-                    debug!(
-                        "{} iteration(s) in a subgraph with inputs {:?} and sources {:?}",
-                        num_iter, input_tuple, model.graph.inputs
-                    );
-
-                    debug!("input_mappings: {:?}", input_mappings);
-
-                    let mut full_results: Vec<Tensor<Fp>> = vec![];
-
-                    for i in 0..num_iter {
-                        // replace the Stacked input with the current chunk iter
-                        for ((mapping, inp), og_input) in
-                            input_mappings.iter().zip(&mut inputs).zip(&orig_inputs)
-                        {
-                            if let InputMapping::Stacked { axis, chunk } = mapping {
-                                let start = i * chunk;
-                                let end = (i + 1) * chunk;
-                                let t = crate::tensor::ops::slice(og_input, axis, &start, &end)?;
-                                *inp = t;
-                            }
-                        }
-
-                        let res = model.forward(&inputs)?;
-                        // recursively get the max lookup inputs for subgraphs
-                        max_lookup_inputs = max_lookup_inputs.max(res.max_lookup_inputs);
-                        min_lookup_inputs = min_lookup_inputs.min(res.min_lookup_inputs);
-
-                        let mut outlets = BTreeMap::new();
-                        for (mappings, outlet_res) in output_mappings.iter().zip(res.outputs) {
-                            for mapping in mappings {
-                                match mapping {
-                                    OutputMapping::Single { outlet, .. } => {
-                                        outlets.insert(outlet, outlet_res.clone());
-                                    }
-                                    OutputMapping::Stacked { outlet, axis, .. } => {
-                                        if !full_results.is_empty() {
-                                            let stacked_res = crate::tensor::ops::concat(
-                                                &[&full_results[*outlet], &outlet_res],
-                                                *axis,
-                                            )?;
-
-                                            outlets.insert(outlet, stacked_res);
-                                        } else {
-                                            outlets.insert(outlet, outlet_res.clone());
-                                        }
-                                    }
-                                }
-                            }
-                        }
-
-                        full_results = outlets.into_values().collect_vec();
-
-                        let output_states = output_state_idx(output_mappings);
-                        let input_states = input_state_idx(&input_mappings);
-
-                        assert_eq!(input_states.len(), output_states.len());
-
-                        for (input_idx, output_idx) in input_states.iter().zip(output_states) {
-                            inputs[*input_idx] = full_results[output_idx].clone();
-                        }
-                    }
-
-                    trace!(
-                        "------------ output subgraph node {}: {:?}",
-                        idx,
-                        full_results
-                            .iter()
-                            .map(|x|
-                            // convert to tensor i32
-                            x.map(crate::fieldutils::felt_to_i32).show())
-                            .collect_vec()
-                    );
-
-                    results.insert(idx, full_results);
-                }
-            }
-        }
-
-        let output_nodes = self.graph.outputs.iter();
-        debug!(
-            "model outputs are nodes: {:?}",
-            output_nodes.clone().collect_vec()
-        );
-        let outputs = output_nodes
-            .map(|(idx, outlet)| {
-                Ok(results.get(&idx).ok_or(GraphError::MissingResults)?[*outlet].clone())
-            })
-            .collect::<Result<Vec<_>, GraphError>>()?;
-
-        let res = ForwardResult {
-            outputs,
-            max_lookup_inputs,
-            min_lookup_inputs,
-        };
-
-        Ok(res)
+        let valtensor_inputs: Vec<ValTensor<Fp>> = model_inputs
+            .iter()
+            .map(|x| x.map(|elem| ValType::Value(Value::known(elem))).into())
+            .collect();
+        let res = self.dummy_layout(&RunArgs::default(), &valtensor_inputs)?;
+        Ok(res.into())
     }
 
     /// Loads an Onnx model from a specified path.
@@ -806,7 +611,7 @@ impl Model {
         for (symbol, value) in run_args.variables.iter() {
             let symbol = model.symbol_table.sym(symbol);
             symbol_values = symbol_values.with(&symbol, *value as i64);
-            info!("set {} to {}", symbol, value);
+            debug!("set {} to {}", symbol, value);
         }
 
         // Note: do not optimize the model, as the layout will depend on underlying hardware
@@ -1066,6 +871,7 @@ impl Model {
                         i,
                         symbol_values,
                         run_args.div_rebasing,
+                        run_args.rebase_frac_zero_constants,
                     )?;
                     if let Some(ref scales) = override_input_scales {
                         if let Some(inp) = n.opkind.get_input() {
@@ -1206,8 +1012,8 @@ impl Model {
         );
         // set scale for HybridOp::RangeCheck and call self.conf_lookup on that op for percentage tolerance case
         let input = &vars.advices[0];
-        let output = &vars.advices[1];
-        let index = &vars.advices[2];
+        let output = &vars.advices[2];
+        let index = &vars.advices[1];
         for op in required_lookups {
             base_gate.configure_lookup(meta, input, output, index, lookup_range, logrows, &op)?;
         }
@@ -1326,7 +1132,7 @@ impl Model {
 
         // Then number of columns in the circuits
         #[cfg(not(target_arch = "wasm32"))]
-        info!(
+        debug!(
             "{} {} {} (coord={}, constants={})",
             "model uses".blue(),
             num_rows.to_string().blue(),
@@ -1368,18 +1174,29 @@ impl Model {
             };
 
             debug!(
-                "laying out {}: {}, row:{}, coord:{}, total_constants: {}",
+                "laying out {}: {}, row:{}, coord:{}, total_constants: {}, max_lookup_inputs: {}, min_lookup_inputs: {}",
                 idx,
                 node.as_str(),
                 region.row(),
                 region.linear_coord(),
-                region.total_constants()
+                region.total_constants(),
+                region.max_lookup_inputs(),
+                region.min_lookup_inputs()
             );
             debug!("dims: {:?}", node.out_dims());
             debug!(
                 "input_dims {:?}",
                 values.iter().map(|v| v.dims()).collect_vec()
             );
+            debug!("output scales: {:?}", node.out_scales());
+            debug!("input indices: {:?}", node.inputs());
+            debug!(
+                "input scales: {:?}",
+                node.inputs()
+                    .iter()
+                    .map(|(idx, outlet)| self.graph.nodes[idx].out_scales()[*outlet])
+                    .collect_vec()
+            );
 
             match &node {
                 NodeType::Node(n) => {
@@ -1522,28 +1339,13 @@ impl Model {
     pub fn dummy_layout(
         &self,
         run_args: &RunArgs,
-        input_shapes: &[Vec<usize>],
-    ) -> Result<(usize, usize, usize), Box<dyn Error>> {
-        info!("calculating num of constraints using dummy model layout...");
+        inputs: &[ValTensor<Fp>],
+    ) -> Result<DummyPassRes, Box<dyn Error>> {
+        debug!("calculating num of constraints using dummy model layout...");
 
         let start_time = instant::Instant::now();
 
         let mut results = BTreeMap::<usize, Vec<ValTensor<Fp>>>::new();
-        let default_value = if !self.visibility.input.is_fixed() {
-            ValType::Value(Value::<Fp>::unknown())
-        } else {
-            ValType::Constant(Fp::ONE)
-        };
-
-        let inputs: Vec<ValTensor<Fp>> = input_shapes
-            .iter()
-            .map(|shape| {
-                let mut t: ValTensor<Fp> =
-                    vec![default_value.clone(); shape.iter().product()].into();
-                t.reshape(shape)?;
-                Ok(t)
-            })
-            .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
 
         for (i, input_idx) in self.graph.inputs.iter().enumerate() {
             results.insert(*input_idx, vec![inputs[i].clone()]);
@@ -1578,12 +1380,12 @@ impl Model {
                 .collect::<Result<Vec<_>, Box<dyn Error>>>()?;
 
             let _ = outputs
-                .into_iter()
+                .iter()
                 .zip(comparator)
                 .map(|(o, c)| {
                     dummy_config.layout(
                         &mut region,
-                        &[o, c],
+                        &[o.clone(), c],
                         Box::new(HybridOp::RangeCheck(run_args.tolerance)),
                     )
                 })
@@ -1599,7 +1401,7 @@ impl Model {
 
         // Then number of columns in the circuits
         #[cfg(not(target_arch = "wasm32"))]
-        info!(
+        debug!(
             "{} {} {} (coord={}, constants={})",
             "model uses".blue(),
             region.row().to_string().blue(),
@@ -1608,11 +1410,26 @@ impl Model {
             region.total_constants().to_string().red()
         );
 
