Gkr2

.github/workflows/build.yml

@@ -34,7 +34,7 @@ jobs:
 
       - uses: Swatinem/rust-cache@v2
         with:
-          prefix-key: "mpi-v5.0.6"
+          prefix-key: "mpi-v5.0.7"
 
       - name: Install MPI for MacOS workflow
         if: matrix.os == 'macos-latest'

.github/workflows/ci.yml

@@ -55,7 +55,7 @@ jobs:
       - name: Cache Rust dependencies
         uses: Swatinem/rust-cache@v2
         with:
-          prefix-key: "mpi-v5.0.6"
+          prefix-key: "mpi-v5.0.7"
 
       - uses: actions/setup-go@v5
         if: matrix.os != '7950x3d'

.github/workflows/nightly_e2e._yml

@@ -34,7 +34,7 @@ jobs:
     - uses: Swatinem/rust-cache@v2
       with:
         # The prefix cache key, this can be changed to start a new cache manually.
-        prefix-key: "mpi-v5.0.6" # update me if brew formula changes to a new version
+        prefix-key: "mpi-v5.0.7" # update me if brew formula changes to a new version
 
     - name: Run tests
       run: |

arith/polynomials/Cargo.toml

@@ -5,6 +5,8 @@ edition = "2021"
 
 [dependencies]
 arith = { path = "../" }
+gkr_field_config = { path = "../../config/gkr_field_config" }
+mpi_config = { path = "../../config/mpi_config" }
 
 ark-std.workspace = true
 criterion.workspace = true

arith/polynomials/src/mle.rs

@@ -1,7 +1,13 @@
-use std::ops::{Index, IndexMut, Mul};
+use std::{
+    cmp,
+    marker::PhantomData,
+    ops::{Index, IndexMut, Mul},
+};
 
-use arith::Field;
+use arith::{Field, SimdField};
 use ark_std::{log2, rand::RngCore};
+use gkr_field_config::GKRFieldConfig;
+use mpi_config::MPIConfig;
 
 use crate::{EqPolynomial, MultilinearExtension, MutableMultilinearExtension};
 
