#include <benchmark/benchmark.h>
#include "cuelim.cuh"

template <MatGF256 (*GpuFunc)(const MatGF256 &, const MatGF256 &, const GF256 &)>
void bench_mul8(benchmark::State &state)
{
    GF256 ff(0b100011101);
    uint_fast32_t seed = 41921095;
    size_t x = state.range(0), y = state.range(1), z = state.range(2);
    MatGF256 A(x, y), B(y, z);
    A.randomize(seed);
    B.randomize(seed);
    for (auto _ : state)
    {
        MatGF256 C(GpuFunc(A, B, ff));
    }
}

BENCHMARK_TEMPLATE(bench_mul8, gpu_mul)->Args({100000, 100000, 100000});