-        Ok((
-            region.row(),
-            region.linear_coord(),
-            region.total_constants(),
-        ))
+        let outputs = outputs
+            .iter()
+            .map(|x| {
+                x.get_felt_evals()
+                    .unwrap_or(Tensor::new(Some(&[Fp::ZERO]), &[1]).unwrap())
+            })
+            .collect();
+
+        let res = DummyPassRes {
+            num_rows: region.row(),
+            linear_coord: region.linear_coord(),
+            total_const_size: region.total_constants(),
+            lookup_ops: region.used_lookups(),
+            range_checks: region.used_range_checks(),
+            max_lookup_inputs: region.max_lookup_inputs(),
+            min_lookup_inputs: region.min_lookup_inputs(),
+            outputs,
+        };
+
+        Ok(res)
     }
 
     /// Retrieves all constants from the model.

src/graph/node.rs

@@ -12,16 +12,12 @@ use crate::circuit::Constant;
 use crate::circuit::Input;
 use crate::circuit::Op;
 use crate::circuit::Unknown;
-use crate::fieldutils::felt_to_i128;
-use crate::fieldutils::i128_to_felt;
 #[cfg(not(target_arch = "wasm32"))]
 use crate::graph::new_op_from_onnx;
 use crate::tensor::Tensor;
 use crate::tensor::TensorError;
 use halo2curves::bn256::Fr as Fp;
 #[cfg(not(target_arch = "wasm32"))]
-use itertools::Itertools;
-#[cfg(not(target_arch = "wasm32"))]
 use log::trace;
 use serde::Deserialize;
 use serde::Serialize;
@@ -94,10 +90,6 @@ impl Op<Fp> for Rescaled {
         Op::<Fp>::out_scale(&*self.inner, in_scales)
     }
 
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        self.inner.required_lookups()
-    }
-
     fn layout(
         &self,
         config: &mut crate::circuit::BaseConfig<Fp>,
@@ -126,14 +118,14 @@ impl Op<Fp> for Rescaled {
 pub struct RebaseScale {
     /// The operation that has to be rescaled.
     pub inner: Box<SupportedOp>,
-    /// the multiplier applied to the node output
-    pub multiplier: f64,
+    /// rebase op
+    pub rebase_op: HybridOp,
     /// scale being rebased to
     pub target_scale: i32,
     /// The original scale of the operation's inputs.
     pub original_scale: i32,
-    /// if true then the operation is a multiplicative division
-    pub div_rebasing: bool,
+    /// multiplier
+    pub multiplier: f64,
 }
 
 impl RebaseScale {
@@ -152,20 +144,27 @@ impl RebaseScale {
             let multiplier =
                 scale_to_multiplier(op_out_scale - global_scale * scale_rebase_multiplier as i32);
             if let Some(op) = inner.get_rebased() {
+                let multiplier = op.multiplier * multiplier;
                 SupportedOp::RebaseScale(RebaseScale {
                     inner: op.inner.clone(),
                     target_scale: op.target_scale,
-                    multiplier: op.multiplier * multiplier,
+                    multiplier,
+                    rebase_op: HybridOp::Div {
+                        denom: crate::circuit::utils::F32((multiplier) as f32),
+                        use_range_check_for_int: !div_rebasing,
+                    },
                     original_scale: op.original_scale,
-                    div_rebasing,
                 })
             } else {
                 SupportedOp::RebaseScale(RebaseScale {
                     inner: Box::new(inner),
                     target_scale: global_scale * scale_rebase_multiplier as i32,
                     multiplier,
+                    rebase_op: HybridOp::Div {
+                        denom: crate::circuit::utils::F32(multiplier as f32),
+                        use_range_check_for_int: !div_rebasing,
+                    },
                     original_scale: op_out_scale,
-                    div_rebasing,
                 })
             }
         } else {
@@ -183,12 +182,16 @@ impl RebaseScale {
         if (op_out_scale < (target_scale)) && !inner.is_constant() && !inner.is_input() {
             let multiplier = scale_to_multiplier(op_out_scale - target_scale);
             if let Some(op) = inner.get_rebased() {
+                let multiplier = op.multiplier * multiplier;
                 SupportedOp::RebaseScale(RebaseScale {
                     inner: op.inner.clone(),
                     target_scale: op.target_scale,
-                    multiplier: op.multiplier * multiplier,
+                    multiplier,
                     original_scale: op.original_scale,
-                    div_rebasing,
+                    rebase_op: HybridOp::Div {
+                        denom: crate::circuit::utils::F32((multiplier) as f32),
+                        use_range_check_for_int: !div_rebasing,
+                    },
                 })
             } else {
                 SupportedOp::RebaseScale(RebaseScale {
@@ -196,22 +199,16 @@ impl RebaseScale {
                     target_scale,
                     multiplier,
                     original_scale: op_out_scale,
-                    div_rebasing,
+                    rebase_op: HybridOp::Div {
+                        denom: crate::circuit::utils::F32(multiplier as f32),
+                        use_range_check_for_int: !div_rebasing,
+                    },
                 })
             }
         } else {
             inner
         }
     }
-
-    /// Calculate the require range bracket for the operation
-    fn range_bracket(&self) -> i128 {
-        if self.div_rebasing {
-            0
-        } else {
-            self.multiplier as i128 - 1
-        }
-    }
 }
 
 impl Op<Fp> for RebaseScale {
@@ -220,28 +217,17 @@ impl Op<Fp> for RebaseScale {
     }
     fn f(&self, x: &[Tensor<Fp>]) -> Result<crate::circuit::ForwardResult<Fp>, TensorError> {
         let mut res = Op::<Fp>::f(&*self.inner, x)?;
-
-        if self.div_rebasing {
-            let ri = res.output.map(felt_to_i128);
-            let rescaled = crate::tensor::ops::nonlinearities::const_div(&ri, self.multiplier);
-            res.output = rescaled.map(i128_to_felt);
-            res.intermediate_lookups.push(ri);
-        } else {
-            let ri = res.output.map(felt_to_i128);
-            let divisor = Tensor::from(vec![self.multiplier as i128].into_iter());
-            let rescaled = crate::tensor::ops::div(&[ri, divisor.clone()])?;
-            res.output = rescaled.map(i128_to_felt);
-            res.intermediate_lookups.extend([-divisor.clone(), divisor]);
-        }
+        let rebase_res = Op::<Fp>::f(&self.rebase_op, &[res.output])?;
+        res.output = rebase_res.output;
 
         Ok(res)
     }
 
     fn as_string(&self) -> String {
         format!(
-            "REBASED (div={:?}, div_r={}) ({})",
+            "REBASED (div={:?}, rebasing_op={}) ({})",
             self.multiplier,
-            self.div_rebasing,
+            <HybridOp as Op<Fp>>::as_string(&self.rebase_op),
             self.inner.as_string()
         )
     }
@@ -250,25 +236,6 @@ impl Op<Fp> for RebaseScale {
         Ok(self.target_scale)
     }
 
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        let mut lookups: Vec<LookupOp> = self.inner.required_lookups();
-        if self.div_rebasing {
-            lookups.push(LookupOp::Div {
-                denom: crate::circuit::utils::F32(self.multiplier as f32),
-            });
-        }
-        lookups
-    }
-
-    fn required_range_checks(&self) -> Vec<crate::circuit::table::Range> {
-        let mut range_checks = self.inner.required_range_checks();
-        if !self.div_rebasing {
-            let bracket = self.range_bracket();
-            range_checks.push((-bracket, bracket));
-        }
-        range_checks
-    }
-
     fn layout(
         &self,
         config: &mut crate::circuit::BaseConfig<Fp>,
@@ -278,25 +245,8 @@ impl Op<Fp> for RebaseScale {
         let original_res = self
             .inner
             .layout(config, region, values)?
-            .ok_or("no layout")?;
-
-        if !self.div_rebasing {
-            Ok(Some(crate::circuit::layouts::div(
-                config,
-                region,
-                &[original_res],
-                Fp::from(self.multiplier as u64),
-            )?))
-        } else {
-            Ok(Some(crate::circuit::layouts::nonlinearity(
-                config,
-                region,
-                &[original_res],
-                &LookupOp::Div {
-                    denom: crate::circuit::utils::F32(self.multiplier as f32),
-                },
-            )?))
-        }
+            .ok_or("no inner layout")?;
+        self.rebase_op.layout(config, region, &[original_res])
     }
 
     fn clone_dyn(&self) -> Box<dyn Op<Fp>> {
@@ -479,14 +429,6 @@ impl Op<Fp> for SupportedOp {
         self
     }
 
