**Summary**
Update onednn from v2.7.3 to v3.1.1.
It is bc-breaking as some APIs are changed on oneDNN side. Changes include:
- PyTorch code where oneDNN is directly called
- Submodule `third_party/ideep` to adapt to oneDNN's new API.
- CMAKE files to fix build issues.
**Test plan**
Building issues and correctness are covered by CI checks.
For performance, we have run TorchBench models to ensure there is no regression. Below is the comparison before and after oneDNN update.
Note:
- Base commit of PyTorch: da322ea
- CPU: Intel(R) Xeon(R) Platinum 8380 CPU @ 2.30GHz (Ice Lake)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97957
Approved by: https://github.com/jgong5, https://github.com/jerryzh168
Re-landing #68111/#74596
## Description
v0.5 PR of this [RFC](https://github.com/pytorch/pytorch/issues/49444).
On the basis of #50256, the below improvements are included:
* The [v0.5 release branch](https://github.com/oneapi-src/oneDNN/releases/tag/graph-v0.5) of the oneDNN Graph API is used
* The fuser now works with the profiling graph executor. We have inserted type check nodes to guard the profiled tensor properties.
### User API:
The optimization pass is disabled by default. Users could enable it by:
```
torch.jit.enable_onednn_fusion(True)
```
`torch.jit.freeze` should be used after tracing (recommended) or scripting a model.
### Performance:
[pytorch/benchmark](https://github.com/pytorch/benchmark) tool is used to compare the performance:
* SkyLake 8180 (1 socket of 28 cores):
* SkyLake 8180 (single thread):
* By mapping hardswish to oneDNN Graph, it’s 8% faster than PyTorch JIT (NNC + OFI)
** We expect performance gain after mapping transpose, contiguous & view to oneDNN graph ops
### Directory structure of the integration code
Fuser-related code is placed under:
```
torch/csrc/jit/codegen/onednn/
```
Optimization pass registration is done in:
```
torch/csrc/jit/passes/onednn_graph_fuser.h
```
CMake for the integration code is in:
```
caffe2/CMakeLists.txt
cmake/public/mkldnn.cmake
cmake/Modules/FindMKLDNN.cmake
```
## Limitations
* In this PR, we only support Pytorch-oneDNN-Graph integration on Linux platform. Support on Windows and MacOS will be enabled as a next step.
* We have only optimized the inference use-case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76622
Approved by: https://github.com/eellison
Summary:
## Description
Preview4 PR of this [RFC](https://github.com/pytorch/pytorch/issues/49444).
On the basis of https://github.com/pytorch/pytorch/pull/50256, the below improvements are included:
- The [preview4 release branch](https://github.com/oneapi-src/oneDNN/releases/tag/graph-v0.4.1) of the oneDNN Graph API is used
- The fuser now works with the profiling graph executor. We have inserted type check nodes to guard the profiled tensor properties.
### User API:
The optimization pass is disabled by default. Users could enable it by:
```
torch.jit.enable_onednn_fusion(True)
```
### Performance:
[pytorch/benchmark](https://github.com/pytorch/benchmark) tool is used to compare the performance:
- SkyLake 8180 (1 socket of 28 cores):
- SkyLake 8180 (single thread):
\* By mapping hardswish to oneDNN Graph, it’s 8% faster than PyTorch JIT (NNC + OFI)
\** We expect performance gain after mapping transpose, contiguous & view to oneDNN graph ops
### Directory structure of the integration code
Fuser-related code are placed under:
```
torch/csrc/jit/codegen/onednn/
```
Optimization pass registration is done in:
```
torch/csrc/jit/passes/onednn_graph_fuser.h
```
CMake for the integration code is:
```
caffe2/CMakeLists.txt
```
## Limitations
- In this PR, we have only supported the optimization on Linux platform. The support on Windows and MacOS will be enabled as the next step.
- We have only optimized the inference use case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68111
Reviewed By: eellison
Differential Revision: D34584878
Pulled By: malfet
fbshipit-source-id: ce817aa8cc9052ee9ed930c9cf66be83449e61a4
(cherry picked from commit cd17683aa7d9c0947df45a1ab53627feff795587)
Summary:
This PR upgrades oneDNN to v2.5.2, and includes some building support for oneDNN v2.5.2.
v2.4 changes:
- Improved performance for future Intel Xeon Scalable processor (code name Sapphire Rapids). The functionality is disabled by default and should be enabled via CPU dispatcher control.
