Merge pull request #37 from okx/dev-dumi

add LDE bench and salt support

Cargo.toml

@@ -3,7 +3,7 @@ members = ["field", "maybe_rayon", "plonky2", "starky", "util", "gen", "u32", "e
 resolver = "2"
 
 [workspace.dependencies]
-cryptography_cuda = { git = "ssh://git@github.com/okx/cryptography_cuda.git", rev = "2a7c42d29ee72d7c2c2da9378ae816384c43cdec" }
+cryptography_cuda = { git = "ssh://git@github.com/okx/cryptography_cuda.git", rev = "547192b2ef42dc7519435059c86f88431b8de999" }
 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/src/fft.rs

@@ -1,8 +1,6 @@
 use alloc::vec::Vec;
 use core::cmp::{max, min};
 
-#[cfg(feature = "cuda")]
-use cryptography_cuda::{ntt, types::NTTInputOutputOrder};
 use plonky2_util::{log2_strict, reverse_index_bits_in_place};
 use unroll::unroll_for_loops;
 
@@ -34,20 +32,6 @@ pub fn fft_root_table<F: Field>(n: usize) -> FftRootTable<F> {
     root_table
 }
 
-#[allow(dead_code)]
-#[cfg(feature = "cuda")]
-fn fft_dispatch_gpu<F: Field>(
-    input: &mut [F],
-    zero_factor: Option<usize>,
-    root_table: Option<&FftRootTable<F>>,
-) {
-    if F::CUDA_SUPPORT {
-        return ntt(0, input, NTTInputOutputOrder::NN);
-    } else {
-        return fft_dispatch_cpu(input, zero_factor, root_table);
-    }
-}
-
 fn fft_dispatch_cpu<F: Field>(
     input: &mut [F],
     zero_factor: Option<usize>,

plonky2/Cargo.toml

@@ -37,7 +37,7 @@ serde = { workspace = true, features = ["rc"] }
 static_assertions = { workspace = true }
 unroll = { workspace = true }
 web-time = { version = "1.0.0", optional = true }
-once_cell = { version = "1.18.0" }
+once_cell = { version = "1.20.2" }
 papi-bindings = { version = "0.5.2" }
 
 # Local dependencies
@@ -80,6 +80,10 @@ harness = false
 name = "ffts"
 harness = false
 
+[[bench]]
+name = "lde"
+harness = false
+
 [[bench]]
 name = "hashing"
 harness = false

plonky2/benches/lde.rs

@@ -0,0 +1,59 @@
+mod allocator;
+
+use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion};
+#[cfg(feature = "cuda")]
+use cryptography_cuda::init_cuda_degree_rs;
+use plonky2::field::extension::Extendable;
+use plonky2::field::goldilocks_field::GoldilocksField;
+use plonky2::field::polynomial::PolynomialCoeffs;
+use plonky2::fri::oracle::PolynomialBatch;
+use plonky2::hash::hash_types::RichField;
+use plonky2::plonk::config::{GenericConfig, PoseidonGoldilocksConfig};
+use plonky2::util::timing::TimingTree;
+use tynm::type_name;
+
+pub(crate) fn bench_batch_lde<
+    F: RichField + Extendable<D>,
+    C: GenericConfig<D, F = F>,
+    const D: usize,
+>(
+    c: &mut Criterion,
+) {
+    const RATE_BITS: usize = 3;
+
+    let mut group = c.benchmark_group(&format!("lde<{}>", type_name::<F>()));
+
+    #[cfg(feature = "cuda")]
+    init_cuda_degree_rs(16);
+
+    for size_log in [13, 14, 15] {
+        let orig_size = 1 << (size_log - RATE_BITS);
+        let lde_size = 1 << size_log;
+        let batch_size = 1 << 4;
+
+        group.bench_with_input(BenchmarkId::from_parameter(lde_size), &lde_size, |b, _| {
+            let polynomials: Vec<PolynomialCoeffs<F>> = (0..batch_size)
+                .into_iter()
+                .map(|_i| PolynomialCoeffs::new(F::rand_vec(orig_size)))
+                .collect();
+            let mut timing = TimingTree::new("lde", log::Level::Error);
+            b.iter(|| {
+                PolynomialBatch::<F, C, D>::from_coeffs(
+                    polynomials.clone(),
+                    RATE_BITS,
+                    false,
+                    1,
+                    &mut timing,
+                    None,
+                )
+            });
+        });
+    }
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    bench_batch_lde::<GoldilocksField, PoseidonGoldilocksConfig, 2>(c);
+}
+
+criterion_group!(benches, criterion_benchmark);
+criterion_main!(benches);

plonky2/src/fri/oracle.rs

@@ -27,6 +27,22 @@ use crate::util::reducing::ReducingFactor;
 use crate::util::timing::TimingTree;
 use crate::util::{log2_strict, reverse_bits, reverse_index_bits_in_place, transpose};
 