-    fn required_lookups(&self) -> Vec<LookupOp> {
-        self.as_op().required_lookups()
-    }
-
-    fn required_range_checks(&self) -> Vec<crate::circuit::table::Range> {
-        self.as_op().required_range_checks()
-    }
-
     fn out_scale(&self, in_scales: Vec<crate::Scale>) -> Result<crate::Scale, Box<dyn Error>> {
         self.as_op().out_scale(in_scales)
     }
@@ -520,15 +462,7 @@ impl Tabled for Node {
 
     fn headers() -> Vec<std::borrow::Cow<'static, str>> {
         let mut headers = Vec::with_capacity(Self::LENGTH);
-        for i in [
-            "idx",
-            "opkind",
-            "out_scale",
-            "inputs",
-            "out_dims",
-            "required_lookups",
-            "required_range_checks",
-        ] {
+        for i in ["idx", "opkind", "out_scale", "inputs", "out_dims"] {
             headers.push(std::borrow::Cow::Borrowed(i));
         }
         headers
@@ -541,18 +475,6 @@ impl Tabled for Node {
         fields.push(std::borrow::Cow::Owned(self.out_scale.to_string()));
         fields.push(std::borrow::Cow::Owned(display_vector(&self.inputs)));
         fields.push(std::borrow::Cow::Owned(display_vector(&self.out_dims)));
-        fields.push(std::borrow::Cow::Owned(format!(
-            "{:?}",
-            self.opkind
-                .required_lookups()
-                .iter()
-                .map(<LookupOp as Op<Fp>>::as_string)
-                .collect_vec()
-        )));
-        fields.push(std::borrow::Cow::Owned(format!(
-            "{:?}",
-            self.opkind.required_range_checks()
-        )));
         fields
     }
 }
@@ -583,9 +505,8 @@ impl Node {
         idx: usize,
         symbol_values: &SymbolValues,
         div_rebasing: bool,
+        rebase_frac_zero_constants: bool,
     ) -> Result<Self, Box<dyn Error>> {
-        use log::warn;
-
         trace!("Create {:?}", node);
         trace!("Create op {:?}", node.op);
 
@@ -623,6 +544,7 @@ impl Node {
             node.clone(),
             &mut inputs,
             symbol_values,
+            rebase_frac_zero_constants,
         )?; // parses the op name
 
         // we can only take the inputs as mutable once -- so we need to collect them first
@@ -678,8 +600,6 @@ impl Node {
                     input_node.bump_scale(out_scale);
                     in_scales[input] = out_scale;
                 }
-            } else {
-                warn!("input {} not found for rescaling, skipping ...", input);
             }
         }
 

src/graph/utilities.rs

@@ -243,6 +243,7 @@ pub fn new_op_from_onnx(
     node: OnnxNode<TypedFact, Box<dyn TypedOp>>,
     inputs: &mut [super::NodeType],
     symbol_values: &SymbolValues,
+    rebase_frac_zero_constants: bool,
 ) -> Result<(SupportedOp, Vec<usize>), Box<dyn std::error::Error>> {
     use crate::circuit::InputType;
 
@@ -261,7 +262,9 @@ pub fn new_op_from_onnx(
             inputs[index].bump_scale(scale);
             c.rebase_scale(scale)?;
             inputs[index].replace_opkind(SupportedOp::Constant(c.clone()));
-            Ok(SupportedOp::Linear(PolyOp::Identity))
+            Ok(SupportedOp::Linear(PolyOp::Identity {
+                out_scale: Some(scale),
+            }))
         } else {
             Ok(default_op)
         }
@@ -282,8 +285,8 @@ pub fn new_op_from_onnx(
                         "shift left".to_string(),
                     )));
                 }
-                SupportedOp::Nonlinear(LookupOp::Div {
-                    denom: crate::circuit::utils::F32(1.0 / 2.0f32.powf(raw_values[0])),
+                SupportedOp::Linear(PolyOp::Identity {
+                    out_scale: Some(input_scales[0] - raw_values[0] as i32),
                 })
             } else {
                 return Err(Box::new(GraphError::OpMismatch(
@@ -304,8 +307,8 @@ pub fn new_op_from_onnx(
                         "shift right".to_string(),
                     )));
                 }
-                SupportedOp::Nonlinear(LookupOp::Div {
-                    denom: crate::circuit::utils::F32(2.0f32.powf(raw_values[0])),
+                SupportedOp::Linear(PolyOp::Identity {
+                    out_scale: Some(input_scales[0] + raw_values[0] as i32),
                 })
             } else {
                 return Err(Box::new(GraphError::OpMismatch(
@@ -436,17 +439,16 @@ pub fn new_op_from_onnx(
             let op = load_op::<ScatterElements>(node.op(), idx, node.op().name().to_string())?;
             let axis = op.axis;
 
-            let mut op =
-                SupportedOp::Hybrid(crate::circuit::ops::hybrid::HybridOp::ScatterElements {
-                    dim: axis,
-                    constant_idx: None,
-                });
+            let mut op = SupportedOp::Linear(crate::circuit::ops::poly::PolyOp::ScatterElements {
+                dim: axis,
+                constant_idx: None,
+            });
 
             // if param_visibility.is_public() {
             if let Some(c) = inputs[1].opkind().get_mutable_constant() {
                 inputs[1].decrement_use();
                 deleted_indices.push(1);
-                op = SupportedOp::Hybrid(crate::circuit::ops::hybrid::HybridOp::ScatterElements {
+                op = SupportedOp::Linear(crate::circuit::ops::poly::PolyOp::ScatterElements {
                     dim: axis,
                     constant_idx: Some(c.raw_values.map(|x| x as usize)),
                 })
@@ -475,17 +477,16 @@ pub fn new_op_from_onnx(
             let op = load_op::<GatherElements>(node.op(), idx, node.op().name().to_string())?;
             let axis = op.axis;
 
-            let mut op =
-                SupportedOp::Hybrid(crate::circuit::ops::hybrid::HybridOp::GatherElements {
-                    dim: axis,
-                    constant_idx: None,
-                });
+            let mut op = SupportedOp::Linear(crate::circuit::ops::poly::PolyOp::GatherElements {
+                dim: axis,
+                constant_idx: None,
+            });
 
             // if param_visibility.is_public() {
             if let Some(c) = inputs[1].opkind().get_mutable_constant() {
                 inputs[1].decrement_use();
                 deleted_indices.push(inputs.len() - 1);
-                op = SupportedOp::Hybrid(crate::circuit::ops::hybrid::HybridOp::GatherElements {
+                op = SupportedOp::Linear(crate::circuit::ops::poly::PolyOp::GatherElements {
                     dim: axis,
                     constant_idx: Some(c.raw_values.map(|x| x as usize)),
                 })
@@ -544,7 +545,7 @@ pub fn new_op_from_onnx(
             // Raw values are always f32
             let raw_value = extract_tensor_value(op.0)?;
             // If bool or a tensor dimension then don't scale
-            let constant_scale = match dt {
+            let mut constant_scale = match dt {
                 DatumType::Bool
                 | DatumType::TDim
                 | DatumType::I64
@@ -559,6 +560,12 @@ pub fn new_op_from_onnx(
                 _ => return Err(Box::new(GraphError::UnsupportedDataType)),
             };
 
+            // if all raw_values are round then set scale to 0
+            let all_round = raw_value.iter().all(|x| (x).fract() == 0.0);
+            if all_round && rebase_frac_zero_constants {
+                constant_scale = 0;
+            }
+
             // Quantize the raw value
             let quantized_value =
                 quantize_tensor(raw_value.clone(), constant_scale, param_visibility)?;
@@ -665,8 +672,10 @@ pub fn new_op_from_onnx(
                 if unit == 0. {
                     SupportedOp::Nonlinear(LookupOp::ReLU)
                 } else {
+                    // get the non-constant index
+                    let non_const_idx = if const_idx == 0 { 1 } else { 0 };
                     SupportedOp::Nonlinear(LookupOp::Max {
-                        scale: scale_to_multiplier(inputs[0].out_scales()[0]).into(),
+                        scale: scale_to_multiplier(inputs[non_const_idx].out_scales()[0]).into(),
                         a: crate::circuit::utils::F32(unit),
                     })
                 }
@@ -707,8 +716,11 @@ pub fn new_op_from_onnx(
                     deleted_indices.push(const_idx);
                 }
 
