### Description
This PR fixes a deadlock bug in EigenNonBlockingThreadPool.h. It only happens on platforms with weakly ordered memory model, such as ARM64.
### Description
1. Remove the onnxruntime::OrtMutex class and replace it with
~absl::Mutex~ std::mutex.
2. After this change, most source files will not include <Windows.h>
indirectly.
### Motivation and Context
To reduce the number of deps we have, and address some Github issues
that are related to build ONNX Runtime from source.
In PR #3000 , I added a custom implementation of std::mutex . It was
mainly because at that time std::mutex's default constructor was not
trivial on Windows. If you had such a mutex as a global var, it could
not be initialized at compile time. Then VC++ team fixed this issue.
Therefore we don't need this custom implementation anymore.
This PR also removes nsync. I ran several models tests on Linux. I
didn't see any perf difference.
This PR also reverts PR #21005 , which is no longer needed since conda
has updated its msvc runtime DLL.
This PR unblocks #22173 and resolves#22092 . We have a lot of open
issues with nsync. This PR can resolve all of them.
This reverts commit 4e15b229a0.
Reason: We are seeing an increase in the number of deadlocks after this
PR. We have a release coming up next week and do not have enough time to
investigate the root cause, hence reverting this PR temporarily.
Moreover, this is causing an increase int he binary size.
### Description
We are seeing an [increase in the number of
deadlocks](https://github.com/microsoft/onnxruntime/pull/22315#issuecomment-2394821893)
after this PR. We have a release coming up next week and do not have
enough time to investigate the root cause, hence reverting this PR
temporarily.
### Motivation and Context
See above.
The purpose of the patch is primarily to save power, but it also has
nice perf benefits (mostly from allowing the system to better distribute
power to cores doing meaningful work).
Changes are twofold:
1) Decrease WorkerLoop spin count dramatically ~10^6 -> ~10^4. The
reality is after ~10^4 spins, if there hasn't been any new work
added its unlikely any new work is imminent so sleep to
preserve power. This aligns more closely with upstream EigenV3.
2) Use exponential backoff for waiting on memory. This saves a bit
more power, and important increases the time between iterations
in WorkerLoop to help accomidate the dramatically lowering spin
counts.
Since the tuning for both the iteration counts / backoff counts are
dramatically different for hybrid/non-hybrid systems, this patch
templates the affected functions and dynamically choses based on
`CPUIDInfo::IsHybrid()`. This seemed like the "lightest weight" way of
getting the change in, although its likely we could incur less dynamic
overhead if we added the template argument to the entirety of
`ThreadPoolTempl`.
Measured performance on an [Intel Meteor Lake
CPU](https://www.intel.com/content/www/us/en/products/sku/237329/intel-core-ultra-7-processor-165u-12m-cache-up-to-4-90-ghz/specifications.html)
across a range of models.
Below are the result of 3 runs with each metric being the
value-before-patch / value-after-patch (so for something like inference
time, lower is better).
<div align="center">
<table>
<tr>
<th>Session creation time cost</th>
<td>0.7179</td>
</tr>
<tr>
<th>First inference time cost</th>
<td>0.7156</td>
</tr>
<tr>
<th>Total inference time cost</th>
<td>1.0146</td>
</tr>
<tr>
<th>Total inference requests</th>
<td>0.8874</td>
</tr>
<tr>
<th>Average inference time cost</th>
<td>0.8800</td>
</tr>
<tr>
<th>Total inference run time</th>
<td>1.0146</td>
</tr>
<tr>
<th>Number of inferences per second</th>
<td>0.8955</td>
</tr>
<tr>
<th>Avg CPU usage</th>
<td>0.9462</td>
</tr>
<tr>
<th>Peak working set size</th>
<td>0.9922</td>
</tr>
<tr>
<th>Runs</th>
<td>1.1552</td>
</tr>
<tr>
<th>Min Latency</th>
<td>0.7283</td>
</tr>
<tr>
<th>Max Latency</th>
<td>0.9258</td>
</tr>
<tr>
<th>P50 Latency</th>
<td>0.9534</td>
</tr>
<tr>
<th>P90 Latency</th>
<td>0.9639</td>
</tr>
<tr>
<th>P95 Latency</th>
<td>0.9659</td>
</tr>
<tr>
<th>P99 Latency</th>
<td>0.9640</td>
</tr>
</table>
</div>
So the net result is a 1.16x improvement in throughput and between
1.08-1.37x improvement in latency.