- Improved binary primitive performance for cases when one of the tensors is broadcasted.
- Improved performance of reduction primitive, reorder, shuffle primitives.
- Improved performance of depthwise convolution forward propagation for processors with Intel AVX5-12 support
- Improved performance of forward inner product primitive for the shapes with minibatch equal to 1 for processors with Intel AVX-512 support
- Improved performance of int8 matmul and inner product primitives for processors with Intel AVX2 and Intel DL Boost support
v2.5 changes:
- Improved performance for future Intel Xeon Scalable processors (code name Sapphire Rapids). The functionality is now enabled by default and requires Linux kernel 5.16.
- Improved performance of matmul primitive for processors with Intel AVX-512 support.
v2.5.2 changes:
- Fixed performance regression in binary primitive with broadcast
- Fixed segmentation fault in depthwise convolution primitive for shapes with huge spatial size for processors with Intel AVX-512 support
Pull Request resolved: https://github.com/pytorch/pytorch/pull/71546
Reviewed By: george-qi
Differential Revision: D33827108
Pulled By: VitalyFedyunin
fbshipit-source-id: 8f5a19b331c82af5b0783f081e061e1034a93952
(cherry picked from commit 9705212fe9b7b0838cc010d040c37d1175be83ce)
Summary:
This PR upgrades oneDNN to [v2.3.3](https://github.com/oneapi-src/oneDNN/releases/tag/v2.3.3) and includes [Graph API preview release](https://github.com/oneapi-src/oneDNN/releases/tag/graph-v0.2) in one package.
- oneDNN will be located at `pytorch/third_party/ideep/mkl-dnn/third_party/oneDNN`
- The version of oneDNN will be [v2.3.3](https://github.com/oneapi-src/oneDNN/releases/tag/v2.3.3)
The main changes on CPU:
- v2.3
- Extended primitive cache to improve primitive descriptor creation performance.
- Improved primitive cache performance in multithreaded configurations.
- Introduced initial optimizations for bfloat16 compute functionality for future Intel Xeon Scalable processor (code name Sapphire Rapids).
- Improved performance of binary primitive and binary post-op for cases with broadcast and mixed source and destination formats.
- Improved performance of reduction primitive
- Improved performance of depthwise convolution primitive with NHWC activations for training cases
- v2.3.1
- Improved int8 GEMM performance for processors with Intel AVX2 and Intel DL Boost support
- Fixed integer overflow for inner product implementation on CPUs
- Fixed out of bounds access in GEMM implementation for Intel SSE 4.1
- v2.3.2
- Fixed performance regression in fp32 inner product primitive for processors with Intel AVX512 support
- v2.3.3
- Reverted check for memory descriptor stride validity for unit dimensions
- Fixed memory leak in CPU GEMM implementation
More changes can be found in https://github.com/oneapi-src/oneDNN/releases.
- The Graph API provides flexible API for aggressive fusion, and the preview2 supports fusion for FP32 inference. See the [Graph API release branch](https://github.com/oneapi-src/oneDNN/tree/dev-graph-preview2) and [spec](https://spec.oneapi.io/onednn-graph/latest/introduction.html) for more details. A separate PR will be submitted to integrate the oneDNN Graph API to Torchscript graph.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/63748
Reviewed By: albanD
Differential Revision: D32153889
Pulled By: malfet
fbshipit-source-id: 536071168ffe312d452f75d54f34c336ca3778c1
Summary:
This PR: (1) enables the use of a system-provided Intel TBB for building PyTorch, (2) removes `tbb:task_scheduler_init` references since it has been removed from TBB a while ago (3) marks the implementation of `_internal_set_num_threads` with a TODO as it requires a revision that fixes its thread allocation logic.
Tested with `test/run_test`; no new tests are introduced since there are no behavioral changes (removal of `tbb::task_scheduler_init` has no impact on the runtime behavior).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61934
Reviewed By: malfet
Differential Revision: D29805416
Pulled By: cbalioglu
fbshipit-source-id: 22042b428b57b8fede9dfcc83878d679a19561dd
Summary:
Since version 1.6, oneDNN has provided limited support for AArch64 builds.
