test(benchmark): add baby bear poseidon2 benchmark

Cargo.Bazel.lock

@@ -1,5 +1,5 @@
 {
- "checksum": "d9a69b7b7c2e8f7447076811098173ebe669c22af14223e2e9dfeb933bf91d26",
+ "checksum": "5d4563b52a4fb67e240f27d8b902c760a53d92fd0b9afc3a6e6c6b6f33bb5236",
  "crates": {
  "addchain 0.2.0": {
  "name": "addchain",
@@ -13491,6 +13491,10 @@
  "id": "ff 0.13.0",
  "target": "ff"
  },
+ {
+ "id": "p3-baby-bear 0.1.3-succinct",
+ "target": "p3_baby_bear"
+ },
  {
  "id": "p3-bn254-fr 0.1.3-succinct",
  "target": "p3_bn254_fr"

benchmark/poseidon2/BUILD.bazel

@@ -41,7 +41,10 @@ tachyon_cc_binary(
  "//benchmark/poseidon:simple_poseidon_benchmark_reporter",
  "//benchmark/poseidon2/horizen",
  "//benchmark/poseidon2/plonky3",
+ "//tachyon/base/containers:contains",
  "//tachyon/c/math/elliptic_curves/bn/bn254:fr",
+ "//tachyon/c/math/finite_fields/baby_bear",
  "//tachyon/math/elliptic_curves/bn/bn254:poseidon2",
+ "//tachyon/math/finite_fields/baby_bear:poseidon2",
  ],
 )

benchmark/poseidon2/README.md

@@ -15,42 +15,134 @@ CPU Caches:
  L2 Unified 4096 KiB (x12)
 ```
 
+Note that Poseidon2 runs 100x per test due to some time results being too small when running a single iteration.
+
+## BN254
+
 ```shell
 bazel run -c opt --//:has_openmp --//:has_rtti --//:has_matplotlib //benchmark/poseidon2:poseidon2_benchmark -- -p bn254_fr --vendor horizen --vendor plonky3 --check_results
 ```
 
+### On Intel i9-13900K
+
+| Trial Number | Tachyon | Horizen | Plonky3 |
+| :----------- | --------- | ------------- | --------- |
+| 0 | 0.000788 | **0.000534** | 0.000876 |
+| 1 | 0.000628 | **0.000585** | 0.00087 |
+| 2 | 0.000624 | **0.000517** | 0.000865 |
+| 3 | 0.000622 | **0.000513** | 0.000866 |
+| 4 | 0.000634 | **0.000603** | 0.000861 |
+| 5 | 0.000628 | **0.000512** | 0.001002 |
+| 6 | 0.000618 | **0.00051** | 0.000853 |
+| 7 | 0.000616 | **0.000553** | 0.000852 |
+| 8 | 0.0007 | **0.000693** | 0.000873 |
+| 9 | 0.000614 | **0.000525** | 0.000937 |
+| avg | 0.0006472 | **0.0005545** | 0.0008855 |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_bn254_ubuntu_i9.png)
+
+### On Mac M3 Pro
+
+| Trial Number | Tachyon | Horizen | Plonky3 |
+| :----------- | --------- | ------------- | --------- |
+| 0 | 0.001053 | **0.000816** | 0.001186 |
+| 1 | 0.001033 | **0.00076** | 0.001177 |
+| 2 | 0.001019 | **0.000726** | 0.001157 |
+| 3 | 0.001012 | **0.000712** | 0.001172 |
+| 4 | 0.001007 | **0.000691** | 0.001152 |
+| 5 | 0.001023 | **0.000684** | 0.001131 |
+| 6 | 0.001051 | **0.000682** | 0.001123 |
+| 7 | 0.001005 | **0.000678** | 0.001116 |
+| 8 | 0.000996 | **0.000687** | 0.001118 |
+| 9 | 0.001003 | **0.00068** | 0.001127 |
+| avg | 0.0010202 | **0.0007116** | 0.0011459 |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_bn254_mac_m3.png)
+
+## Baby Bear
+
+Note: Horizen and Plonky3 compute values with a different internal matrix, requiring them to be compared with Tachyon separately.
