Method run

bench/benchmark-blocking-sizes.cpp:153–229

}

void benchmark_t::run()
{
  size_triple_t productsizes(compact_product_size);

  if (use_default_block_size) {
    eigen_use_specific_block_size = false;
  } else {
    // feed eigen with our custom blocking params
    eigen_use_specific_block_size = true;
    size_triple_t blocksizes(compact_block_size);
    eigen_block_size_k = blocksizes.k;
    eigen_block_size_m = blocksizes.m;
    eigen_block_size_n = blocksizes.n;
  }

  // set up the matrix pool

  const size_t combined_three_matrices_sizes =
    sizeof(Scalar) *
      (productsizes.k * productsizes.m +
       productsizes.k * productsizes.n +
       productsizes.m * productsizes.n);

  // 64 M is large enough that nobody has a cache bigger than that,
  // while still being small enough that everybody has this much RAM,
  // so conveniently we don't need to special-case platforms here.
  const size_t unlikely_large_cache_size = 64 << 20;

  const size_t working_set_size =
    min_working_set_size ? min_working_set_size : unlikely_large_cache_size;

  const size_t matrix_pool_size =
    1 + working_set_size / combined_three_matrices_sizes;

  MatrixType *lhs = new MatrixType[matrix_pool_size];
  MatrixType *rhs = new MatrixType[matrix_pool_size];
  MatrixType *dst = new MatrixType[matrix_pool_size];
  
  for (size_t i = 0; i < matrix_pool_size; i++) {
    lhs[i] = MatrixType::Zero(productsizes.m, productsizes.k);
    rhs[i] = MatrixType::Zero(productsizes.k, productsizes.n);
    dst[i] = MatrixType::Zero(productsizes.m, productsizes.n);
  }

  // main benchmark loop

  int iters_at_a_time = 1;
  float time_per_iter = 0.0f;
  size_t matrix_index = 0;
  while (true) {

    double starttime = timer.getCpuTime();
    for (int i = 0; i < iters_at_a_time; i++) {
      dst[matrix_index].noalias() = lhs[matrix_index] * rhs[matrix_index];
      matrix_index++;
      if (matrix_index == matrix_pool_size) {
        matrix_index = 0;
      }