This minor change is to detect an AArch64 CPU and permit the use of
`USE_MKLDNN` in that case.
Build flags for oneDNN are also modified accordingly.
Note: oneDNN on AArch64, by default, will use oneDNN's reference C++ kernels.
These are not optimised for AArch64, but oneDNN v1.7 onwards provides support
for a limited set of primitives based Arm Compute Library.
See: https://github.com/oneapi-src/oneDNN/pull/795
and: https://github.com/oneapi-src/oneDNN/pull/820
for more details. Support for ACL-based oneDNN primitives in PyTorch
will require some further modification,
Fixes #{issue number}
Pull Request resolved: https://github.com/pytorch/pytorch/pull/50400
Reviewed By: izdeby
Differential Revision: D25886589
Pulled By: malfet
fbshipit-source-id: 2c81277a28ad4528c2d2211381e7c6692d952bc1
Summary:
Fixes https://github.com/pytorch/pytorch/issues/45838
`ARCH_OPT_FLAGS` was the old name of `MKLDNN_ARCH_OPT_FLAGS`, which has been renamed in [this commit](2a011ff02e (diff-a0abcbf647ed740b80615fb5b1614a44L97)), but not updated in pytorch.
As its default value will be set to sse4.1, some kernels are going to fail on the legacy arch that does not support SSE4.1. This patch was to make this flag effective.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46082
Reviewed By: glaringlee
Differential Revision: D24252149
Pulled By: agolynski
fbshipit-source-id: 7079deed373d664763c5888feb28795e5235caa8
Summary:
## Motivation
This PR upgrades MKL-DNN from v0.20 to DNNL v1.2 and resolves https://github.com/pytorch/pytorch/issues/30300.
DNNL (Deep Neural Network Library) is the new brand of MKL-DNN, which improves performance, quality, and usability over the old version.
This PR focuses on the migration of all existing functionalities, including minor fixes, performance improvement and code clean up. It serves as the cornerstone of our future efforts to accommodate new features like OpenCL support, BF16 training, INT8 inference, etc. and to let the Pytorch community derive more benefits from the Intel Architecture.
<br>
## What's included?
Even DNNL has many breaking changes to the API, we managed to absorb most of them in ideep. This PR contains minimalist changes to the integration code in pytorch. Below is a summary of the changes:
<br>
**General:**
1. Replace op-level allocator with global-registered allocator
```
// before
ideep::sum::compute<AllocForMKLDNN>(scales, {x, y}, z);
// after
ideep::sum::compute(scales, {x, y}, z);
```
The allocator is now being registeted at `aten/src/ATen/native/mkldnn/IDeepRegistration.cpp`. Thereafter all tensors derived from the `cpu_engine` (by default) will use the c10 allocator.
```
RegisterEngineAllocator cpu_alloc(
ideep::engine::cpu_engine(),
[](size_t size) {
return c10::GetAllocator(c10::DeviceType::CPU)->raw_allocate(size);
},
[](void* p) {
c10::GetAllocator(c10::DeviceType::CPU)->raw_deallocate(p);
}
);
```
------
2. Simplify group convolution
We had such a scenario in convolution where ideep tensor shape mismatched aten tensor: when `groups > 1`, DNNL expects weights tensors to be 5-d with an extra group dimension, e.g. `goihw` instead of `oihw` in 2d conv case.
As shown below, a lot of extra checks came with this difference in shape before. Now we've completely hidden this difference in ideep and all tensors are going to align with pytorch's definition. So we could safely remove these checks from both aten and c2 integration code.
```
// aten/src/ATen/native/mkldnn/Conv.cpp
if (w.ndims() == x.ndims() + 1) {
AT_ASSERTM(
groups > 1,
"Only group _mkldnn_conv2d weights could have been reordered to 5d");
kernel_size[0] = w.get_dim(0) * w.get_dim(1);
std::copy_n(
w.get_dims().cbegin() + 2, x.ndims() - 1, kernel_size.begin() + 1);
} else {
std::copy_n(w.get_dims().cbegin(), x.ndims(), kernel_size.begin());
}
```
------
3. Enable DNNL built-in cache
Previously, we stored DNNL jitted kernels along with intermediate buffers inside ideep using an LRU cache. Now we are switching to the newly added DNNL built-in cache, and **no longer** caching buffers in order to reduce memory footprint.