+
+### Horizen
+
+```shell
+bazel run -c opt --//:has_openmp --//:has_rtti --//:has_matplotlib //benchmark/poseidon2:poseidon2_benchmark -- -p baby_bear --vendor horizen --check_results
+```
+
 #### On Intel i9-13900K
 
-| Repetition | Tachyon | Horizen | Plonky3 |
-| :--------- | ------- | ----------- | ------- |
-| 0 | 8e-06 | **7e-06** | 1e-05 |
-| 1 | 7e-06 | **5e-06** | 8e-06 |
-| 2 | 5e-06 | **4e-06** | 8e-06 |
-| 3 | 6e-06 | **4e-06** | 8e-06 |
-| 4 | 5e-06 | **3e-06** | 7e-06 |
-| 5 | 6e-06 | **3e-06** | 7e-06 |
-| 6 | 5e-06 | **3e-06** | 7e-06 |
-| 7 | 6e-06 | **3e-06** | 7e-06 |
-| 8 | 5e-06 | **4e-06** | 7e-06 |
-| 9 | 5e-06 | **3e-06** | 7e-06 |
-| avg | 5.8e-06 | **3.9e-06** | 7.6e-06 |
-
-![image](/benchmark/poseidon2/poseidon2_benchmark_ubuntu_i9.png)
+| Trial Number | Tachyon | Horizen |
+| :----------- | ------------- | --------- |
+| 0 | **0.000127** | 0.000381 |
+| 1 | **0.000126** | 0.00036 |
+| 2 | **0.000125** | 0.00037 |
+| 3 | **0.000125** | 0.000356 |
+| 4 | **0.000125** | 0.000354 |
+| 5 | **0.000125** | 0.000354 |
+| 6 | **0.000125** | 0.000354 |
+| 7 | **0.000125** | 0.00036 |
+| 8 | **0.000125** | 0.000359 |
+| 9 | **0.000125** | 0.000353 |
+| avg | **0.0001253** | 0.0003601 |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_baby_bear_horizen_ubuntu_i9.png)
+
+#### On Mac M3 Pro
+
+| Trial Number | Tachyon | Horizen |
+| :----------- | ------------- | --------- |
+| 0 | **0.000191** | 0.000203 |
+| 1 | **0.000191** | 0.0002 |
+| 2 | **0.000189** | 0.0002 |
+| 3 | **0.000188** | 0.0002 |
+| 4 | **0.000194** | 0.000199 |
+| 5 | **0.000188** | 0.000199 |
+| 6 | **0.000189** | 0.000199 |
+| 7 | **0.000189** | 0.000199 |
+| 8 | **0.000188** | 0.0002 |
+| 9 | **0.000188** | 0.000199 |
+| avg | **0.0001895** | 0.0001998 |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_baby_bear_horizen_mac_m3.png)
+
+### Plonky3
+
+```shell
+bazel run -c opt --//:has_openmp --//:has_rtti --//:has_matplotlib //benchmark/poseidon2:poseidon2_benchmark -- -p baby_bear --vendor plonky3 --check_results
+```
+
+#### On Intel i9-13900K
+
+| Trial Number | Tachyon | Plonky3 |
+| :----------- | --------- | ------------ |
+| 0 | 0.000112 | **6.6e-05** |
+| 1 | 0.000111 | **6.5e-05** |
+| 2 | 0.000111 | **6.6e-05** |
+| 3 | 0.000111 | **6.6e-05** |
+| 4 | 0.00011 | **6.6e-05** |
+| 5 | 0.000116 | **6.6e-05** |
+| 6 | 0.00011 | **6.5e-05** |
+| 7 | 0.000109 | **6.6e-05** |
+| 8 | 0.00011 | **6.6e-05** |
+| 9 | 0.000109 | **6.5e-05** |
+| avg | 0.0001109 | **6.57e-05** |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_baby_bear_plonky3_ubuntu_i9.png)
 
 #### On Mac M3 Pro
 
-| Repetition | Tachyon | Horizen  | Plonky3 |
-| :--------- | ------- | ----------- | -------- |
-| 0 | 1.3e-05 | **1.2e-05** | 1.5e-05 |
-| 1 | 1e-05 | **8e-06** | 1.1e-05 |
-| 2 | 9e-06 | **7e-06** | 1e-05  |
-| 3 | 9e-06 | **7e-06** | 1e-05  |
-| 4 | 9e-06 | **7e-06** | 1e-05  |
-| 5 | 9e-06 | **7e-06** | 1e-05  |
-| 6 | 9e-06 | **7e-06** | 1e-05  |
-| 7 | 9e-06 | **7e-06** | 1e-05  |
-| 8 | 9e-06 | **7e-06** | 1e-05  |
-| 9 | 9e-06 | **7e-06** | 1e-05  |
-| avg | 9.5e-06 | **7.6e-06** | 1.06e-05 |
-
-![image](/benchmark/poseidon2/poseidon2_benchmark_mac_m3.png)
+| Trial Number | Tachyon  | Plonky3   |
+| :----------- | --------- | ------------- |
