feat: implement RPO hash using SVE instructionss

.gitignore

@@ -8,3 +8,6 @@ Cargo.lock
 
 # These are backup files generated by rustfmt
 **/*.rs.bk
+
+# Generated by cmake
+cmake-build-*

Cargo.toml

@@ -32,6 +32,7 @@ name = "store"
 harness = false
 
 [features]
+arch-arm64-sve = ["dep:cc"]
 default = ["blake3/default", "std", "winter_crypto/default", "winter_math/default", "winter_utils/default"]
 std = ["blake3/std", "winter_crypto/std", "winter_math/std", "winter_utils/std", "rand_utils"]
 serde = ["winter_math/serde", "dep:serde", "serde/alloc"]
@@ -49,3 +50,6 @@ rand_utils = { version = "0.6", package = "winter-rand-utils", optional = true }
 criterion = { version = "0.5", features = ["html_reports"] }
 proptest = "1.1.0"
 rand_utils = { version = "0.6", package = "winter-rand-utils" }
+
+[build-dependencies]
+cc = { version = "1.0.79", optional = true }

arch/arm64-sve/CMakeLists.txt

@@ -0,0 +1,10 @@
+cmake_minimum_required(VERSION 3.0)
+project(rpo_sve C)
+
+set(CMAKE_C_STANDARD 23)
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=armv8-a+sve -Wall -Wextra -pedantic -g -O3")
+
+add_library(rpo_sve library.c rpo_hash.h)
+
+add_executable(rpo_test test.c)
+target_link_libraries(rpo_test rpo_sve)

arch/arm64-sve/library.c

@@ -0,0 +1,78 @@
+#include <stddef.h>
+#include <arm_sve.h>
+#include "library.h"
+#include "rpo_hash.h"
+
+// The STATE_WIDTH of RPO hash is 12x u64 elements.
+// The current generation of SVE-enabled processors - Neoverse V1
+// (e.g. AWS Graviton3) have 256-bit vector registers (4x u64)
+// This allows us to split the state into 3 vectors of 4 elements
+// and process all 3 independent of each other.
+
+// We see the biggest performance gains by leveraging both
+// vector and scalar operations on parts of the state array.
+// Due to high latency of vector operations, the processor is able
+// to reorder and pipeline scalar instructions while we wait for
+// vector results. This effectively gives us some 'free' scalar
+// operations and masks vector latency.
+//
+// This also means that we can fully saturate all four arithmetic
+// units of the processor (2x scalar, 2x SIMD)
+//
+// THIS ANALYSIS NEEDS TO BE PERFORMED AGAIN ONCE PROCESSORS
+// GAIN WIDER REGISTERS. It's quite possible that with 8x u64
+// vectors processing 2 partially filled vectors might
+// be easier and faster than dealing with scalar operations
+// on the remainder of the array.
+//
+// FOR NOW THIS IS ONLY ENABLED ON 4x u64 VECTORS! It falls back
+// to the regular, already highly-optimized scalar version
+// if the conditions are not met.
+
+bool add_constants_and_apply_sbox(uint64_t state[STATE_WIDTH], uint64_t constants[STATE_WIDTH]) {
+    const uint64_t vl = svcntd();   // number of u64 numbers in one SVE vector
+
+    if (vl != 4) {
+        return false;
+    }
+
+    svbool_t ptrue = svptrue_b64();
+
+    svuint64_t state1 = svld1(ptrue, state + 0*vl);
+    svuint64_t state2 = svld1(ptrue, state + 1*vl);
+
+    svuint64_t const1 = svld1(ptrue, constants + 0*vl);
+    svuint64_t const2 = svld1(ptrue, constants + 1*vl);
+
+    add_constants(ptrue, &state1, &const1, &state2, &const2, state+8, constants+8);
+    apply_sbox(ptrue, &state1, &state2, state+8);
+
+    svst1(ptrue, state + 0*vl, state1);
+    svst1(ptrue, state + 1*vl, state2);
+
+    return true;
+}
+
+bool add_constants_and_apply_inv_sbox(uint64_t state[STATE_WIDTH], uint64_t constants[STATE_WIDTH]) {
+    const uint64_t vl = svcntd();   // number of u64 numbers in one SVE vector
+
+    if (vl != 4) {
+        return false;
+    }
+
+    svbool_t ptrue = svptrue_b64();
+
+    svuint64_t state1 = svld1(ptrue, state + 0 * vl);
+    svuint64_t state2 = svld1(ptrue, state + 1 * vl);
+
+    svuint64_t const1 = svld1(ptrue, constants + 0 * vl);
+    svuint64_t const2 = svld1(ptrue, constants + 1 * vl);
+
+    add_constants(ptrue, &state1, &const1, &state2, &const2, state + 8, constants + 8);
+    apply_inv_sbox(ptrue, &state1, &state2, state + 8);
+
+    svst1(ptrue, state + 0 * vl, state1);
+    svst1(ptrue, state + 1 * vl, state2);
+
+    return true;
+}

