Method VlogOccupancyInfo

tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:542–579

This is a non-essential operation; if there's a failure, proceed without logging an error. It's nearly certain that in case of failures, we'd never get here in the first place; these are very low-impact routines.

// logging an error. It's nearly certain that in case of failures, we'd never
// get here in the first place; these are very low-impact routines.
void GpuExecutor::VlogOccupancyInfo(const KernelBase& kernel,
                                    const ThreadDim& thread_dims,
                                    const BlockDim& block_dims) {
  VLOG(2) << "Computing kernel occupancy for kernel "
          << kernel.demangled_name();
  VLOG(2) << "Thread dimensions (" << thread_dims.x << ", " << thread_dims.y
          << ", " << thread_dims.z << ")";

  int regs_per_thread;
  if (!kernel.metadata().registers_per_thread(&regs_per_thread)) {
    return;
  }

  int smem_per_block;
  if (!kernel.metadata().shared_memory_bytes(&smem_per_block)) {
    return;
  }

  const DeviceDescription &device_description =
      kernel.parent()->GetDeviceDescription();

  const GpuKernel* cuda_kernel = AsGpuKernel(&kernel);
  CUfunction cufunc = cuda_kernel->AsGpuFunctionHandle();

  int blocks_per_sm = CalculateOccupancy(device_description, regs_per_thread,
                                         smem_per_block, thread_dims, cufunc);
  VLOG(2) << "Resident blocks per SM is " << blocks_per_sm;

  int suggested_threads =
      CompareOccupancy(&blocks_per_sm, device_description, regs_per_thread,
                       smem_per_block, thread_dims, cufunc);
  if (suggested_threads != 0) {
    VLOG(2) << "The cuda occupancy calculator recommends using "
            << suggested_threads
            << " threads per block to achieve an occupancy of " << blocks_per_sm
            << " blocks per SM.";
  }
}

// Compute and return maximum blocks per core (occupancy) based on the
// device description, some kernel characteristics and the number of threads per