| /external/tensorflow/tensorflow/stream_executor/cuda/ |
| D | cuda_gpu_executor.cc | 420 bool GpuExecutor::Launch(Stream* stream, const ThreadDim& thread_dims, in Launch() argument
434 VlogOccupancyInfo(kernel, thread_dims, block_dims); in Launch()
449 block_dims.z, thread_dims.x, thread_dims.y, in Launch()
450 thread_dims.z, args.number_of_shared_bytes(), in Launch()
455 << " args; thread dim: " << thread_dims.ToString() in Launch()
467 const ThreadDim& thread_dims, in VlogOccupancyInfo() argument
471 VLOG(2) << "Thread dimensions (" << thread_dims.x << ", " << thread_dims.y in VlogOccupancyInfo()
472 << ", " << thread_dims.z << ")"; in VlogOccupancyInfo()
491 smem_per_block, thread_dims, cufunc); in VlogOccupancyInfo()
496 smem_per_block, thread_dims, cufunc); in VlogOccupancyInfo()
[all …]
|
| /external/tensorflow/tensorflow/stream_executor/gpu/ |
| D | gpu_executor.h | 72 bool Launch(Stream* stream, const ThreadDim& thread_dims,
80 const ThreadDim& thread_dims, GpuFunctionHandle func);
87 const ThreadDim& thread_dims, GpuFunctionHandle func);
269 void VlogOccupancyInfo(const KernelBase& kernel, const ThreadDim& thread_dims,
|
| /external/tensorflow/tensorflow/stream_executor/rocm/ |
| D | rocm_gpu_executor.cc | 286 bool GpuExecutor::Launch(Stream* stream, const ThreadDim& thread_dims, in Launch() argument
300 VlogOccupancyInfo(kernel, thread_dims, block_dims); in Launch()
332 block_dims.z, thread_dims.x, thread_dims.y, thread_dims.z, in Launch()
336 << "; thread dim: " << thread_dims.ToString() in Launch()
347 const ThreadDim& thread_dims, in CalculateOccupancy() argument
357 const ThreadDim& thread_dims, in CompareOccupancy() argument
420 const ThreadDim& thread_dims, in VlogOccupancyInfo() argument
|
| /external/tensorflow/tensorflow/stream_executor/ |
| D | trace_listener.h | 49 virtual void LaunchSubmit(Stream* stream, const ThreadDim& thread_dims, in LaunchSubmit() argument
|
| D | stream_executor_pimpl.h | 452 bool Launch(Stream *stream, const ThreadDim &thread_dims,
867 inline Stream &Stream::ThenLaunch(ThreadDim thread_dims, BlockDim block_dims, in ThenLaunch() argument
882 parent_->Launch(this, thread_dims, block_dims, kernel, kernel_args); in ThenLaunch()
|
| D | stream_executor_pimpl.cc | 479 bool StreamExecutor::Launch(Stream *stream, const ThreadDim &thread_dims, in Launch() argument
483 SubmitTrace(&TraceListener::LaunchSubmit, stream, thread_dims, block_dims, in Launch()
486 return implementation_->Launch(stream, thread_dims, block_dims, kernel, args); in Launch()
|
| D | stream_executor_internal.h | 191 virtual bool Launch(Stream *stream, const ThreadDim &thread_dims, in Launch() argument
|
| D | stream.h | 171 Stream &ThenLaunch(ThreadDim thread_dims, BlockDim block_dims,
|
| /external/tensorflow/tensorflow/compiler/xla/service/interpreter/ |
| D | executor.h | 65 bool Launch(Stream *stream, const ThreadDim &thread_dims, in Launch() argument
|
| /external/tensorflow/tensorflow/stream_executor/host/ |
| D | host_gpu_executor.h | 57 bool Launch(Stream *stream, const ThreadDim &thread_dims, in Launch() argument
|