platform/upstream/pytorch.git
5 years agoAllow building libraries with setuptools that dont have abi suffix (#14130)
Peter Goldsborough [Wed, 28 Nov 2018 01:33:54 +0000 (17:33 -0800)]
Allow building libraries with setuptools that dont have abi suffix (#14130)

Summary:
When using `setuptools` to build a Python extension, setuptools will automatically add an ABI suffix like `cpython-37m-x86_64-linux-gnu` to the shared library name when using Python 3. This is required for extensions meant to be imported as Python modules. When we use setuptools to build shared libraries not meant as Python modules, for example libraries that define and register TorchScript custom ops, having your library called `my_ops.cpython-37m-x86_64-linux-gnu.so` is a bit annoying compared to just `my_ops.so`, especially since you have to reference the library name when loading it with `torch.ops.load_library` in Python.

This PR fixes this by adding a `with_options` class method to the `torch.utils.cpp_extension.BuildExtension` which allows configuring the `BuildExtension`. In this case, the first option we add is `no_python_abi_suffix`, which we then use in `get_ext_filename` (override from `setuptools.build_ext`) to throw away the ABI suffix.

I've added a test `setup.py` in a `no_python_abi_suffix_test` folder.

Fixes https://github.com/pytorch/pytorch/issues/14188

t-vi fmassa soumith
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14130

Differential Revision: D13216575

Pulled By: goldsborough

fbshipit-source-id: 67dc345c1278a1a4ee4ca907d848bc1fb4956cfa

5 years agoFix clang tidy errors
Wanchao Liang [Wed, 28 Nov 2018 01:28:55 +0000 (17:28 -0800)]
Fix clang tidy errors

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14427

Differential Revision: D13222381

Pulled By: wanchaol

fbshipit-source-id: d90d210a810e95bf0eb404f9c1c304f4e6a3f61e

5 years agoHandling of pretty-printing methods (#14378)
Zachary DeVito [Wed, 28 Nov 2018 01:08:09 +0000 (17:08 -0800)]
Handling of pretty-printing methods (#14378)

Summary:
Stacked on #14176, review only the last commit.
* Print parameters to methods as self.weight rather than as extra inputs.
* Print entire set of methods out as a single string
* Update test code to test the module-at-a-time export/import
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14378

Differential Revision: D13198463

Pulled By: zdevito

fbshipit-source-id: 3fab02e8239cfd6f40d6ab6399047bd02cf0a8c8

5 years agoEliminate necessity of HIPify on AccumulateType.h (#14412)
Edward Yang [Wed, 28 Nov 2018 00:36:09 +0000 (16:36 -0800)]
Eliminate necessity of HIPify on AccumulateType.h (#14412)

Summary:
I'd like to NOT HIPify files that are not in a cuda/
directory, so hand-HIPify AccumulateType.h

Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14412

Differential Revision: D13221801

Pulled By: ezyang

fbshipit-source-id: d1927cfc956e50a6a5e67168ac0e1ce56ecd1e0b

5 years agowhen BUILD_CAFFE2_OPS is OFF, torch-python needs a direct dep on nccl (#14430)
andersj [Tue, 27 Nov 2018 23:51:17 +0000 (15:51 -0800)]
when BUILD_CAFFE2_OPS is OFF, torch-python needs a direct dep on nccl (#14430)

Summary:
https://github.com/pytorch/pytorch/issues/14431 tracks supporting this with CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14430

Differential Revision: D13224079

Pulled By: anderspapitto

fbshipit-source-id: 47d7900d25910ed61585b93f9003acd1b2630a9f

5 years agoSpeed-up "advanced" indexing operations (#13420)
Sam Gross [Tue, 27 Nov 2018 23:18:39 +0000 (15:18 -0800)]
Speed-up "advanced" indexing operations (#13420)

Summary:
This speeds-up "advanced" indexing (indexing a tensor by a tensor)
on CPU and GPU. There's still a bunch of work to do, including
speeding up indexing by a byte (boolean) mask and speeding up the derivative
calculation for advanced indexing.

Here's some speed comparisons to indexing on master using a little [benchmark script](https://gist.github.com/colesbury/c369db72aad594e5e032c8fda557d909) with 16 OpenMP threads and on a P100. The test cases are listed as (input shape -> output shape).

| Test case             | CPU (old vs. new)   | CUDA (old vs. new)     |
|-----------------------|---------------------|------------------------|
| 1024x1024 -> 512x1024 | 225 us vs. **57 us**  | 297 us vs. **47 us** |
| 1024x1024 -> 1024x512 | 208 us vs. **153 us** | 335 us vs. **54 us** |
| 50x50 -> 20000x50     | 617 us vs. **77 us**  | 239 us vs. **54 us** |
| 50x50 -> 50x20000     | 575 us vs. **236 us** | 262 us vs. **58 us** |
| 2x5x10 -> 10          | 65 us  vs. **18 us**  | 612 us vs. **93 us** |

See #11647
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13420

Reviewed By: soumith

Differential Revision: D13088936

Pulled By: colesbury

fbshipit-source-id: 0a5c2ee9aa54e15f96d06692d1694c3b24b924e2

5 years agoResubmit: Set the correct engine name for position weighted pooling when fp16 is...
Jiyan Yang [Tue, 27 Nov 2018 22:49:28 +0000 (14:49 -0800)]
Resubmit: Set the correct engine name for position weighted pooling when fp16 is used for training

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/13768

Reviewed By: xianjiec

Differential Revision: D12996103

fbshipit-source-id: 5ca4cda4210f68ece2b5d6eced8cf52ee91fb36f

5 years agoWindows local build: restore original working dir after activating VC environment...
Will Feng [Tue, 27 Nov 2018 22:13:48 +0000 (14:13 -0800)]
Windows local build: restore original working dir after activating VC environment (#14416)

Summary:
`call "C:\\Program Files (x86)\\Microsoft Visual Studio\\2017\\Community\\VC\\Auxiliary\\Build\\vcvarsall.bat" x64` seems to change the working dir to `C:\Users\Administrator\source`, and we need to cd back to the PyTorch directory before running `git submodule update --init --recursive`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14416

Differential Revision: D13222269

Pulled By: yf225

fbshipit-source-id: a0eb3311fb11713b1bb8f52cd13e2c21d5ca9c7b

5 years agocondition blob in while_op test changes data type (#14279)
Jerry Zhang [Tue, 27 Nov 2018 22:10:41 +0000 (14:10 -0800)]
condition blob in while_op test changes data type (#14279)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14279

att

Reviewed By: smessmer

Differential Revision: D13144472

fbshipit-source-id: af4d920a3148c648d1a428a5bcd56da19ea8c38c

5 years agoAdd test of ONNX_ATEN (#14259)
zrphercule [Tue, 27 Nov 2018 21:49:21 +0000 (13:49 -0800)]
Add test of ONNX_ATEN (#14259)

Summary:
In #14239 we fixed ONNX_ATEN.
In order to make sure its correctness in the future, we should add related test case.
We use torch.fmod() to test ONNX_ATEN.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14259

Differential Revision: D13204610

Pulled By: zrphercule

fbshipit-source-id: e4660c346e5edd201f1458b7d74d7dfac49b94c7

5 years agoAllowing TaskGroups to carry remote nets (#14342)
Hassan Eslami [Tue, 27 Nov 2018 21:31:59 +0000 (13:31 -0800)]
Allowing TaskGroups to carry remote nets (#14342)

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

5 years agoAdd scaffolding for HIP backend in ATen/core. (#14285)
Edward Yang [Tue, 27 Nov 2018 21:14:12 +0000 (13:14 -0800)]
Add scaffolding for HIP backend in ATen/core. (#14285)

Summary:
This code doesn't actually do anything, but it will be the
groundwork necessary to change PyTorch's HIPIFY pass from reusing
CUDA identifiers directly, to actually switching to using HIP
identifiers (moving us closer to a world where we can compile
both HIP and CUDA PyTorch side-by-side.)

Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14285

Differential Revision: D13158851

Pulled By: ezyang

fbshipit-source-id: df2462daa5d0d4112455b67bd3067d60ba55cda5

5 years agoDocument device_guard in native_functions.yaml (#14235)
Edward Yang [Tue, 27 Nov 2018 21:12:25 +0000 (13:12 -0800)]
Document device_guard in native_functions.yaml (#14235)

Summary:
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14235

Differential Revision: D13145780

Pulled By: ezyang

fbshipit-source-id: 0e93bf009ad492551bcdcada0357f2fef529e67d

5 years agoRevert D13192228: [pytorch][PR] [jit] Add boolean dispatch for function overloading
David Riazati [Tue, 27 Nov 2018 21:12:14 +0000 (13:12 -0800)]
Revert D13192228: [pytorch][PR] [jit] Add boolean dispatch for function overloading

Differential Revision:
D13192228

Original commit changeset: fce33c400c1f

fbshipit-source-id: 75c9991dc7097f9513c6c89d16eff2de6e287c3b

5 years agoRemove fake dependencies from TensorImpl to caffe2 (#14141)
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Remove fake dependencies from TensorImpl to caffe2 (#14141)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14141

These includes weren't actually used, let's remove them.

Reviewed By: ezyang

Differential Revision: D13113129

fbshipit-source-id: 816995e280b81bf99002772ea8aea458bdfcd2c7

5 years agoFix include paths for TensorTypeId.h and TensorTypeIdRegistration.h
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Fix include paths for TensorTypeId.h and TensorTypeIdRegistration.h

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14070

Reviewed By: ezyang

Differential Revision: D13081610

fbshipit-source-id: 685994a15a2cd15e9e5447cf77671343de5dd278

5 years agoMove TensorTypeId to c10/core
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Move TensorTypeId to c10/core

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14327

Reviewed By: ezyang

Differential Revision: D13131338

fbshipit-source-id: c4682cb6ed6fe4cd1636e09d918eef6e90c836f1

5 years agoFix include paths for Storage.h and StorageImpl.h
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Fix include paths for Storage.h and StorageImpl.h

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14062

Reviewed By: ezyang

Differential Revision: D13081603

fbshipit-source-id: c272b715ef2f513d21d1c3f34fbf79eec6946441

5 years agoMove Storage and StorageImpl to c10
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Move Storage and StorageImpl to c10

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14061

Reviewed By: ezyang

Differential Revision: D13081608

fbshipit-source-id: 1ea2d32e9ec9293b6ffa4b9e76c674cca55d5a1c

5 years agoFix include paths for Allocator.h
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Fix include paths for Allocator.h

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14060

Reviewed By: ezyang

Differential Revision: D13081605

fbshipit-source-id: 02f23af174c0f0c38fb0163c2dfef3873ff5635d

5 years agoMove Allocator.h to c10
Sebastian Messmer [Tue, 27 Nov 2018 20:43:24 +0000 (12:43 -0800)]
Move Allocator.h to c10

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14059

Reviewed By: ezyang

Differential Revision: D13081606

fbshipit-source-id: d6ad59ad4e3d363268cd4307b6c999a168681246

5 years agoMove UniqueVoidPtr to c10
Sebastian Messmer [Tue, 27 Nov 2018 20:43:22 +0000 (12:43 -0800)]
Move UniqueVoidPtr to c10

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14058

Reviewed By: dzhulgakov

Differential Revision: D13081602

fbshipit-source-id: e91ccf9fba9a7a02f99ed90b7a3a0fe7afd56832

5 years agoMove ScalarTypeUtils.h to c10
Sebastian Messmer [Tue, 27 Nov 2018 20:43:22 +0000 (12:43 -0800)]
Move ScalarTypeUtils.h to c10

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14024

Reviewed By: ezyang

Differential Revision: D13081604

fbshipit-source-id: d7a09610f64eb2e9dd831bbb3c85f20691251594

5 years agoFix include paths for Scalar.h and ScalarType.h
Sebastian Messmer [Tue, 27 Nov 2018 20:43:22 +0000 (12:43 -0800)]
Fix include paths for Scalar.h and ScalarType.h

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14023

Reviewed By: ezyang

Differential Revision: D13081609

fbshipit-source-id: c27eeafa381b39e043f0261ea7f6f634ee8bc238

5 years agoMove Scalar and ScalarType to c10/core
Sebastian Messmer [Tue, 27 Nov 2018 20:43:22 +0000 (12:43 -0800)]
Move Scalar and ScalarType to c10/core

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14022

Reviewed By: ezyang

Differential Revision: D13015236

fbshipit-source-id: 92aac4e342d85f75a31837b2943fa5b80f0c35c9

5 years agoTrace in-place ops (#14254)
Michael Suo [Tue, 27 Nov 2018 20:38:28 +0000 (12:38 -0800)]
Trace in-place ops (#14254)

Summary:
This PR adds a `try_outplace` option to the tracer. When `try_outplace` is true, the tracer will attempt to out-of-place ops (similar to how things are done today). When it's false, the correct in-place op is emitted.

I made `try_outplace` false by default, but flipped it to true for ONNX export utils. zdevito jamesr66a, anywhere else I should preserve the existing behavior?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14254

Reviewed By: eellison

Differential Revision: D13166691

Pulled By: suo

fbshipit-source-id: ce39fdf73ac39811c55100e567466d53108e856b

5 years agoFixed torch.multiprocessing.spawn for not being able to spawn like dataloader workers...
Teng Li [Tue, 27 Nov 2018 20:32:56 +0000 (12:32 -0800)]
Fixed torch.multiprocessing.spawn for not being able to spawn like dataloader workers (#14391)

Summary:
Should fix: https://github.com/pytorch/pytorch/issues/14390

Now imagenet example works fine with multiprocessing and more than 1 dataloader worker
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14391

Reviewed By: calebho

Differential Revision: D13209800

Pulled By: teng-li

fbshipit-source-id: e8abc0fb38d4436cf3474dcbba0e28f4290e4d29

5 years agoTensor construction: combine Resize+mutable_data - 4/4 (#13856)
Jerry Zhang [Tue, 27 Nov 2018 20:31:17 +0000 (12:31 -0800)]
Tensor construction: combine Resize+mutable_data - 4/4 (#13856)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13856

Codemod generated with clangr shard mode, 25 files per diff,
motivation: https://github.com/pytorch/pytorch/pull/12407

Reviewed By: smessmer

Differential Revision: D13007310

fbshipit-source-id: 941f064ef8934bb17fbfb706e6ed3db173b5d268

5 years agoPrint default values and introduce ir view classes (#14176)
Zachary DeVito [Tue, 27 Nov 2018 19:46:17 +0000 (11:46 -0800)]
Print default values and introduce ir view classes (#14176)

Summary:
[Stacked commit, only review the last commit]

This PR adds support for printing default values in python printing as well as the logic
for parsing default values back in using the parser. For simplicity, this PR simply
creates a subgraph of the constant expressions and then runs that graph to generate the defaults.
A more lightweight approach should be possible later, but would require more machinery.

