GetRocBlasGemmTypeStringAndOps Class — pytorch Architecture

Source Code

aten/src/ATen/cuda/tunable/GemmRocblas.h lines 174–203

template <typename T>
auto GetRocBlasGemmTypeStringAndOps() {
  rocblas_handle handle = (rocblas_handle)at::cuda::getCurrentCUDABlasHandle();
  int solution_size;
  auto input_output_type = RocBlasDataTypeFor<T>();
  auto compute_type = RocBlasComputeTypeFor<T>();
  // Get the number of available solutions
  TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
                                                            input_output_type,
                                                            input_output_type,
                                                            compute_type,
                                                            rocblas_gemm_flags_none,
                                                            nullptr,
                                                            &solution_size));
  std::vector<int> solutions(solution_size);
  // Get the list of available solutions
  TORCH_ROCBLAS_CHECK(rocblas_gemm_ex_get_solutions_by_type(handle,
                                                            input_output_type,
                                                            input_output_type,
                                                            compute_type,
                                                            rocblas_gemm_flags_none,
                                                            solutions.data(),
                                                            &solution_size));
  std::vector<std::pair<std::string, std::unique_ptr<Callable<GemmParams<T>>>>> ret;
  for (size_t i = 0; i < solutions.size(); ++i) {
    auto callable = std::make_unique<RocblasGemmOp<T>>(solutions[i]);
    ret.emplace_back(std::make_pair(fmt::sprintf("Gemm_Rocblas_%d", solutions[i]), std::move(callable)));
  }
  return ret;
}