### Description
Re-implement stacktrace. The new implementation doesn't directly use
Windows API, hence can avoid problems regarding to
initialize/uninitialize the dbghelp library.
### Motivation and Context
- Fix some warnings from Xcode build (`-Wshorten-64-to-32`).
- Enable `-Wshorten-64-to-32` warning if available. Currently it's not fully enabled for `onnxruntime_test_all` and `onnxruntime_providers_xnnpack` yet.
- Some clean up in build.py including setting CMake generator more consistently.
### Description
Run clang-format in CI. Formatted all c/c++, objective-c/c++ files.
Excluded
```
'onnxruntime/core/mlas/**',
'onnxruntime/contrib_ops/cuda/bert/tensorrt_fused_multihead_attention/**',
```
because they contain assembly or is data heavy
### Motivation and Context
Coding style consistency
### Description
Detect and report thread creation failure on Windows.
Do not throw out of constructor after the thread is created,
the thread handle is lost and cannot be joined, resulting in a deadlock.
Make setting a thread priority on Linux consistent with windows.
Set thread priority in the thread itself. Log failure properly,
but do not exit the thread.
### Motivation and Context
Address issues https://github.com/microsoft/onnxruntime/issues/13291
And
https://github.com/microsoft/onnxruntime/issues/13285#issuecomment-1278063223
* Reserve the first core for the main thread
Currently in "auto affinity" mode the worker threads are affinized to cores 0..(N-1), leaving the very last core for the main thread. This patch preserves core #0 for the main thread, and affinizes the worker threads to cores 1..N.
* Avoid unneeded spin_pause in thread pool's worker threads
Remove unneeded PAUSE instruction (0.1-0.2 usec latency) after a worker thread finds a task to execute.
* MLAS/x86: optimize QLinearConv on hybrid CPUs
Existing 4x task granularity for task partitioning on hybrid CPUs is
not sufficient to compensate the difference of VNNI instructions
throughput
between performance and efficient cores. This patch...
* Increases granularity for QLinearConv by 2x, to have 2x more tasks
with 2x
smaller output count
* Limits QLinearConv task count from above, to avoid output count per
task
getting smaller than kernel's capability
* Remove hardcoded task count for QLineConv as it limited scaling on
16+ cores CPUs
* MLAS/x86: optimize QLinearConv on hybrid CPUs
Existing 4x task granularity for task partitioning on hybrid CPUs is not sufficient to compensate the difference of VNNI instructions
throughput between performance and efficient cores. This patch...
* Increases granularity for QLinearConv by 2x, to have 2x more tasks
with 2x smaller output count
* Limits QLinearConv task count from above, to avoid output count per
task getting smaller than kernel's capability
* Remove hardcoded task count for QLineConv as it limited scaling on
16+ cores CP
* Addressing comments
* combining x86 ARM branches in qlinearconv threaded job partition
* revert first core assignment
Co-authored-by: Saurabh <saurabh.tangri@intel.com>
Co-authored-by: Chen Fu <fuchen@microsoft.com>
* add restrictions for hybrid cpus
* add unit test to mock hybrid cpu
* attach hybrid flag
* add mocking interface to CpuInfo
* make is_hybrid
* make mock function const
* add force_hybrid for thread pool
* remove header
Switched the code to C++17. To build ONNX Runtime on old distros like CentOS 7, you need to install a newer GCC from additionary repos. If you build onnxruntime with the newer GCC, typically the result binary can't be distributed to other places because it depends on the new GCC's runtime libraries, something that the stock OS doesn't have. But on RHEL/CentOS, it can be better. We use Red Hat devtoolset 8/9/10 with CentOS7 building our code. The new library features(like std::filesystem) that not exists in the old C++ runtime will be statically linked into the applications with some restrictions:
1. GCC has dual ABI, but we can only use the old one. It means std::string is still copy-on-write and std::list::size() is still O(n). Also, if you build onnxruntime on CentOS 7 and link it with some binaries that were built on CentOS 8 or Ubuntu with the new ABI and export C++ symbols directly(instead of using a C API), the it won't work.