+                // get the non-constant index
+                let non_const_idx = if const_idx == 0 { 1 } else { 0 };
+
                 SupportedOp::Nonlinear(LookupOp::Min {
-                    scale: scale_to_multiplier(inputs[0].out_scales()[0]).into(),
+                    scale: scale_to_multiplier(inputs[non_const_idx].out_scales()[0]).into(),
                     a: crate::circuit::utils::F32(unit),
                 })
             } else {
@@ -717,16 +729,12 @@ pub fn new_op_from_onnx(
         }
         "Recip" => {
             let in_scale = inputs[0].out_scales()[0];
+            let max_scale = std::cmp::max(scales.get_max(), in_scale);
             // If the input scale is larger than the params scale
-            let scale_diff = std::cmp::max(scales.input, scales.params) - inputs[0].out_scales()[0];
-            let additional_scale = if scale_diff > 0 {
-                scale_to_multiplier(scale_diff)
-            } else {
-                1.0
-            };
-
-            SupportedOp::Nonlinear(LookupOp::Recip {
-                scale: (scale_to_multiplier(in_scale).powf(2.0) * additional_scale).into(),
+            SupportedOp::Hybrid(HybridOp::Recip {
+                input_scale: (scale_to_multiplier(in_scale) as f32).into(),
+                output_scale: (scale_to_multiplier(max_scale) as f32).into(),
+                use_range_check_for_int: false,
             })
         }
 
@@ -751,7 +759,9 @@ pub fn new_op_from_onnx(
         "Scan" => {
             return Err("scan should never be analyzed explicitly".into());
         }
-        "QuantizeLinearU8" | "DequantizeLinearF32" => SupportedOp::Linear(PolyOp::Identity),
+        "QuantizeLinearU8" | "DequantizeLinearF32" => {
+            SupportedOp::Linear(PolyOp::Identity { out_scale: None })
+        }
         "Abs" => SupportedOp::Nonlinear(LookupOp::Abs),
         "Neg" => SupportedOp::Linear(PolyOp::Neg),
         "Sigmoid" => SupportedOp::Nonlinear(LookupOp::Sigmoid {
@@ -856,11 +866,11 @@ pub fn new_op_from_onnx(
                             }),
                         )?
                     } else {
-                        SupportedOp::Linear(PolyOp::Identity)
+                        SupportedOp::Linear(PolyOp::Identity { out_scale: None })
                     }
                 }
                 DatumType::F16 | DatumType::F32 | DatumType::F64 => {
-                    SupportedOp::Linear(PolyOp::Identity)
+                    SupportedOp::Linear(PolyOp::Identity { out_scale: None })
                 }
                 _ => return Err(Box::new(GraphError::UnsupportedDataType)),
             }
@@ -885,12 +895,15 @@ pub fn new_op_from_onnx(
                 let const_idx = const_idx[0];
                 if let Some(c) = inputs[const_idx].opkind().get_mutable_constant() {
                     if c.raw_values.len() == 1 && c.raw_values[0] < 1. {
-                        inputs[const_idx].decrement_use();
-                        deleted_indices.push(const_idx);
-                        op = SupportedOp::Nonlinear(LookupOp::Div {
-                            // we invert the constant for division
-                            denom: crate::circuit::utils::F32(1. / c.raw_values[0]),
-                        })
+                        // if not divisible by 2 then we need to add a range check
+                        let raw_values = 1.0 / c.raw_values[0];
+                        if raw_values.log2().fract() == 0.0 {
+                            inputs[const_idx].decrement_use();
+                            deleted_indices.push(const_idx);
+                            op = SupportedOp::Linear(PolyOp::Identity {
+                                out_scale: Some(input_scales[0] + raw_values.log2() as i32),
+                            });
+                        }
                     }
                 }
             }

src/graph/vars.rs

@@ -237,6 +237,11 @@ impl VarScales {
         std::cmp::max(self.input, self.params)
     }
 
+    ///
+    pub fn get_min(&self) -> crate::Scale {
+        std::cmp::min(self.input, self.params)
+    }
+
     /// Place in [VarScales] struct.
     pub fn from_args(args: &RunArgs) -> Result<Self, Box<dyn Error>> {
         Ok(Self {

src/lib.rs

@@ -7,7 +7,6 @@
     overflowing_literals,
     path_statements,
     patterns_in_fns_without_body,
-    private_in_public,
     unconditional_recursion,
     unused,
     unused_allocation,
@@ -111,8 +110,11 @@ pub struct RunArgs {
     #[arg(long, default_value = "private")]
     pub param_visibility: Visibility,
     #[arg(long, default_value = "false")]
-    /// Multiplicative division
+    /// Rebase the scale using lookup table for division instead of using a range check
     pub div_rebasing: bool,
+    /// Should constants with 0.0 fraction be rebased to scale 0
+    #[arg(long, default_value = "false")]
+    pub rebase_frac_zero_constants: bool,
     /// check mode (safe, unsafe, etc)
     #[arg(long, default_value = "unsafe")]
     pub check_mode: CheckMode,
@@ -133,6 +135,7 @@ impl Default for RunArgs {
             output_visibility: Visibility::Public,
             param_visibility: Visibility::Private,
             div_rebasing: false,
+            rebase_frac_zero_constants: false,
             check_mode: CheckMode::UNSAFE,
         }
     }

src/pfsys/mod.rs

@@ -40,8 +40,24 @@ use std::ops::Deref;
 use std::path::PathBuf;
 use thiserror::Error as thisError;
 
+// not wasm
+#[cfg(not(target_arch = "wasm32"))]
+use lazy_static::lazy_static;
+
 use halo2curves::bn256::{Bn256, Fr, G1Affine};
 
+#[cfg(not(target_arch = "wasm32"))]
+// Buf writer capacity
+lazy_static! {
+    static ref EZKL_BUF_CAPACITY: usize = std::env::var("EZKL_BUF_CAPACITY")
+        .unwrap_or("8000".to_string())
+        .parse()
+        .unwrap();
+}
+
+#[cfg(target_arch = "wasm32")]
+const EZKL_BUF_CAPACITY: &usize = &8000;
+
 #[allow(missing_docs)]
 #[derive(
     ValueEnum, Copy, Clone, Default, Debug, PartialEq, Eq, Deserialize, Serialize, PartialOrd,
@@ -315,7 +331,7 @@ where
     /// Saves the Proof to a specified `proof_path`.
     pub fn save(&self, proof_path: &PathBuf) -> Result<(), Box<dyn Error>> {
         let file = std::fs::File::create(proof_path)?;
-        let mut writer = BufWriter::new(file);
+        let mut writer = BufWriter::with_capacity(*EZKL_BUF_CAPACITY, file);
         serde_json::to_writer(&mut writer, &self)?;
         Ok(())
     }
@@ -555,6 +571,7 @@ where
             verifier_params,
             pk.get_vk(),
             strategy,
+            verifier_params.n(),
         )?;
     }
     let elapsed = now.elapsed();
@@ -642,6 +659,7 @@ pub fn verify_proof_circuit<
     params: &'params Scheme::ParamsVerifier,
     vk: &VerifyingKey<Scheme::Curve>,
     strategy: Strategy,
+    orig_n: u64,
 ) -> Result<Strategy::Output, halo2_proofs::plonk::Error>
 where
     Scheme::Scalar: SerdeObject
@@ -662,7 +680,7 @@ where
     trace!("instances {:?}", instances);
 
     let mut transcript = TranscriptReadBuffer::init(Cursor::new(snark.proof.clone()));
-    verify_proof::<Scheme, V, _, TR, _>(params, vk, strategy, instances, &mut transcript)
+    verify_proof::<Scheme, V, _, TR, _>(params, vk, strategy, instances, &mut transcript, orig_n)
 }
 