arch/arm64-sve/library.h

@@ -0,0 +1,12 @@
+#ifndef CRYPTO_LIBRARY_H
+#define CRYPTO_LIBRARY_H
+
+#include <stdint.h>
+#include <stdbool.h>
+
+#define STATE_WIDTH 12
+
+bool add_constants_and_apply_sbox(uint64_t state[STATE_WIDTH], uint64_t constants[STATE_WIDTH]);
+bool add_constants_and_apply_inv_sbox(uint64_t state[STATE_WIDTH], uint64_t constants[STATE_WIDTH]);
+
+#endif //CRYPTO_LIBRARY_H

arch/arm64-sve/rpo_hash.h

@@ -0,0 +1,221 @@
+#ifndef RPO_SVE_RPO_HASH_H
+#define RPO_SVE_RPO_HASH_H
+
+#include <arm_sve.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <string.h>
+
+#define COPY(NAME, VIN1, VIN2, SIN3)                \
+    svuint64_t NAME ## _1 = VIN1;                   \
+    svuint64_t NAME ## _2 = VIN2;                   \
+    uint64_t NAME ## _3[4];                         \
+    memcpy(NAME ## _3, SIN3, 4 * sizeof(uint64_t))
+
+#define MULTIPLY(PRED, DEST, OP)                    \
+    mul(PRED, &DEST ## _1, &OP ## _1, &DEST ## _2, &OP ## _2, DEST ## _3, OP ## _3)
+
+#define SQUARE(PRED, NAME)                          \
+    sq(PRED, &NAME ## _1, &NAME ## _2, NAME ## _3)
+
+#define SQUARE_DEST(PRED, DEST, SRC)                \
+    COPY(DEST, SRC ## _1, SRC ## _2, SRC ## _3);    \
+    SQUARE(PRED, DEST);
+
+#define POW_ACC(PRED, NAME, CNT, TAIL)              \
+    for (size_t i = 0; i < CNT; i++) {              \
+        SQUARE(PRED, NAME);                         \
+    }                                               \
+    MULTIPLY(PRED, NAME, TAIL);
+
+#define POW_ACC_DEST(PRED, DEST, CNT, HEAD, TAIL)   \
+    COPY(DEST, HEAD ## _1, HEAD ## _2, HEAD ## _3); \
+    POW_ACC(PRED, DEST, CNT, TAIL)
+
+extern inline void add_constants(
+    svbool_t pg,
+    svuint64_t *state1,
+    svuint64_t *const1,
+    svuint64_t *state2,
+    svuint64_t *const2,
+    uint64_t *state3,
+    uint64_t *const3
+) {
+    uint64_t Ms = 0xFFFFFFFF00000001ull;
+    svuint64_t Mv = svindex_u64(Ms, 0);
+
+    uint64_t p_1 = Ms - const3[0];
+    uint64_t p_2 = Ms - const3[1];
+    uint64_t p_3 = Ms - const3[2];
+    uint64_t p_4 = Ms - const3[3];
+
+    uint64_t x_1, x_2, x_3, x_4;
+    uint32_t adj_1 = -__builtin_sub_overflow(state3[0], p_1, &x_1);
+    uint32_t adj_2 = -__builtin_sub_overflow(state3[1], p_2, &x_2);
+    uint32_t adj_3 = -__builtin_sub_overflow(state3[2], p_3, &x_3);
+    uint32_t adj_4 = -__builtin_sub_overflow(state3[3], p_4, &x_4);
+
+    state3[0] = x_1 - (uint64_t)adj_1;