This change will be mainly reflected in lower memory usage from memory profiling results. On the code side, we removed couple of lines of `op_key_` that depended on the ideep cache before.
------
4. Use 64-bit integer to denote dimensions
We changed the type of `ideep::dims` from `vector<int32_t>` to `vector<int64_t>`. This renders ideep dims no longer compatible with 32-bit dims used by caffe2. So we use something like `{stride_.begin(), stride_.end()}` to cast parameter `stride_` into a int64 vector.
<br>
**Misc changes in each commit:**
**Commit:** change build options
Some build options were slightly changed, mainly to avoid name collisions with other projects that include DNNL as a subproject. In addition, DNNL built-in cache is enabled by option `DNNL_ENABLE_PRIMITIVE_CACHE`.
Old | New
-- | --
WITH_EXAMPLE | MKLDNN_BUILD_EXAMPLES
WITH_TEST | MKLDNN_BUILD_TESTS
MKLDNN_THREADING | MKLDNN_CPU_RUNTIME
MKLDNN_USE_MKL | N/A (not use MKL anymore)
------
**Commit:** aten reintegration
- aten/src/ATen/native/mkldnn/BinaryOps.cpp
Implement binary ops using new operation `binary` provided by DNNL
- aten/src/ATen/native/mkldnn/Conv.cpp
Clean up group convolution checks
Simplify conv backward integration
- aten/src/ATen/native/mkldnn/MKLDNNConversions.cpp
Simplify prepacking convolution weights
- test/test_mkldnn.py
Fixed an issue in conv2d unit test: it didn't check conv results between mkldnn and aten implementation before. Instead, it compared the mkldnn with mkldnn as the default cpu path will also go into mkldnn. Now we use `torch.backends.mkldnn.flags` to fix this issue
- torch/utils/mkldnn.py
Prepack weight tensor on module `__init__` to achieve better performance significantly
------
**Commit:** caffe2 reintegration
- caffe2/ideep/ideep_utils.h
Clean up unused type definitions
- caffe2/ideep/operators/adam_op.cc & caffe2/ideep/operators/momentum_sgd_op.cc
Unify tensor initialization with `ideep::tensor::init`. Obsolete `ideep::tensor::reinit`
- caffe2/ideep/operators/conv_op.cc & caffe2/ideep/operators/quantization/int8_conv_op.cc
Clean up group convolution checks
Revamp convolution API
- caffe2/ideep/operators/conv_transpose_op.cc
Clean up group convolution checks
Clean up deconv workaround code
------
**Commit:** custom allocator
- Register c10 allocator as mentioned above
<br><br>
## Performance
We tested inference on some common models based on user scenarios, and most performance numbers are either better than or on par with DNNL 0.20.