 /// Loads a [VerifyingKey] at `path`.
@@ -678,7 +696,7 @@ where
     info!("loading verification key from {:?}", path);
     let f =
         File::open(path.clone()).map_err(|_| format!("failed to load vk at {}", path.display()))?;
-    let mut reader = BufReader::new(f);
+    let mut reader = BufReader::with_capacity(*EZKL_BUF_CAPACITY, f);
     VerifyingKey::<Scheme::Curve>::read::<_, C>(
         &mut reader,
         halo2_proofs::SerdeFormat::RawBytes,
@@ -700,7 +718,7 @@ where
     info!("loading proving key from {:?}", path);
     let f =
         File::open(path.clone()).map_err(|_| format!("failed to load pk at {}", path.display()))?;
-    let mut reader = BufReader::new(f);
+    let mut reader = BufReader::with_capacity(*EZKL_BUF_CAPACITY, f);
     ProvingKey::<Scheme::Curve>::read::<_, C>(
         &mut reader,
         halo2_proofs::SerdeFormat::RawBytes,
@@ -720,7 +738,7 @@ where
 {
     info!("saving proving key 💾");
     let f = File::create(path)?;
-    let mut writer = BufWriter::new(f);
+    let mut writer = BufWriter::with_capacity(*EZKL_BUF_CAPACITY, f);
     vk.write(&mut writer, halo2_proofs::SerdeFormat::RawBytes)?;
     writer.flush()?;
     Ok(())
@@ -737,7 +755,7 @@ where
 {
     info!("saving verification key 💾");
     let f = File::create(path)?;
-    let mut writer = BufWriter::new(f);
+    let mut writer = BufWriter::with_capacity(*EZKL_BUF_CAPACITY, f);
     vk.write(&mut writer, halo2_proofs::SerdeFormat::RawBytes)?;
     writer.flush()?;
     Ok(())
@@ -750,7 +768,7 @@ pub fn save_params<Scheme: CommitmentScheme>(
 ) -> Result<(), io::Error> {
     info!("saving parameters 💾");
     let f = File::create(path)?;
-    let mut writer = BufWriter::new(f);
+    let mut writer = BufWriter::with_capacity(*EZKL_BUF_CAPACITY, f);
     params.write(&mut writer)?;
     writer.flush()?;
     Ok(())
@@ -840,6 +858,7 @@ pub(crate) fn verify_proof_circuit_kzg<
     proof: Snark<Fr, G1Affine>,
     vk: &VerifyingKey<G1Affine>,
     strategy: Strategy,
+    orig_n: u64,
 ) -> Result<Strategy::Output, halo2_proofs::plonk::Error> {
     match proof.transcript_type {
         TranscriptType::EVM => verify_proof_circuit::<
@@ -849,7 +868,7 @@ pub(crate) fn verify_proof_circuit_kzg<
             _,
             _,
             EvmTranscript<G1Affine, _, _, _>,
-        >(&proof, params, vk, strategy),
+        >(&proof, params, vk, strategy, orig_n),
         TranscriptType::Poseidon => verify_proof_circuit::<
             Fr,
             VerifierSHPLONK<'_, Bn256>,
@@ -857,7 +876,7 @@ pub(crate) fn verify_proof_circuit_kzg<
             _,
             _,
             PoseidonTranscript<NativeLoader, _>,
-        >(&proof, params, vk, strategy),
+        >(&proof, params, vk, strategy, orig_n),
     }
 }
 

src/python.rs

@@ -15,7 +15,7 @@ use crate::graph::{
 use crate::pfsys::evm::aggregation::AggregationCircuit;
 use crate::pfsys::{
     load_pk, load_vk, save_params, save_vk, srs::gen_srs as ezkl_gen_srs, srs::load_srs, ProofType,
-    Snark, TranscriptType,
+    TranscriptType,
 };
 use crate::RunArgs;
 use ethers::types::H160;
@@ -162,6 +162,8 @@ struct PyRunArgs {
     #[pyo3(get, set)]
     pub div_rebasing: bool,
     #[pyo3(get, set)]
+    pub rebase_frac_zero_constants: bool,
+    #[pyo3(get, set)]
     pub check_mode: CheckMode,
 }
 
@@ -190,6 +192,7 @@ impl From<PyRunArgs> for RunArgs {
             param_visibility: py_run_args.param_visibility,
             variables: py_run_args.variables,
             div_rebasing: py_run_args.div_rebasing,
+            rebase_frac_zero_constants: py_run_args.rebase_frac_zero_constants,
             check_mode: py_run_args.check_mode,
         }
     }
@@ -210,6 +213,7 @@ impl Into<PyRunArgs> for RunArgs {
             param_visibility: self.param_visibility,
             variables: self.variables,
             div_rebasing: self.div_rebasing,
+            rebase_frac_zero_constants: self.rebase_frac_zero_constants,
             check_mode: self.check_mode,
         }
     }
@@ -685,14 +689,23 @@ fn prove(
     settings_path=PathBuf::from(DEFAULT_SETTINGS),
     vk_path=PathBuf::from(DEFAULT_VK),
     srs_path=None,
+    non_reduced_srs=Some(DEFAULT_USE_REDUCED_SRS_FOR_VERIFICATION.parse::<bool>().unwrap()),
 ))]
 fn verify(
     proof_path: PathBuf,
     settings_path: PathBuf,
     vk_path: PathBuf,
     srs_path: Option<PathBuf>,
+    non_reduced_srs: Option<bool>,
 ) -> Result<bool, PyErr> {
-    crate::execute::verify(proof_path, settings_path, vk_path, srs_path).map_err(|e| {
+    crate::execute::verify(
+        proof_path,
+        settings_path,
+        vk_path,
+        srs_path,
+        non_reduced_srs,
+    )
+    .map_err(|e| {
         let err_str = format!("Failed to run verify: {}", e);
         PyRuntimeError::new_err(err_str)
     })?;
@@ -1093,16 +1106,6 @@ fn create_evm_verifier_aggr(
     Ok(true)
 }
 
-/// print hex representation of a proof
-#[pyfunction(signature = (proof_path))]
-fn print_proof_hex(proof_path: PathBuf) -> Result<String, PyErr> {
-    let proof = Snark::load::<KZGCommitmentScheme<Bn256>>(&proof_path)
-        .map_err(|_| PyIOError::new_err("Failed to load proof"))?;
-
-    let hex_str = hex::encode(proof.proof);
-    Ok(format!("0x{}", hex_str))
-}
-
 // Python Module
 #[pymodule]
 fn ezkl(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
@@ -1141,7 +1144,6 @@ fn ezkl(_py: Python<'_>, m: &PyModule) -> PyResult<()> {
     m.add_function(wrap_pyfunction!(deploy_vk_evm, m)?)?;
     m.add_function(wrap_pyfunction!(deploy_da_evm, m)?)?;
     m.add_function(wrap_pyfunction!(verify_evm, m)?)?;
-    m.add_function(wrap_pyfunction!(print_proof_hex, m)?)?;
     m.add_function(wrap_pyfunction!(setup_test_evm_witness, m)?)?;
     m.add_function(wrap_pyfunction!(create_evm_verifier_aggr, m)?)?;
     m.add_function(wrap_pyfunction!(create_evm_data_attestation, m)?)?;

src/tensor/mod.rs

@@ -580,16 +580,16 @@ impl<T: Clone + TensorType> Tensor<T> {
     /// use ezkl::tensor::Tensor;
     /// let mut a = Tensor::<i32>::new(Some(&[1,2,3,4,5,6]), &[2, 3]).unwrap();
     /// let expected = Tensor::<i32>::new(Some(&[1, 2, 3, 4, 5, 6, 0, 0]), &[8]).unwrap();
-    /// assert_eq!(a.pad_to_zero_rem(4).unwrap(), expected);
+    /// assert_eq!(a.pad_to_zero_rem(4, 0).unwrap(), expected);
     ///
     /// let expected = Tensor::<i32>::new(Some(&[1, 2, 3, 4, 5, 6, 0, 0, 0]), &[9]).unwrap();
-    /// assert_eq!(a.pad_to_zero_rem(9).unwrap(), expected);
+    /// assert_eq!(a.pad_to_zero_rem(9, 0).unwrap(), expected);
     /// ```
-    pub fn pad_to_zero_rem(&self, n: usize) -> Result<Tensor<T>, TensorError> {
+    pub fn pad_to_zero_rem(&self, n: usize, pad: T) -> Result<Tensor<T>, TensorError> {
         let mut inner = self.inner.clone();
         let remainder = self.len() % n;
         if remainder != 0 {
-            inner.resize(self.len() + n - remainder, T::zero().unwrap());
+            inner.resize(self.len() + n - remainder, pad);
         }
         Tensor::new(Some(&inner), &[inner.len()])
     }

src/tensor/ops.rs

@@ -243,7 +243,7 @@ pub fn and<
 /// Some(&[1, 0, 1, 0, 1, 0]),
 /// &[2, 3],
 /// ).unwrap();
-/// let result = equals(&a, &b).unwrap().0;
+/// let result = equals(&a, &b).unwrap();
 /// let expected = Tensor::<i128>::new(Some(&[1, 0, 1, 0, 1, 1]), &[2, 3]).unwrap();
 /// assert_eq!(result, expected);
 /// ```
@@ -260,7 +260,7 @@ pub fn equals<
 >(
     a: &Tensor<T>,
     b: &Tensor<T>,
-) -> Result<(Tensor<T>, Vec<Tensor<T>>), TensorError> {
+) -> Result<Tensor<T>, TensorError> {
     let a = a.clone();
     let b = b.clone();
 
