platform/upstream/pytorch.git
5 years agoRemove nomscheduler (#17693)
Duc Ngo [Wed, 6 Mar 2019 18:31:00 +0000 (10:31 -0800)]
Remove nomscheduler (#17693)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17693

Remove nomscheduler tool

Reviewed By: yinghai

Differential Revision: D14328168

fbshipit-source-id: 674d0e18596a4dc2bbb6b8d321f4066c4fc454ab

5 years agoindex operation support for torch.HalfTensor (#17645)
bhushan [Wed, 6 Mar 2019 18:28:49 +0000 (10:28 -0800)]
index operation support for torch.HalfTensor (#17645)

Summary:
- Test cases added
1. indexing for half tensor
2. setting for half tensor

fixes #17161
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17645

Differential Revision: D14302069

Pulled By: ezyang

fbshipit-source-id: 100f141c07046f200c904e27c5882a9417bccda0

5 years agoRevert D14160172: Implement a Caffe2 standalone LSTM operator
Soumith Chintala [Wed, 6 Mar 2019 16:41:42 +0000 (08:41 -0800)]
Revert D14160172: Implement a Caffe2 standalone LSTM operator

Differential Revision:
D14160172

Original commit changeset: c33e3f9e8aea

fbshipit-source-id: cffe35d93f0ac75ca93aa98a3b82af3d372f2fc1

5 years agofix typo in hub doc
Tongzhou Wang [Wed, 6 Mar 2019 07:14:25 +0000 (23:14 -0800)]
fix typo in hub doc

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17705

Differential Revision: D14338380

Pulled By: ailzhang

fbshipit-source-id: d53eece30bede88a642e718ee6f829ba29c7d1c4

5 years agofix dropout AD & rename range to rangelist (#17691)
Ailing Zhang [Wed, 6 Mar 2019 04:47:02 +0000 (20:47 -0800)]
fix dropout AD & rename range to rangelist (#17691)

Summary:
fixes #17669
Address apaszke 's comments in #17523
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17691

Differential Revision: D14328083

Pulled By: ailzhang

fbshipit-source-id: 9ec4a54f13bfd1aaf4b1821dd00c31793ac07a44

5 years agoenable use of MIOpen for depthwise convolutions (#17685)
Chaitanya Sri Krishna Lolla [Wed, 6 Mar 2019 02:41:20 +0000 (18:41 -0800)]
enable use of MIOpen for depthwise convolutions (#17685)

Summary:
* added miopen conv mode to be used for setConvDescriptor
* added miopen depthwise convolutions
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17685

Differential Revision: D14327811

Pulled By: bddppq

fbshipit-source-id: d5bdc1abafd5f39694fadf3f9275b9d880c5b115

5 years agoImplement a Caffe2 standalone LSTM operator (#17461)
Ahmed Aly [Wed, 6 Mar 2019 01:31:51 +0000 (17:31 -0800)]
Implement a Caffe2 standalone LSTM operator (#17461)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17461

Implementing a standalone LSTM Operator in Caffe2 adopted from this Aten implementation: diffusion/FBS/browse/master/fbcode/caffe2/aten/src/ATen/native/RNN.cpp. The most tricky thing in this exercise was that caffe2::Tensor has no copy constructor that made it necessary to implement a custom templated copy constructor for the different Tensor containers used in the code. Also there was no way to use off-the-shelf C2 operators in my code easily so I had to copy some code that is doing basic matmul, cat, split, transpose and linear as utility functions.

Two things missing:

- Profiling this implementation against the current ONNXified LSTM op
- Make this operator available to use in PyTorch

Reviewed By: dzhulgakov

Differential Revision: D14160172

fbshipit-source-id: c33e3f9e8aeae578b64d97593cb031a251216029

5 years agoFix nll_loss crash on cpu where ignore_index is out of bounds (#17328)
Soumith Chintala [Tue, 5 Mar 2019 22:26:20 +0000 (14:26 -0800)]
Fix nll_loss crash on cpu where ignore_index is out of bounds (#17328)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/15508
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17328

Differential Revision: D14322629

Pulled By: soumith

fbshipit-source-id: 7d02f372be78794782c18affcfc109ce30b1e91c

5 years agoAdd '--hip-clang-launch' to favor <<<>>>-based launch. (#17686)
Johannes M Dieterich [Tue, 5 Mar 2019 20:49:25 +0000 (12:49 -0800)]
Add '--hip-clang-launch' to favor <<<>>>-based launch. (#17686)

Summary:
hip-clang uses triple chevron kernel dispatch syntax. Add an option to the hipification script to skip translating triple chevron to hipLaunchKernelGGL.

