Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13756
This implements general Gather operator for arbitrary axis, sharing the code with BatchGather.
- CPU gather & batch gather logic is now shared through caffe2::gather_helper, for any axis.
- Shared CUDA kernel moved to gather_op.cuh, for any axis.
- Gradients of axis > 0 delegate to BatchGatherGradientOp which now has axis argument.
- BatchGatherOp doc strings updated to have correct rank (q + (r -1)) and output.
- Added tests for axis == 2.
GatherOp supports index wrapping for axis == 0 by default, which was earlier for ONNX.
This diff also extends it to work in Cuda kernel. Added "wrap_indices" argument which specifies
wheather this wrapping should be done; set it to true if you'd like wrapping for any axis.
TBD: Update gradients to support negative indices (separate diff).
TBD: Once we have operator versioning, we'd like to update GatherOp to NOT support axis 0 wrapping
by default, but rather do it only if wrap_indices is set.
Reviewed By: dzhulgakov
Differential Revision: D12983815
fbshipit-source-id: 8add9d67b47fe8c5ba7a335f581ca0530b205cd7
Summary:
Goal of this PR is to unify cuda and hip device types in caffe2 python front end.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14221
Differential Revision: D13148564
Pulled By: bddppq
fbshipit-source-id: ef9bd2c7d238200165f217097ac5727e686d887b
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:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14342
Sometimes, when we are creating a TaskGroup, we are in fact creating a TaskGroup for a distributed job. In some cases, we may want to register a few nets as "remote" to a TaskGroup. The remote net should have sufficient attributes on where they should be executed later on.
This diff adds the remote net attribute to the TaskGroup class. It exposes two minimal functionalities: adding a remote net, and getting all remote nets added to a TaskGroup.
Reviewed By: d4l3k
Differential Revision: D13188320
fbshipit-source-id: efe947aec30817e9512a5e18be985713b9356bdc
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14143
ConvTranspose has a per-operator attribute rename, which meant that the
global attribute rename for kernels => kernel_shape was not applied.
Changing the behavior so that the global renames always apply, but per-op
renames can override those for specific attributes.
Note: The python frontend path isn't actually used for ConvTranspose, but I
thought it would be good to make it consistent.
Reviewed By: yinghai
Differential Revision: D13113395
fbshipit-source-id: cd3f124b4b5c753a506d297138b7d002b51bfb38
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14196
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13641
FeedTensor function used to take a pointer to Tensor and feed the content using Resize
and mutable_data, but since Tensor is a pointer now, we can just return a Tensor instead.
Reviewed By: dzhulgakov
Differential Revision: D13091163
fbshipit-source-id: 9abf2fd320baca76e050530c500dd29f8e2d0211
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14214
This is to pick up the residual task of T36325466 to make sure that input/output binding of c2 Onnxifi op is positional.
Reviewed By: dzhulgakov
Differential Revision: D13134470
fbshipit-source-id: d1b916dade65c79133b86507cd54ea5166fa6810
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:
xw285cornell
- To make hip files to have unique filename extension we change hip files from _hip.cc to .hip (it's the only blessing option other than .cu in hipcc 3d51a1fb01/bin/hipcc (L552)).
- Change to use host compiler to compile .cc|.cpp files. Previously we use hcc to compile them which is unnecessary
- Change the hipify script to not replace "gpu" with "hip" in the filename of the generated hipified files. Previously we do this because hcc has a bug when linking files that have same filename. We have now changed to use host linker to do linking so this is unnecessary anymore.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14036
Reviewed By: xw285cornell
Differential Revision: D13091813
Pulled By: bddppq
fbshipit-source-id: ea3d887751d8abb39d75f5d5104aa66ce66b9ee0
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13795
Add ability for dot string generation for a single subgraph and python bindings (which is pretty useful for model exploration in Python)
Restructure DotGenerator class a bit to make it easy to implement this feature
Reviewed By: bwasti
Differential Revision: D13010512
fbshipit-source-id: 825665438394b7e6968ab6da167b477af82a7b62
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14074
expose inducesEdges and addNode to python's NNSubgraph. This make it easy to manually construct a NNSubgraph in python
Reviewed By: bwasti
Differential Revision: D13092885
fbshipit-source-id: a94ed0b318162e27e3a4b5a4954eb6d169da7405
Summary:
Currently after performing export it gives two entries of externel_input
of input data in predict_net proto because it extends the externel_input
twice once seperately using input blob and one it is extendind all the entries
of external_input from proto in which input blob is already included
Signed-off-by: Parth Raichura <parth.raichura@softnautics.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12979
Differential Revision: D12916349
Pulled By: soumith
fbshipit-source-id: 4d4a1c68c0936f8de3f4e380aea1393fe193cd2d
Summary:
Avoid false failure by checking for the presence of the test data in setup.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13627
Differential Revision: D13090324
Pulled By: ezyang
fbshipit-source-id: e85571943d168c0007212d7b1a5b99ffa0c39235
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13825
async_polling was an intermediate step towards async_scheduling and is not used
Reviewed By: yinghai
Differential Revision: D13019059
fbshipit-source-id: eee6ba53e7f476ddb481afba3bf1768303864d32
Summary:
This pull request contains changes for:
1. Removing ConvTranspose related changes from caffe2/operators/hip/conv_op_miopen.cc
2. Adding the file caffe2/operators/hip/conv_transpose_op_miopen.cc
3. Modifying the tests to run convTranspose op using MIOpen engine
Differential Revision: D13055099
Pulled By: bddppq
fbshipit-source-id: ca284f8f9a073005b22013c375cc958257815865
Summary: Currently Lambdarank applies exponential emphasis on relevance, i.e., g=2^rel when calculating dcg, this diff adds options that supports g=rel in the loss function.