@@ -268,7 +268,7 @@ pub fn equals<
 
     let result = nonlinearities::kronecker_delta(&diff);
 
-    Ok((result, vec![diff]))
+    Ok(result)
 }
 
 /// Greater than operation.
@@ -289,7 +289,7 @@ pub fn equals<
 /// ).unwrap();
 /// let result = greater(&a, &b).unwrap();
 /// let expected = Tensor::<i128>::new(Some(&[0, 1, 1, 0, 0, 0]), &[2, 3]).unwrap();
-/// assert_eq!(result.0, expected);
+/// assert_eq!(result, expected);
 /// ```
 pub fn greater<
     T: TensorType
@@ -302,7 +302,7 @@ pub fn greater<
 >(
     a: &Tensor<T>,
     b: &Tensor<T>,
-) -> Result<(Tensor<T>, Vec<Tensor<T>>), TensorError> {
+) -> Result<Tensor<T>, TensorError> {
     let mask_inter = (a.clone() - b.clone())?;
     let mask = mask_inter.map(|x| {
         if x > T::zero().ok_or(TensorError::Unsupported).unwrap() {
@@ -311,7 +311,7 @@ pub fn greater<
             T::zero().ok_or(TensorError::Unsupported).unwrap()
         }
     });
-    Ok((mask, vec![mask_inter]))
+    Ok(mask)
 }
 
 /// Greater equals than operation.
@@ -332,7 +332,7 @@ pub fn greater<
 /// ).unwrap();
 /// let result = greater_equal(&a, &b).unwrap();
 /// let expected = Tensor::<i128>::new(Some(&[1, 1, 1, 1, 0, 0]), &[2, 3]).unwrap();
-/// assert_eq!(result.0, expected);
+/// assert_eq!(result, expected);
 /// ```
 pub fn greater_equal<
     T: TensorType
@@ -345,7 +345,7 @@ pub fn greater_equal<
 >(
     a: &Tensor<T>,
     b: &Tensor<T>,
-) -> Result<(Tensor<T>, Vec<Tensor<T>>), TensorError> {
+) -> Result<Tensor<T>, TensorError> {
     let mask_inter = (a.clone() - b.clone())?;
     let mask = mask_inter.map(|x| {
         if x >= T::zero().ok_or(TensorError::Unsupported).unwrap() {
@@ -354,7 +354,7 @@ pub fn greater_equal<
             T::zero().ok_or(TensorError::Unsupported).unwrap()
         }
     });
-    Ok((mask, vec![mask_inter]))
+    Ok(mask)
 }
 
 /// Less than to operation.
@@ -375,7 +375,7 @@ pub fn greater_equal<
 /// ).unwrap();
 /// let result = less(&a, &b).unwrap();
 /// let expected = Tensor::<i128>::new(Some(&[0, 1, 0, 0, 0, 1]), &[2, 3]).unwrap();
-/// assert_eq!(result.0, expected);
+/// assert_eq!(result, expected);
 /// ```
 ///
 pub fn less<
@@ -389,7 +389,7 @@ pub fn less<
 >(
     a: &Tensor<T>,
     b: &Tensor<T>,
-) -> Result<(Tensor<T>, Vec<Tensor<T>>), TensorError> {
+) -> Result<Tensor<T>, TensorError> {
     // a < b <=> b > a
     greater(b, a)
 }
@@ -412,7 +412,7 @@ pub fn less<
 /// ).unwrap();
 /// let result = less_equal(&a, &b).unwrap();
 /// let expected = Tensor::<i128>::new(Some(&[1, 1, 0, 1, 1, 1]), &[2, 3]).unwrap();
-/// assert_eq!(result.0, expected);
+/// assert_eq!(result, expected);
 /// ```
 ///
 pub fn less_equal<
@@ -426,7 +426,7 @@ pub fn less_equal<
 >(
     a: &Tensor<T>,
     b: &Tensor<T>,
-) -> Result<(Tensor<T>, Vec<Tensor<T>>), TensorError> {
+) -> Result<Tensor<T>, TensorError> {
     // a < b <=> b > a
     greater_equal(b, a)
 }
@@ -992,45 +992,6 @@ pub fn mult<T: TensorType + Mul<Output = T> + std::marker::Send + std::marker::S
     Ok(output)
 }
 
-/// Divides multiple tensors.
-/// # Arguments
-/// * `t` - Tensors
-/// # Examples
-/// ```
-/// use ezkl::tensor::Tensor;
-/// use ezkl::tensor::ops::div;
-/// let x = Tensor::<i128>::new(
-///    Some(&[2, 1, 2, 1, 1, 1]),
-/// &[2, 3],
-/// ).unwrap();
-/// let k = Tensor::<i128>::new(
-///   Some(&[2, 3, 2, 1, 1, 1]),
-/// &[2, 3],
-/// ).unwrap();
-/// let result = div(&[x, k]).unwrap();
-/// let expected = Tensor::<i128>::new(Some(&[1, 0, 1, 1, 1, 1]), &[2, 3]).unwrap();
-/// assert_eq!(result, expected);
-/// ```
-pub fn div<
-    T: TensorType
-        + Div<Output = T>
-        + Mul<Output = T>
-        + From<u64>
-        + std::marker::Send
-        + std::marker::Sync,
->(
-    t: &[Tensor<T>],
-) -> Result<Tensor<T>, TensorError> {
-    // calculate value of output
-    let mut output: Tensor<T> = t[0].clone();
-
-    for e in t[1..].iter() {
-        output = (output / e.clone())?;
-    }
-
-    Ok(output)
-}
-
 /// Rescale a tensor with a const integer (similar to const_mult).
 /// # Arguments
 ///
@@ -2339,12 +2300,12 @@ pub fn deconv<
 ///     Some(&[5, 2, 3, 0, 4, -1, 3, 1, 6]),
 ///     &[1, 1, 3, 3],
 /// ).unwrap();
-/// let pooled = sumpool(&x, [(0, 0); 2], (1, 1), (2, 2), false).unwrap().0;
+/// let pooled = sumpool(&x, [(0, 0); 2], (1, 1), (2, 2), false).unwrap();
 /// let expected: Tensor<i128> = Tensor::<i128>::new(Some(&[11, 8, 8, 10]), &[1, 1, 2, 2]).unwrap();
 /// assert_eq!(pooled, expected);
 ///
 /// // This time with normalization
-/// let pooled = sumpool(&x, [(0, 0); 2], (1, 1), (2, 2), true).unwrap().0;
+/// let pooled = sumpool(&x, [(0, 0); 2], (1, 1), (2, 2), true).unwrap();
 /// let expected: Tensor<i128> = Tensor::<i128>::new(Some(&[3, 2, 2, 3]), &[1, 1, 2, 2]).unwrap();
 /// assert_eq!(pooled, expected);
 /// ```
@@ -2354,7 +2315,7 @@ pub fn sumpool(
     stride: (usize, usize),
     kernel_shape: (usize, usize),
     normalize: bool,
-) -> Result<(Tensor<i128>, Vec<Tensor<i128>>), TensorError> {
+) -> Result<Tensor<i128>, TensorError> {
     let image_dims = image.dims();
     let batch_size = image_dims[0];
     let image_channels = image_dims[1];
@@ -2384,15 +2345,12 @@ pub fn sumpool(
     let mut combined = res.combine()?;
     combined.reshape(&[&[batch_size, image_channels], shape].concat())?;
 
-    let mut inter = vec![];
-
     if normalize {
-        inter.push(combined.clone());
         let norm = kernel.len();
         combined = nonlinearities::const_div(&combined, norm as f64);
     }
 
-    Ok((combined, inter))
+    Ok(combined)
 }
 
 /// Applies 2D max pooling over a 4D tensor of shape B x C x H x W.
@@ -3087,11 +3045,7 @@ pub mod nonlinearities {
     }
 
     /// softmax layout
-    pub fn softmax_axes(
-        a: &Tensor<i128>,
-        scale: f64,
-        axes: &[usize],
-    ) -> (Tensor<i128>, Vec<Tensor<i128>>) {
+    pub fn softmax_axes(a: &Tensor<i128>, scale: f64, axes: &[usize]) -> Tensor<i128> {
         // we want this to be as small as possible so we set the output scale to 1
         let dims = a.dims();
 
