Function GetGpu2DLaunchConfig

tensorflow/core/util/gpu_launch_config.h:253–279

CREATE_CUDA_TYPE_ALIAS(Gpu2DLaunchConfig, Cuda2DLaunchConfig);

inline Gpu2DLaunchConfig GetGpu2DLaunchConfig(int xdim, int ydim,
                                              const Eigen::GpuDevice& d) {
  Gpu2DLaunchConfig config;

  if (xdim <= 0 || ydim <= 0) {
    return config;
  }

  const int kThreadsPerBlock = 256;
  int block_cols = std::min(xdim, kThreadsPerBlock);
  // ok to round down here and just do more loops in the kernel
  int block_rows = std::max(kThreadsPerBlock / block_cols, 1);

  const int physical_thread_count =
      d.getNumGpuMultiProcessors() * d.maxGpuThreadsPerMultiProcessor();

  const int max_blocks = std::max(physical_thread_count / kThreadsPerBlock, 1);

  config.virtual_thread_count = dim3(xdim, ydim, 1);
  config.thread_per_block = dim3(block_cols, block_rows, 1);

  int grid_x = std::min(DivUp(xdim, block_cols), max_blocks);

  config.block_count = dim3(
      grid_x, std::min(max_blocks / grid_x, std::max(ydim / block_rows, 1)), 1);
  return config;
}
#ifndef TENSORFLOW_USE_ROCM
inline Cuda2DLaunchConfig GetCuda2DLaunchConfig(int xdim, int ydim,
                                                const Eigen::GpuDevice& d) {