Function GetGpuLaunchConfig

tensorflow/core/util/gpu_launch_config.h:127–144

Calculate the GPU launch config we should use for a kernel launch. This is assuming the kernel is quite simple and will largely be memory-limited. REQUIRES: work_element_count > 0.

// memory-limited.
// REQUIRES: work_element_count > 0.
inline GpuLaunchConfig GetGpuLaunchConfig(int work_element_count,
                                          const Eigen::GpuDevice& d) {
  CHECK_GT(work_element_count, 0);
  GpuLaunchConfig config;
  const int virtual_thread_count = work_element_count;
  const int physical_thread_count = std::min(
      d.getNumGpuMultiProcessors() * d.maxGpuThreadsPerMultiProcessor(),
      virtual_thread_count);
  const int thread_per_block = std::min(1024, d.maxGpuThreadsPerBlock());
  const int block_count =
      std::min(DivUp(physical_thread_count, thread_per_block),
               d.getNumGpuMultiProcessors());

  config.virtual_thread_count = virtual_thread_count;
  config.thread_per_block = thread_per_block;
  config.block_count = block_count;
  return config;
}
#ifndef TENSORFLOW_USE_ROCM
inline CudaLaunchConfig GetCudaLaunchConfig(int work_element_count,
                                            const Eigen::GpuDevice& d) {