Summary: I found out that without exporting to public format IDEEP transpose operator in the middle of convolution net produces incorrect results (probably reading some out-of-bound memory). Exporting to public format might not be the most efficient solution, but at least it ensures correct behavior.
Test Plan: Running ConvFusion followed by transpose should give identical results on CPU and IDEEP
Reviewed By: bwasti
Differential Revision: D22970872
fbshipit-source-id: 1ddca16233e3d7d35a367c93e72d70632d28e1ef
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37101Fixes#36954.
The basic concept is to streamline the process of rethrowing
c10::Error with extra error information. This is in a few
steps:
- I completely remodeled the Error data type and the internal
invariants. Instead of manually adding in newlines, the
message stack formatting process is responsible for inserting
newlines and spacing as necessary. Call sites are then
modified to respect the new API model.
- TORCH_RETHROW macro is added, which adds context to an error
message and then rethrows it.
New internal assert failure looks like:
```
0 INTERNAL ASSERT FAILED at ../c10/test/util/exception_test.cpp:64, please report a bug to PyTorch.
Exception raised from TestBody at ../c10/test/util/exception_test.cpp:64 (most recent call first):
frame #0: <unknown function> + 0x6aab9 (0x7ff611d3aab9 in /data/users/ezyang/pytorch-tmp/build/lib/libc10.so)
frame #1: ...
```
Error message with context looks like:
```
This is an error
This is context 1
This is context 2
```
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D21202891
Pulled By: ezyang
fbshipit-source-id: 361cadd16bc52e5886dba08e79277771ada76169
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:
In facebookincubator/gloo#212, a libuv based Gloo transport was introduced,
which allows us to use Gloo on macOS (and later perhaps also Windows). This
commit updates CMake code to enable building with USE_DISTRIBUTED=1 on macOS.
A few notes:
* The Caffe2 ops are not compiled, for they depend on `gloo::transport::tcp`.
* The process group implementation uses `gloo::transport::tcp` on Linux (because of `epoll(2)` on Linux and `gloo::transport::uv` on macOS).
* The TCP store works but sometimes crashes on process termination.
* The distributed tests are not yet run.
* The nightly builds don't use `USE_DISTRIBUTED=1`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25260
Reviewed By: mrshenli
Differential Revision: D17202381
Pulled By: pietern
fbshipit-source-id: ca80a82e78a05b4154271d2fb0ed31c8d9f26a7c
Summary:
The input shape checkers in conv/int8_conv operator is aims to avoid the issue when running with mkldnn winograd, the weigths has to be reordered each time if input shape changed.
However, the checkers result to big performance regression due to frequent reorder.
Meanwhile, in mkldnn-bridge, such case has been already fixed by correcting the prop_kind.
Therefore, we have to remove the useless checker to fix the performance regression.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19608
Differential Revision: D15061169
Pulled By: yinghai
fbshipit-source-id: 649a43ae6fce989e84939210f6dffb143ec3d350
Summary:
The mkldnn-bridge is upgraded in this PR to support DNNLOWP operators.
Meanwhile, APIs have been updated in caffe2 to use latest version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16308
Differential Revision: D14697018
Pulled By: yinghai
fbshipit-source-id: ca952589098accb08295fd5aa92924c61e74d69c
Summary:
Since we are going to add ideep to ATen, and ATen is always compiled, it makes sense to have the registration in ATen rather than C2.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18335
Reviewed By: bddppq
Differential Revision: D14578652
Pulled By: gchanan
fbshipit-source-id: 4d77fcfc21a362b21d5291a127498aa722548873
Summary:
For MKL-DNN,the filter data will be reorderd to primitive format, it takes a lot of time.
So the patch provide a method to convert filter format before training.
And "OptimizeForIdeep" will be changed to "OptimizeForMkldnn" in this patch.
This patch depends on https://github.com/pytorch/pytorch/pull/12866
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15171
Differential Revision: D14590741
Pulled By: yinghai
fbshipit-source-id: 07971c9977edac3c8eec08ca2c39cda639683492
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18123
the motivation of this fix is to resolve things like:
for(auto i = 0; i < N; i++) where N is bigger than int32
These instances of comparison were found by enabling -Wsign-compare
There are way too many things to fix, so issuing this as a series of fixes
The plan is to fix all these issues and then enable this flag into Caffe2 to catch future instances
Reviewed By: ZolotukhinM
Differential Revision: D14497094
fbshipit-source-id: bca3927a2188bd33a508fa503ba221c220cdaefe
Summary:
fallback operators to CPU for onnx support
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15270
Differential Revision: D14099496
Pulled By: yinghai
fbshipit-source-id: 52b744aa5917700a802bdf19f7007cdcaa6e640a
Summary:
Impl ExpandDims op and fallback to CPU if needed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15264
Differential Revision: D13808797
Pulled By: yinghai
fbshipit-source-id: 7795ec303a46e85f84e5490273db0ec76e8b9374
Summary:
Add winograd conv method. Users can select the direct conv or winograd conv in the model file.
