experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Adjust logging in the algorithm picker. Print more information under VLOG(1) instead of relying on the crasher flag, because they are indepdent intentions.

Browse Source 
PiperOrigin-RevId: 249725221
This commit is contained in:
Tim Shen 2019-05-23 15:16:04 -07:00 committed by TensorFlower Gardener
parent c02e3f9af9
commit 310a4a4419
1 changed files with 9 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
tensorflow/compiler/xla/service/gpu/cudnn_conv_algorithm_picker.cc
View File

@ -178,8 +178,8 @@ struct ConvCacheStats {
int64 cache_misses = 0; int64 cache_misses = 0;
void LogStats() { void LogStats() {
VLOG(1) << "Cache hits: " << cache_hits; VLOG(2) << "Cache hits: " << cache_hits;
VLOG(1) << "Cache misses: " << cache_misses; VLOG(2) << "Cache misses: " << cache_misses;
} }
}; };
@ -425,6 +425,8 @@ StatusOr<AutotuneResult> CudnnConvAlgorithmPicker::PickBestAlgorithmNoCache(
<< AlgorithmToString(first_algorithm) << " vs " << AlgorithmToString(first_algorithm) << " vs "
<< AlgorithmToString(alg); << AlgorithmToString(alg);
PrintPlatformInfo(&stream); PrintPlatformInfo(&stream);
VLOG(1) << "Full module on failure: \n"
<< instr->GetModule()->ToString();
auto* fail = result.mutable_failure(); auto* fail = result.mutable_failure();
fail->set_kind(AutotuneResult::WRONG_RESULT); fail->set_kind(AutotuneResult::WRONG_RESULT);
auto* reference_conv = fail->mutable_reference_conv(); auto* reference_conv = fail->mutable_reference_conv();
@ -462,12 +464,10 @@ StatusOr<AutotuneResult> CudnnConvAlgorithmPicker::PickBestAlgorithmNoCache(
*log.mutable_cudnn_version() = GetCudnnVersion(stream_exec_); *log.mutable_cudnn_version() = GetCudnnVersion(stream_exec_);
log.set_device_pci_bus_id( log.set_device_pci_bus_id(
stream_exec_->GetDeviceDescription().pci_bus_id()); stream_exec_->GetDeviceDescription().pci_bus_id());
VLOG(1) << "Autotuning result: " << log.ShortDebugString();
// If we crash on checking failure, we are in a testing/benchmark mode, thus // If we crash on checking failure, we are in a testing/benchmark mode, thus
// print more information instead of logging to the logger. // omitting logging through the logger.
if (crash_on_checking_failure) { if (!crash_on_checking_failure) {
LOG(INFO) << "Autotuning result: " << log.ShortDebugString();
} else {
VLOG(2) << "Autotuning result:\n" << log.DebugString();
tensorflow::Logger::Singleton()->LogProto(log); tensorflow::Logger::Singleton()->LogProto(log);
} }
} }
@ -527,7 +527,7 @@ StatusOr<bool> CudnnConvAlgorithmPicker::RunOnInstruction(
} }
auto best_algo = std::move(best_algo_or).ValueOrDie(); auto best_algo = std::move(best_algo_or).ValueOrDie();
VLOG(1) << "Setting cudnn conv to use algorithm " VLOG(2) << "Setting cudnn conv to use algorithm "
<< best_algo.conv().algorithm() << " and " << best_algo.conv().algorithm() << " and "
<< NumBytesToString(best_algo.scratch_bytes()) << NumBytesToString(best_algo.scratch_bytes())
<< " of scratch memory: " << instr->ToString() << " of scratch memory: " << instr->ToString()
@ -548,7 +548,7 @@ StatusOr<bool> CudnnConvAlgorithmPicker::RunOnInstruction(
HloInstruction* new_call = computation->AddInstruction( HloInstruction* new_call = computation->AddInstruction(
instr->CloneWithNewOperands(new_call_shape, instr->operands())); instr->CloneWithNewOperands(new_call_shape, instr->operands()));
VLOG(1) << "Replacing convolution " << instr->ToString() << " with " VLOG(2) << "Replacing convolution " << instr->ToString() << " with "
<< new_call->ToString(); << new_call->ToString();
TF_RETURN_IF_ERROR(new_call->set_backend_config(backend_config)); TF_RETURN_IF_ERROR(new_call->set_backend_config(backend_config));