@@ -205,3 +211,125 @@ impl<F: Field> MutableMultilinearExtension<F> for MultiLinearPoly<F> {
         }
     }
 }
+
+#[derive(Debug, Clone, Default)]
+pub struct MultiLinearPolyExpander<C: GKRFieldConfig> {
+    _config: PhantomData<C>,
+}
+
+/// Some dedicated mle implementations for GKRFieldConfig
+/// Take into consideration the simd challenge and the mpi challenge
+///
+/// This is more efficient than the generic implementation by avoiding
+/// unnecessary conversions between field types
+impl<C: GKRFieldConfig> MultiLinearPolyExpander<C> {
+    pub fn new() -> Self {
+        Self {
+            _config: PhantomData,
+        }
+    }
+
+    #[inline]
+    pub fn eval_circuit_vals_at_challenge(
+        evals: &[C::SimdCircuitField],
+        x: &[C::ChallengeField],
+        scratch: &mut [C::Field],
+    ) -> C::Field {
+        assert_eq!(1 << x.len(), evals.len());
+        assert!(scratch.len() >= evals.len());
+
+        if x.is_empty() {
+            C::simd_circuit_field_into_field(&evals[0])
+        } else {
+            for i in 0..(evals.len() >> 1) {
+                scratch[i] = C::field_add_simd_circuit_field(
+                    &C::simd_circuit_field_mul_challenge_field(
+                        &(evals[i * 2 + 1] - evals[i * 2]),
+                        &x[0],
+                    ),
+                    &evals[i * 2],
+                );
+            }
+
+            let mut cur_eval_size = evals.len() >> 2;
+            for r in x.iter().skip(1) {
+                for i in 0..cur_eval_size {
+                    scratch[i] = scratch[i * 2] + (scratch[i * 2 + 1] - scratch[i * 2]).scale(r);
+                }
+                cur_eval_size >>= 1;
+            }
+            scratch[0]
+        }
+    }
+
+    /// This assumes each mpi core hold their own evals, and collectively
+    /// compute the global evaluation.
+    /// Mostly used by the prover run with `mpiexec`
+    #[inline]
+    pub fn collectively_eval_circuit_vals_at_expander_challenge(
+        local_evals: &[C::SimdCircuitField],
+        x: &[C::ChallengeField],
+        x_simd: &[C::ChallengeField],
+        x_mpi: &[C::ChallengeField],
+        scratch_field: &mut [C::Field],
+        scratch_challenge_field: &mut [C::ChallengeField],
+        mpi_config: &MPIConfig,
+    ) -> C::ChallengeField {
+        assert!(scratch_challenge_field.len() >= 1 << cmp::max(x_simd.len(), x_mpi.len()));
+
+        let local_simd = Self::eval_circuit_vals_at_challenge(local_evals, x, scratch_field);
+        let local_simd_unpacked = local_simd.unpack();
+        let local_v = MultiLinearPoly::evaluate_with_buffer(
+            &local_simd_unpacked,
+            x_simd,
+            scratch_challenge_field,
+        );
+
+        if mpi_config.is_root() {
+            let mut claimed_v_gathering_buffer =
+                vec![C::ChallengeField::zero(); mpi_config.world_size()];
+            mpi_config.gather_vec(&vec![local_v], &mut claimed_v_gathering_buffer);
+            MultiLinearPoly::evaluate_with_buffer(
+                &claimed_v_gathering_buffer,
+                x_mpi,
+                scratch_challenge_field,
+            )
+        } else {
+            mpi_config.gather_vec(&vec![local_v], &mut vec![]);
+            C::ChallengeField::zero()
+        }
+    }
+
+    /// This assumes only a single core holds all the evals, and evaluate it locally
+    /// mostly used by the verifier
+    #[inline]
+    pub fn single_core_eval_circuit_vals_at_expander_challenge(
+        global_vals: &[C::SimdCircuitField],
+        x: &[C::ChallengeField],
+        x_simd: &[C::ChallengeField],
+        x_mpi: &[C::ChallengeField],
+    ) -> C::ChallengeField {
+        let local_poly_size = global_vals.len() >> x_mpi.len();
+        assert_eq!(local_poly_size, 1 << x.len());
+
+        let mut scratch_field = vec![C::Field::default(); local_poly_size];
+        let mut scratch_challenge_field =
+            vec![C::ChallengeField::default(); 1 << cmp::max(x_simd.len(), x_mpi.len())];
+        let local_evals = global_vals
+            .chunks(local_poly_size)
+            .map(|local_vals| {
+                let local_simd =
+                    Self::eval_circuit_vals_at_challenge(local_vals, x, &mut scratch_field);
+                let local_simd_unpacked = local_simd.unpack();
+                MultiLinearPoly::evaluate_with_buffer(
+                    &local_simd_unpacked,
+                    x_simd,
+                    &mut scratch_challenge_field,
+                )
+            })
+            .collect::<Vec<C::ChallengeField>>();
+
+        let mut scratch = vec![C::ChallengeField::default(); local_evals.len()];
+        MultiLinearPoly::evaluate_with_buffer(&local_evals, x_mpi, &mut scratch)
+    }
+}

config/gkr_field_config/src/traits.rs

@@ -73,37 +73,4 @@ pub trait GKRFieldConfig: Default + Debug + Clone + Send + Sync + 'static {
     fn get_field_pack_size() -> usize {
         Self::SimdCircuitField::PACK_SIZE
     }
-
-    /// Evaluate the circuit values at the challenge
-    #[inline]
-    fn eval_circuit_vals_at_challenge(
-        evals: &[Self::SimdCircuitField],
-        x: &[Self::ChallengeField],
-        scratch: &mut [Self::Field],
-    ) -> Self::Field {
-        assert_eq!(1 << x.len(), evals.len());
-
-        if x.is_empty() {
-            Self::simd_circuit_field_into_field(&evals[0])
-        } else {
-            for i in 0..(evals.len() >> 1) {
-                scratch[i] = Self::field_add_simd_circuit_field(
-                    &Self::simd_circuit_field_mul_challenge_field(
-                        &(evals[i * 2 + 1] - evals[i * 2]),
-                        &x[0],
-                    ),
-                    &evals[i * 2],
-                );
-            }
-
-            let mut cur_eval_size = evals.len() >> 2;
-            for r in x.iter().skip(1) {
-                for i in 0..cur_eval_size {
-                    scratch[i] = scratch[i * 2] + (scratch[i * 2 + 1] - scratch[i * 2]).scale(r);
-                }
-                cur_eval_size >>= 1;
-            }
-            scratch[0]
-        }
-    }
 }