We close the origin pr https://github.com/pytorch/pytorch/pull/12154 and create this new one for better rebasing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15196
Differential Revision: D13463721
Pulled By: yinghai
fbshipit-source-id: c5cd5c8aa7622ae7e52aeabd3dbb8ffb99b9b4ee
Summary:
1. Add some gloo communication operators into related fallback list;
2. Work around to avoid compiling errors while using fallback operator whose CPU operator inherits from 'OperatorBase' directly like PrefetchOperator;
3. Add new cpu context support for some python module files and resnet50 training example file.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11330
Reviewed By: yinghai
Differential Revision: D13624519
Pulled By: wesolwsk
fbshipit-source-id: ce39d57ddb8cd7786db2e873bfe954069d972f4f
Summary:
Implementation LeakyRelu operator for mkl-dnn,the speed-up of a single operation is up to 10X on BDW.
Implementation rashape operator for mkl-dnn,it will resolve occasionally crash issue which use fallback reshape operator.
Implementation CreateBlobQueue and SafeEnqueueBlobs operators,it will resolve crash issue which use fallback operators.
Fallback CreateBlobsQueueDBOp,TensorProtosDBInput,CloseBlobsQueue operators.
Implement adam operator for mkl-dnn,the speed-up of a single operator is up to 6X on BDW.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11696
Reviewed By: yinghai
Differential Revision: D10100438
Pulled By: wesolwsk
fbshipit-source-id: 0b6e06897cc11e0a8e349d80a870b1e72e47f10d
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15418
Previously we are using Resize + ShareData.
Instead, we'll create a function on Tensor that clones itself with same storage.
Suppose we want `t` to `ShareData` with `t0`, Previous:
```
Tensor t(dims, CPU);
t.Resize(t0.sizes());
t.ShareData(t0);
```
Now:
```
Tensor t = t0.Alias();
```
Reviewed By: dzhulgakov
Differential Revision: D13507609
fbshipit-source-id: 6e4275d02f4c3356cbce91127f1b01111dc86b9f
Summary:
support 0 size in any of the tensor dimensions in mkldnn
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15295
Differential Revision: D13573747
Pulled By: yinghai
fbshipit-source-id: 5bf7a0b9e2567e80f44981a7823be5407fc94e53
Summary:
the speed-up of a single operation is up to 3X .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15106
Differential Revision: D13429596
Pulled By: bddppq
fbshipit-source-id: f8d987cafeac9bef9c3daf7e43ede8c6a4ee2ce5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14268
Removes the need for Context in Tensor by doing simple dispatch for CopyBytes. It'd eventually be subsumed by Roy Li's changes of proper copy_ op, but before that is done, let's get a clear logic of how copies are implemented and clean up some craft in CopyFrom implementation.
Note, that with these changes, one can probably can get rid of Context::CopyFromCPU/CopyToCPU, but it's a matter for follow up diffs.
This diff doesn't change the API of Tensor yet, but relies on the fact that passing `Context` to CopyFrom makes copy async if the device is CUDA and doesn't have any effect otherwise (that's how Context methods are implemented).
This doesn't change semantics of copy async implementation - as before it blindly calls cudaMemcpyAsync which probably means that it can be misused if invoked separately outside of operator body. I'll leave it for the follow up copy_ unification.
For Extend() we always do async copy - it makes sense as it's an in-place device-device operation and only any further op would be observable.
Note: there are now three ways of invoking copy in C2 code - templated CopyBytes, virtual CopyFromCPU/etc, and double-dispatch free method here. Hopefully we can get rid of the second one.
Also, please advise whether it's c10-worthy :)
Reviewed By: ezyang
Differential Revision: D13117987
fbshipit-source-id: a6772d6dcf3effaf06717da3a656fc9873b310b5
Summary:
Add "axis" and "axis_w" arguments in FC to support customized axix to reduce dim.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12971
Reviewed By: bddppq
Differential Revision: D12850675
Pulled By: yinghai
fbshipit-source-id: f1cde163201bd7add53b8475329db1f038a73019
Summary:
It is a operator to copy blob from ideep device to ideep device.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12820
Reviewed By: ezyang
Differential Revision: D10850956
Pulled By: yinghai
fbshipit-source-id: f25bff6238cefe847eb98277979fa59139bff843