+| 0  | 0.000169 | **0.000106** |
+| 1  | 0.000167 | **0.000105** |
+| 2  | 0.000166 | **0.000105** |
+| 3  | 0.000169 | **0.000105** |
+| 4  | 0.000167 | **0.000105** |
+| 5  | 0.00017 | **0.000105** |
+| 6  | 0.000168 | **0.000105** |
+| 7  | 0.000167 | **0.000105** |
+| 8  | 0.000168 | **0.000105** |
+| 9  | 0.000168 | **0.000105** |
+| avg  | 0.0001679 | **0.0001051** |
+
+![image](/benchmark/poseidon2/poseidon2_benchmark_baby_bear_plonky3_mac_m3.png)****

benchmark/poseidon2/horizen/src/lib.rs

@@ -1,22 +1,40 @@
-use std::time::Instant;
-use tachyon_rs::math::elliptic_curves::bn::bn254::Fr as CppFr;
+use std::{sync::Arc, time::Instant};
+use tachyon_rs::math::{
+ elliptic_curves::bn::bn254::Fr as CppBn254Fr, finite_fields::baby_bear::BabyBear as CppBabyBear,
+};
 use zkhash::{
- fields::bn256::FpBN256,
- poseidon2::{poseidon2::Poseidon2, poseidon2_instance_bn256::POSEIDON2_BN256_PARAMS},
+ ark_ff::PrimeField,
+ poseidon2::{
+ poseidon2::Poseidon2, poseidon2_instance_babybear::POSEIDON2_BABYBEAR_16_PARAMS,
+ poseidon2_instance_bn256::POSEIDON2_BN256_PARAMS, poseidon2_params::Poseidon2Params,
+ },
 };
 
-#[no_mangle]
-pub extern "C" fn run_poseidon_horizen_bn254_fr(duration: *mut u64) -> *mut CppFr {
- let poseidon = Poseidon2::new(&POSEIDON2_BN256_PARAMS);
+fn run_poseidon2<F: PrimeField + std::convert::From<i32>, R>(
+ duration: *mut u64,
+ params: &Arc<Poseidon2Params<F>>,
+) -> *mut R {
+ let poseidon2 = Poseidon2::new(params);
 
- let t = poseidon.get_t();
- let input: Vec<FpBN256> = (0..t).map(|_i| FpBN256::from(0)).collect();
+ let t = poseidon2.get_t();
+ let mut input: Vec<F> = (0..t).map(|_| F::from(0)).collect();
 
  let start = Instant::now();
- let state = poseidon.permutation(&input);
+ for _ in 0..100 {
+ input = poseidon2.permutation(&input);
+ }
  unsafe {
  duration.write(start.elapsed().as_micros() as u64);
  }
+ Box::into_raw(Box::new(input[1] as F)) as *mut R
+}
 
- Box::into_raw(Box::new(state[1])) as *mut CppFr
+#[no_mangle]
+pub extern "C" fn run_poseidon2_horizen_baby_bear(duration: *mut u64) -> *mut CppBabyBear {
+ run_poseidon2::<_, CppBabyBear>(duration, &POSEIDON2_BABYBEAR_16_PARAMS)
+}
+
+#[no_mangle]
+pub extern "C" fn run_poseidon2_horizen_bn254_fr(duration: *mut u64) -> *mut CppBn254Fr {
+ run_poseidon2::<_, CppBn254Fr>(duration, &POSEIDON2_BN256_PARAMS)
 }

benchmark/poseidon2/plonky3/Cargo.toml

@@ -17,6 +17,7 @@ publish = false
 
 [dependencies]
 ff = { version = "0.13", features = ["derive", "derive_bits"] }
+p3-baby-bear = "0.1.3-succinct"
 p3-bn254-fr = "0.1.3-succinct"
 p3-field = "0.1.3-succinct"
 p3-poseidon2 = "0.1.3-succinct"

benchmark/poseidon2/plonky3/src/lib.rs

@@ -1,13 +1,19 @@
+use core::fmt;
 use ff::PrimeField;
+use p3_baby_bear::{BabyBear, DiffusionMatrixBabyBear};
 use p3_bn254_fr::{Bn254Fr, DiffusionMatrixBN254, FFBn254Fr};
 use p3_field::AbstractField;
-use p3_poseidon2::{Poseidon2, Poseidon2ExternalMatrixHL};
+use p3_poseidon2::{DiffusionPermutation, Poseidon2, Poseidon2ExternalMatrixHL};
 use p3_symmetric::Permutation;
 use std::time::Instant;
-use tachyon_rs::math::elliptic_curves::bn::bn254::Fr as CppFr;
-use zkhash::ark_ff::{BigInteger, PrimeField as ark_PrimeField};
-use zkhash::fields::bn256::FpBN256 as ark_FpBN256;
-use zkhash::poseidon2::poseidon2_instance_bn256::RC3;
