Update cryptography_cuda reference, fix AVX2 issues

CHANGELOG.md

@@ -5,8 +5,17 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## Unreleased
+## [Unreleased]
 
+## [0.2.3] - 2024-04-16
+
+- Code refactoring ([#1558](https://github.com/0xPolygonZero/plonky2/pull/1558))
+- Simplify types: remove option from CTL filters ([#1567](https://github.com/0xPolygonZero/plonky2/pull/1567))
+- Add stdarch_x86_avx512 feature ([#1566](https://github.com/0xPolygonZero/plonky2/pull/1566))
+
+## [0.2.2] - 2024-03-21
+
+### Changed
 - Fix CTLs with exactly two looking tables ([#1555](https://github.com/0xPolygonZero/plonky2/pull/1555))
 - Make Starks without constraints provable ([#1552](https://github.com/0xPolygonZero/plonky2/pull/1552))
 

Cargo.toml

@@ -3,7 +3,7 @@ members = ["field", "maybe_rayon", "plonky2", "starky", "util", "gen", "u32", "e
 resolver = "2"
 
 [workspace.dependencies]
-cryptography_cuda = { git = "ssh://[email protected]/okx/cryptography_cuda.git", rev = "173510160183f3299f4765b30bd4f2c1685353f9" }
+cryptography_cuda = { git = "ssh://[email protected]/okx/cryptography_cuda.git", rev = "2a7c42d29ee72d7c2c2da9378ae816384c43cdec" }
 ahash = { version = "0.8.7", default-features = false, features = [
     "compile-time-rng",
 ] } # NOTE: Be sure to keep this version the same as the dependency in `hashbrown`.

field/Cargo.toml

@@ -1,13 +1,8 @@
 [package]
 name = "plonky2_field"
 description = "Finite field arithmetic"
-version = "0.2.0"
-authors = [
-    "Daniel Lubarov <[email protected]>",
-    "William Borgeaud <[email protected]>",
-    "Jacqueline Nabaglo <[email protected]>",
-    "Hamish Ivey-Law <[email protected]>",
-]
+version = "0.2.2"
+authors = ["Daniel Lubarov <[email protected]>", "William Borgeaud <[email protected]>", "Jacqueline Nabaglo <[email protected]>", "Hamish Ivey-Law <[email protected]>"]
 edition.workspace = true
 license.workspace = true
 homepage.workspace = true

field/src/fft.rs

@@ -53,6 +53,7 @@ fn fft_dispatch_cpu<F: Field>(
     zero_factor: Option<usize>,
     root_table: Option<&FftRootTable<F>>,
 ) {
+    /*
     if root_table.is_some() {
         return fft_classic(input, zero_factor.unwrap_or(0), root_table.unwrap());
     } else {
@@ -68,6 +69,11 @@ fn fft_dispatch_cpu<F: Field>(
 
         return fft_classic(input, zero_factor.unwrap_or(0), computed.as_ref());
     };
+    */
+    let computed_root_table = root_table.is_none().then(|| fft_root_table(input.len()));
+    let used_root_table = root_table.or(computed_root_table.as_ref()).unwrap();
+
+    fft_classic(input, zero_factor.unwrap_or(0), used_root_table);
 }
 
 #[inline]

field/src/lib.rs

@@ -4,6 +4,7 @@
 #![deny(rustdoc::broken_intra_doc_links)]
 #![deny(missing_debug_implementations)]
 #![feature(specialization)]
+#![cfg_attr(target_arch = "x86_64", feature(stdarch_x86_avx512))]
 #![cfg_attr(not(test), no_std)]
 #![cfg(not(test))]
 extern crate alloc;

maybe_rayon/src/lib.rs

@@ -1,5 +1,3 @@
-#![cfg_attr(not(std), no_std)]
-
 #[cfg(not(feature = "parallel"))]
 extern crate alloc;
 

plonky2/Cargo.toml

@@ -1,12 +1,8 @@
 [package]
 name = "plonky2"
 description = "Recursive SNARKs based on PLONK and FRI"
-version = "0.2.0"
-authors = [
-    "Daniel Lubarov <[email protected]>",
-    "William Borgeaud <[email protected]>",
-    "Nicholas Ward <[email protected]>",
-]
+version = "0.2.2"
+authors = ["Daniel Lubarov <[email protected]>", "William Borgeaud <[email protected]>", "Nicholas Ward <[email protected]>"]
 readme = "README.md"
 edition.workspace = true
 license.workspace = true
@@ -45,7 +41,7 @@ once_cell = { version = "1.18.0" }
 papi-bindings = { version = "0.5.2" }
 
 # Local dependencies
-plonky2_field = { version = "0.2.0", path = "../field", default-features = false }
+plonky2_field = { version = "0.2.2", path = "../field", default-features = false }
 plonky2_maybe_rayon = { version = "0.2.0", path = "../maybe_rayon", default-features = false }
 plonky2_util = { version = "0.2.0", path = "../util", default-features = false }
 cryptography_cuda = { workspace = true, optional = true }

plonky2/src/gates/lookup.rs

@@ -5,7 +5,6 @@ use alloc::{
     vec,
     vec::Vec,
 };
-use core::usize;
 
 use itertools::Itertools;
 use keccak_hash::keccak;

plonky2/src/gates/lookup_table.rs

@@ -6,7 +6,6 @@ use alloc::{
     vec,
     vec::Vec,
 };
-use core::usize;
 #[cfg(feature = "std")]
 use std::sync::Arc;
 

plonky2/src/hash/arch/x86_64/goldilocks_avx2.rs

@@ -81,6 +81,7 @@ pub fn reduce_avx_128_64(c_h: &__m256i, c_l: &__m256i) -> __m256i {
 }
 
 // Here we suppose c_h < 2^32
+#[allow(dead_code)]
 #[inline(always)]
 pub fn reduce_avx_96_64(c_h: &__m256i, c_l: &__m256i) -> __m256i {
     unsafe {
@@ -162,6 +163,7 @@ pub fn mult_avx(a: &__m256i, b: &__m256i) -> __m256i {
 }
 
 // Multiply two 64bit numbers with the assumption that the product does not averflow.
+#[allow(dead_code)]
 #[inline]
 pub unsafe fn mul64_no_overflow(a: &__m256i, b: &__m256i) -> __m256i {
     let r = _mm256_mul_epu32(*a, *b);
@@ -282,6 +284,7 @@ where
     }
 }
 
+#[allow(dead_code)]
 #[inline(always)]
 pub fn sbox_avx_m256i(s0: &__m256i, s1: &__m256i, s2: &__m256i) -> (__m256i, __m256i, __m256i) {
     // x^2