To make reading code in the printer easier, this also add ir_views.h.
Similar to tree_views.h these classes can provide views of some commonly used IR nodes
that have complicated structure and common operations on that structure.

Currently it has only read-only views for prim::If and prim::Loop,
but we should eventually add helpers to manipulate If/Loop nodes as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14176

Differential Revision: D13198455

Pulled By: zdevito

fbshipit-source-id: dc99ab9692804ccaedb60a55040c0b89ac7a6a6d

5 years agoAdd Type support to the fuser, fuse more (#14336)
Thomas Viehmann [Tue, 27 Nov 2018 19:30:41 +0000 (11:30 -0800)]
Add Type support to the fuser, fuse more (#14336)

Summary:
This adds scalar type support to the fuser, both internally (instead of auto / assuming float) and for the inputs/outputs.
We can now fuse things with input / output of arbitrary scalar type, in particular comparisons and where work well. So it fixes #13384 by returning the right type tensor (and adds a test where byte and double tensors are returned).
The type inference is done by re-calling PropagateTensorShapeOnNode in the compilation, I would venture that it isn't prohibitively expensive compared to the actual compilation. (Propagation was fixed for where to return the second argument's type and amended to handle FusedConcat.)
I'm not sure how to add a check for the code generated by the fuser, but I am not sure we absolutely need to (we'd see if it is invalid / produces wrong results).

Thanks in particular to apaszke, fmassa, mruberry for advice and encouragement! All the errors are my own.

I have discussed order of PRs briefly with mruberry, if this goes in before he submits the PR, he graciously agreed to rebasing his, but I'd happily rebase, too.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14336

Differential Revision: D13202620

Pulled By: soumith

fbshipit-source-id: 855159e261fa15f21aca3053bfc05fb3f720a8ef

5 years agoUpdating submodules
svcscm [Tue, 27 Nov 2018 19:20:46 +0000 (11:20 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: e63160e97550942931bacaa860d91d591d2e1712

5 years agoAdd boolean dispatch for function overloading (#14081)
David Riazati [Tue, 27 Nov 2018 18:49:14 +0000 (10:49 -0800)]
Add boolean dispatch for function overloading (#14081)

Summary:
This PR allows to overload functions based on the value of a parameter (so long as it is a constant). See `max_pool1d` for an example usage.

This is the first step in enabling the use of `max_pool` functions for the standard library that can return `Tensor` or `Tuple[Tensor, Tensor]` based on the `return_indices` flag. This will give the JIT identical results to the Python versions of the functions.

Depends on #14232 for `Optional[BroadcastingList[T]]`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14081

Differential Revision: D13192228

Pulled By: driazati

fbshipit-source-id: fce33c400c1fd06e59747d98507c5fdcd8d4c113

5 years agoBarrier synchronizes with prior work before completing (#14386)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:06 +0000 (10:41 -0800)]
Barrier synchronizes with prior work before completing (#14386)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14386

See #13573, #14142, and #14271 for discussion.

This change updates ProcessGroupGloo to ensure that all prior
operations have completed before executing the barrier.

Reviewed By: manojkris

Differential Revision: D13205022

fbshipit-source-id: 673e7e6ca357dc843874d6dd8da590832e1de7fa

5 years agoMake ProcessGroup::Work::wait() throw (#14298)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:06 +0000 (10:41 -0800)]
Make ProcessGroup::Work::wait() throw (#14298)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14298

This is a breaking API change for users of the C++ c10d API. The work
object defined wait() to return a boolean. If the work completed
successfully it would return true, if it didn't it would return false.
It was then up to the user to call the exception() function to figure
out what went wrong. This has proven suboptimal as it allows users to
forget about failure handling and errors may be ignored.

The work class is semantically very similar to std::future, where a
call to get() may throw if the underlying std::promise has set an
exception. This commit changes the semantic of the work class to be
similar to this and turns wait() into a void function that throws if
the work completes with an exception.

The exception() function can still be used to retrieve the exception
if isSuccess() returns false, but now returns an std::exception_ptr
instead of a reference to a std::exception.

Reviewed By: manojkris

Differential Revision: D13158475

fbshipit-source-id: 9cd8569b9e7cbddc867a5f34c6fd0b7be85581b8

5 years agoAdd option structs and timeout field (#14297)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:04 +0000 (10:41 -0800)]
Add option structs and timeout field (#14297)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14297

Adds option structs for allgather and barrier such that we have one
for every collective. Add timeout member field to every one of these
such that we can support per operation timeouts.

Use default constructed options struct for every collective process
group function exposed to Python.

Reviewed By: manojkris

Differential Revision: D13158474

fbshipit-source-id: 3d28977de2f2bd6fc2f42ba3108b63a429338906

5 years agoRefer to all work with ProcessGroup prefix (#14296)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:04 +0000 (10:41 -0800)]
Refer to all work with ProcessGroup prefix (#14296)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14296

There was mixed usage of "ProcessGroup::Work" and just "Work".
Adding prefix for readability/consistency.

Reviewed By: manojkris

Differential Revision: D13128977

fbshipit-source-id: a54a8784fa91cd6023c723cb83e9f626fb896a30

5 years agoRemove algorithm caching in ProcessGroupGloo (#14295)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:04 +0000 (10:41 -0800)]
Remove algorithm caching in ProcessGroupGloo (#14295)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14295

This is no longer used after moving to Gloo new style algorithms.

Closes #11912.

Reviewed By: manojkris

Differential Revision: D13111781

fbshipit-source-id: 53e347080e29d847cd9da36f2d93af047930690c

5 years agoUse new style barrier support in c10d/gloo (#14294)
Pieter Noordhuis [Tue, 27 Nov 2018 18:41:04 +0000 (10:41 -0800)]
Use new style barrier support in c10d/gloo (#14294)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14294

This is the final collective to be ported to the new style where there
is no longer a need to keep a cached algorithm instance around. There
is a follow up change incoming to remove the algorithm caching
functionality in ProcessGroupGloo.

Reviewed By: manojkris

Differential Revision: D13111509

fbshipit-source-id: f3ea0d955a62029fc4e7cfc09055e4957e0943ac

5 years agofix doc for sparse.addmm (#14403)
Wei Yang [Tue, 27 Nov 2018 18:22:24 +0000 (10:22 -0800)]
fix doc for sparse.addmm (#14403)

Summary:
- fixing the doc issue in sparse.addmm

================ before change ==================
![image](https://user-images.githubusercontent.com/38509346/49063994-2f10fe80-f1ce-11e8-9ccc-54241bc45f0b.png)
![image](https://user-images.githubusercontent.com/38509346/49064064-641d5100-f1ce-11e8-865a-7227be7156ef.png)

================ post change ==================
![image](https://user-images.githubusercontent.com/38509346/49064078-76978a80-f1ce-11e8-8f38-f1f8ac9ce63b.png)
![image](https://user-images.githubusercontent.com/38509346/49064085-7bf4d500-f1ce-11e8-8a0d-bf9e5460d21f.png)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14403

Differential Revision: D13216582

Pulled By: weiyangfb

fbshipit-source-id: 52e0a20c6b341c37cfb31f281be3afe2a52ca532

5 years agoper-group and per-channel quantization (#14340)
Jongsoo Park [Tue, 27 Nov 2018 18:05:28 +0000 (10:05 -0800)]
per-group and per-channel quantization (#14340)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14340

Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/25

Per-group and per-channel quantization in fbgemm
This diff also cleans up explicit template instantiation using macro expansion
This diff also changes randFill interface which was easy to make mistakes of generating integer random numbers for floating point vectors.