Once we switch to hip-clang, this option will be default and subsequently removed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17686

Differential Revision: D14327810

Pulled By: bddppq

fbshipit-source-id: 5e1512325077dd3ebb8fb9b5bf35fd1f8d9a4dc3

5 years agoImprove caching allocator for Pascal and newer GPUs. (#17120)
Sam Gross [Tue, 5 Mar 2019 17:38:23 +0000 (09:38 -0800)]
Improve caching allocator for Pascal and newer GPUs. (#17120)

Summary:
```
NVIDIA changed the CUDA allocation behavior on Pascal GPUs. The
page size increased from 1MB to 2MB and allocations larger than 1MB
are now always page-aligned. Previously, allocations larger than 1MB
were aligned to 128KB boundaries.

This interacted poorly with the caching allocator. The remaining
memory in a page could only be filled by small cudaMalloc calls, but
the caching allocator never cudaMalloc's a chunk smaller than 1MB.
This behavior could also cause a large discrepancy between the memory
usage reported by nvidia-smi and the memory usage reported by
PyTorch, because nvidia-smi counts a partially used page as "full",
while PyTorch only counts the actual memory requested.

This PR makes a few changes to the caching allocator to better support
Pascal and Volta GPUs:

 - All cudaMalloc calls are now multiples of 2MB (the page size)
 - Requests between 1-10MB allocate (and split) a 20MB block to
   reduce wasted space due to rounding
 - Small requests are now packed into 2MB blocks (instead of 1MB)

This improves Mask R-CNN memory usage by 10-20% in internal tests on
Volta GPUs. Maxwell performance seems to be largely unchanged, but
it's possible that some use cases suffer slightly.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17120

Differential Revision: D14301536

Pulled By: colesbury

fbshipit-source-id: a8282315ea8f7b8ca149b5066fdeaecd0d404edf

5 years agoTurn the Half::from_bits into a constexpr function to avoid unresolve… (#17661)
Davide Libenzi [Tue, 5 Mar 2019 15:24:27 +0000 (07:24 -0800)]
Turn the Half::from_bits into a constexpr function to avoid unresolve… (#17661)

Summary:
…d symbol errors when building in DEBUG mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17661

Differential Revision: D14319610

Pulled By: soumith

fbshipit-source-id: 6c508a37155e29260f403d7174f343aa1ff32385

5 years agoRemove Expect Files from python / tracing / script interop
Elias Ellison [Tue, 5 Mar 2019 06:38:41 +0000 (22:38 -0800)]
Remove Expect Files from python / tracing / script interop

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17622

Differential Revision: D14308307

Pulled By: eellison

fbshipit-source-id: bda249d38ac2570000a12b0ca328c26233ecefe8

5 years agoEnable apex on Windows
peterjc123 [Tue, 5 Mar 2019 05:50:53 +0000 (21:50 -0800)]
Enable apex on Windows

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17675

Differential Revision: D14320473

Pulled By: soumith

fbshipit-source-id: cb696984f5196f9b8b50722b4fe927bb6407c322

5 years agobump docker build to upgrade magma to 2.5.0 (#17674)
Soumith Chintala [Tue, 5 Mar 2019 04:28:06 +0000 (20:28 -0800)]
bump docker build to upgrade magma to 2.5.0 (#17674)

Summary:
upgrades magma in docker build.

vishwakftw
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17674

Differential Revision: D14320187

Pulled By: soumith

fbshipit-source-id: 7887f65fb703b802fc6231408b55ad9c4039882b

5 years agorefactor caffe2 operator constructors - 1/9 (#17082)
Sebastian Messmer [Mon, 4 Mar 2019 23:56:21 +0000 (15:56 -0800)]
refactor caffe2 operator constructors - 1/9 (#17082)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17082

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078498

fbshipit-source-id: f7f65d6d81c7942293f53fdaa61f756d8b7360c1

5 years agoExpose cuda kernel for caffe2::GenerateProposals
Sebastian Messmer [Mon, 4 Mar 2019 22:53:55 +0000 (14:53 -0800)]
Expose cuda kernel for caffe2::GenerateProposals

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17066

Reviewed By: ezyang, wat3rBro

Differential Revision: D14071130

fbshipit-source-id: 6fe26503f6069c36ec31d6c09b549b932d5db242

5 years agoprint warnings when DNNLOWP_16 or DNNLOWP_ROWWISE_16 engine is used (#17176)
Jongsoo Park [Mon, 4 Mar 2019 22:25:19 +0000 (14:25 -0800)]
print warnings when DNNLOWP_16 or DNNLOWP_ROWWISE_16 engine is used (#17176)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17176

As title

Reviewed By: csummersea

Differential Revision: D14111616

fbshipit-source-id: 1282cb2452c4ad385fd2dc6d3f8c19e9fec715ff

5 years agoFix XOutput/XOutputTensor for ivalue based c2 operators (#17599)
Sebastian Messmer [Mon, 4 Mar 2019 22:17:11 +0000 (14:17 -0800)]