@@ -3099,8 +3053,6 @@ pub mod nonlinearities {
             return softmax(a, scale);
         }
 
-        let mut intermediate_values = vec![];
-
         let cartesian_coord = dims[..dims.len() - 1]
             .iter()
             .map(|x| 0..*x)
@@ -3123,8 +3075,7 @@ pub mod nonlinearities {
 
             let res = softmax(&softmax_input, scale);
 
-            outputs.push(res.0);
-            intermediate_values.extend(res.1);
+            outputs.push(res);
         }
 
         let mut res = Tensor::new(Some(&outputs), &[outputs.len()])
@@ -3132,7 +3083,7 @@ pub mod nonlinearities {
             .combine()
             .unwrap();
         res.reshape(dims).unwrap();
-        (res, intermediate_values)
+        res
     }
 
     /// Applies softmax
@@ -3149,24 +3100,20 @@ pub mod nonlinearities {
     ///     Some(&[2, 2, 3, 2, 2, 0]),
     ///     &[2, 3],
     /// ).unwrap();
-    /// let result = softmax(&x, 128.0).0;
+    /// let result = softmax(&x, 128.0);
     /// // doubles the scale of the input
     /// let expected = Tensor::<i128>::new(Some(&[2730, 2730, 2751, 2730, 2730, 2688]), &[2, 3]).unwrap();
     /// assert_eq!(result, expected);
     /// ```
-    pub fn softmax(a: &Tensor<i128>, scale: f64) -> (Tensor<i128>, Vec<Tensor<i128>>) {
+    pub fn softmax(a: &Tensor<i128>, scale: f64) -> Tensor<i128> {
         // the more accurate calculation is commented out and we implement as below so it matches the steps in layout
-        let mut intermediate_values = vec![];
-
-        intermediate_values.push(a.clone());
 
         let exp = exp(a, scale);
 
         let sum = sum(&exp).unwrap();
-        intermediate_values.push(sum.clone());
-        let inv_denom = recip(&sum, scale.powf(2.0));
+        let inv_denom = recip(&sum, scale, scale);
 
-        ((exp * inv_denom).unwrap(), intermediate_values)
+        (exp * inv_denom).unwrap()
     }
 
     /// Applies range_check_percent
@@ -3201,7 +3148,7 @@ pub mod nonlinearities {
         // the more accurate calculation is commented out and we implement as below so it matches the steps in layout
         let scale = input_scale * output_scale;
         let diff: Tensor<i128> = sub(t).unwrap();
-        let recip = recip(&t[0], scale as f64);
+        let recip = recip(&t[0], input_scale as f64, output_scale as f64);
         let product = mult(&[diff, recip]).unwrap();
         let _tol = ((tol / 100.0) * scale as f32).round() as f64;
         let upper_bound = greater_than(&product, _tol);
@@ -3812,14 +3759,15 @@ pub mod nonlinearities {
     ///     &[2, 3],
     /// ).unwrap();
     /// let k = 2_f64;
-    /// let result = recip(&x, k);
+    /// let result = recip(&x, 1.0, k);
     /// let expected = Tensor::<i128>::new(Some(&[1, 2, 1, 0, 2, 2]), &[2, 3]).unwrap();
     /// assert_eq!(result, expected);
     /// ```
-    pub fn recip(a: &Tensor<i128>, scale: f64) -> Tensor<i128> {
+    pub fn recip(a: &Tensor<i128>, input_scale: f64, out_scale: f64) -> Tensor<i128> {
         a.par_enum_map(|_, a_i| {
-            let denom = (1_f64) / (a_i as f64 + f64::EPSILON);
-            let d_inv_x = scale * denom;
+            let rescaled = (a_i as f64) / input_scale;
+            let denom = (1_f64) / (rescaled + f64::EPSILON);
+            let d_inv_x = out_scale * denom;
             Ok::<_, TensorError>(d_inv_x.round() as i128)
         })
         .unwrap()

src/tensor/val.rs

@@ -454,12 +454,12 @@ impl<F: PrimeField + TensorType + PartialOrd> ValTensor<F> {
     }
 
     /// Calls `pad_to_zero_rem` on the inner tensor.
-    pub fn pad_to_zero_rem(&mut self, n: usize) -> Result<(), Box<dyn Error>> {
+    pub fn pad_to_zero_rem(&mut self, n: usize, pad: ValType<F>) -> Result<(), Box<dyn Error>> {
         match self {
             ValTensor::Value {
                 inner: v, dims: d, ..
             } => {
-                *v = v.pad_to_zero_rem(n)?;
+                *v = v.pad_to_zero_rem(n, pad)?;
                 *d = v.dims().to_vec();
             }
             ValTensor::Instance { .. } => {
@@ -871,3 +871,30 @@ impl<F: PrimeField + TensorType + PartialOrd> ValTensor<F> {
         }
     }
 }
+
+impl<F: PrimeField + TensorType + PartialOrd> ValTensor<F> {
+    /// inverts the inner values
+    pub fn inverse(&self) -> Result<ValTensor<F>, Box<dyn Error>> {
+        let mut cloned_self = self.clone();
+
+        match &mut cloned_self {
+            ValTensor::Value {
+                inner: v, dims: d, ..
+            } => {
+                *v = v.map(|x| match x {
+                    ValType::AssignedValue(v) => ValType::AssignedValue(v.invert()),
+                    ValType::PrevAssigned(v) | ValType::AssignedConstant(v, ..) => {
+                        ValType::AssignedValue(v.value_field().invert())
+                    }
+                    ValType::Value(v) => ValType::Value(v.map(|x| x.invert().unwrap_or(F::ZERO))),
+                    ValType::Constant(v) => ValType::Constant(v.invert().unwrap_or(F::ZERO)),
+                });
+                *d = v.dims().to_vec();
+            }
+            ValTensor::Instance { .. } => {
+                return Err(Box::new(TensorError::WrongMethod));
+            }
+        };
+        Ok(cloned_self)
+    }
+}

src/wasm.rs

@@ -311,13 +311,19 @@ pub fn verify(
     let vk = VerifyingKey::<G1Affine>::read::<_, GraphCircuit>(
         &mut reader,
         halo2_proofs::SerdeFormat::RawBytes,
-        circuit_settings,
+        circuit_settings.clone(),
     )
     .map_err(|e| JsError::new(&format!("Failed to deserialize vk: {}", e)))?;
 
     let strategy = KZGSingleStrategy::new(params.verifier_params());
 
-    let result = verify_proof_circuit_kzg(params.verifier_params(), snark, &vk, strategy);
+    let result = verify_proof_circuit_kzg(
+        params.verifier_params(),
+        snark,
+        &vk,
+        strategy,
+        1 << circuit_settings.run_args.logrows,
+    );
 
     match result {
         Ok(_) => Ok(true),
@@ -387,15 +393,6 @@ pub fn prove(
         .into_bytes())
 }
 
-/// print hex representation of a proof
-#[wasm_bindgen]
-#[allow(non_snake_case)]
-pub fn printProofHex(proof: wasm_bindgen::Clamped<Vec<u8>>) -> Result<String, JsError> {
-    let proof: crate::pfsys::Snark<Fr, G1Affine> = serde_json::from_slice(&proof[..])
-        .map_err(|e| JsError::new(&format!("Failed to deserialize proof: {}", e)))?;
-    let hex_str = hex::encode(proof.proof);
-    Ok(format!("0x{}", hex_str))
-}
 // VALIDATION FUNCTIONS
 
 /// Witness file validation

tests/integration_tests.rs

@@ -182,95 +182,96 @@ mod native_tests {
         "mnist_gan",
     ];
 