Using this in DNNLOWP operators will be done in a separate diff.

Reviewed By: dskhudia

Differential Revision: D13176386

fbshipit-source-id: e46c53e31e21520bded71b8ed86e8b19e010e2dd

5 years agoAdd variable_factories.h to cppdocs (#14381)
Peter Goldsborough [Tue, 27 Nov 2018 18:04:57 +0000 (10:04 -0800)]
Add variable_factories.h to cppdocs (#14381)

Summary:
This will document `torch::from_blob` and such.

soumith ezyang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14381

Differential Revision: D13216560

Pulled By: goldsborough

fbshipit-source-id: 112f60e45e4d38a8a9983fa71e9cc56bc1a73465

5 years agoUse integer math to compute output size of pooling operations (#14405)
Jan SchlĂ¼ter [Tue, 27 Nov 2018 17:36:11 +0000 (09:36 -0800)]
Use integer math to compute output size of pooling operations (#14405)

Summary:
As reported in #13386, the pooling operations can return wrong results for large inputs. The root of the problem is that while the output shape is initially being computed with integer operations, it is converted to float32 for division by the stride and applying either a `ceil` or a `floor` depending on the `ceil_mode`. Since even moderately large integers (the smallest being 16,777,217) cannot be expressed exactly in float32, this leads to wrong result shapes.

This PR relies purely on integer operations to perform the shape computation, including the ceil/floor distinction. Since I could not stand all that duplicated code, I pulled it out into a `pooling_shape.h` header, similar to the existing `linear_upsampling.h` header. I hope this is acceptable, let me know if you'd like to see it solved differently. I've also added tests to `test_nn.py` that fail without my changes and pass with my changes. They cover `{max,avg}_pool{1,2,3}d()` for CPU and GPU.

Fixes #13386.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14405

Differential Revision: D13215260

Pulled By: soumith

fbshipit-source-id: 802588ce6cba8db6c346448c3b3c0dac14d12b2d

5 years agoDelete legacy THCStream (long live THCStream). (#14246)
Edward Yang [Tue, 27 Nov 2018 16:23:34 +0000 (08:23 -0800)]
Delete legacy THCStream (long live THCStream). (#14246)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14246

This commit systematically eliminates THCStream entirely from THC, replacing it
with at::cuda::CUDAStream.  In places where the previous pointer type showed up
in a public API signature, those functions are now only available to C++
clients.  (It would not be too difficult to make a C-compatible version of
CUDAStream, as it's really just a simple struct, but we leave this for
future work.)

All functions in THC that referred to THCStream were expunged in favor of their
modern counterparts.

One annoyance was that I didn't feel like redoing how the torch.cuda.Stream
binding code worked, but I really wanted to get rid of the stored THCStream*
pointer.  So I repurposed the bit-packing code I implemented for Stream hashing,
and used that to (reversibly) store streams in a uint64_t cdata field.  A perhaps
more future proof solution would be to get rid of cdata entirely, and store the
device and stream ID directly.

Billing of changes:
- All CUDAStream_ pointer API functions are now hidden and anonymously
  namespaced (instead of being in the impl namespace).  All use sites
  rewritten to use the modern C++ API.  Since CUDAStreamInternals is no
  longer part of the public API, the CUDAStreamInternals constructor and
  internals() method have been removed, and replaced with anonymous
  functions in the C++ file.
- device_index() returns DeviceIndex rather than int64_t now
- Stream and CUDAStream now have pack/unpack methods.  (CUDAStream checks
  that the unpacked bit-pattern is for a CUDA device.)
- THCStream.h header is removed entirely
- Most THCStream handling functions in THC API are removed

Reviewed By: gchanan

Differential Revision: D13121531

fbshipit-source-id: 48873262cc0a37c3eec75a7ba1c93c800da40222

5 years agoAdd hash functions for Stream, CUDAStream; fix Device hash function (#14191)
Edward Yang [Tue, 27 Nov 2018 16:23:34 +0000 (08:23 -0800)]
Add hash functions for Stream, CUDAStream; fix Device hash function (#14191)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14191

Previously, Device's hash function only worked for CPU and CUDA.  Now
it works for everything.

Implementing the bit concatenation was a bit tricky, and I got it wrong the
first time. See Note [Hazard when concatenating signed integers]

Reviewed By: smessmer

Differential Revision: D13119624

fbshipit-source-id: 36bfa139cfc739bb0624f52aaf466438c2428207

5 years agoImplement NaN-propagating max/min on Vec256.
Owen Anderson [Tue, 27 Nov 2018 06:41:56 +0000 (22:41 -0800)]
Implement NaN-propagating max/min on Vec256.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/13399

Differential Revision: D13199957

Pulled By: resistor

fbshipit-source-id: 1565e079b13c5d4f42f2033830a7c997b7d824bc

5 years agoUpdating submodules
svcscm [Tue, 27 Nov 2018 03:35:44 +0000 (19:35 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: 210f7eec65bea5e31817fb56dec27b0ab8af797a

5 years agoRemove unused executors, part 3 (#14199)
Ilia Cherniavskii [Tue, 27 Nov 2018 03:07:07 +0000 (19:07 -0800)]
Remove unused executors, part 3 (#14199)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14199

Remove legacy code for dag, async_dag

Reviewed By: salexspb

Differential Revision: D13019102

fbshipit-source-id: ff07e45304d9af4be0375215f4b642c4b0edb12d

5 years agoRemove unused executors, part 2 (#14115)
Ilia Cherniavskii [Tue, 27 Nov 2018 03:07:07 +0000 (19:07 -0800)]
Remove unused executors, part 2 (#14115)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14115

Remove legacy implementation of prof_dag

Reviewed By: salexspb

Differential Revision: D13019096

fbshipit-source-id: 4f2bf676444d84eaa2cc1effcc3ebdc764e0a016

5 years agoRemove unused executors, part 1 (#14117)
Ilia Cherniavskii [Tue, 27 Nov 2018 03:07:06 +0000 (19:07 -0800)]
Remove unused executors, part 1 (#14117)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14117

Removing unused legacy executors (htrace)

Reviewed By: salexspb

Differential Revision: D13019078

fbshipit-source-id: 19d0ed1b47a22cc17c27fdd15d748ced54806132

5 years agoDelete OPENMP_STUB translation. (#14286)
Edward Yang [Tue, 27 Nov 2018 03:06:06 +0000 (19:06 -0800)]
Delete OPENMP_STUB translation. (#14286)

Summary:
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14286

Differential Revision: D13205356

Pulled By: ezyang

fbshipit-source-id: 08e9821e4b32f8d7f3c41906e481f280ee6cf2e3

5 years agobackward for sparse.addmm(D, S, D, alpha, beta) -> D (#13345)
Wei Yang [Tue, 27 Nov 2018 01:43:21 +0000 (17:43 -0800)]
backward for sparse.addmm(D, S, D, alpha, beta) -> D (#13345)

Summary:
- introduce `sparse.addmm()` with backward for sparse matrix input for https://github.com/pytorch/pytorch/issues/12308
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13345