2. We still can't use std::optional. It is a limitation coming from macOS. We will solve it when we got macOS 11 build machines. It won't be too long.
3. Please avoid to use C++17 in CUDA files(*.cu). Also, the *.h files that they include(like core/framework/float16.h). This is Because CUDA 10.2 doesn't support C++17. You are welcome to use the new features in any *.cc files.
* prepare for C# to configure provider options
* add c# code
* revert modification
* Add update provider info configuration in trt ep side
* fix bugs
* fix bug for compiler error C2259
* Add c# test
* fix bug
* fix bug
* Properly deal with string
* Add c# api for accepting trt provider options
* fix bug
* Modify C# test
* add shared lib test
* Add get provider options functionality
* clean up
* clean up
* fix bug
* fix bugs for CI
* Fix bugs for CI and documentation
* Move TRT EP provider options related functions out of C API
* revert
* fix bug
* refactor
* add check for provider options string
* code refactor
* fix CI bug
* Fix CI bugs
* clean up
* fix bug
* Fix bug for Post Analysis
* fix accidental bug
* Add API_IMPL_BEGIN/API_IMPL_END
* clean up
* code refactor
* code refactor
* fix CI fail
* fix bug
* use string append
* Change the code to better handle strncpy and string append
[ PR previously merged as https://github.com//pull/7372, then reverted pending investigation of lost-wake-up issue seen with ParallelExecutor. Issue was a missing test for new work pushed to thread concurrent with a worker blocking. Change from 7372 is the addition of: https://github.com/microsoft/onnxruntime/blob/tiharr/dev-sticky-4/include/onnxruntime/core/platform/EigenNonBlockingThreadPool.h#L1473-L1492 ]
Description: This change updates the heuristics used when a thread selects which worker threads to push work to on entering a parallel loop. Previously, worker threads would maintain a best-effort bitmap of "good worker hints" indicating the threads that were likely to be spinning waiting for work. This change uses a simpler heuristic where a thread records which workers ran its previous loop, and then re-submits its next loop to those same workers. The aim is to retain affinity between a thread and a set of workers, and to avoid maintaining the "good worker hints" bitmaps.
Motivation and Context: Profiling suggested that maintaining the "good worker hints" was taking unexpected time, particularly on NUMA systems. In addition, when running many concurrent workloads, the hints did not provide a way to help retain locality of workers and hence data in caches. Testing to confirm no regressions on microbenchmark (./build/Linux/Release/onnxruntime_benchmark --benchmark_filter=BM_ThreadPoolParallelFor) and on Linux mobilenet_v1_1.0_224.onnx, comparing p50 and p99 with vs without this change:
1 concurrent:
p50 0.0172s vs 0.0181s
p99 0.0204s vs 0.0216s
2 concurrent:
p50 0.0172s vs 0.0181s
p99 0.0213s vs 0.0221s
* Check whether nvcc supports -Wstrict-aliasing before adding the compiler flag in CMakeList.txt.
* Removed reinterpret_cast to not cause strict aliasing violation errors or require -Wno-strict-aliasing when it is not available.