+    state3[1] = x_2 - (uint64_t)adj_2;
+    state3[2] = x_3 - (uint64_t)adj_3;
+    state3[3] = x_4 - (uint64_t)adj_4;
+
+    svuint64_t p1 = svsub_x(pg, Mv, *const1);
+    svuint64_t p2 = svsub_x(pg, Mv, *const2);
+
+    svuint64_t x1 = svsub_x(pg, *state1, p1);
+    svuint64_t x2 = svsub_x(pg, *state2, p2);
+
+    svbool_t pt1 = svcmplt_u64(pg, *state1, p1);
+    svbool_t pt2 = svcmplt_u64(pg, *state2, p2);
+
+    *state1 = svsub_m(pt1, x1, (uint32_t)-1);
+    *state2 = svsub_m(pt2, x2, (uint32_t)-1);
+}
+
+extern inline void mul(
+    svbool_t pg,
+    svuint64_t *r1,
+    const svuint64_t *op1,
+    svuint64_t *r2,
+    const svuint64_t *op2,
+    uint64_t *r3,
+    const uint64_t *op3
+) {
+    __uint128_t x_1 = r3[0];
+    __uint128_t x_2 = r3[1];
+    __uint128_t x_3 = r3[2];
+    __uint128_t x_4 = r3[3];
+
+    x_1 *= (__uint128_t) op3[0];
+    x_2 *= (__uint128_t) op3[1];
+    x_3 *= (__uint128_t) op3[2];
+    x_4 *= (__uint128_t) op3[3];
+
+    uint64_t x0_1 = x_1;
+    uint64_t x0_2 = x_2;
+    uint64_t x0_3 = x_3;
+    uint64_t x0_4 = x_4;
+
+    svuint64_t l1 = svmul_x(pg, *r1, *op1);
+    svuint64_t l2 = svmul_x(pg, *r2, *op2);
+
+    uint64_t x1_1 = (x_1 >> 64);
+    uint64_t x1_2 = (x_2 >> 64);
+    uint64_t x1_3 = (x_3 >> 64);
+    uint64_t x1_4 = (x_4 >> 64);
+
+    uint64_t a_1, a_2, a_3, a_4;
+    uint64_t e_1 = __builtin_add_overflow(x0_1, (x0_1 << 32), &a_1);
+    uint64_t e_2 = __builtin_add_overflow(x0_2, (x0_2 << 32), &a_2);
+    uint64_t e_3 = __builtin_add_overflow(x0_3, (x0_3 << 32), &a_3);
+    uint64_t e_4 = __builtin_add_overflow(x0_4, (x0_4 << 32), &a_4);
+
+    svuint64_t ls1 = svlsl_x(pg, l1, 32);
+    svuint64_t ls2 = svlsl_x(pg, l2, 32);
+
+    svuint64_t a1 = svadd_x(pg, l1, ls1);
+    svuint64_t a2 = svadd_x(pg, l2, ls2);
+
+    svbool_t e1 = svcmplt(pg, a1, l1);
+    svbool_t e2 = svcmplt(pg, a2, l2);
+
+    svuint64_t as1 = svlsr_x(pg, a1, 32);
+    svuint64_t as2 = svlsr_x(pg, a2, 32);
+
+    svuint64_t b1 = svsub_x(pg, a1, as1);
+    svuint64_t b2 = svsub_x(pg, a2, as2);
+
+    b1 = svsub_m(e1, b1, 1);
+    b2 = svsub_m(e2, b2, 1);
+
+    uint64_t b_1 = a_1 - (a_1 >> 32) - e_1;
+    uint64_t b_2 = a_2 - (a_2 >> 32) - e_2;
+    uint64_t b_3 = a_3 - (a_3 >> 32) - e_3;
+    uint64_t b_4 = a_4 - (a_4 >> 32) - e_4;
+
+    uint64_t r_1, r_2, r_3, r_4;
+    uint32_t c_1 = __builtin_sub_overflow(x1_1, b_1, &r_1);
+    uint32_t c_2 = __builtin_sub_overflow(x1_2, b_2, &r_2);