Reviewed By: itomatik
Differential Revision: D9891514
fbshipit-source-id: 64730d467a665670edd37e6dc1c077987991d1a8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13641
FeedTensor function used to take a pointer to Tensor and feed the content using Resize
and mutable_data, but since Tensor is a pointer now, we can just return a Tensor instead.
Reviewed By: ezyang
Differential Revision: D12873145
fbshipit-source-id: 653735c20d611ff6ac9e380d8b3c721cb396a28f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13798
The semantics of C2 and ONNX Concat is a bit different. C2 concat accepts "add_axis" arg and will raise the dim if so. It's equivalent of attaching a Reshape after plain concat in ONNX.
Reviewed By: rdzhabarov
Differential Revision: D13012867
fbshipit-source-id: da23e555bae709fd2a373b04dcb9db4e984ae315
Summary:
There was a bug in the uniqueness check that only made the first run
unique
Reviewed By: duc0
Differential Revision: D13013504
fbshipit-source-id: ecf7526d0fafd7968f1301734123f93968efef46
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13812
Original commit changeset: 2cf95bdc5ed8
Looks like in iOS, `uint64_t` is not the same as `size_t`. :( Fixed it here.
Reviewed By: houseroad
Differential Revision: D13017390
fbshipit-source-id: d33854ce341225aba372fb945c3704edc14f9411
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13745
We need to support types beside `int64` and `float`.
Reviewed By: bddppq, rdzhabarov
Differential Revision: D12967258
fbshipit-source-id: 688076e6f504b2bf24bba89714df87a678c5638a
Summary:
Add a markdown document summarizing the coverage of serialized operator tests. This currently only takes into account what has been covered by the tests with respect to the entire registry of c2 operators.
Next, we will break down the coverage by which operators have unit tests associated with them, which have hypothesis tests, and which have tests more specifically calling assertReferenceChecks.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13703
Reviewed By: dzhulgakov
Differential Revision: D12970810
Pulled By: ajyu
fbshipit-source-id: 4f0cd057b1cf734371333e24d26cbab630a170e1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13377
* Enable junk fill for the default CPU allocator. The first diff only enables this for the tests. A second diff will change the default of zero-fill to false.
* Fix tests to use 64-bit counters that IterOp and LearningRateOp demands.
* Fix kernels that uses uninitialized memory.
Reviewed By: salexspb
Differential Revision: D10866512
fbshipit-source-id: 17860e77e63a203edf46d0da0335608f77884821
Summary:
I was hitting this error:
caffe2/caffe2/operators/stats_put_ops.h:66:25: runtime error: 9.22337e+18 is outside the range of representable values of type 'long'
So, the assignment from int64_t to float loses some precision and because of that we overflow.
Reproduced this issue with this diff D12945013
Reviewed By: mlappelbaum, jdshi-fb
Differential Revision: D12927086
fbshipit-source-id: 7eae7fe25ab49d5ac15279335bd5b1fa89d6e683
Summary: Adding Fetching Real number representation for int8 tensor in workpace.py
Reviewed By: harouwu
Differential Revision: D12936556
fbshipit-source-id: f8756a37bce21c93d44d52faf5da9c9bd6473f4a
Summary:
We updated the description of upsample_op in onnx: https://github.com/onnx/onnx/pull/1467
Therefore, we need to support the new upsample_op in caffe2-onnx backend as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13272
Reviewed By: houseroad
Differential Revision: D12833656
Pulled By: zrphercule
fbshipit-source-id: 21af5282abaae12d2d044e4018a2b152aff79917
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12733
Conv in NHWC layout only works for 2D images. This has been a pain point when implementing quantized 3D convolution because we need NHWC layout for best performance (note that NHWC layout in general gives better performance in CPU not just for quantized operators). For example, our quantized ops have a functionality to measure quantized error operator by operator but this needs running a shadow fp32 operator, but this is not easy when there's no 3D conv in NHWC layout is available (currently we're doing layout conversion on the fly for the shadow fp32 operator which is error prone). Some of Caffe2 frameworks like brew generates error when we try to create a 3D conv op in NHWC layout. This was also a blocker for using aibench because aibench is using brew.
i-am-not-moving-c2-to-c10
Reviewed By: houseroad
Differential Revision: D10333829
fbshipit-source-id: 2d203ee1db833cd3f9d39353219e3894b46c4389
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13554
D10233252 broke ROCM test.
We don't have group conv in NHWC for hip yet and this diff omits related tests.
Reviewed By: hyuen
Differential Revision: D12917880
fbshipit-source-id: 9baf36a8cb061ee8cf393b2c438a2d1460ce5cd8
