From c391dd633e5c0266a31ac119e4fc8423e437e38b Mon Sep 17 00:00:00 2001
From: Nikola Samardzic <nsamar@csail.mit.edu>
Date: Thu, 15 Jun 2023 20:32:58 +0000
Subject: [PATCH] [`bench.cpp`] Add benchmarks for `mul_avx`, `add_avx`, and
 `mul33c_avx`

---
 benchs/bench.cpp | 53 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 53 insertions(+)

diff --git a/benchs/bench.cpp b/benchs/bench.cpp
index c84f32b..2cc1ab3 100644
--- a/benchs/bench.cpp
+++ b/benchs/bench.cpp
@@ -1,6 +1,7 @@
 #include <benchmark/benchmark.h>
 #include <iostream>
 
+#include "../src/goldilocks_cubic_extension.hpp"
 #include "../src/goldilocks_base_field.hpp"
 #include "../src/poseidon_goldilocks.hpp"
 #include "../src/poseidon_goldilocks_avx.hpp"
@@ -23,6 +24,46 @@
 #define NPHASES_LDE 2
 #define NBLOCKS 1
 
+static void escape(void *p) {
+    asm volatile("" : : "g"(p): "memory");
+}
+
+static void BENCH_ADD_AVX(benchmark::State &state) {
+    __m256i a_;
+    __m256i b_;
+
+    // Benchmark
+    for (auto _ : state)
+    {
+         Goldilocks::add_avx(a_, a_, b_);
+         escape(&a_);
+    }
+}
+
+static void BENCH_MUL_AVX(benchmark::State &state) {
+    Goldilocks::Element a_ = Goldilocks::fromU64(9);
+    Goldilocks::Element b_ = Goldilocks::fromU64(10);
+
+    // Benchmark
+    for (auto _ : state)
+    {
+        Goldilocks::mul_avx(&a_, &a_, &b_);
+        escape(&a_);
+    }
+}
+
+static void BENCH_MUL33C_AVX(benchmark::State &state) {
+    Goldilocks::Element a_[12];
+    Goldilocks::Element b_[12];
+
+    // Benchmark
+    for (auto _ : state)
+    {
+        Goldilocks3::mul33c_avx(a_, a_, b_);
+        escape(a_);
+    }
+}
+
 static void POSEIDON_BENCH_FULL(benchmark::State &state)
 {
     uint64_t input_size = (uint64_t)NUM_HASHES * (uint64_t)SPONGE_WIDTH;
@@ -886,6 +927,18 @@ static void EXTENDEDPOL_BENCH(benchmark::State &state)
     free(c);
 }
 
+BENCHMARK(BENCH_ADD_AVX)
+    ->Unit(benchmark::kNanosecond)
+    ->UseRealTime();
+
+BENCHMARK(BENCH_MUL_AVX)
+    ->Unit(benchmark::kNanosecond)
+    ->UseRealTime();
+
+BENCHMARK(BENCH_MUL33C_AVX)
+    ->Unit(benchmark::kNanosecond)
+    ->UseRealTime();
+
 BENCHMARK(POSEIDON_BENCH_FULL)
     ->Unit(benchmark::kMicrosecond)
     ->DenseRange(omp_get_max_threads() / 2, omp_get_max_threads(), omp_get_max_threads() / 2)