config/mpi_config/src/lib.rs

@@ -258,7 +258,7 @@ impl MPIConfig {
             ret
         } else {
             self.gather_vec(local_vec, &mut vec![]);
-            vec![]
+            vec![F::ZERO; local_vec.len()]
         }
     }
 

crosslayer_prototype/src/gkr.rs

@@ -1,6 +1,6 @@
 use arith::{Field, SimdField};
 use gkr_field_config::GKRFieldConfig;
-use polynomials::MultiLinearPoly;
+use polynomials::{MultiLinearPoly, MultiLinearPolyExpander};
 use transcript::Transcript;
 
 use crate::sumcheck::{sumcheck_prove_gather_layer, sumcheck_prove_scatter_layer};
@@ -18,7 +18,11 @@ pub fn prove_gkr<C: GKRFieldConfig, T: Transcript<C::ChallengeField>>(
         .generate_challenge_field_elements(final_layer_vals.len().trailing_zeros() as usize);
     let r_simd = transcript
         .generate_challenge_field_elements(C::get_field_pack_size().trailing_zeros() as usize);
-    let output_claim = C::eval_circuit_vals_at_challenge(final_layer_vals, &rz0, &mut sp.v_evals);
+    let output_claim = MultiLinearPolyExpander::<C>::eval_circuit_vals_at_challenge(
+        final_layer_vals,
+        &rz0,
+        &mut sp.v_evals,
+    );
     let output_claim = MultiLinearPoly::<C::ChallengeField>::evaluate_with_buffer(
         &output_claim.unpack(),
         &r_simd,

gkr/src/prover/gkr.rs

@@ -1,10 +1,9 @@
 //! This module implements the core GKR IOP.
 
-use arith::{Field, SimdField};
 use circuit::Circuit;
 use gkr_field_config::GKRFieldConfig;
 use mpi_config::MPIConfig;
-use polynomials::MultiLinearPoly;
+use polynomials::MultiLinearPolyExpander;
 use sumcheck::{sumcheck_prove_gkr_layer, ProverScratchPad};
 use transcript::Transcript;
 use utils::timer::Timer;
@@ -44,27 +43,16 @@ pub fn gkr_prove<C: GKRFieldConfig, T: Transcript<C::ChallengeField>>(
     let mut alpha = None;
 
     let output_vals = &circuit.layers.last().unwrap().output_vals;
-
-    let claimed_v_simd = C::eval_circuit_vals_at_challenge(output_vals, &rz0, &mut sp.hg_evals);
-    let claimed_v_local = MultiLinearPoly::<C::ChallengeField>::evaluate_with_buffer(
-        &claimed_v_simd.unpack(),
-        &r_simd,
-        &mut sp.eq_evals_at_r_simd0,
-    );
-
-    let claimed_v = if mpi_config.is_root() {
-        let mut claimed_v_gathering_buffer =
-            vec![C::ChallengeField::zero(); mpi_config.world_size()];
-        mpi_config.gather_vec(&vec![claimed_v_local], &mut claimed_v_gathering_buffer);
-        MultiLinearPoly::evaluate_with_buffer(
-            &claimed_v_gathering_buffer,
+    let claimed_v =
+        MultiLinearPolyExpander::<C>::collectively_eval_circuit_vals_at_expander_challenge(
+            output_vals,
+            &rz0,
+            &r_simd,
             &r_mpi,
-            &mut sp.eq_evals_at_r_mpi0,
-        )
-    } else {
-        mpi_config.gather_vec(&vec![claimed_v_local], &mut vec![]);
-        C::ChallengeField::zero()
-    };
+            &mut sp.hg_evals,
+            &mut sp.eq_evals_first_half, // confusing name here..
+            mpi_config,
+        );
 
