1#include <benchmark/benchmark.h>
  2
  3#include <cstdint>
  4
  5#include "spqlios/q120/q120_arithmetic.h"
  6
  7#define ARGS Arg(128)->Arg(4096)->Arg(10000)
  8
  9template <typeof(q120_vec_mat1col_product_baa_ref) f>
 10void benchmark_baa(benchmark::State& state) {
 11  const uint64_t ell = state.range(0);
 12  q120_mat1col_product_baa_precomp* precomp = q120_new_vec_mat1col_product_baa_precomp();
 13
 14  uint64_t* a = new uint64_t[ell * 4];
 15  uint64_t* b = new uint64_t[ell * 4];
 16  uint64_t* c = new uint64_t[4];
 17  for (uint64_t i = 0; i < 4 * ell; i++) {
 18    a[i] = rand();
 19    b[i] = rand();
 20  }
 21  for (auto _ : state) {
 22    f(precomp, ell, (q120b*)c, (q120a*)a, (q120a*)b);
 23  }
 24  delete[] c;
 25  delete[] b;
 26  delete[] a;
 27  q120_delete_vec_mat1col_product_baa_precomp(precomp);
 28}
 29
 30BENCHMARK(benchmark_baa<q120_vec_mat1col_product_baa_ref>)->Name("q120_vec_mat1col_product_baa_ref")->ARGS;
 31BENCHMARK(benchmark_baa<q120_vec_mat1col_product_baa_avx2>)->Name("q120_vec_mat1col_product_baa_avx2")->ARGS;
 32
 33template <typeof(q120_vec_mat1col_product_bbb_ref) f>
 34void benchmark_bbb(benchmark::State& state) {
 35  const uint64_t ell = state.range(0);
 36  q120_mat1col_product_bbb_precomp* precomp = q120_new_vec_mat1col_product_bbb_precomp();
 37
 38  uint64_t* a = new uint64_t[ell * 4];
 39  uint64_t* b = new uint64_t[ell * 4];
 40  uint64_t* c = new uint64_t[4];
 41  for (uint64_t i = 0; i < 4 * ell; i++) {
 42    a[i] = rand();
 43    b[i] = rand();
 44  }
 45  for (auto _ : state) {
 46    f(precomp, ell, (q120b*)c, (q120b*)a, (q120b*)b);
 47  }
 48  delete[] c;
 49  delete[] b;
 50  delete[] a;
 51  q120_delete_vec_mat1col_product_bbb_precomp(precomp);
 52}
 53
 54BENCHMARK(benchmark_bbb<q120_vec_mat1col_product_bbb_ref>)->Name("q120_vec_mat1col_product_bbb_ref")->ARGS;
 55BENCHMARK(benchmark_bbb<q120_vec_mat1col_product_bbb_avx2>)->Name("q120_vec_mat1col_product_bbb_avx2")->ARGS;
 56
 57template <typeof(q120_vec_mat1col_product_bbc_ref) f>
 58void benchmark_bbc(benchmark::State& state) {
 59  const uint64_t ell = state.range(0);
 60  q120_mat1col_product_bbc_precomp* precomp = q120_new_vec_mat1col_product_bbc_precomp();
 61
 62  uint64_t* a = new uint64_t[ell * 4];
 63  uint64_t* b = new uint64_t[ell * 4];
 64  uint64_t* c = new uint64_t[4];
 65  for (uint64_t i = 0; i < 4 * ell; i++) {
 66    a[i] = rand();
 67    b[i] = rand();
 68  }
 69  for (auto _ : state) {
 70    f(precomp, ell, (q120b*)c, (q120b*)a, (q120c*)b);
 71  }
 72  delete[] c;
 73  delete[] b;
 74  delete[] a;
 75  q120_delete_vec_mat1col_product_bbc_precomp(precomp);
 76}
 77
 78BENCHMARK(benchmark_bbc<q120_vec_mat1col_product_bbc_ref>)->Name("q120_vec_mat1col_product_bbc_ref")->ARGS;
 79BENCHMARK(benchmark_bbc<q120_vec_mat1col_product_bbc_avx2>)->Name("q120_vec_mat1col_product_bbc_avx2")->ARGS;
 80
 81EXPORT void q120x2_vec_mat2cols_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
 82                                                 q120b* const res, const q120b* const x, const q120c* const y);
 83EXPORT void q120x2_vec_mat1col_product_bbc_avx2(q120_mat1col_product_bbc_precomp* precomp, const uint64_t ell,
 84                                                q120b* const res, const q120b* const x, const q120c* const y);
 85
 86template <typeof(q120_vec_mat1col_product_bbc_ref) f>
 87void benchmark_x2c2_bbc(benchmark::State& state) {
 88  const uint64_t ell = state.range(0);
 89  q120_mat1col_product_bbc_precomp* precomp = q120_new_vec_mat1col_product_bbc_precomp();
 90
 91  uint64_t* a = new uint64_t[ell * 8];
 92  uint64_t* b = new uint64_t[ell * 16];
 93  uint64_t* c = new uint64_t[16];
 94  for (uint64_t i = 0; i < 8 * ell; i++) {
 95    a[i] = rand();
 96  }
 97  for (uint64_t i = 0; i < 16 * ell; i++) {
 98    b[i] = rand();
 99  }
100  for (auto _ : state) {
101    f(precomp, ell, (q120b*)c, (q120b*)a, (q120c*)b);
102  }
103  delete[] c;
104  delete[] b;
105  delete[] a;
106  q120_delete_vec_mat1col_product_bbc_precomp(precomp);
107}
108
109BENCHMARK(benchmark_x2c2_bbc<q120x2_vec_mat2cols_product_bbc_avx2>)->Name("q120x2_vec_mat2col_product_bbc_avx2")->ARGS;
110
111template <typeof(q120_vec_mat1col_product_bbc_ref) f>
112void benchmark_x2c1_bbc(benchmark::State& state) {
113  const uint64_t ell = state.range(0);
114  q120_mat1col_product_bbc_precomp* precomp = q120_new_vec_mat1col_product_bbc_precomp();
115
116  uint64_t* a = new uint64_t[ell * 8];
117  uint64_t* b = new uint64_t[ell * 8];
118  uint64_t* c = new uint64_t[8];
119  for (uint64_t i = 0; i < 8 * ell; i++) {
120    a[i] = rand();
121  }
122  for (uint64_t i = 0; i < 8 * ell; i++) {
123    b[i] = rand();
124  }
125  for (auto _ : state) {
126    f(precomp, ell, (q120b*)c, (q120b*)a, (q120c*)b);
127  }
128  delete[] c;
129  delete[] b;
130  delete[] a;
131  q120_delete_vec_mat1col_product_bbc_precomp(precomp);
132}
133
134BENCHMARK(benchmark_x2c1_bbc<q120x2_vec_mat1col_product_bbc_avx2>)->Name("q120x2_vec_mat1col_product_bbc_avx2")->ARGS;
135
136BENCHMARK_MAIN();