Fix XOutput/XOutputTensor for ivalue based c2 operators (#17599)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17599

XOutput/XOutputTensor was broken for ivalue based operators. This diff fixes that.

Reviewed By: ezyang

Differential Revision: D14274003

fbshipit-source-id: b99f020244c66c4e2551dbd32ae0f665cc91b338

5 years agoFix InputSize/OutputSize for ivalue based operators (#17579)
Sebastian Messmer [Mon, 4 Mar 2019 22:17:11 +0000 (14:17 -0800)]
Fix InputSize/OutputSize for ivalue based operators (#17579)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17579

These methods previously just returned 0 when it was not a legacy operator,
making it impossible to convert some operators.

Reviewed By: dzhulgakov

Differential Revision: D14253094

fbshipit-source-id: 72bfdcf6da291a4ab80d1e0ceb20984b86edc408

5 years agoFix clamp fusion on missing limits (#17533)
Wanchao Liang [Mon, 4 Mar 2019 21:04:53 +0000 (13:04 -0800)]
Fix clamp fusion on missing limits (#17533)

Summary:
Fixes #17449

Context: before #17186, we don't fuse `clamp` for the case when `min/max` are missing inputs, because they are `prim::None` node, after #17186, we make None a `prim::Constant` node which enables the fusion for `clamp`. But codegen.cpp does not handle the case when `prim::Constant` is not a Double/Int/Bool, this PR makes it so that missing inputs are handled correctly, it is done in the following way:

1. emit nothing when you see `type? = prim::Constant()`
2. when emitRHS, do special casing for aten::clamp
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17533

Differential Revision: D14238450

Pulled By: wanchaol

fbshipit-source-id: 61a272154754b13e89021bb86002927f02cde19c

5 years agoint32 indexing for Tensor Iterator Reduction (#17428)
Jie [Mon, 4 Mar 2019 21:02:40 +0000 (13:02 -0800)]
int32 indexing for Tensor Iterator Reduction (#17428)

Summary:
1. Enabling int32 indexing for cases where TI cannot accumulate in output due to
incompatible data types (e.g. Welford).
2. Updating Welford kernel to use int32 instead of int64 indexing on GPU.

This change improves performance for torch.var / torch.std

Implementation:
1. Allocated extra buffer to handle accumulation between sub Tensor Iterators.
2. Removed int64 indexing in gpu_reduce_kernel
3. WelfordOps now supports index type / combination typeas a template parameter.
While GPU uses int32_t and float, CPU implementation uses int64_t and double.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17428

Differential Revision: D14264608

Pulled By: umanwizard

fbshipit-source-id: 3eb54451de925b469dbc1127e5ea7443c4431036

5 years agoRemoved all usages of TH_Index_Base (#17591)
Iurii Zdebskyi [Mon, 4 Mar 2019 20:43:28 +0000 (12:43 -0800)]
Removed all usages of TH_Index_Base (#17591)

Summary:
TH_Index_Base is hard coded to 0 and can be removed from the code base.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17591

Differential Revision: D14269273

Pulled By: izdeby

fbshipit-source-id: d844e261f4af7297bad8a81e7d6dcf0a391b94e6

5 years agoPyTorch/Caffe2 tensor interop in Python (#17190)
Dmytro Dzhulgakov [Mon, 4 Mar 2019 19:30:43 +0000 (11:30 -0800)]
PyTorch/Caffe2 tensor interop in Python (#17190)

Summary:
Because of two separate python extensions with different pybind
instances I have to go through void* conversion. Since it's hidden from
user, it's fine.

New APIs added on C2 side:
- workspace.FetchTorch('blob')
- workspace.Workspace.current.blobs['blob'].to_torch()
- workspace.FeedBlob('blob', pytorch_tensor)

Works on CPU an GPU.

The only glitches are with resizing because of variable/tensor split.
But data sharing works properly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17190

Reviewed By: ezyang

Differential Revision: D14163882

Pulled By: dzhulgakov

fbshipit-source-id: d18e5b8fcae026f393c842a1149e972515732de2

5 years agoFixed typo in aten/src/ATen/native_parse.py (#17641)
wkcn [Mon, 4 Mar 2019 18:08:04 +0000 (10:08 -0800)]
Fixed typo in aten/src/ATen/native_parse.py (#17641)

Summary:
Hi, there.
There is a typo in aten/src/ATen/native_parse.py, and I fix it.
`std::aray` -> `std::array`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17641

Differential Revision: D14301981

Pulled By: ezyang

fbshipit-source-id: a37859cdedcbf6c29333b954486dfa086d6c2176

5 years agoRemove GPU dependency from ProfileObserver (#17592)
Martin Schatz [Mon, 4 Mar 2019 17:55:05 +0000 (09:55 -0800)]
Remove GPU dependency from ProfileObserver (#17592)

Summary:
Remove GPU dependency and register ProfileObserver.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17592

Reviewed By: ezyang

Differential Revision: D14265801

Pulled By: mdschatz

fbshipit-source-id: f98c0c32653c64a8b087c58ece4f864dfbe1d4b8