Differential Revision: D13094070

Pulled By: weiyangfb

fbshipit-source-id: 136c08c3ca9bafb20577b60dd43d31c3e5cd5461

5 years agoSwitch Int8ChannelShuffle operator to QNNPACK (#14362)
Marat Dukhan [Tue, 27 Nov 2018 01:41:13 +0000 (17:41 -0800)]
Switch Int8ChannelShuffle operator to QNNPACK (#14362)

Summary:
1.8-2.2X better performance on ARM devices
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14362

Reviewed By: jerryzh168

Differential Revision: D13192312

Pulled By: Maratyszcza

fbshipit-source-id: 0d3dff067e300c7d741c42615b61246cbf09a829

5 years agoFixed file init_method write/read race (#14388)
Teng Li [Tue, 27 Nov 2018 01:05:17 +0000 (17:05 -0800)]
Fixed file init_method write/read race (#14388)

Summary:
This should fix the race among multiple processes: https://github.com/pytorch/pytorch/issues/13750

Essentially, the reader is trying to open the file, and will error out if it doesn't exist, we here factor in the timeout option of FileStore to apply a timeout for creating a file (should always be created anyway unless something is wrong), and more importantly, waiting for the file to be created.

Tested on both NFS and local drive, the race disappears when 8 concurrent processes do distributed training.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14388

Differential Revision: D13207178

Pulled By: teng-li

fbshipit-source-id: d3d5d62c4c8f01c0522bf1653c8986155c54ff80

5 years agoFix dataloader iterator test (#14045)
Peter Goldsborough [Tue, 27 Nov 2018 01:04:51 +0000 (17:04 -0800)]
Fix dataloader iterator test (#14045)

Summary:
I noticed the test `DataLoaderTest.CanDereferenceIteratorMultipleTimes` doesn't test proper progression of the iterator. I also added a test for using `std::copy`.

Fixes https://github.com/pytorch/pytorch/issues/14276

ebetica ezyang apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14045

Differential Revision: D13092187

Pulled By: goldsborough

fbshipit-source-id: 57698ec00fa7b914b159677a4ab38b6b25c2860b

5 years agoFixed c10d test (#14389)
Teng Li [Tue, 27 Nov 2018 00:44:11 +0000 (16:44 -0800)]
Fixed c10d test (#14389)

Summary:
Most likely a typo.