+#[cfg(all(feature = "cuda", any(test, doctest)))]
+pub static GPU_INIT: once_cell::sync::Lazy<std::sync::Arc<std::sync::Mutex<u64>>> =
+    once_cell::sync::Lazy::new(|| std::sync::Arc::new(std::sync::Mutex::new(0)));
+
+#[cfg(all(feature = "cuda", any(test, doctest)))]
+fn init_gpu() {
+    use cryptography_cuda::init_cuda_rs;
+
+    let mut init = GPU_INIT.lock().unwrap();
+    if *init == 0 {
+        println!("Init GPU!");
+        init_cuda_rs();
+        *init = 1;
+    }
+}
+
 /// Four (~64 bit) field elements gives ~128 bit security.
 pub const SALT_SIZE: usize = 4;
 
@@ -192,10 +208,17 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
         timing: &mut TimingTree,
         fft_root_table: Option<&FftRootTable<F>>,
     ) -> Self {
+        let pols = polynomials.len();
         let degree = polynomials[0].len();
         let log_n = log2_strict(degree);
 
-        if log_n + rate_bits > 1 && polynomials.len() > 0 {
+        #[cfg(any(test, doctest))]
+        init_gpu();
+
+        if log_n + rate_bits > 1
+            && polynomials.len() > 0
+            && pols * (1 << (log_n + rate_bits)) < (1 << 31)
+        {
             let _num_gpus: usize = std::env::var("NUM_OF_GPUS")
                 .expect("NUM_OF_GPUS should be set")
                 .parse()
@@ -232,17 +255,17 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
     }
 
     #[cfg(feature = "cuda")]
-    pub fn from_coeffs_gpu(
+    fn from_coeffs_gpu(
         polynomials: &[PolynomialCoeffs<F>],
         rate_bits: usize,
-        _blinding: bool,
+        blinding: bool,
         cap_height: usize,
         timing: &mut TimingTree,
         _fft_root_table: Option<&FftRootTable<F>>,
         log_n: usize,
         _degree: usize,
     ) -> MerkleTree<F, <C as GenericConfig<D>>::Hasher> {
-        // let salt_size = if blinding { SALT_SIZE } else { 0 };
+        let salt_size = if blinding { SALT_SIZE } else { 0 };
         // println!("salt_size: {:?}", salt_size);
         let output_domain_size = log_n + rate_bits;
 
@@ -255,8 +278,9 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
         let total_num_of_fft = polynomials.len();
         // println!("total_num_of_fft: {:?}", total_num_of_fft);
 
+        let num_of_cols = total_num_of_fft + salt_size; // if blinding, extend by salt_size
         let total_num_input_elements = total_num_of_fft * (1 << log_n);
-        let total_num_output_elements = total_num_of_fft * (1 << output_domain_size);
+        let total_num_output_elements = num_of_cols * (1 << output_domain_size);
 
         let mut gpu_input: Vec<F> = polynomials
             .into_iter()
@@ -270,6 +294,7 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
         cfg_lde.are_outputs_on_device = true;
         cfg_lde.with_coset = true;
         cfg_lde.is_multi_gpu = true;
+        cfg_lde.salt_size = salt_size as u32;
 
         let mut device_output_data: HostOrDeviceSlice<'_, F> =
             HostOrDeviceSlice::cuda_malloc(0 as i32, total_num_output_elements).unwrap();
@@ -302,7 +327,7 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
         }
 
         let mut cfg_trans = TransposeConfig::default();
-        cfg_trans.batches = total_num_of_fft as u32;
+        cfg_trans.batches = num_of_cols as u32;
         cfg_trans.are_inputs_on_device = true;
         cfg_trans.are_outputs_on_device = true;
 