5 years agoDon't make factory methods create a tensor and then immediately copy it (#17565)
Brennan Vincent [Mon, 4 Mar 2019 06:13:27 +0000 (22:13 -0800)]
Don't make factory methods create a tensor and then immediately copy it (#17565)

Summary:
Create a `make_variable` override that moves out of a tensor instead of going through `shallow_copy_and_detach`. Call this override from factory methods like `empty` that create a brand new tensor, do nothing with it, and then copy it into a variable.

Will update this with actual numbers, but it seems to get rid of around 20-40% of the overhead of calling `torch.empty(0)`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17565

Differential Revision: D14266130

Pulled By: umanwizard

fbshipit-source-id: f57d5f2ca3f80ee8ee96d50f905e852fd10db941

5 years agoFixed typo in torch/functional.py w/r/t broadcast_tensors (#17642)
Jack Richter-Powell [Sun, 3 Mar 2019 18:05:36 +0000 (10:05 -0800)]
Fixed typo in torch/functional.py w/r/t broadcast_tensors (#17642)

Summary:
In reference to #17574
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17642

Differential Revision: D14297177

Pulled By: ezyang

fbshipit-source-id: 968176ea3b46a0153da0fd9e6b40db314d29e51c

5 years agoChange fake tqdm constructor to match real tqdm (#17636)
Bryan He [Sun, 3 Mar 2019 09:01:26 +0000 (01:01 -0800)]
Change fake tqdm constructor to match real tqdm (#17636)

Summary:
Currently, the fake tqdm implementation requires an input (whereas real tqdm does not).

This caused a problem in torchvision (https://github.com/pytorch/vision/pull/770), and seems likely to cause minor irritations elsewhere.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17636

Differential Revision: D14296530

Pulled By: ezyang

fbshipit-source-id: bc077d898773c93dab34c985a7b30525a43e558a

5 years agoMark native_functions as matched if uncaptured by JIT (#17631)
Christian Puhrsch [Sun, 3 Mar 2019 02:14:02 +0000 (18:14 -0800)]
Mark native_functions as matched if uncaptured by JIT (#17631)

Summary:
Various functions aren't used by the JIT, so they're jit-compliant w.r.t. their schema by default.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17631

Differential Revision: D14295559

Pulled By: cpuhrsch

fbshipit-source-id: a2ecdcb5df47eb67c54ec642d88d42e985515142

5 years agoBan std::array from native_functions.yaml
Christian Puhrsch [Sat, 2 Mar 2019 03:18:47 +0000 (19:18 -0800)]
Ban std::array from native_functions.yaml

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17629

Differential Revision: D14292941

Pulled By: cpuhrsch

fbshipit-source-id: 3c3eed57a5505a4e1da3aea682092677ab0e73e3

5 years agoRemove more usages of BoolTensor and IndexTensor from native_functions.yaml
Christian Puhrsch [Sat, 2 Mar 2019 03:12:08 +0000 (19:12 -0800)]
Remove more usages of BoolTensor and IndexTensor from native_functions.yaml

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16468

Differential Revision: D14095405

Pulled By: cpuhrsch

fbshipit-source-id: ea4d6bb7a4e81c05fe9861190ddbf52201612bbf

5 years agoImplement kthvalue in ATen (#17544)
Thomas Viehmann [Sat, 2 Mar 2019 02:57:02 +0000 (18:57 -0800)]
Implement kthvalue in ATen (#17544)

Summary:
The CPU version is based on the TH version.
The GPU version is based on #8406 by Pararth Shah (thank you).

CPU quickselect based on that in TH's THTensorMoreMath.cpp, but with C++ (quickselectnoindex will be achieved by a different swap)
CPU kthvalue is based on the THTensor function in the same file.
The dim_apply function is a C++ replacement for TH_TENSOR_DIM_APPLYx macros.
The CUDA kernel uses functions adapted from the THCTensorSortK implementation.
In particular radixSelect is from THCTensorTopK.cuh.
The CUDA launcher code replaces a bunch of macros with C++. It will be re-used in one of the following patches.

Plan for further PRs:
- This
- Sort
- TopK + Mode + Median in any order
- Rip out THC stuff.

There may be utility functions / structs in the SortingCommon.cuh that come into
relevance only with sort.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17544

Differential Revision: D14286934

Pulled By: ezyang

fbshipit-source-id: 35dbea050b097e88777ac5fa5c0f499d5e23c738

5 years agoChange vml.h to support sizes greater than 2**32 - 1
Christian Puhrsch [Sat, 2 Mar 2019 00:53:23 +0000 (16:53 -0800)]
Change vml.h to support sizes greater than 2**32 - 1

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17280

Differential Revision: D14154997

Pulled By: cpuhrsch

fbshipit-source-id: c19b15d18da59c9ee87e82765d3244d2a4ef6729

5 years agomsvc_fixes (#17201)
Grigory Arutyunov [Fri, 1 Mar 2019 23:07:18 +0000 (15:07 -0800)]
msvc_fixes (#17201)

Summary:
Fixing MSVC errors