* wait for dispatch done in RunParallelSection
* pass worker_fn by value
* cancel move
* only move work_fn when it is lastly referred
Co-authored-by: Randy Shuai <rashuai@microsoft.com>
* add async dispatch
* minor renamings
* build py38
* restore yml
* fix sync up issue between dispatch thread and main
* fix comments
* refactor SummonWorker and rename to RunInParallelInternal
* Simplified version of WebAssembly support to keep most of existing data structures and add cmake using Ninja and emcmake
* Clean up CMakeLists.txt and add an example to create and compute a kernel
* Load a model from bytes and remove graph building steps
* Add all cpu and contrib ops with mlas library
* WebAssembly build with Onnxruntime C/CXX API
* Use protobuf cmakefile directory instead of adding every necessary source file
* Fix invalid output at example
* add missing files
* Change an example to use Teams model and support ort mobile format
* add API for javascript
* fix input releasing in _ort_run()
* update API
* Let onnxruntime cmake build WebAssembly with option '--wasm'
* allow one-step building for wasm
* Make build script working on Linux and MacOS
* Fix broken build from Windows command
* Enable unit test on building WebAssembly
* Resolve comments
* update build flags
* wasm conv improvement from: 1) GemmV; 2) Depthwise direct convolution 3x3; 3) Direct convolution 3x3
* Cleaned mlas unittest.
* use glob
* update comments
* Update baseline due to loss scale fix (#6948)
* fix stream sync issue (#6954)
* Enable type reduction in EyeLike, Mod, random.cc CPU kernels. (#6960)
* Update EyeLike CPU kernel.
* Update Mod CPU kernel.
* Update Multinomial CPU kernel.
* Slight improvement to Pad CPU kernel binary size.
* Update RandomNormal[Like], RandomUniform[Like] CPU kernels.
* Fix warning from setting multiple MSVC warning level options. (#6917)
Fix warning from setting multiple MSVC warning level options. Replace an existing /Wn flag instead of always appending a new one.
* MLAS: quantized GEMM update (#6916)
Various updates to the int8_t GEMMs:
1) Add ARM64 udot kernel to take advantage of dot product instructions available in newer cores. Some models run 4x faster than the stock implementation we used before.
2) Refactor the x64 kernels to share common code for AVX2(u8u8/u8s8/avxvnni) vs AVX512(u8u8/u8s8/avx512vnni) to reduce binary size.
3) Extend kernels to support per-column zero points for matrix B. This is not currently wired to an operator.
* Implement QLinearAveragePool with unit tests. (#6896)
Implement QLinearAveragePool with unit tests.
* Attention fusion detect num_heads and hidden_size automatically (#6920)
* fixed type to experimental session constructor (#6950)
* fixed type to experimental session constructor
Co-authored-by: David Medine <david.medine@brainproducts.com>
* Update onnxruntime_perf_test.exe to accept free dimension overrides (#6962)
Co-authored-by: Ori Levari <orlevari@microsoft.com>
* Fix possible fd leak in NNAPI (#6966)
* Release buffers for prepacked tensors (#6820)
Unsolved problems:
1. One test failure was caused by a bug in Cudnn rnn kernels, when they can allocate a buffer and partially initialize it, the garbage data near tail of the buffer caused problem in some of the hardware. To attack this problem in a broader sense, should we add code in our allocators, and during a memory fuzzing test, fill an allocated buffer with garbage before returning to the caller?
2. Prepacking is used more widely than we know. For instance, Cudnn rnn kernels also cache their weights. They mix several weight tensors together into a single buffer, and never touch the original weight tensor anymore. This is the same idea with pre-pack, but they didn't override the virtual function, and they never tried to release those weight tensors, leading to memory waste. It also seems to me that there are some other kernels have similar behavior. Wonder how much memory we can save if we try to cleanup those too.
3. Turning off memory pattern planning does increase memory fragmentation, leading to out of memory error in some training test cases. Perhaps we can revisit the idea of pushing kernels-creation stage earlier, and then during initializer deserialization, we only avoid tracing those that will be prepacked.
* Enable type reduction for Range, ReverseSequence, ScatterND, Split, and Unique CPU kernels. (#6963)
* add CI
* fix test in ci
* fix flags for nsync in wasm build
* add copyright banner
* fix wasm source glob
* add missing exports
* resolve comments
* Perf gain by make packb wide to 4 from 16 on GEMM for WASM.
Remove no need direct conv in previous perf tuning.
* fix buildbreak introduced from latest master merge
* fix buildbreak in mlasi.h
* resolve all comments except MLAS
* rewrite packb related 3 functions for WASM_SCALAR seperately rather than using #ifdef in each.
and other changes according to PR feedback in mlas.
* More complete scalar path in sgemm from Tracy.