     for i in (0..layer_num).rev() {
         let timer = Timer::new(

gkr/src/prover/gkr_square.rs

@@ -3,15 +3,29 @@
 
 use ark_std::{end_timer, start_timer};
 use circuit::Circuit;
-use gkr_field_config::GKRFieldConfig;
+use gkr_field_config::{FieldType, GKRFieldConfig};
+use mpi_config::MPIConfig;
+use polynomials::MultiLinearPolyExpander;
 use sumcheck::{sumcheck_prove_gkr_square_layer, ProverScratchPad};
 use transcript::Transcript;
 
+#[allow(clippy::type_complexity)]
 pub fn gkr_square_prove<C: GKRFieldConfig, T: Transcript<C::ChallengeField>>(
     circuit: &Circuit<C>,
     sp: &mut ProverScratchPad<C>,
     transcript: &mut T,
-) -> (C::Field, Vec<C::ChallengeField>) {
+    mpi_config: &MPIConfig,
+) -> (
+    C::ChallengeField,
+    Vec<C::ChallengeField>,
+    Vec<C::ChallengeField>,
+    Vec<C::ChallengeField>,
+) {
+    assert_ne!(
+        C::FIELD_TYPE,
+        FieldType::GF2,
+        "GF2 is not supported in GKR^2"
+    );
     let timer = start_timer!(|| "gkr^2 prove");
     let layer_num = circuit.layers.len();
 
@@ -20,11 +34,41 @@ pub fn gkr_square_prove<C: GKRFieldConfig, T: Transcript<C::ChallengeField>>(
         rz0.push(transcript.generate_challenge_field_element());
     }
 
-    let circuit_output = &circuit.layers.last().unwrap().output_vals;
-    let claimed_v = C::eval_circuit_vals_at_challenge(circuit_output, &rz0, &mut sp.hg_evals);
+    let mut r_simd = vec![];
+    for _i in 0..C::get_field_pack_size().trailing_zeros() {
+        r_simd.push(transcript.generate_challenge_field_element());
+    }
+    log::trace!("Initial r_simd: {:?}", r_simd);
+
+    let mut r_mpi = vec![];
+    for _ in 0..mpi_config.world_size().trailing_zeros() {
+        r_mpi.push(transcript.generate_challenge_field_element());
+    }
+
+    let output_vals = &circuit.layers.last().unwrap().output_vals;
+    let claimed_v =
+        MultiLinearPolyExpander::<C>::collectively_eval_circuit_vals_at_expander_challenge(
+            output_vals,
+            &rz0,
+            &r_simd,
+            &r_mpi,
+            &mut sp.hg_evals,
+            &mut sp.eq_evals_first_half, // confusing name here..
+            mpi_config,
+        );
+
+    log::trace!("Claimed v: {:?}", claimed_v);
 
     for i in (0..layer_num).rev() {
-        rz0 = sumcheck_prove_gkr_square_layer(&circuit.layers[i], &rz0, transcript, sp);
+        (rz0, r_simd, r_mpi) = sumcheck_prove_gkr_square_layer(
+            &circuit.layers[i],
+            &rz0,
+            &r_simd,
+            &r_mpi,
+            transcript,
+            sp,
+            mpi_config,
+        );
 
         log::trace!("Layer {} proved", i);
         log::trace!("rz0.0: {:?}", rz0[0]);
@@ -33,5 +77,5 @@ pub fn gkr_square_prove<C: GKRFieldConfig, T: Transcript<C::ChallengeField>>(
     }
 
     end_timer!(timer);
-    (claimed_v, rz0)
+    (claimed_v, rz0, r_simd, r_mpi)
 }