-    const ACCURACY_CAL_TESTS: [&str; 5] = [
+    const ACCURACY_CAL_TESTS: [&str; 6] = [
         "accuracy",
         "1l_mlp",
         "4l_relu_conv_fc",
         "1l_elu",
         "1l_prelu",
+        "1l_tiny_div",
     ];
 
     const TESTS: [&str; 77] = [
-        "1l_mlp",
+        "1l_mlp", //0
         "1l_slice",
         "1l_concat",
         "1l_flatten",
         // "1l_average",
         "1l_div",
-        "1l_pad",
+        "1l_pad", // 5
         "1l_reshape",
         "1l_eltwise_div",
         "1l_sigmoid",
         "1l_sqrt",
-        "1l_softmax",
+        "1l_softmax", //10
         // "1l_instance_norm",
         "1l_batch_norm",
         "1l_prelu",
         "1l_leakyrelu",
         "1l_gelu_noappx",
         // "1l_gelu_tanh_appx",
-        "1l_relu",
+        "1l_relu", //15
         "1l_downsample",
         "1l_tanh",
         "2l_relu_sigmoid_small",
         "2l_relu_fc",
-        "2l_relu_small",
+        "2l_relu_small", //20
         "2l_relu_sigmoid",
         "1l_conv",
         "2l_sigmoid_small",
         "2l_relu_sigmoid_conv",
-        "3l_relu_conv_fc",
+        "3l_relu_conv_fc", //25
         "4l_relu_conv_fc",
         "1l_erf",
         "1l_var",
-        "1l_elu", //30
-        "min",
+        "1l_elu",
+        "min", //30
         "max",
         "1l_max_pool",
         "1l_conv_transpose",
-        "1l_upsample", //35
-        "1l_identity",
+        "1l_upsample",
+        "1l_identity", //35
         "idolmodel",
         "trig",
         "prelu_gmm",
-        "lstm", //40
-        "rnn",
+        "lstm",
+        "rnn", //40
         "quantize_dequantize",
         "1l_where",
         "boolean",
         "boolean_identity",
-        "decision_tree", // "variable_cnn",
+        "decision_tree", // 45
         "random_forest",
         "gradient_boosted_trees",
         "1l_topk",
-        "xgboost", //50
-        "lightgbm",
+        "xgboost",
+        "lightgbm", //50
         "hummingbird_decision_tree",
         "oh_decision_tree",
         "linear_svc",
         "gather_elements",
-        "less",
+        "less", //55
         "xgboost_reg",
         "1l_powf",
         "scatter_elements",
-        "1l_linear", //60
-        "linear_regression",
+        "1l_linear",
+        "linear_regression", //60
         "sklearn_mlp",
         "1l_mean",
         "rounding_ops",
         // "mean_as_constrain",
         "arange",
-        "layernorm",
+        "layernorm", //65
         "bitwise_ops",
         "blackman_window",
-        "softsign", //70
+        "softsign", //68
         "softplus",
-        "selu",
+        "selu", //70
         "hard_sigmoid",
         "log_softmax",
         "eye",
         "ltsf",
-        "remainder",
+        "remainder", //75
         "bitshift",
     ];
 
@@ -489,7 +490,7 @@ mod native_tests {
                 test_dir.close().unwrap();
             }
 
-            seq!(N in 0..=4 {
+            seq!(N in 0..=5 {
             #(#[test_case(ACCURACY_CAL_TESTS[N])])*
             fn mock_accuracy_cal_tests(test: &str) {
                 crate::native_tests::init_binary();
@@ -559,7 +560,7 @@ mod native_tests {
                 crate::native_tests::setup_py_env();
                 let test_dir = TempDir::new(test).unwrap();
                 let path = test_dir.path().to_str().unwrap(); crate::native_tests::mv_test_(path, test);
-                accuracy_measurement(path, test.to_string(), "private", "private", "public", 1, "resources", 18.0, false);
+                accuracy_measurement(path, test.to_string(), "private", "private", "public", 1, "resources", 3.1, false);
                 test_dir.close().unwrap();
             }
 
@@ -1744,6 +1745,30 @@ mod native_tests {
             .status()
             .expect("failed to execute process");
         assert!(status.success());
+
+        // load settings file
+        let settings =
+            std::fs::read_to_string(settings_path.clone()).expect("failed to read settings file");
+
+        let graph_settings = serde_json::from_str::<GraphSettings>(&settings)
+            .expect("failed to parse settings file");
+
+        // get_srs for the graph_settings_num_instances
+        let _ = download_srs(graph_settings.log2_total_instances());
+
+        let status = Command::new(format!("{}/release/ezkl", *CARGO_TARGET_DIR))
+            .args([
+                "verify",
+                format!("--settings-path={}", settings_path).as_str(),
+                "--proof-path",
+                &format!("{}/{}/proof.pf", test_dir, example_name),
+                "--vk-path",
+                &format!("{}/{}/key.vk", test_dir, example_name),
+                "--reduced-srs=true",
+            ])
+            .status()
+            .expect("failed to execute process");
+        assert!(status.success());
     }
 
     // prove-serialize-verify, the usual full path
@@ -2032,7 +2057,7 @@ mod native_tests {
             1,
             "resources",
             // we need the accuracy
-            Some(vec![7, 8]),
+            Some(vec![4]),
             1,
             false,
         );

tests/output_comparison.py

@@ -78,14 +78,20 @@ def compare_outputs(zk_output, onnx_output):
 
     zip_object = zip(np.array(zk_output).flatten(),
                      np.array(onnx_output).flatten())
-    for list1_i, list2_i in zip_object:
+    for (i, (list1_i, list2_i)) in enumerate(zip_object):
         if list1_i == 0.0 and list2_i == 0.0:
             res.append(0)
         else:
             diff = list1_i - list2_i
             res.append(100 * (diff) / (list2_i))
+            # iterate and print the diffs  if they are greater than 0.0
+            if abs(diff) > 0.0:
+                print("------- index: ", i)
+                print("------- diff: ", diff)
+                print("------- zk_output: ", list1_i)
+                print("------- onnx_output: ", list2_i)
+
 
-    print("res: ", res)
 
     return np.mean(np.abs(res))
 

tests/py_integration_tests.rs

@@ -131,7 +131,7 @@ mod py_tests {
         "simple_demo_aggregated_proofs.ipynb",
         "ezkl_demo.ipynb", // 10
         "lstm.ipynb",
-        "set_membership.ipynb",
+        "set_membership.ipynb", // 12
         "decision_tree.ipynb",
         "random_forest.ipynb",
         "gradient_boosted_trees.ipynb", // 15

tests/python/binding_tests.py

@@ -113,13 +113,13 @@ def test_calibrate_over_user_range():
     data_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'input.json'
     )
     model_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'network.onnx'
     )
     output_path = os.path.join(
@@ -147,13 +147,13 @@ def test_calibrate():
     data_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'input.json'
     )
     model_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'network.onnx'
     )
     output_path = os.path.join(
@@ -183,7 +183,7 @@ def test_model_compile():
     model_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'network.onnx'
     )
     compiled_model_path = os.path.join(
@@ -205,7 +205,7 @@ def test_forward():
     data_path = os.path.join(
         examples_path,
         'onnx',
-        '1l_average',
+        '1l_relu',
         'input.json'
     )
     model_path = os.path.join(
@@ -392,9 +392,7 @@ def test_prove_evm():
     assert res['transcript_type'] == 'EVM'
     assert os.path.isfile(proof_path)
 
-    res = ezkl.print_proof_hex(proof_path)
-    # to figure out a better way of testing print_proof_hex
-    assert type(res) == str
+
 
 
 def test_create_evm_verifier():

tests/wasm.rs

@@ -9,8 +9,8 @@ mod wasm32 {
     use ezkl::pfsys;
     use ezkl::wasm::{
         bufferToVecOfstring, compiledCircuitValidation, encodeVerifierCalldata, genPk, genVk,
-        genWitness, inputValidation, pkValidation, poseidonHash, printProofHex, proofValidation,
-        prove, settingsValidation, srsValidation, stringToFelt, stringToFloat, stringToInt,
+        genWitness, inputValidation, pkValidation, poseidonHash, proofValidation, prove,
+        settingsValidation, srsValidation, stringToFelt, stringToFloat, stringToInt,
         u8_array_to_u128_le, verify, vkValidation, witnessValidation,
     };
     use halo2_solidity_verifier::encode_calldata;
@@ -258,15 +258,6 @@ mod wasm32 {
         assert!(value);
     }
 
-    #[wasm_bindgen_test]
-    async fn print_proof_hex_test() {
-        let proof = printProofHex(wasm_bindgen::Clamped(PROOF.to_vec()))
-            .map_err(|_| "failed")
-            .unwrap();
-
-        assert!(proof.len() > 0);
-    }
-
     #[wasm_bindgen_test]
     async fn verify_validations() {
         // Run witness validation on network (should fail)

tests/wasm/settings.json

@@ -23,6 +23,7 @@
         "output_visibility": "Public",
         "param_visibility": "Private",
         "div_rebasing": false,
+        "rebase_frac_zero_constants": false,
         "check_mode": "UNSAFE"
     },
     "num_rows": 16,