* Fix edge case handling in depthwise conv2d kernel 3x3. where:
*) support input W==1 and H==1
*) recalc in accurate pad_right and pad_bottom
*) support hidden pad_right == 2 or pad_bottom == 2 when W == 1 or H==1 and no pad left/top
* Add more test coverage for conv depthwise from Tracy.
Fix one typo according to PR.
* resolve comments
* replace typedef by using
* do not use throw in OrtRun()
* output error message
Co-authored-by: Sunghoon <35605090+hanbitmyths@users.noreply.github.com>
Co-authored-by: Lei Zhang <zhang.huanning@hotmail.com>
Co-authored-by: Wei-Sheng Chin <wschin@outlook.com>
Co-authored-by: Tianlei Wu <tlwu@microsoft.com>
Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com>
Co-authored-by: Tracy Sharpe <42477615+tracysh@users.noreply.github.com>
Co-authored-by: David Medine <david.eric.medine@gmail.com>
Co-authored-by: David Medine <david.medine@brainproducts.com>
Co-authored-by: Ori Levari <ori.levari@microsoft.com>
Co-authored-by: Ori Levari <orlevari@microsoft.com>
Co-authored-by: Guoyu Wang <62914304+gwang-msft@users.noreply.github.com>
Co-authored-by: Chen Fu <chenfucs@gmail.com>
Description: This change adds alignment and padding to avoid false sharing on fields in the thread pool. It also adds a new microbenchmark to profile thread-pool performance over short loops.
Motivation and Context
MobileNet on a 2*12-core system showed a performance gap between the ORT thread pool and OpenMP. One cause appeared to be false sharing on fields in the thread pool: ThreadPoolParallelSection::tasks_finished (which the main thread spins on waiting for workers to complete a loop), and the RunQueue::front_ and back_ fields (used respectively by the worker thread and the main thread).
The additional micro-benchmark BM_ThreadPoolSimpleParallelFor tests performance of loops of different sizes at different thread counts. The results below are on a machine with 2*14-core processors (E5-2690 v4) running with 1, 14, 15, and 28 threads. For each test, the microbenchmark has N threads run a loop with N iterations; hence a perfect result is for the time taken to be constant as additional threads are added (although we will also see power management effects helping at very low thread counts). The loop durations (100000, 10000, 1000) correspond roughly to 200us, 20us, and 2us on this machine.
Before change:
BM_ThreadPoolSimpleParallelFor/1/1/100000/real_time 17153 us 17154 us 32
BM_ThreadPoolSimpleParallelFor/14/14/100000/real_time 22553 us 22553 us 30
BM_ThreadPoolSimpleParallelFor/15/15/100000/real_time 21521 us 21521 us 29
BM_ThreadPoolSimpleParallelFor/28/28/100000/real_time 24111 us 24111 us 24
BM_ThreadPoolSimpleParallelFor/1/1/10000/real_time 1719 us 1719 us 407
BM_ThreadPoolSimpleParallelFor/14/14/10000/real_time 3409 us 3409 us 200
BM_ThreadPoolSimpleParallelFor/15/15/10000/real_time 3541 us 3541 us 201
BM_ThreadPoolSimpleParallelFor/28/28/10000/real_time 4576 us 4576 us 151
BM_ThreadPoolSimpleParallelFor/1/1/1000/real_time 174 us 174 us 4017
BM_ThreadPoolSimpleParallelFor/14/14/1000/real_time 1586 us 1586 us 402
BM_ThreadPoolSimpleParallelFor/15/15/1000/real_time 1586 us 1586 us 397
BM_ThreadPoolSimpleParallelFor/28/28/1000/real_time 2864 us 2864 us 232
After change:
BM_ThreadPoolSimpleParallelFor/1/1/100000/real_time 17160 us 17160 us 33
BM_ThreadPoolSimpleParallelFor/14/14/100000/real_time 20989 us 20989 us 31
BM_ThreadPoolSimpleParallelFor/15/15/100000/real_time 22286 us 22286 us 31
BM_ThreadPoolSimpleParallelFor/28/28/100000/real_time 24631 us 24631 us 25
BM_ThreadPoolSimpleParallelFor/1/1/10000/real_time 1718 us 1718 us 407
BM_ThreadPoolSimpleParallelFor/14/14/10000/real_time 2868 us 2868 us 242
BM_ThreadPoolSimpleParallelFor/15/15/10000/real_time 2907 us 2907 us 240
BM_ThreadPoolSimpleParallelFor/28/28/10000/real_time 3872 us 3872 us 186
BM_ThreadPoolSimpleParallelFor/1/1/1000/real_time 175 us 175 us 3938
BM_ThreadPoolSimpleParallelFor/14/14/1000/real_time 933 us 933 us 659
BM_ThreadPoolSimpleParallelFor/15/15/1000/real_time 912 us 912 us 591
BM_ThreadPoolSimpleParallelFor/28/28/1000/real_time 1976 us 1976 us 317
This is a small perf / clean-up change. It removes the Env::Task abstraction which wraps a single std::function field, and adds at least one virtual method call overhead when creating a Task and when executing it. The POSIX and Windows implementations are now identical.