```
  D:\pytorch-scripts\caffe2_builders\v141\pytorch\aten\src\THC/THCReduce.cuh(144): error C4002: too many actual paramet
ers for macro 'C10_LAUNCH_BOUNDS_1' [D:\pytorch-scripts\caffe2_builders\v141\pytorch\build\Debug\caffe2\caffe2_gpu.vcxp
roj]
  D:\pytorch-scripts\caffe2_builders\v141\pytorch\aten\src\THC/THCReduce.cuh(259): error C4002: too many actual paramet
ers for macro 'C10_LAUNCH_BOUNDS_1' [D:\pytorch-scripts\caffe2_builders\v141\pytorch\build\Debug\caffe2\caffe2_gpu.vcxp
roj]
  D:/pytorch-scripts/caffe2_builders/v141/pytorch/aten/src/THCUNN/SpatialDilatedMaxPooling.cu(51): error C4002: too man
y actual parameters for macro 'C10_LAUNCH_BOUNDS_1' [D:\pytorch-scripts\caffe2_builders\v141\pytorch\build\Debug\caffe2
\caffe2_gpu.vcxproj]
```

on variadic C10_LAUNCH_BOUNDS as well as Debug linking issues with at::Half in pool_op_cudnn.cc like this one

```
pool_op_cudnn.obj : error LNK2019: unresolved external symbol "public: bool __cdecl caffe2::MaxPoolFunctor<class caff
e2::CUDAContext>::GlobalPoolingBackward<struct c10::Half,2>(int,int,int,struct c10::Half const *,struct c10::Half const
 ,struct c10::Half const ,struct c10::Half ,class caffe2::CUDAContext )const " (??$GlobalPoolingBackward@UHalf@c10@
@$01@?$MaxPoolFunctor@VCUDAContext@caffe2@@caffe2@QEBA_NHHHPEBUHalf@c10@00PEAU23@PEAVCUDAContext@1@Z) referenced in
 function "public: bool __cdecl caffe2::`anonymous namespace'::CuDNNMaxPoolFunctor::GlobalPoolingBackward<struct c10::H
alf,2>(int,int,int,struct c10::Half const ,struct c10::Half const ,struct c10::Half const ,struct c10::Half ,class
caffe2::CUDAContext *)const " (??$GlobalPoolingBackward@UHalf@c10@@$01@CuDNNMaxPoolFunctor@?A0xb936404a@caffe2@QEBA_NH
HHPEBUHalf@c10@00PEAU34@PEAVCUDAContext@2@Z) [D:\pytorch-scripts\caffe2_builders\v141\pytorch\build\Debug\caffe2\caff
e2_gpu.vcxproj]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17201

Differential Revision: D14165732

Pulled By: ezyang

fbshipit-source-id: 875fd9a5b2db6f83fc483f6d750d2c011260eb8b

5 years agoHipify fixes for Masquerade logic (#17598)
Jithun Nair [Fri, 1 Mar 2019 23:00:30 +0000 (15:00 -0800)]
Hipify fixes for Masquerade logic (#17598)

Summary:
ezyang Please review.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17598

Differential Revision: D14287724

Pulled By: ezyang

fbshipit-source-id: 46e5083854a827370bb4c81b82e5a4ede511e473

5 years agoRename prim::Undefined to prim::AutogradZero (#17611)
Wanchao Liang [Fri, 1 Mar 2019 23:00:01 +0000 (15:00 -0800)]
Rename prim::Undefined to prim::AutogradZero (#17611)

Summary:
supersedes #17245
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17611

Differential Revision: D14283581

Pulled By: wanchaol

fbshipit-source-id: 8022d02b8a021ea2fee9a18a2c8920eb123200c5

5 years agoAdd python test for extension backend tensor.device (#17602)
Roy Li [Fri, 1 Mar 2019 22:18:58 +0000 (14:18 -0800)]
Add python test for extension backend tensor.device (#17602)

Summary:
Adding a test for #17361
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17602

Differential Revision: D14287373

Pulled By: li-roy

fbshipit-source-id: 544ecf17eb310aed22ba0ea5f86f46b8e3bb69b5

5 years agoRevert D13935403: Call c10 cuda op from test_torch
Edward Yang [Fri, 1 Mar 2019 22:14:02 +0000 (14:14 -0800)]
Revert D13935403: Call c10 cuda op from test_torch

Differential Revision:
D13935403

Original commit changeset: b2915ec8a366

fbshipit-source-id: 0f3409d5c102d719bc1f0483695aee93e7d613c9

5 years agoadd command line option to use hive filler; add README (#17619)
Amy Yang [Fri, 1 Mar 2019 21:53:11 +0000 (13:53 -0800)]
add command line option to use hive filler; add README (#17619)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17619

--filler hive --iter -1 will let debugger exhaust all batches from a hive partition before exiting.
add README that summarizes command line options and usage.

Reviewed By: yinghai

Differential Revision: D14220166

fbshipit-source-id: daa23b7e8a9184481c6d7b67acf1599e5c99d74a

5 years agoRemove TH(CU)NN Sparse Linear (#17610)
Thomas Viehmann [Fri, 1 Mar 2019 20:32:47 +0000 (12:32 -0800)]
Remove TH(CU)NN Sparse Linear (#17610)

Summary:
Sparse Linear in TH(CU)NN implements sparse linear layers without
using sparse matrices.
It is currently not documented in PyTorch and there is no functional or
module interface. This means it is unused from a PyTorch point of view.