@@ -327,10 +352,14 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
             MerkleTree::new_from_gpu_leaves(
                 &device_transpose_data,
                 1 << output_domain_size,
-                total_num_of_fft,
+                num_of_cols,
                 cap_height
             )
         );
+
+        drop(device_transpose_data);
+        drop(device_output_data);
+
         mt
     }
 
@@ -340,6 +369,9 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
         blinding: bool,
         fft_root_table: Option<&FftRootTable<F>>,
     ) -> Vec<Vec<F>> {
+        #[cfg(all(feature = "cuda", any(test, doctest)))]
+        init_gpu();
+
         let degree = polynomials[0].len();
         #[cfg(all(feature = "cuda", feature = "batch"))]
         let log_n = log2_strict(degree) + rate_bits;
@@ -443,11 +475,11 @@ impl<F: RichField + Extendable<D>, C: GenericConfig<D, F = F>, const D: usize>
                     println!("collect data from gpu used: {:?}", start.elapsed());
                     r
                 })
-                // .chain(
-                //     (0..salt_size)
-                //         .into_par_iter()
-                //         .map(|_| F::rand_vec(degree << rate_bits)),
-                // )
+                .chain(
+                    (0..salt_size)
+                        .into_par_iter()
+                        .map(|_| F::rand_vec(degree << rate_bits)),
+                )
                 .collect();
             println!("real lde elapsed: {:?}", start_lde.elapsed());
             return ret;

plonky2/src/gates/gate.rs

@@ -8,7 +8,7 @@ use core::ops::Range;
 use std::sync::Arc;
 
 use hashbrown::HashMap;
-use serde::{ Serialize, Serializer};
+use serde::{Serialize, Serializer};
 
 use crate::field::batch_util::batch_multiply_inplace;
 use crate::field::extension::{Extendable, FieldExtension};

plonky2/src/gates/low_degree_interpolation.rs

@@ -84,11 +84,7 @@ impl<F: RichField + Extendable<D>, const D: usize> Gate<F, D> for LowDegreeInter
     fn id(&self) -> String {
         format!("{self:?}<D={D}>")
     }
-    fn serialize(
-        &self,
-        dst: &mut Vec<u8>,
-        _common_data: &CommonCircuitData<F, D>,
-    ) -> IoResult<()> {
+    fn serialize(&self, dst: &mut Vec<u8>, _common_data: &CommonCircuitData<F, D>) -> IoResult<()> {
         dst.write_usize(self.subgroup_bits)?;
         Ok(())
     }

plonky2/src/hash/arch/x86_64/goldilocks_avx512.rs

@@ -7,11 +7,26 @@ const MSB_: i64 = 0x8000000000000000u64 as i64;
 const P8_: i64 = 0xFFFFFFFF00000001u64 as i64;
 const P8_N_: i64 = 0xFFFFFFFF;
 
+#[allow(non_snake_case)]
+#[repr(align(64))]
+struct FieldConstants {
+    MSB_V: [i64; 8],
+    P8_V: [i64; 8],
+    P8_N_V: [i64; 8],
+}
+
+const FC: FieldConstants = FieldConstants {
+    MSB_V: [MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_],
+    P8_V: [P8_, P8_, P8_, P8_, P8_, P8_, P8_, P8_],
+    P8_N_V: [P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_],
+};
+
 #[allow(dead_code)]
 #[inline(always)]
 pub fn shift_avx512(a: &__m512i) -> __m512i {
     unsafe {
-        let msb = _mm512_set_epi64(MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_);
+        // let msb = _mm512_set_epi64(MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_, MSB_);
+        let msb = _mm512_load_si512(FC.MSB_V.as_ptr().cast::<i32>());
         _mm512_xor_si512(*a, msb)
     }
 }
@@ -20,27 +35,31 @@ pub fn shift_avx512(a: &__m512i) -> __m512i {
 #[inline(always)]
 pub fn to_canonical_avx512(a: &__m512i) -> __m512i {
     unsafe {
-        let p8 = _mm512_set_epi64(P8_, P8_, P8_, P8_, P8_, P8_, P8_, P8_);
-        let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        // let p8 = _mm512_set_epi64(P8_, P8_, P8_, P8_, P8_, P8_, P8_, P8_);
+        // let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        let p8 = _mm512_load_si512(FC.P8_V.as_ptr().cast::<i32>());
+        let p8_n = _mm512_load_si512(FC.P8_N_V.as_ptr().cast::<i32>());
         let result_mask = _mm512_cmpge_epu64_mask(*a, p8);
         _mm512_mask_add_epi64(*a, result_mask, *a, p8_n)
     }
 }
 
