| /external/tensorflow/tensorflow/python/eager/ |
| D | profiler_test.py | 45 profile_result = profiler.stop()
47 profile_pb.ParseFromString(profile_result)
62 profile_result = profile_pb.SerializeToString()
63 profiler.save(logdir, profile_result)
|
| /external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
| D | gpu_conv_algorithm_picker.cc | 465 se::dnn::ProfileResult profile_result; in PickBestAlgorithmNoCacheCuda() local
471 options.profile_result = &profile_result; in PickBestAlgorithmNoCacheCuda()
481 if (!profile_result.is_valid()) { in PickBestAlgorithmNoCacheCuda()
494 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in PickBestAlgorithmNoCacheCuda()
699 auto profile_result = algorithms[0]; in PickBestAlgorithmNoCacheRocm() local
702 result.mutable_conv()->set_algorithm(profile_result.algorithm().algo_id()); in PickBestAlgorithmNoCacheRocm()
704 profile_result.algorithm().tensor_ops_enabled()); in PickBestAlgorithmNoCacheRocm()
706 result.set_scratch_bytes(profile_result.scratch_size()); in PickBestAlgorithmNoCacheRocm()
708 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in PickBestAlgorithmNoCacheRocm()
718 se::dnn::ProfileResult profile_result; in PickBestAlgorithmNoCacheRocm() local
[all …]
|
| D | gemm_thunk.cc | 177 se::blas::ProfileResult *profile_result, in RunGemm() argument
280 /*output_profile_result=*/profile_result); in RunGemm()
286 /*output_profile_result=*/profile_result); in RunGemm()
292 /*output_profile_result=*/profile_result); in RunGemm()
297 /*output_profile_result=*/profile_result); in RunGemm()
302 /*output_profile_result=*/profile_result); in RunGemm()
|
| D | gemm_algorithm_picker.cc | 128 se::blas::ProfileResult profile_result; in DoUncachedGemmAutotune() local
138 /*profile_result=*/&profile_result, algorithm) in DoUncachedGemmAutotune()
141 if (!profile_result.is_valid()) { in DoUncachedGemmAutotune()
151 << profile_result.elapsed_time_in_ms() << "ms" << std::endl; in DoUncachedGemmAutotune()
154 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in DoUncachedGemmAutotune()
|
| D | gpu_conv_runner.cc | 92 algorithm, options.profile_result); in RunGpuConvForward()
135 scratch_allocator, algorithm, options.profile_result); in RunGpuConvForwardActivation()
170 algorithm, options.profile_result); in RunGpuConvInternalImpl()
182 algorithm, options.profile_result); in RunGpuConvInternalImpl()
|
| D | gemm_thunk.h | 85 se::blas::ProfileResult* profile_result = nullptr,
|
| D | gpu_conv_runner.h | 36 se::dnn::ProfileResult* profile_result = nullptr; member
|
| /external/tensorflow/tensorflow/core/kernels/ |
| D | conv_grad_input_ops.cc | 419 ProfileResult profile_result; in operator ()() local
423 AlgorithmConfig(profile_algorithm), &profile_result); in operator ()()
424 if (cudnn_launch_status.ok() && profile_result.is_valid()) { in operator ()()
435 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
458 auto profile_result = algorithms[0]; in operator ()()
462 profile_result.algorithm().algo_id()); in operator ()()
464 profile_result.algorithm().tensor_ops_enabled()); in operator ()()
466 result.set_scratch_bytes(profile_result.scratch_size()); in operator ()()
468 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
472 ProfileResult profile_result; in operator ()()
[all …]
|
| D | conv_ops_3d.cc | 539 ProfileResult profile_result; in launch() local
543 AlgorithmConfig(profile_algorithm), &profile_result); in launch()
545 if (profile_result.is_valid()) { in launch()
557 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in launch()
579 auto profile_result = algorithms[0]; in launch()
583 profile_result.algorithm().algo_id()); in launch()
585 profile_result.algorithm().tensor_ops_enabled()); in launch()