The reason for removing it is twofold:
- The module uses sort, which I would like to move to ATen.
- When we implement a SparseLinear layer, we would want to do it
  using sparse tensors, so it's not all that useful, anyway.

I checked this on slack with soumith, I hope the above is an accurate
representation. All bad ideas are my own.

This is part of the ongoing work to move
sort/topk/mode/median/kthvalue to ATen.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17610

Differential Revision: D14280663

Pulled By: gchanan

fbshipit-source-id: 289231d2c20626855ce2ceecd4f204b460c32378

5 years agoCorrect docstring of vision/init functions
ZhuBaohe [Fri, 1 Mar 2019 19:26:54 +0000 (11:26 -0800)]
Correct docstring of vision/init functions

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17351

Differential Revision: D14276355

Pulled By: soumith

fbshipit-source-id: 9b572b6a04eeb1e44cd93961edac76ed10f7b24e

5 years agoCall c10 cuda op from test_torch
Sebastian Messmer [Fri, 1 Mar 2019 18:51:34 +0000 (10:51 -0800)]
Call c10 cuda op from test_torch

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16692

Reviewed By: ezyang

Differential Revision: D13935403

fbshipit-source-id: b2915ec8a3664bb6e918ed357908cc33d8f9449a

5 years agoRevert #17191 and #17215 that no longer apply on Windows (#17567)
peter [Fri, 1 Mar 2019 18:33:58 +0000 (10:33 -0800)]
Revert #17191 and #17215 that no longer apply on Windows (#17567)

Summary:
They are previously merged to resolve #17051. However, since it was resolved by the upstream, and it was causing some issues like https://github.com/abjer/tsds/issues/8, I think it's time to revert these changes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17567

Differential Revision: D14265241

Pulled By: kostmo

fbshipit-source-id: 7fa2b7dd4ebc5148681acb439cf82d983898694e

5 years agousertype -> class (#17528)
Michael Suo [Fri, 1 Mar 2019 18:00:19 +0000 (10:00 -0800)]
usertype -> class (#17528)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17528

as title. register_prim_ops is messy because someone ruined clang-format, but I figured it's okay to include here since this is such a mechanical change

Reviewed By: driazati

Differential Revision: D14236943

fbshipit-source-id: c2b22845837b7f830015510e48ec2ee5202fa407

5 years agoalias analysis refactor take 2 (#17594)
Michael Suo [Fri, 1 Mar 2019 18:00:19 +0000 (10:00 -0800)]
alias analysis refactor take 2 (#17594)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17594

The original version of this broke things because a concurrent change raced with it in CI.

Reviewed By: ezyang

Differential Revision: D14266663

fbshipit-source-id: e8ac5dfcb7349b4f2c425d9f0eabbfc964314063

5 years agoFix the missing Windows CPU job in the build status section (#17608)
peter [Fri, 1 Mar 2019 17:56:43 +0000 (09:56 -0800)]
Fix the missing Windows CPU job in the build status section (#17608)

Summary:
It will be better to split the CPU job on CI. But unluckily, we are out of Windows machines.
cc, davidbrownellWork yf225
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17608

Differential Revision: D14281393

Pulled By: soumith

fbshipit-source-id: ae9a6140b7207ce56cfb2da3d812bc3fe060764a

5 years agoUpdate magma to 2.5.0 for Windows
peter [Fri, 1 Mar 2019 17:45:38 +0000 (09:45 -0800)]
Update magma to 2.5.0 for Windows

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17607

Differential Revision: D14281291

Pulled By: yf225

fbshipit-source-id: 51209c5540932871e45e54ba6d61b3b7d264aa8c

5 years agoAdding support for 0-d tensor for transpose (.t()) (#17535)
bhushan [Fri, 1 Mar 2019 16:38:06 +0000 (08:38 -0800)]
Adding support for 0-d tensor for transpose (.t()) (#17535)

Summary:
- Test updates
1. test_torch: added 0-d test case and t_() test cases
2. test_jit  : updated error message for TestAsync.test_async_script_error

- Updating documentation for torch.t()
Adding information regarding new support of 0-D and 1-D tenso

Fixes #17520
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17535

Differential Revision: D14269984

Pulled By: gchanan

fbshipit-source-id: 38b723f31484be939261c88edb33575d242eca65

5 years agoUpdating submodules
svcscm [Fri, 1 Mar 2019 09:33:59 +0000 (01:33 -0800)]
Updating submodules

Reviewed By: yns88

fbshipit-source-id: 05fafcfb34c76f425ac5c8ef24a5f920641c2cf7

5 years agoMark cudaGetLastError return value unused in C10_CUDA_CHECK
Junjie Bai [Fri, 1 Mar 2019 08:02:56 +0000 (00:02 -0800)]
Mark cudaGetLastError return value unused in C10_CUDA_CHECK

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17605

Reviewed By: xw285cornell

Differential Revision: D14277586

Pulled By: bddppq

fbshipit-source-id: 38879208f2ab83cf39d8a8a61b288cd09fcafd9a

5 years agoadd dropout during eval (#17549)
Huan Gui [Fri, 1 Mar 2019 07:17:35 +0000 (23:17 -0800)]
add dropout during eval (#17549)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17549

Currently Dropout is only enabled in training, we enable the option of having dropout in Eval.