 #[inline(always)]
-pub fn add_avx512_b_c(a: &__m512i, b: &__m512i) -> __m512i {
+pub fn add_avx512(a: &__m512i, b: &__m512i) -> __m512i {
     unsafe {
-        let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        // let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        let p8_n = _mm512_load_si512(FC.P8_N_V.as_ptr().cast::<i32>());
         let c0 = _mm512_add_epi64(*a, *b);
         let result_mask = _mm512_cmpgt_epu64_mask(*a, c0);
         _mm512_mask_add_epi64(c0, result_mask, c0, p8_n)
     }
 }
 
 #[inline(always)]
-pub fn sub_avx512_b_c(a: &__m512i, b: &__m512i) -> __m512i {
+pub fn sub_avx512(a: &__m512i, b: &__m512i) -> __m512i {
     unsafe {
-        let p8 = _mm512_set_epi64(P8_, P8_, P8_, P8_, P8_, P8_, P8_, P8_);
+        // let p8 = _mm512_set_epi64(P8_, P8_, P8_, P8_, P8_, P8_, P8_, P8_);
+        let p8 = _mm512_load_si512(FC.P8_V.as_ptr().cast::<i32>());
         let c0 = _mm512_sub_epi64(*a, *b);
         let result_mask = _mm512_cmpgt_epu64_mask(*b, *a);
         _mm512_mask_add_epi64(c0, result_mask, c0, p8)
@@ -50,11 +69,25 @@ pub fn sub_avx512_b_c(a: &__m512i, b: &__m512i) -> __m512i {
 #[inline(always)]
 pub fn reduce_avx512_128_64(c_h: &__m512i, c_l: &__m512i) -> __m512i {
     unsafe {
-        let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        // let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        let p8_n = _mm512_load_si512(FC.P8_N_V.as_ptr().cast::<i32>());
         let c_hh = _mm512_srli_epi64(*c_h, 32);
-        let c1 = sub_avx512_b_c(c_l, &c_hh);
+        let c1 = sub_avx512(c_l, &c_hh);
         let c2 = _mm512_mul_epu32(*c_h, p8_n);
-        add_avx512_b_c(&c1, &c2)
+        add_avx512(&c1, &c2)
+    }
+}
+
+// Here we suppose c_h < 2^32
+#[inline(always)]
+pub fn reduce_avx512_96_64(c_h: &__m512i, c_l: &__m512i) -> __m512i {
+    unsafe {
+        let msb = _mm512_load_si512(FC.MSB_V.as_ptr().cast::<i32>());
+        let p_n = _mm512_load_si512(FC.P8_N_V.as_ptr().cast::<i32>());
+        let c_ls = _mm512_xor_si512(*c_l, msb);
+        let c2 = _mm512_mul_epu32(*c_h, p_n);
+        let c_s = add_avx512(&c_ls, &c2);
+        _mm512_xor_si512(c_s, msb)
     }
 }
 
@@ -69,7 +102,8 @@ pub fn mult_avx512_128(a: &__m512i, b: &__m512i) -> (__m512i, __m512i) {
         let c_ll = _mm512_mul_epu32(*a, *b);
         let c_ll_h = _mm512_srli_epi64(c_ll, 32);
         let r0 = _mm512_add_epi64(c_hl, c_ll_h);
-        let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        // let p8_n = _mm512_set_epi64(P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_, P8_N_);
+        let p8_n = _mm512_load_si512(FC.P8_N_V.as_ptr().cast::<i32>());
         let r0_l = _mm512_and_si512(r0, p8_n);
         let r0_h = _mm512_srli_epi64(r0, 32);
         let r1 = _mm512_add_epi64(c_lh, r0_l);

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

@@ -10,7 +10,7 @@ pub mod goldilocks_avx512;
 pub mod poseidon2_goldilocks_avx2;
 #[cfg(target_feature = "avx2")]
 pub mod poseidon_bn128_avx2;
-#[cfg(all(target_feature = "avx2", not(target_feature = "avx512dq")))]
+#[cfg(target_feature = "avx2")]
 pub mod poseidon_goldilocks_avx2;
 #[cfg(all(target_feature = "avx2", target_feature = "avx512dq"))]
 pub mod poseidon_goldilocks_avx512;