587 result.set_scratch_bytes(profile_result.scratch_size()); in launch()
589 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in launch()
593 ProfileResult profile_result; in launch()
[all …]
|
| D | conv_grad_filter_ops.cc | 1014 ProfileResult profile_result; in operator ()() local
1018 AlgorithmConfig(profile_algorithm), &profile_result); in operator ()()
1019 if (cudnn_launch_status.ok() && profile_result.is_valid()) { in operator ()()
1031 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
1056 auto profile_result = algorithms[0]; in operator ()()
1060 profile_result.algorithm().algo_id()); in operator ()()
1062 profile_result.algorithm().tensor_ops_enabled()); in operator ()()
1064 result.set_scratch_bytes(profile_result.scratch_size()); in operator ()()
1066 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
1070 ProfileResult profile_result; in operator ()()
[all …]
|
| D | conv_ops.cc | 1026 ProfileResult profile_result; in operator ()() local
1030 AlgorithmConfig(profile_algorithm), &profile_result); in operator ()()
1031 if (cudnn_launch_status.ok() && profile_result.is_valid()) { in operator ()()
1043 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
1068 auto profile_result = algorithms[0]; in operator ()()
1072 profile_result.algorithm().algo_id()); in operator ()()
1074 profile_result.algorithm().tensor_ops_enabled()); in operator ()()
1076 result.set_scratch_bytes(profile_result.scratch_size()); in operator ()()
1078 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in operator ()()
1082 ProfileResult profile_result; in operator ()()
[all …]
|
| D | conv_grad_ops_3d.cc | 1453 ProfileResult profile_result; in Compute() local
1457 AlgorithmConfig(profile_algorithm), &profile_result); in Compute()
1459 if (profile_result.is_valid()) { in Compute()
1471 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in Compute()
1491 ProfileResult profile_result; in Compute()
1496 &profile_result); in Compute()
1498 if (profile_result.is_valid()) { in Compute()
1506 absl::Milliseconds(profile_result.elapsed_time_in_ms())); in Compute()
1937 ProfileResult profile_result; in Compute() local
1941 AlgorithmConfig(profile_algorithm), &profile_result); in Compute()
[all …]
|
| D | conv_ops_fused_impl.h | 403 se::dnn::ProfileResult profile_result;
407 output_ptr_rz, &profile_result);
409 if (cudnn_launch_status.ok() && profile_result.is_valid()) {
420 absl::Milliseconds(profile_result.elapsed_time_in_ms()));
672 se::dnn::ProfileResult* profile_result) -> Status {
682 scratch_allocator, algorithm_config, profile_result);
|
| /external/tensorflow/tensorflow/stream_executor/rocm/ |
| D | rocm_dnn.cc | 506 dnn::ProfileResult profile_result; in GetProfileResultFromConvSolution() local
507 profile_result.set_algorithm( in GetProfileResultFromConvSolution()
509 profile_result.set_elapsed_time_in_ms(solution.time); in GetProfileResultFromConvSolution()
510 profile_result.set_scratch_size(solution.workspace_size); in GetProfileResultFromConvSolution()
511 return profile_result; in GetProfileResultFromConvSolution()
516 dnn::ProfileResult profile_result; in GetProfileResultFromConvAlgoPerf() local
519 profile_result.set_algorithm( in GetProfileResultFromConvAlgoPerf()
523 profile_result.set_algorithm( in GetProfileResultFromConvAlgoPerf()
527 profile_result.set_algorithm( in GetProfileResultFromConvAlgoPerf()
535 profile_result.set_elapsed_time_in_ms(algorithm.time); in GetProfileResultFromConvAlgoPerf()
[all …]
|
| /external/tensorflow/tensorflow/stream_executor/ |
| D | stream.cc | 3243 Args... args, blas::ProfileResult *profile_result) { in operator ()()
3245 bool record_error = profile_result == nullptr; in operator ()()
3246 return Runner.Run(stream, blas_func, record_error, args..., profile_result); in operator ()()
|