+    uint32_t c_3 = __builtin_sub_overflow(x1_3, b_3, &r_3);
+    uint32_t c_4 = __builtin_sub_overflow(x1_4, b_4, &r_4);
+
+    svuint64_t h1 = svmulh_x(pg, *r1, *op1);
+    svuint64_t h2 = svmulh_x(pg, *r2, *op2);
+
+    svuint64_t tr1 = svsub_x(pg, h1, b1);
+    svuint64_t tr2 = svsub_x(pg, h2, b2);
+
+    svbool_t c1 = svcmplt_u64(pg, h1, b1);
+    svbool_t c2 = svcmplt_u64(pg, h2, b2);
+
+    *r1 = svsub_m(c1, tr1, (uint32_t) -1);
+    *r2 = svsub_m(c2, tr2, (uint32_t) -1);
+
+    uint32_t minus1_1 = 0 - c_1;
+    uint32_t minus1_2 = 0 - c_2;
+    uint32_t minus1_3 = 0 - c_3;
+    uint32_t minus1_4 = 0 - c_4;
+
+    r3[0] = r_1 - (uint64_t)minus1_1;
+    r3[1] = r_2 - (uint64_t)minus1_2;
+    r3[2] = r_3 - (uint64_t)minus1_3;
+    r3[3] = r_4 - (uint64_t)minus1_4;
+}
+
+extern inline void sq(svbool_t pg, svuint64_t *a, svuint64_t *b, uint64_t *c) {
+    mul(pg, a, a, b, b, c, c);
+}
+
+extern inline void apply_sbox(
+    svbool_t pg,
+    svuint64_t *state1,
+    svuint64_t *state2,
+    uint64_t *state3
+) {
+    COPY(x, *state1, *state2, state3);                // copy input to x
+    SQUARE(pg, x);                                    // x contains input^2
+    mul(pg, state1, &x_1, state2, &x_2, state3, x_3); // state contains input^3
+    SQUARE(pg, x);                                    // x contains input^4
+    mul(pg, state1, &x_1, state2, &x_2, state3, x_3); // state contains input^7
+}
+
+extern inline void apply_inv_sbox(
+    svbool_t pg,
+    svuint64_t *state_1,
+    svuint64_t *state_2,
+    uint64_t *state_3
+) {
+    // base^10
+    COPY(t1, *state_1, *state_2, state_3);
+    SQUARE(pg, t1);
+
+    // base^100
+    SQUARE_DEST(pg, t2, t1);
+
+    // base^100100
+    POW_ACC_DEST(pg, t3, 3, t2, t2);
+
+    // base^100100100100
+    POW_ACC_DEST(pg, t4, 6, t3, t3);
+
+    // compute base^100100100100100100100100
+    POW_ACC_DEST(pg, t5, 12, t4, t4);
+
+    // compute base^100100100100100100100100100100
+    POW_ACC_DEST(pg, t6, 6, t5, t3);
+
+    // compute base^1001001001001001001001001001000100100100100100100100100100100
+    POW_ACC_DEST(pg, t7, 31, t6, t6);
+
+    // compute base^1001001001001001001001001001000110110110110110110110110110110111
+    SQUARE(pg, t7);
+    MULTIPLY(pg, t7, t6);
+    SQUARE(pg, t7);
+    SQUARE(pg, t7);
+    MULTIPLY(pg, t7, t1);
+    MULTIPLY(pg, t7, t2);
+    mul(pg, state_1, &t7_1, state_2, &t7_2, state_3, t7_3);
+}
+
+#endif //RPO_SVE_RPO_HASH_H

