[WIP] Pattern matching for uniform_quantize/dequantize ops with convolution custom calls #1
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
justkw
pushed a commit
that referenced
this pull request
May 14, 2024
…d phase to Initialize() Imported from GitHub PR openxla#12228 The first time that a NormThunk is executed, it will build a cudnn execution plan. This build step can hang if a NCCL collective is running at the same time. To fix this, I've moved the build step to take place during thunk initialization. We only observe this hang when using cudnn 9. Here's a backtrace from the hang that will be fixed: ``` Thread 585 (Thread 0x7fb9391ff640 (LWP 41364) "main.py"): #0 0x00007fd3d17cffd9 in ?? () from /lib/x86_64-linux-gnu/libc.so.6 #1 0x00007fd3d17da24f in pthread_rwlock_wrlock () from /lib/x86_64-linux-gnu/libc.so.6 #2 0x00007fd070967dfe in ?? () from /lib/x86_64-linux-gnu/libcuda.so.1 #3 0x00007fd0709c928a in ?? () from /lib/x86_64-linux-gnu/libcuda.so.1 #4 0x00007f1970d76102 in ?? () from /lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.0 openxla#5 0x00007f1970f2c999 in ?? () from /lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.0 openxla#6 0x00007f1970a7d4ab in ?? () from /lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.0 openxla#7 0x00007f1970d0a9cb in ?? () from /lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.1.0 openxla#8 0x00007fce60b2a98c in cudnn::backend::ExecutionPlan::finalize_internal() () from /lib/x86_64-linux-gnu/libcudnn_graph.so.9.1.0 openxla#9 0x00007fce60aefbb1 in cudnn::backend::Descriptor::finalize() () from /lib/x86_64-linux-gnu/libcudnn_graph.so.9.1.0 openxla#10 0x00007fce60b15bec in cudnnBackendFinalize () from /lib/x86_64-linux-gnu/libcudnn_graph.so.9.1.0 #11 0x00007fd2521b8f39 in cudnn_frontend::ExecutionPlanBuilder_v8::build() () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so #12 0x00007fd2521734ba in stream_executor::gpu::(anonymous namespace)::GetExecPlanFromHeuristics(cudnn_frontend::OperationGraph_v8&&, stream_executor::gpu::(anonymous namespace)::CudnnHandle const&, bool) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#13 0x00007fd25216ff9b in stream_executor::gpu::CudnnSupport::NormRunnerFromDesc(stream_executor::Stream*, stream_executor::dnn::AlgorithmDesc const&, stream_executor::dnn::NormKind, double, stream_executor::dnn::TensorDescriptor const&, stream_executor::dnn::TensorDescriptor const&, stream_executor::dnn::TensorDescriptor const&, std::optional<stream_executor::dnn::TensorDescriptor>, std::optional<stream_executor::dnn::TensorDescriptor>, std::optional<stream_executor::dnn::TensorDescriptor>, std::optional<stream_executor::dnn::TensorDescriptor>, std::optional<stream_executor::dnn::TensorDescriptor>, std::optional<stream_executor::dnn::TensorDescriptor>) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so #14 0x00007fd24e36b88b in stream_executor::dnn::NormOp::RunnerFromAlgorithmDesc(stream_executor::dnn::AlgorithmDesc const&, stream_executor::dnn::NormOp::Config, stream_executor::Stream*) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#15 0x00007fd24e36ae37 in stream_executor::dnn::LazyOpRunner<stream_executor::dnn::NormOp>::GetOrCreateRunner(stream_executor::dnn::NormOp::Config, stream_executor::Stream*)::{lambda()#1}::operator()() const () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#16 0x00007fd24e36adbc in void absl::lts_20230802::base_internal::CallOnceImpl<stream_executor::dnn::LazyOpRunner<stream_executor::dnn::NormOp>::GetOrCreateRunner(stream_executor::dnn::NormOp::Config, stream_executor::Stream*)::{lambda()#1}>(std::atomic<unsigned int>*, absl::lts_20230802::base_internal::SchedulingMode, stream_executor::dnn::LazyOpRunner<stream_executor::dnn::NormOp>::GetOrCreateRunner(stream_executor::dnn::NormOp::Config, stream_executor::Stream*)::{lambda()#1}&&) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#17 0x00007fd24e36a9bd in stream_executor::dnn::LazyOpRunner<stream_executor::dnn::NormOp>::GetOrCreateRunner(stream_executor::dnn::NormOp::Config, stream_executor::Stream*) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#18 0x00007fd24e369d29 in xla::gpu::RunGpuNorm(xla::gpu::GpuNormConfig const&, stream_executor::DeviceMemoryBase const&, stream_executor::DeviceMemoryBase const&, stream_executor::DeviceMemoryBase const&, std::optional<stream_executor::DeviceMemoryBase>, std::optional<stream_executor::DeviceMemoryBase>, std::optional<stream_executor::DeviceMemoryBase>, std::optional<stream_executor::DeviceMemoryBase>, std::optional<stream_executor::DeviceMemoryBase>, std::optional<stream_executor::DeviceMemoryBase>, stream_executor::DeviceMemoryBase const&, stream_executor::Stream*, xla::gpu::RunNormOptions) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so openxla#19 0x00007fd24e368be6 in xla::gpu::NormThunk::ExecuteOnStream(xla::gpu::Thunk::ExecuteParams const&) () from /usr/local/lib/python3.10/dist-packages/jaxlib/xla_extension.so ``` Copybara import of the project: -- f535330 by Trevor Morris <[email protected]>: Fix hang with cudnn layer norm by moving cudnn init to Initialize() Merging this change closes openxla#12228 COPYBARA_INTEGRATE_REVIEW=openxla#12228 from trevor-m:tmorris-norm-init f535330 PiperOrigin-RevId: 633220207
justkw
pushed a commit
that referenced
this pull request
Aug 1, 2024
name old cpu/op new cpu/op delta BM_SelectAndScatterF32/128/process_time 889µs ± 1% 740µs ± 3% -16.70% BM_SelectAndScatterF32/256/process_time 3.64ms ± 2% 3.00ms ± 1% -17.64% BM_SelectAndScatterF32/512/process_time 15.3ms ± 1% 13.1ms ± 3% -14.61% PiperOrigin-RevId: 658063846
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
These are draft changes to pattern match uniform_quantize/dequantize with convolution custom calls.
This will match the patterns:
into
onednn_custom_call[int8 in, f32 out]
into
onednn_custom_call[int8 in, int8 out] -> onednn_custom_call[int8 in, f32 out]