ratio: new / old | Latency (batch=1 4T) | Throughput (batch=64 56T)
-- | -- | --
pytorch resnet18 | 121.4% | 99.7%
pytorch resnet50 | 123.1% | 106.9%
pytorch resnext101_32x8d | 116.3% | 100.1%
pytorch resnext50_32x4d | 141.9% | 104.4%
pytorch mobilenet_v2 | 163.0% | 105.8%
caffe2 alexnet | 303.0% | 99.2%
caffe2 googlenet-v3 | 101.1% | 99.2%
caffe2 inception-v1 | 102.2% | 101.7%
caffe2 mobilenet-v1 | 356.1% | 253.7%
caffe2 resnet101 | 100.4% | 99.8%
caffe2 resnet152 | 99.8% | 99.8%
caffe2 shufflenet | 141.1% | 69.0% †
caffe2 squeezenet | 98.5% | 99.2%
caffe2 vgg16 | 136.8% | 100.6%
caffe2 googlenet-v3 int8 | 100.0% | 100.7%
caffe2 mobilenet-v1 int8 | 779.2% | 943.0%
caffe2 resnet50 int8 | 99.5% | 95.5%
_Configuration:
Platform: Skylake 8180
Latency Test: 4 threads, warmup 30, iteration 500, batch size 1
Throughput Test: 56 threads, warmup 30, iteration 200, batch size 64_
† Shufflenet is one of the few models that require temp buffers during inference. The performance degradation is an expected issue since we no longer cache any buffer in the ideep. As for the solution, we suggest users opt for caching allocator like **jemalloc** as a drop-in replacement for system allocator in such heavy workloads.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32422
Test Plan:
Perf results: https://our.intern.facebook.com/intern/fblearner/details/177790608?tab=Experiment%20Results
10% improvement for ResNext with avx512, neutral on avx2
More results: https://fb.quip.com/ob10AL0bCDXW#NNNACAUoHJP
Reviewed By: yinghai
Differential Revision: D20381325
Pulled By: dzhulgakov
fbshipit-source-id: 803b906fd89ed8b723c5fcab55039efe3e4bcb77
Summary:
1. upgrade MKL-DNN to v0.20.3
2. allow user to change the capability of primitive cache in mkldnn-bridge by environment value LRU_CACHE_CAPACITY
3. support to fill all tensor elements by one scalar
4. fix the link issue if build with private MKLML other than pre-installed MKL
5. add rnn support in mkldnn-bridge
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22910
Differential Revision: D16365998
Pulled By: VitalyFedyunin
fbshipit-source-id: b8d2bb454cbfbcd4b8983b1a8fa3b83e55ad01c3
Summary:
Currently once user has set `USE_NATIVE_ARCH` to OFF, they will never be able to turn it on for MKLDNN again by simply changing `USE_NATIVE_ARCH`. This commit fixes this issue.
Following up 09ba4df031ed51e05724bb490d4d6fc52b3b1ac6
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23608
Differential Revision: D16599600
Pulled By: ezyang
fbshipit-source-id: 88bbec1b1504b5deba63e56f78632937d003a1f6
Summary:
Currently there is no way to build MKLDNN more optimized than sse4. This commit let MKLDNN build respect USE_NATIVE_ARCH.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23445
Differential Revision: D16542275
Pulled By: ezyang
fbshipit-source-id: 550976531d6a52db9128c0e3d4589a33715feee2
Summary:
Illegal instruction is encountered in pre-built package in MKL-DNN. https://github.com/pytorch/pytorch/issues/23231
To avoid such binary compatibility issue, the HostOpts option in MKL-DNN is disabled in order to build MKL-DNN for generic arch.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23292
Differential Revision: D16488773
Pulled By: soumith
fbshipit-source-id: 9e13c76fb9cb9338103cb767d7463c10891d294a
Summary:
MKL-DNN is the main library for computation when we use ideep device. It can use kernels implemented by different algorithms (including JIT, CBLAS, etc.) for computation. We add the "USE_MKLDNN_CBLAS" (default OFF) build option so that users can decide whether to use CBLAS computation methods or not.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19014
Differential Revision: D16094090
Pulled By: ezyang
fbshipit-source-id: 3f0b1d1a59a327ea0d1456e2752f2edd78d96ccc
Summary:
Upgrade MKl-DNN to 0.17 and static build MKL-DNN to fix the potentail build error due to old mkldnn version in host system.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15504
Differential Revision: D13547885
Pulled By: soumith
fbshipit-source-id: 46f790a3d9289c1e153e51c62be17c5206ea8f9a
Summary:
The motivational of this PR is to enforce mkldnn to use the same omp version of caffe2 framework.
Meanwhile, do not change other assumptions within mkldnn.
Previously, the MKL_cmake_included is set in caffe2 in order to disable omp seeking in mkldnn.
But, with such change, mkldnn has no chance to adapt for mkl found by caffe2.
Then, some building flags of mkl will be not set in mkldnn.
For example, USE_MKL, USE_CBLAS, etc.
In this PR, we enforce set the MKLIOMP5LIB for mkldnn according to caffe2, and tell the mkl root path in MKLROOT for mkldnn. Then, mkldnn is built as expected.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13449
Differential Revision: D12899504
Pulled By: yinghai
fbshipit-source-id: 22a196bd00b4ef0a11d350a32c049304613edf52