arch/arm64-sve/test.c

@@ -0,0 +1,27 @@
+#include <stdio.h>
+#include "library.h"
+
+void print_array(size_t len, uint64_t arr[len]);
+
+int main() {
+    uint64_t C[STATE_WIDTH] = {1, 1, 1, 1 ,1, 1, 1, 1 ,1, 1, 1, 1};
+    uint64_t T[STATE_WIDTH] = {1, 2, 3, 4, 1, 2, 3, 4,1, 2, 3, 4};
+
+    add_constants_and_apply_sbox(T, C);
+    add_constants_and_apply_inv_sbox(T, C);
+
+    print_array(STATE_WIDTH, T);
+
+    return 0;
+}
+
+void print_array(size_t len, uint64_t arr[len])
+{
+    printf("[");
+    for (size_t i = 0; i < len; i++)
+    {
+        printf("%lu ", arr[i]);
+    }
+
+    printf("]\n");
+}

build.rs

@@ -0,0 +1,17 @@
+fn main() {
+    #[cfg(feature = "arch-arm64-sve")]
+    compile_arch_arm64_sve();
+}
+
+#[cfg(feature = "arch-arm64-sve")]
+fn compile_arch_arm64_sve() {
+    println!("cargo:rerun-if-changed=arch/arm64-sve/library.c");
+    println!("cargo:rerun-if-changed=arch/arm64-sve/library.h");
+    println!("cargo:rerun-if-changed=arch/arm64-sve/rpo_hash.h");
+
+    cc::Build::new()
+        .file("arch/arm64-sve/library.c")
+        .flag("-march=armv8-a+sve")
+        .flag("-O3")
+        .compile("rpo_sve");
+}

src/hash/rpo/mod.rs

@@ -10,6 +10,19 @@ use mds_freq::mds_multiply_freq;
 #[cfg(test)]
 mod tests;
 
+#[cfg(feature = "arch-arm64-sve")]
+#[link(name = "rpo_sve", kind = "static")]
+extern "C" {
+    fn add_constants_and_apply_sbox(
+        state: *mut std::ffi::c_ulong,
+        constants: *const std::ffi::c_ulong,
+    ) -> bool;
+    fn add_constants_and_apply_inv_sbox(
+        state: *mut std::ffi::c_ulong,
+        constants: *const std::ffi::c_ulong,
+    ) -> bool;
+}
+
 // CONSTANTS
 // ================================================================================================
 
@@ -345,18 +358,65 @@ impl Rpo256 {
     pub fn apply_round(state: &mut [Felt; STATE_WIDTH], round: usize) {
         // apply first half of RPO round
         Self::apply_mds(state);
-        Self::add_constants(state, &ARK1[round]);
-        Self::apply_sbox(state);
+        if !Self::optimized_add_constants_and_apply_sbox(state, &ARK1[round]) {
+            Self::add_constants(state, &ARK1[round]);
+            Self::apply_sbox(state);
+        }
 
         // apply second half of RPO round
         Self::apply_mds(state);
-        Self::add_constants(state, &ARK2[round]);
-        Self::apply_inv_sbox(state);
+        if !Self::optimized_add_constants_and_apply_inv_sbox(state, &ARK2[round]) {
+            Self::add_constants(state, &ARK2[round]);
+            Self::apply_inv_sbox(state);
+        }
     }
 
     // HELPER FUNCTIONS
     // --------------------------------------------------------------------------------------------
 
+    #[inline(always)]
+    #[cfg(feature = "arch-arm64-sve")]
+    fn optimized_add_constants_and_apply_sbox(
+        state: &mut [Felt; STATE_WIDTH],
+        ark: &[Felt; STATE_WIDTH],
+    ) -> bool {
+        unsafe {
+            add_constants_and_apply_sbox(state.as_mut_ptr() as *mut u64, ark.as_ptr() as *const u64)
+        }
+    }
+
+    #[inline(always)]
+    #[cfg(not(feature = "arch-arm64-sve"))]
+    fn optimized_add_constants_and_apply_sbox(
+        _state: &mut [Felt; STATE_WIDTH],
+        _ark: &[Felt; STATE_WIDTH],
+    ) -> bool {
+        false
+    }
+
+    #[inline(always)]
+    #[cfg(feature = "arch-arm64-sve")]
+    fn optimized_add_constants_and_apply_inv_sbox(
+        state: &mut [Felt; STATE_WIDTH],
+        ark: &[Felt; STATE_WIDTH],
+    ) -> bool {
+        unsafe {
+            add_constants_and_apply_inv_sbox(
+                state.as_mut_ptr() as *mut u64,
+                ark.as_ptr() as *const u64,
+            )
+        }
+    }
+
+    #[inline(always)]
+    #[cfg(not(feature = "arch-arm64-sve"))]
+    fn optimized_add_constants_and_apply_inv_sbox(
+        _state: &mut [Felt; STATE_WIDTH],
+        _ark: &[Felt; STATE_WIDTH],
+    ) -> bool {
+        false
+    }
+
     #[inline(always)]
     fn apply_mds(state: &mut [Felt; STATE_WIDTH]) {
         let mut result = [ZERO; STATE_WIDTH];