Tested on 8-GPU machine

```
tengli@learnfair062:~/pytorch/test$ python test_c10d.py ProcessGroupNCCLTest.test_barrier
.
----------------------------------------------------------------------
Ran 1 test in 29.341s

OK
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14389

Differential Revision: D13207207

Pulled By: teng-li

fbshipit-source-id: aaffe14237076fe19d94e2fa4d9c093397f07bb9

5 years agofix typo in `torch.sum` documentation (#14250)
Brennan Vincent [Tue, 27 Nov 2018 00:34:47 +0000 (16:34 -0800)]
fix typo in `torch.sum` documentation (#14250)

Summary:
Notice that an extra colon was added to `:attr:`, so in https://pytorch.org/docs/stable/torch.html#torch.sum , `dim` shows up as ":attr::_dim_". This patch fixes the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14250

Reviewed By: soumith

Differential Revision: D13146363

Pulled By: umanwizard

fbshipit-source-id: f7d03dcb0973aae248b56ab407ba8489f2b1fe36

5 years agoMore JIT type hierarchy refinement (#14127)
Wanchao Liang [Tue, 27 Nov 2018 00:21:08 +0000 (16:21 -0800)]
More JIT type hierarchy refinement (#14127)

Summary:
JIT type system hierarchy refinement and refactors:

1. Make NumberType be the base type of IntType FloatType
2. Make single type container like OptionalType and FutureType share SingleElementType base type
3. Some refactors to make it more robust, e.g. adding python_str() for some types so that we have proper python_print serialization format
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14127

Differential Revision: D13112657

Pulled By: wanchaol

fbshipit-source-id: 335c5b25977be2e0a462c7e4a6649c1b653ccb4f

5 years agochanging some rpath stuff (#14304)
Jesse Hellemn [Mon, 26 Nov 2018 23:55:40 +0000 (15:55 -0800)]
changing some rpath stuff (#14304)

Summary:
See if anything breaks
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14304

Differential Revision: D13201418

Pulled By: pjh5

fbshipit-source-id: ac2101b61a23bda37329d4d923c3d9d120e718bf

5 years agoFix caffe2 => onnx exporter for ConvTranspose (#14143)
Kevin Chen [Mon, 26 Nov 2018 23:49:36 +0000 (15:49 -0800)]
Fix caffe2 => onnx exporter for ConvTranspose (#14143)

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

5 years agoRevert D13166669: [pytorch][PR] Allow dataloader to accept a custom memory pinning...
Will Feng [Mon, 26 Nov 2018 22:51:57 +0000 (14:51 -0800)]
Revert D13166669: [pytorch][PR] Allow dataloader to accept a custom memory pinning function

Differential Revision:
D13166669

Original commit changeset: ca965f9841d4

fbshipit-source-id: 0836b4f50f73ba01c97491a719660f02e36f20ad

5 years agoremove CAFFE2_API from IdWrapper (#14044)
andersj [Mon, 26 Nov 2018 22:05:05 +0000 (14:05 -0800)]
remove CAFFE2_API from IdWrapper (#14044)

Summary:
it doesn't really make sense on a template class. Also it breaks if
you try to build in debug on Windows, so this will save someone some
frustration in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14044

Differential Revision: D13202960

Pulled By: anderspapitto

fbshipit-source-id: 617d78366993d5ecc2ba1f23bb90010f10df41f3

5 years agoFeedTensor returns a Tensor (#14196)
Jerry Zhang [Mon, 26 Nov 2018 21:03:40 +0000 (13:03 -0800)]
FeedTensor returns a Tensor (#14196)

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

5 years agoAllow graph fuser to move chunks past multiple nodes. (#14055)
Richard Zou [Mon, 26 Nov 2018 20:28:44 +0000 (12:28 -0800)]
Allow graph fuser to move chunks past multiple nodes. (#14055)

Summary:
Fixes #12290. Also speeds up JIT LSTM forward pass from 8.8ms to 7.8ms; previously, each JIT lstm cell used 2 fused kernels. Now, it only uses one fused kernel (which is how many kernels cudnn uses).

Explanation:

Let f, g, h be fusible ops.
```
x = f(v, w)
z = g(x, y)
a, b = chunk(z)
c = h(a, b)
```
becomes (before this PR):
```
x = f(v, w)
x', y' = broadcast_tensors([x, y])
ax, bx = chunk(x')
ay, by = chunk(y')
a = g(ax, ay)
b = g(bx, by)
c = h(a, b)
```
The graph fuser then puts g, g, and h into one FusionGroup and is unable
to move `x = f(v, w)` into the FusionGroup.

This PR lets the graph fuser move `x = f(v, w)` into the FusionGroup.
It does this by abstracting the broadcast_tensors + multiple chunk nodes
into one intermediate `prim::BroadcastingChunk[chunks, dim]` node.

A `BroadcastingChunk[chunks, dim](*inputs)` node is equivalent to:
- broadcasting all of *inputs
- chunk-ing each broadcasted input into `chunks` chunks along dim `dim`.

Abstracting the broadcasting chunk behavior away, it is now a lot easier
for the graph fuser to move (broadcast + chunk) past an operation. After
this PR, the above graph becomes:
```
x = f(v, w)
ax, bx, ay, by = BroadcastingChunk(x, y)
a = g(ax, ay)
b = g(bx, by)
c = h(a, b)
```
Now, to move `x = f(v, w)` after the BroadcastingChunk, one just needs
to add f's operands to the BroadcastingChunk:
```
ay, by, av, bv, aw, bw = BroadcastingChunk(y, v, w)
ax = f(av, aw)
by = f(bv, bw)
a = g(ax, ay)
b = g(bx, by)
c = h(a, b)
```

cc apaszke mruberry zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14055

Differential Revision: D13159259

Pulled By: zou3519

fbshipit-source-id: 134e9e645c950384d9be6a06a883a10e17a73d7d

5 years agoUpdating submodules
svcscm [Mon, 26 Nov 2018 20:10:45 +0000 (12:10 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: b4d74bf58b5536a0de654dfe73d41b5e1126eec6

5 years agoRemoving Caffe2-specific conda infra
Jesse Hellemn [Mon, 26 Nov 2018 20:08:42 +0000 (12:08 -0800)]
Removing Caffe2-specific conda infra

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/11961

Differential Revision: D10045909

Pulled By: pjh5

fbshipit-source-id: e9c12124897ee586aeb8b6654b31e4b81687199a

5 years agofix tensor advanced indexing with assignment (#14311)
Michael Suo [Mon, 26 Nov 2018 20:02:09 +0000 (12:02 -0800)]
fix tensor advanced indexing with assignment (#14311)

Summary:
Fix a mishandling of `foo[a] = b` when `a` was a tensor. We were assigning to a copy of `foo`, not a view of it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14311

Differential Revision: D13196109

Pulled By: suo

fbshipit-source-id: c929401fda7c4a27622d3fe2b11278b08a7f17f1

5 years agoremove unnecessary zero_point argument from constructors (#14323)
Jongsoo Park [Mon, 26 Nov 2018 19:45:55 +0000 (11:45 -0800)]
remove unnecessary zero_point argument from constructors (#14323)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14323

Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/24

As title says.

Reviewed By: dskhudia

Differential Revision: D13167073

fbshipit-source-id: 6d6c526fd6e29a14e97f71a0881f28ada8703107

5 years agoUpdating submodules
svcscm [Mon, 26 Nov 2018 19:24:40 +0000 (11:24 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: 06e234f1a0217a268712832f21cb06b7109538a6

5 years agoFix -Wreturn-std-move (#14113)
Peter Goldsborough [Mon, 26 Nov 2018 19:13:52 +0000 (11:13 -0800)]
Fix -Wreturn-std-move (#14113)

Summary:
On clang-7 (internal) a warning, `-Wreturn-std-move`, is being emitted and raised to an error via `-Werror` for the code this PR fixes. The reason is that `autograd::make_variable` returns an `autograd::Variable`, so returning it from a function that returns `at::Tensor` disallows the compiler from eliding the return value (RVO). So let's explicitly convert the `autograd::Variable` to an `at::Tensor` before returning it.

ezyang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14113

Differential Revision: D13105638

Pulled By: goldsborough

fbshipit-source-id: 6e1dc31c6512e105ab2a389d18807422ee29283c

5 years agominimize code compiled with avx2 and header includes from them (#14313)
Jongsoo Park [Mon, 26 Nov 2018 19:07:15 +0000 (11:07 -0800)]
minimize code compiled with avx2 and header includes from them (#14313)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14313

Pull Request resolved: https://github.com/pytorch/FBGEMM/pull/22

This diff is an attempt to minimize code compiled with avx2.

Reviewed By: dskhudia

Differential Revision: D13166591

fbshipit-source-id: 2be241141f6d7478b86a422953791e237ff10268

5 years agoAdd proper from_blob overloads (#13982)
Peter Goldsborough [Mon, 26 Nov 2018 18:12:50 +0000 (10:12 -0800)]
Add proper from_blob overloads (#13982)

Summary:
There was an overload for `torch::from_blob` missing that allowed passing strides.

ezyang soumith
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13982

Differential Revision: D13108089

Pulled By: goldsborough

fbshipit-source-id: b87594ec0bf55b35d106b4438bc18b2ce9fc8f71

5 years agoallow concatenating "hybrid" (sparse/dense) tensors along their dense dimensions...
Brennan Vincent [Mon, 26 Nov 2018 18:03:24 +0000 (10:03 -0800)]
allow concatenating "hybrid" (sparse/dense) tensors along their dense dimensions (#13761)

Summary:
Follow-up to #13577

The idea is to take each values tensor, concatenate it with zeros before and after itself (along the dimension corresponding to the one we're catting the tensors along), to get a tensor corresponding to the values for that tensor in the result. Then we concatenate all of those together to get the final values tensor. (Hopefully, this will be more clear from the example in the comments).

The indices are more straightforward: since we aren't concatenating along a sparse dimension, they don't change at all, so all we need to do are concatenate the indices from the different tensors together.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13761

Differential Revision: D13160343

Pulled By: umanwizard

fbshipit-source-id: 13d7adecd369e0eebdf5bce3d90a51029b66bd1d

5 years agoAllow torch.utils.cpp_extension.load to load shared libraries that aren't Python...
Peter Goldsborough [Mon, 26 Nov 2018 17:37:04 +0000 (09:37 -0800)]
Allow torch.utils.cpp_extension.load to load shared libraries that aren't Python modules (#13941)

Summary:
For custom TorchScript operators, `torch.ops.load_library` must be used and passed the path to the shared library containing the custom ops. Our C++ extensions stuff generally is meant to build a Python module and import it. This PR changes `torch.utils.cpp_extension.load` to have an option to just return the shared library path instead of importing it as a Python module, so you can then pass it to `torch.ops.load_library`. This means folks can re-use `torch.utils.cpp_extension.load` and `torch.utils.cpp_extension.load_inline` to even write their custom ops inline. I think t-vi  and fmassa will appreciate this.