This is to follow [1]. This functionality would be used for uncertainty estimation in exploration project.

[1] Gal, Yarin, and Zoubin Ghahramani. "Dropout as a bayesian approximation: Representing model uncertainty in deep learning." international conference on machine learning. 2016.

Reviewed By: Wakeupbuddy

Differential Revision: D14216216

fbshipit-source-id: 87c8c9cc522a82df467b685805f0775c86923d8b

5 years agoAdjust launch_bounds annotation for AMD hardware. (#17555)
Johannes M Dieterich [Fri, 1 Mar 2019 06:53:34 +0000 (22:53 -0800)]
Adjust launch_bounds annotation for AMD hardware. (#17555)

Summary:
The max pooling backwards kernel is currently annotated with launch bounds (256,8).

Adjust the number of waves to 4 (4 times 64 is 256) for ROCm. This improves training performance for torchvision models by up to 15% (AlexNet) on a gfx906 GPU.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17555

Differential Revision: D14277744

Pulled By: bddppq

fbshipit-source-id: 2a62088f7b8a87d1e350c432bf655288967c7883

5 years agoFix verbose compiler warning in flat_hash_map (#17562)
Sebastian Messmer [Fri, 1 Mar 2019 00:26:49 +0000 (16:26 -0800)]
Fix verbose compiler warning in flat_hash_map (#17562)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17562

fixes https://github.com/pytorch/pytorch/issues/17332

Reviewed By: ezyang

Differential Revision: D14254499

fbshipit-source-id: 9d5d7408c2ce510ac20cd438c6514dc2bbe3a854

5 years agoFix diagnostic pragmas (#17561)
Sebastian Messmer [Fri, 1 Mar 2019 00:26:49 +0000 (16:26 -0800)]
Fix diagnostic pragmas (#17561)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17561

The push at the top of the file was missing a corresponding pop

Reviewed By: ezyang

Differential Revision: D14254500

fbshipit-source-id: ff20359b563d6d6dcc68273dc754ab31aa8fad12

5 years agoAllow dispatch based on tensor list args (#17522)
Sebastian Messmer [Fri, 1 Mar 2019 00:25:37 +0000 (16:25 -0800)]
Allow dispatch based on tensor list args (#17522)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17522

Dispatch is still based on the first tensor arg, but that first "tensor arg" is now allowed to be a tensor list.
That is, the first argument that is either Tensor or TensorList will be the deciding factor for dispatch.
If it is a TensorList, then that TensorList must not be empty or dispatch will fail.

Reviewed By: ezyang

Differential Revision: D14235840

fbshipit-source-id: 266c18912d56ce77aa84306c5605c4191f3d882b

5 years agoAllow exposing caffe2 operators with variable number of input tensors to c10 (#17491)
Sebastian Messmer [Fri, 1 Mar 2019 00:25:37 +0000 (16:25 -0800)]
Allow exposing caffe2 operators with variable number of input tensors to c10 (#17491)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17491

Before, there was no way to expose a caffe2 operator that had a variable number of inputs.
Now, this is allowed by giving the operator one tensor list input.
Note that the tensor list must be the first input, and that any other tensor inputs will be ignored and inaccessible in this case.

Reviewed By: ezyang

Differential Revision: D14220705

fbshipit-source-id: 7f921bfb581caf46b229888c409bbcc40f7dda80

5 years agoblacklist fft algorithms for strided dgrad (#17016)
Syed Tousif Ahmed [Fri, 1 Mar 2019 00:17:37 +0000 (16:17 -0800)]
blacklist fft algorithms for strided dgrad (#17016)

Summary:
Applies https://github.com/pytorch/pytorch/pull/16626 from v1.0.1
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17016

Differential Revision: D14270100

Pulled By: ezyang

fbshipit-source-id: 1137899dd1551d33d16f39e8dde76cad8192af46

5 years agoRevert D14078519: [codemod][caffe2] [clangr] refactor caffe2 operator constructors...
Sebastian Messmer [Thu, 28 Feb 2019 22:58:12 +0000 (14:58 -0800)]
Revert D14078519: [codemod][caffe2] [clangr] refactor caffe2 operator constructors - 5/9

Differential Revision:
D14078519

Original commit changeset: b0ca31a52e4a

fbshipit-source-id: 713ae108d3dd6f33abdbf98a5f213e57e2b64642

5 years agoAdd generic list/dict custom op bindings (#17587)
David Riazati [Thu, 28 Feb 2019 22:43:05 +0000 (14:43 -0800)]
Add generic list/dict custom op bindings (#17587)