+use tachyon_rs::math::{
+ elliptic_curves::bn::bn254::Fr as CppBn254Fr, finite_fields::baby_bear::BabyBear as CppBabyBear,
+};
+use zkhash::ark_ff::{BigInteger, Field, PrimeField as ark_PrimeField};
+use zkhash::fields::{babybear::FpBabyBear as ark_FpBabyBear, bn256::FpBN256 as ark_FpBN256};
+use zkhash::poseidon2::{
+ poseidon2_instance_babybear::RC16 as BabyBearRC16, poseidon2_instance_bn256::RC3 as BN256RC3,
+};
 
 fn bn254_from_ark_ff(input: ark_FpBN256) -> Bn254Fr {
  let bytes = input.into_bigint().to_bytes_le();
@@ -29,20 +35,32 @@ fn bn254_from_ark_ff(input: ark_FpBN256) -> Bn254Fr {
  }
 }
 
-#[no_mangle]
-pub extern "C" fn run_poseidon_plonky3_bn254_fr(duration: *mut u64) -> *mut CppFr {
- const WIDTH: usize = 3;
- const D: u64 = 5;
- const ROUNDS_F: usize = 8;
- const ROUNDS_P: usize = 56;
+fn baby_bear_from_ark_ff(input: ark_FpBabyBear) -> BabyBear {
+ BabyBear::from_canonical_u32(input.into_bigint().0[0] as u32)
+}
 
+fn run_poseidon2<
+ const WIDTH: usize,
+ const D: u64,
+ const ROUNDS_F: usize,
+ const ROUNDS_P: usize,
+ NativeF: p3_field::PrimeField + fmt::Debug,
+ F: Field,
+ R,
+ DiffusionMatrix: DiffusionPermutation<NativeF, WIDTH>,
+>(
+ duration: *mut u64,
+ rc: &Vec<Vec<F>>,
+ from_ark_ff: &dyn Fn(F) -> NativeF,
+ diffusion_matrix: DiffusionMatrix,
+) -> *mut R {
  // Copy over round constants from zkhash.
- let mut round_constants: Vec<[Bn254Fr; WIDTH]> = RC3
+ let mut round_constants: Vec<[NativeF; WIDTH]> = rc
  .iter()
  .map(|vec| {
  vec.iter()
  .cloned()
- .map(bn254_from_ark_ff)
+ .map(from_ark_ff)
  .collect::<Vec<_>>()
  .try_into()
  .unwrap()
@@ -56,27 +74,48 @@ pub extern "C" fn run_poseidon_plonky3_bn254_fr(duration: *mut u64) -> *mut CppF
  .collect::<Vec<_>>();
  let external_round_constants = round_constants;
 
- let poseidon =
- Poseidon2::<Bn254Fr, Poseidon2ExternalMatrixHL, DiffusionMatrixBN254, WIDTH, D>::new(
- ROUNDS_F,
- external_round_constants,
- Poseidon2ExternalMatrixHL,
- ROUNDS_P,
- internal_round_constants,
- DiffusionMatrixBN254,
- );
+ let poseidon2 = Poseidon2::<NativeF, Poseidon2ExternalMatrixHL, DiffusionMatrix, WIDTH, D>::new(
+ ROUNDS_F,
+ external_round_constants,
+ Poseidon2ExternalMatrixHL,
+ ROUNDS_P,
+ internal_round_constants,
+ diffusion_matrix,
+ );
 
- let mut input = (0..3)
- .map(|_i| Bn254Fr::zero())
+ let mut input = (0..WIDTH)
+ .map(|_i| NativeF::zero())
  .collect::<Vec<_>>()
  .try_into()
  .unwrap();
 
  let start = Instant::now();
- poseidon.permute_mut(&mut input);
+ for _ in 0..100 {
+ poseidon2.permute_mut(&mut input);
+ }
  unsafe {
  duration.write(start.elapsed().as_micros() as u64);
  }
 
- Box::into_raw(Box::new(input[1])) as *mut CppFr
+ Box::into_raw(Box::new(input[1])) as *mut R
+}
+
+#[no_mangle]
+pub extern "C" fn run_poseidon2_plonky3_baby_bear(duration: *mut u64) -> *mut CppBabyBear {
+ run_poseidon2::<16, 7, 8, 13, BabyBear, ark_FpBabyBear, CppBabyBear, DiffusionMatrixBabyBear>(
+ duration,
+ &BabyBearRC16,
+ &baby_bear_from_ark_ff,
+ DiffusionMatrixBabyBear,
+ )
+}
+
+#[no_mangle]
+pub extern "C" fn run_poseidon2_plonky3_bn254_fr(duration: *mut u64) -> *mut CppBn254Fr {
+ run_poseidon2::<3, 5, 8, 56, Bn254Fr, ark_FpBN256, CppBn254Fr, DiffusionMatrixBN254>(
+ duration,
+ &BN256RC3,
+ &bn254_from_ark_ff,
+ DiffusionMatrixBN254,
+ )
 }