chore: avx512 acceleration

algebraic/Cargo.toml

@@ -48,3 +48,4 @@ wasm-bindgen-test = "0.3"
 [features]
 default = ["franklin-crypto/multicore", "wasmer/default"]
 wasm = ["wasmer/js-default"]
+avx512 = []

algebraic/src/arch/x86_64/avx512_field_gl.rs

@@ -2,7 +2,7 @@
 //! https://github.com/0xPolygonZero/plonky2/blob/main/field/src/arch/x86_64/avx512_goldilocks_field.rs
 //!
 //! How to build/run/test:
-//! RUSTFLAGS='-C target-feature=+avx512f,+avx512bw,+avx512cd,+avx512dq,+avx512vl' cargo build --release
+//! RUSTFLAGS='-C target-feature=+avx512f,+avx512bw,+avx512cd,+avx512dq,+avx512vl' cargo build --features "avx512" --release
 use crate::ff::*;
 use crate::field_gl::{Fr, FrRepr as GoldilocksField};
 use crate::packed::PackedField;
@@ -27,14 +27,22 @@ const WIDTH: usize = 8;
 
 impl Avx512GoldilocksField {
     #[inline]
-    fn new(x: __m512i) -> Self {
+    pub fn new(x: __m512i) -> Self {
         unsafe { transmute(x) }
     }
     #[inline]
-    fn get(&self) -> __m512i {
+    pub fn get(&self) -> __m512i {
         unsafe { transmute(*self) }
     }
     #[inline]
+    pub fn interleave2(x: __m512i, y: __m512i) -> (__m512i, __m512i) {
+        unsafe { interleave2(x, y) }
+    }
+    #[inline]
+    pub fn reduce(x: __m512i, y: __m512i) -> Avx512GoldilocksField {
+        Self::new(unsafe { reduce128((x, y)) })
+    }
+    #[inline]
     pub fn square(&self) -> Avx512GoldilocksField {
         Self::new(unsafe { square(self.get()) })
     }
@@ -271,7 +279,8 @@ unsafe fn sub_no_double_overflow_64_64(x: __m512i, y: __m512i) -> __m512i {
 
 #[inline]
 unsafe fn add(x: __m512i, y: __m512i) -> __m512i {
-    add_no_double_overflow_64_64(x, canonicalize(y))
+    let res_s = add_no_double_overflow_64_64(x, canonicalize(y));
+    canonicalize(res_s)
 }
 
 #[inline]
@@ -356,7 +365,8 @@ unsafe fn reduce128(x: (__m512i, __m512i)) -> __m512i {
     let hi_hi0 = _mm512_srli_epi64::<32>(hi0);
     let lo1 = sub_no_double_overflow_64_64(lo0, hi_hi0);
     let t1 = _mm512_mul_epu32(hi0, EPSILON);
-    let lo2 = add_no_double_overflow_64_64(lo1, t1);
+    let _lo2 = add_no_double_overflow_64_64(lo1, t1);
+    let lo2 = canonicalize(_lo2);
     lo2
 }
 
@@ -412,7 +422,7 @@ mod tests {
 
     fn test_vals_a() -> [GoldilocksField; 8] {
         [
-            GoldilocksField([14479013849828404771u64]),
+            GoldilocksField([18446744069414584320u64]),
             GoldilocksField([9087029921428221768u64]),
             GoldilocksField([2441288194761790662u64]),
             GoldilocksField([5646033492608483824u64]),
@@ -424,7 +434,7 @@ mod tests {
     }
     fn test_vals_b() -> [GoldilocksField; 8] {
         [
-            GoldilocksField([17891926589593242302u64]),
+            GoldilocksField([18446744069414584320u64]),
             GoldilocksField([11009798273260028228u64]),
             GoldilocksField([2028722748960791447u64]),
             GoldilocksField([7929433601095175579u64]),

algebraic/src/lib.rs

@@ -1,4 +1,5 @@
 #![allow(clippy::unit_arg)]
+#![cfg_attr(feature = "avx512", feature(stdsimd))]
 
 #[macro_use]
 extern crate serde;

starky/Cargo.toml

@@ -60,3 +60,7 @@ harness = false
 [[bench]]
 name = "poseidon"
 harness = false
+
+[features]
+default = []
+avx512 = ["algebraic/avx512"]

starky/README.md

@@ -102,8 +102,15 @@ Because the nExtBits reaches up to 24 so this step would be very slow. Consider
 
 ```
 cargo bench --bench merklehash -- --profile-time=5
+
+# or with AVX enabled like this 
+
+RUSTFLAGS='-C target-feature=+avx512f,+avx512bw,+avx512cd,+avx512dq,+avx512vl' cargo bench --features avx512 -- merklehash
+RUSTFLAGS="-C target-feature=+avx2" cargo bench -- merklehash
+
 ```
 
+
 * https://www.jibbow.com/posts/criterion-flamegraphs/
 
 

starky/benches/merklehash.rs

@@ -2,7 +2,7 @@
 use criterion::*;
 use plonky::field_gl::Fr as FGL;
 use rayon::prelude::*;
-use starky::merklehash_bn128::MerkleTreeBN128;
+use starky::merklehash::MerkleTreeGL;
 use starky::traits::MerkleTree;
 mod perf;
 
@@ -12,7 +12,7 @@ fn run_merklehash(pols: Vec<FGL>) {
     let n_pols = 10;
 
     let now = std::time::Instant::now();
-    let mut tree: MerkleTreeBN128 = MerkleTree::new();
+    let mut tree: MerkleTreeGL = MerkleTree::new();
     tree.merkelize(pols, n_pols, n).unwrap();
     log::trace!("time cost: {}", now.elapsed().as_secs());
     let (group_elements, mp) = tree.get_group_proof(idx).unwrap();

starky/src/arch/x86_64/avx2_poseidon_gl.rs

@@ -1,6 +1,6 @@
 #![allow(non_snake_case)]
 use crate::constant::POSEIDON_CONSTANTS_OPT_AVX2;
-use crate::poseidon_constants_avx as constants;
+use crate::poseidon_constants_avx2 as constants;
 use algebraic::arch::x86_64::avx2_field_gl::Avx2GoldilocksField;
 use algebraic::packed::PackedField;
 use core::arch::x86_64::*;