soumith
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13941

Differential Revision: D13110592

Pulled By: goldsborough

fbshipit-source-id: 37756307dbf80a81d2ed550e67c8743dca01dc20

5 years agoBatch more matrix multiplies (#13456)
Adam Paszke [Mon, 26 Nov 2018 17:18:43 +0000 (09:18 -0800)]
Batch more matrix multiplies (#13456)

Summary:
This handles the input pre-multiplication in RNNs, yielding pretty significant speedups in backward times. This pass depends on loop unrolling, so we'll batch only as many elements as the unrolling factor allows.

cc mruberry ngimel zou3519 zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13456

Differential Revision: D12920339

Pulled By: zou3519

fbshipit-source-id: 5bcd6d259c054a6dea02ae09a9fdf9f030856443

5 years agoEnable native wrappers for the remainder of nn functions.
Gregory Chanan [Mon, 26 Nov 2018 15:56:43 +0000 (07:56 -0800)]
Enable native wrappers for the remainder of nn functions.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14290

Differential Revision: D13162562

Pulled By: gchanan

fbshipit-source-id: 615e1727988bfeeade48f9b38162333a2e298f7b

5 years agoAdd Recency Weighted into SparseLookup (#14291)
Huan Gui [Sat, 24 Nov 2018 10:41:25 +0000 (02:41 -0800)]
Add Recency Weighted into SparseLookup (#14291)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14291

Add RecencyWeighted into SparseLookup.

Reviewed By: Wakeupbuddy

Differential Revision: D13147738

fbshipit-source-id: de5dc3aaee8ce7d41c6d30d2ff47e9786a7fa4da

5 years agoquote NUMPY_INCLUDE_DIR (#14341)
Shuichi KITAGUCHI [Sat, 24 Nov 2018 05:32:10 +0000 (21:32 -0800)]
quote NUMPY_INCLUDE_DIR (#14341)

Summary:
when NUMPY_INCLUDE_DIR contains space character (e.g. "C:\Program Files (x86)\Microsoft Visual Studio\..."), cmake cannot receive correct path name.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14341

Differential Revision: D13188408

Pulled By: soumith

fbshipit-source-id: b62127d90e53da94fe6af5d3bdd2ea4fd6546210

5 years agoshape analysis fix (#14325)
Michael Suo [Fri, 23 Nov 2018 19:22:22 +0000 (11:22 -0800)]
shape analysis fix (#14325)

Summary:
This PR is deceptively large because of an indenting change. The actual change is small; I will highlight it inline
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14325

Differential Revision: D13183296

Pulled By: suo

fbshipit-source-id: fcbf6d5317954694ec83e6b8cc1c989f2d8ac298

5 years agoSome minor fixes for Windows build script (#14218)
peter [Fri, 23 Nov 2018 16:15:28 +0000 (08:15 -0800)]
Some minor fixes for Windows build script (#14218)

Summary:
1. Fix execution failure when some of the paths are not defined
2. Users can now optionally override install dir by setting `CMAKE_INSTALL_PREFIX`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14218

Differential Revision: D13180350

Pulled By: soumith

fbshipit-source-id: 8c9680d1285dbf08b49380af1ebfa43ede99babc

5 years agoAllow dataloader to accept a custom memory pinning function (#14171)
Michael Carilli [Fri, 23 Nov 2018 16:08:35 +0000 (08:08 -0800)]
Allow dataloader to accept a custom memory pinning function (#14171)

Summary:
Currently, the `pin_memory_batch` function in the dataloader will return a batch comprised of any unrecognized type without pinning the data, because it doesn't know how.

This behavior was preventing us from overlapping data prefetching in Mask-RCNN, whose custom `collate_fn` returns a custom batch type.

The present PR adds the ability for the user to pass a `pin_fn` alongside any custom `collate_fn` to handle such custom types.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14171

Differential Revision: D13166669

Pulled By: soumith

fbshipit-source-id: ca965f9841d4a259b3ca4413c8bd0d8743d433ab

5 years agoOption to preserve bitwise accuracy of gradient checkpointed vs non-checkpointed...
Michael Carilli [Fri, 23 Nov 2018 16:07:51 +0000 (08:07 -0800)]
Option to preserve bitwise accuracy of gradient checkpointed vs non-checkpointed dropout (#14253)

Summary:
This issue was noticed, and fix proposed, by raulpuric.

Checkpointing is implemented by rerunning a forward-pass segment for each checkpointed segment during backward.  This can result in the RNG state advancing more than it would without checkpointing, which can cause checkpoints that include dropout invocations to lose end-to-end bitwise accuracy as compared to non-checkpointed passes.

The present PR contains optional logic to juggle the RNG states such that checkpointed passes containing dropout achieve bitwise accuracy with non-checkpointed equivalents.**  The user requests this behavior by supplying `preserve_rng_state=True` to `torch.utils.checkpoint` or `torch.utils.checkpoint_sequential`.

Currently, `preserve_rng_state=True` may incur a moderate performance hit because restoring MTGP states can be expensive.  However, restoring Philox states is dirt cheap, so syed-ahmed's [RNG refactor](https://github.com/pytorch/pytorch/pull/13070#discussion_r235179882), once merged, will make this option more or less free.

I'm a little wary of the [def checkpoint(function, *args, preserve_rng_state=False):](https://github.com/pytorch/pytorch/pull/14253/files#diff-58da227fc9b1d56752b7dfad90428fe0R75) argument-passing method (specifically, putting a kwarg after a variable argument list).  Python 3 seems happy with it.
Edit:  It appears Python 2.7 is NOT happy with a [kwarg after *args](https://travis-ci.org/pytorch/pytorch/builds/457706518?utm_source=github_status&utm_medium=notification).  `preserve_rng_state` also needs to be communicated in a way that doesn't break any existing usage.  I'm open to suggestions (a global flag perhaps)?

**Batchnorm may still be an issue, but that's a battle for another day.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14253

Differential Revision: D13166665

Pulled By: soumith

fbshipit-source-id: 240cddab57ceaccba038b0276151342344eeecd7

5 years agoUpdating submodules
svcscm [Fri, 23 Nov 2018 05:58:28 +0000 (21:58 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: e92b0c24a56b588dcf30542692cb4bdc2d474825

5 years agoRemove individual "using c10:xxx" statements (#13168)
Sebastian Messmer [Thu, 22 Nov 2018 19:55:07 +0000 (11:55 -0800)]
Remove individual "using c10:xxx" statements (#13168)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13168

We now have a "using namespace c10" in the at and caffe2 namespaces, we don't need the individual ones anymore

Reviewed By: ezyang

Differential Revision: D11669870

fbshipit-source-id: fc2bb1008e533906914188da4b6eb30e7db6acc1

5 years agoMake sure we bind input/output of Onnxifi op positionally (#14214)
Yinghai Lu [Thu, 22 Nov 2018 08:28:51 +0000 (00:28 -0800)]
Make sure we bind input/output of Onnxifi op positionally (#14214)

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

5 years agoConvert gumbel_softmax, lp pooling weak functions and modules (#14232)
Wanchao Liang [Thu, 22 Nov 2018 07:42:24 +0000 (23:42 -0800)]
Convert gumbel_softmax, lp pooling weak functions and modules (#14232)

Summary:
1. Support `Optional[BroadcastingList1[int]]` like type annotation to accept a int or a list[int]
2. Convert gumbel_softmax, lp pooling weak functions and modules
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14232

Differential Revision: D13164506

Pulled By: wanchaol

fbshipit-source-id: 6c2a2b9a0613bfe907dbb5934122656ce2b05700

5 years agoUse ADL to find toString (#14021)
Sebastian Messmer [Thu, 22 Nov 2018 07:04:43 +0000 (23:04 -0800)]
Use ADL to find toString (#14021)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14021

I'm planning to move at::Scalar to c10, and there's a at::toString(Scalar) defined.
Unfortunately, we call it by specifying at::toString() instead of relying on ADL.
This diff changes that to prepare the actual move.

Reviewed By: ezyang

Differential Revision: D13015239

fbshipit-source-id: f2a09f43a96bc5ef20ec2c4c88f7790fd5a04870

5 years agoFix include paths for intrusive_ptr (#13692)
Sebastian Messmer [Thu, 22 Nov 2018 07:04:42 +0000 (23:04 -0800)]
Fix include paths for intrusive_ptr (#13692)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13692

This now lives in c10/util, not ATen/core anymore.

Reviewed By: ezyang

Differential Revision: D12937091

fbshipit-source-id: ea2d420a15e7941a38d0b4c75e20ca18437c73f8

5 years agoMove intrusive_ptr to c10/util
Sebastian Messmer [Thu, 22 Nov 2018 07:04:42 +0000 (23:04 -0800)]
Move intrusive_ptr to c10/util

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/13691

Reviewed By: ezyang

Differential Revision: D12937090

fbshipit-source-id: fe9d21d5f7ea4e78e7e38ac60db13814a9971ed9

5 years agoignore generated caffe2 docs and virtualenvs
Joel Marcey [Thu, 22 Nov 2018 06:28:20 +0000 (22:28 -0800)]
ignore generated caffe2 docs and virtualenvs

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14309

Reviewed By: soumith

Differential Revision: D13166626

Pulled By: JoelMarcey

fbshipit-source-id: 4f11228d8b5da85cec222bf11282722a7319581b

5 years agoUpdating submodules
svcscm [Thu, 22 Nov 2018 05:59:40 +0000 (21:59 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: 20976d595e68a08d746d8806fd0205d810656366

5 years agoremoving quantization utility functions moved to fbgemm (#14301)
Jongsoo Park [Thu, 22 Nov 2018 05:36:16 +0000 (21:36 -0800)]
removing quantization utility functions moved to fbgemm (#14301)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14301

This diff removes quantization utility functions copied to fbgemm

Reviewed By: Maratyszcza

Differential Revision: D13159299

fbshipit-source-id: a7f3cd2af0aa241a8578d532a70a157da70d9289

5 years agoCuda version comparison with CUDA_VERSION_STRING (#14302)
Achal Shah [Thu, 22 Nov 2018 05:00:22 +0000 (21:00 -0800)]
Cuda version comparison with CUDA_VERSION_STRING (#14302)

Summary:
Cuda headers include cuda version in form of major.minor. But when we do find_package(cuda). CUDA_VERSION variable includes patch number as well which fails following condition.

`
if(NOT ${cuda_version_from_header} STREQUAL ${CUDA_VERSION})
`

**For example:**
I have cuda 10.0 installed. My nvcc output looks like this
`Cuda compilation tools, release 10.0, **V10.0.130**
`

If I compile my application with caffe2. It gives me following error:

```
CMake Error at /usr/share/cmake/Caffe2/public/cuda.cmake:59 (message):
  FindCUDA says CUDA version is (usually determined by nvcc), but the CUDA
  headers say the version is 10.0.  This often occurs when you set both
  CUDA_HOME and CUDA_NVCC_EXECUTABLE to non-standard locations, without also
  setting PATH to point to the correct nvcc.  Perhaps, try re-running this
  command again with PATH=/usr/local/cuda/bin:$PATH.  See above log messages
  for more diagnostics, and see
  https://github.com/pytorch/pytorch/issues/8092 for more details.
```

**In this case, it got failed because**
cuda_version_from_header = 10.0
CUDA_VERSION = 10.0.130 (Came from NVCC)

`if(NOT ${cuda_version_from_header} STREQUAL ${CUDA_VERSION})
`

**Fix:**
We should compare header version with **major.minor format** which is given by CUDA_VERSION_STRING
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14302

Differential Revision: D13166485

Pulled By: soumith

fbshipit-source-id: 1b74e756a76c4cc5aa09978f5850f763ed5469b6

5 years agoUpdating submodules
svcscm [Thu, 22 Nov 2018 04:51:26 +0000 (20:51 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: ee60b4dddf688608ef80043b1dc336d120a045d0

5 years agoUpdating submodules
svcscm [Thu, 22 Nov 2018 04:29:22 +0000 (20:29 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: 366c29d09bec53459e2a4890c7fe8d10f45ff5c3

5 years agoRobust NCCL barrier improvement to cover all devices combinations (#14271)
Teng Li [Thu, 22 Nov 2018 02:21:55 +0000 (18:21 -0800)]
Robust NCCL barrier improvement to cover all devices combinations (#14271)

Summary:
This covers the very edgy case when we run the same NCCL process group with multiple GPU combinations instead of the last GPU combination. We always keep track of what GPUs have been used previously in the NCCL process group and barrier() itself will synchronize on each GPU's NCCL stream.

Test covered as well. Tested on 8-GPU machine
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14271

Differential Revision: D13164993

Pulled By: teng-li

fbshipit-source-id: 81e04352740ea50b5e943369e74cfcba40bb61c1

5 years agoalias analysis (#14018)
Michael Suo [Thu, 22 Nov 2018 01:46:46 +0000 (17:46 -0800)]
alias analysis (#14018)

Summary:
First draft of an alias analysis pass. It's a big PR unfortunately; a rough table of contents/suggested order of review:
1. `AliasAnalysis` pass, which traverses the graph and builds an `AliasDb`. The basic strategy is to assign alias information to every value of mutable type (list/tuple/tensor), and use the alias annotations of each node's schema to assign alias info to the outputs based on the alias info the inputs. Nodes that aren't explicitly schematized have hand-written analysis rules.

2. Integration of aliasing information into `moveBefore/AfterTopologicallyValid()`. Basically, we pass in an alias DB when we ask for moveBefore/After. Similar to how we can boil down dependency analysis to "what nodes use this node", we can boil down mutability analysis to "what nodes write to an alias set input/output'd by this node".

3. Integration of alias analysis to optimization passes that need it. Right now, it is `GraphFuser`, `CreateAutodiffSubgraphs`, constant prop, and CSE. Not sure if any others need it.

- Testing; still figuring out the best way to do this.
- Eventually we want to integrate the alias db into the graph, but we shouldn't do that until we can guarantee that the information can stay up to date with mutations.
- Do the same thing `python_printer` did for operators and force people to register alias analyzers if they can't schematize their op.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14018

Differential Revision: D13144906

Pulled By: suo

fbshipit-source-id: 1bc964f9121a504c237cef6dfeea6b233694de6a

5 years agoRemove extra include
Ilia Cherniavskii [Thu, 22 Nov 2018 01:19:37 +0000 (17:19 -0800)]
Remove extra include

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/14206

Reviewed By: dzhulgakov

Differential Revision: D13131318

fbshipit-source-id: 559b55b8d98cdf6b7d1d3e31237c5473edc5e462

5 years agoRemoved redundant allreduce options in DDP (#14208)
Teng Li [Thu, 22 Nov 2018 00:54:36 +0000 (16:54 -0800)]
Removed redundant allreduce options in DDP (#14208)

Summary:
This somehow is not cleaned up after the C++ migration. Unused and can be removed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14208

Differential Revision: D13132492

Pulled By: teng-li

fbshipit-source-id: 0f05b6368174664ebb2560c037347c8eb45f7c38

5 years agoAdd list inequality operator (#14129)
David Riazati [Thu, 22 Nov 2018 00:30:43 +0000 (16:30 -0800)]
Add list inequality operator (#14129)

Summary:
This PR adds `aten::neq` for list inequality comparisons and converts
`nll_loss` to weak script
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14129

Differential Revision: D13123894

Pulled By: driazati

fbshipit-source-id: 8c1edf7c163217ec00eb653f95d196db3998613f

5 years agoAdd onnxifi support to SparseLengthsWeightedSum (#14210)
Yinghai Lu [Wed, 21 Nov 2018 23:43:10 +0000 (15:43 -0800)]
Add onnxifi support to SparseLengthsWeightedSum (#14210)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14210

We left `SparseLengthsWeightedSum` as benchmark is not testing it due to fp16 filler issue. It was flushed out by unit tests. Hence we add the support here.

Reviewed By: bddppq

Differential Revision: D13132320

fbshipit-source-id: b21c30c185c9e1fbf3980641bc3cdc39e85af2e1