This PR updates the ThreadPool API to support multi-loop parallel sections. As with the OpenMP "parallel" construct, this allows per-loop work to be amortized over a series of loops. For ORT, it also promotes locality between successive loops in the sense that iteration X of one loop will tend to run on the same worker thread as iteration X of preceding loops.
The change was developed while optimizing the implementation of a model that performed better with OpenMP. Profiling indicated that OpenMP was providing lower loop entry/exit costs and that, via OpenMP's static scheduling, it was leading to a lower L2 miss rate in the series of parallel loops used in GRU.
The main changes are:
- Addition of ThreadPool::ParallelSection and underlying support in the modified Eigen thread pool.
- In EigenNonBlockingThreadPool.h, refactoring the RunInParallel method to support two variants: one that takes an existing parallel section object created by the caller, and another (used by default) that creates its own parallel section.
- Simplify ThreadPool::LoopCounter (used by worker threads to claim loop iterations), basing it an ID supplied by the underlying Eigen thread pool for affinity in a series of loops.
- Fix a possible perf issue where a loop with iterations scheduled in batches would have more threads than batches available.
- Use of parallel sections in the GRU operator.
- Additional test cases in threadpool_test.h.
- Additional comments at the top of threadpool.h and EigenNonBlockingThreadPool.h.
Description: This change makes three changes to the ThreadPool class to clean up issues identified during performance analysis and optimization. (1) It uses mm_pause intrinsics in spin loops, helping avoid consuming pipeline resources while waiting. (2) It re-organizes the spin-then-steal loop for work distribution to start out spinning as intended, rather than to start out trying to steal. (3) It updates the ThreadPool class's API to be consistent in the use of static methods for public functions. The PR includes minor doc updates and corresponding changes to test cases.
Motivation and Context
The change helps ensure consistency in behavior between the OpenMP and Eigen-based implementations. Unlike the instance methods, the static methods abstract over the different ways in which threading can be implemented; they will map onto the OpenMP or Eigen-based implementations when threading is used. When threading is not used they will run work sequentially.
* Add session option and global thread pool option to set denormal as zero.
* Revert unneccessary changes.
* Add cpuinfo submodule
* Add more comments
* Remove cpuinfo submodule dependency and check only SSE3 support for ftz and daz inspired by Tensorflow
* Preserve API order in C api
* Clean up and utilize SSE3 detection logic from existeing cpuid_info.h
* Keep the same order with header file
* Fix build issue with Linux pipeline, which has old g++ compiler
* Fix broken build on Linux and remove a duplicated unit test
* Remove reformatting at eigen thread pool
* Remove flatbuffers which is not intentionally added
* Revert "Remove flatbuffers which is not intentionally added"
This reverts commit 9f509a9aaaa3c7832d88854c82fd26b234770b7f.
* Remove flatbuffers which is not intentionally added
* Resolve comments
- Put details on APIs
- Add a log for ftz/daz initialization
- Add clang
- Fix typo
* Remove unnecessary header include
* Resolve comments