Summary:
Fixes #17017

Sandcastle refuses to land #17037, so trying fresh here
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17587

Differential Revision: D14265402

Pulled By: driazati

fbshipit-source-id: b942721aa9360ac6b3862f552ac95529eb0cf52c

5 years agorefactor caffe2 operator constructors - 8/9 (#17089)
Sebastian Messmer [Thu, 28 Feb 2019 22:12:37 +0000 (14:12 -0800)]
refactor caffe2 operator constructors - 8/9 (#17089)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17089

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078539

fbshipit-source-id: 9ca196af4af7f26fc82e6cf82b35d478d0597752

5 years agorefactor caffe2 operator constructors - 6/9 (#17087)
Sebastian Messmer [Thu, 28 Feb 2019 22:12:14 +0000 (14:12 -0800)]
refactor caffe2 operator constructors - 6/9 (#17087)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17087

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078525

fbshipit-source-id: 7cc03b30b0d4eb99818e35406be4119b27bdb1bc

5 years agorefactor caffe2 operator constructors - 2/9 (#17083)
Sebastian Messmer [Thu, 28 Feb 2019 22:06:51 +0000 (14:06 -0800)]
refactor caffe2 operator constructors - 2/9 (#17083)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17083

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078504

fbshipit-source-id: 34dddb035eee2fca3150e47c57489614b91b6725

5 years agorefactor caffe2 operator constructors - 7/9 (#17088)
Sebastian Messmer [Thu, 28 Feb 2019 22:04:06 +0000 (14:04 -0800)]
refactor caffe2 operator constructors - 7/9 (#17088)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17088

clangr codemod

also manually moved the constructor of a class from the .cpp file to the .h file.

Reviewed By: ezyang

Differential Revision: D14078531

fbshipit-source-id: 2adb4ac0ce523742da6cce3bc3b6c177b816c299

5 years agorefactor caffe2 operator constructors - 4/9 (#17085)
Sebastian Messmer [Thu, 28 Feb 2019 22:03:34 +0000 (14:03 -0800)]
refactor caffe2 operator constructors - 4/9 (#17085)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17085

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078515

fbshipit-source-id: aaa48ae10892e3f47063f2133e026fea46f3240b

5 years agorefactor caffe2 operator constructors - 3/9 (#17084)
Sebastian Messmer [Thu, 28 Feb 2019 22:01:30 +0000 (14:01 -0800)]
refactor caffe2 operator constructors - 3/9 (#17084)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17084

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078507

fbshipit-source-id: ed02d772890b30196302b6830f541f054b7e95c8

5 years agoMake HIPStream also masquerade as CUDA. (#17469)
Edward Yang [Thu, 28 Feb 2019 21:32:22 +0000 (13:32 -0800)]
Make HIPStream also masquerade as CUDA. (#17469)

Summary:
HIPGuard interfaces that interacted with HIPStream were previously
totally busted (because the streams had the wrong device type).
This fixes it, following along the same lines of MasqueardingAsCUDA.

Along the way I beefed up the explanatory comment.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>
cc jithunnair-amd iotamudelta bddppq
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17469

Differential Revision: D14243396

Pulled By: ezyang

fbshipit-source-id: 972455753a62f8584ba9ab194f9c785db7bb9bde

5 years agoFix Python device type property for XLA and MSNPU
Alex Şuhan [Thu, 28 Feb 2019 21:28:17 +0000 (13:28 -0800)]
Fix Python device type property for XLA and MSNPU

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17361

Differential Revision: D14243546

Pulled By: soumith

fbshipit-source-id: b7498968f72e3d97de5bf6e5b44c5a59b6913acb

5 years agoRely on numel() == 1 to check if distribution parameters are scalar. (#17503)
Morgan Funtowicz [Thu, 28 Feb 2019 21:27:27 +0000 (13:27 -0800)]
Rely on numel() == 1 to check if distribution parameters are scalar. (#17503)

Summary:
As discussed here #16952, this PR aims at improving the __repr__ for distribution when the provided parameters are torch.Tensor with only one element.

Currently, __repr__() relies on dim() == 0 leading to the following behaviour :

```
>>> torch.distributions.Normal(torch.tensor([1.0]), torch.tensor([0.1]))
Normal(loc: torch.Size([1]), scale: torch.Size([1]))
```

With this PR, the output looks like the following:
```
>>> torch.distributions.Normal(torch.tensor([1.0]), torch.tensor([0.1]))
Normal(loc: 1.0, scale: 0.10000000149011612)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17503

Differential Revision: D14245439

Pulled By: soumith

fbshipit-source-id: a440998905fd60cf2ac9a94f75706021dd9ce5bf

5 years agofix reordering of inlines (#17557)
Zachary DeVito [Thu, 28 Feb 2019 21:06:10 +0000 (13:06 -0800)]
fix reordering of inlines (#17557)

Summary:
See comment inside of code. This fixes a bug where sometimes we would try to avoid printing long lines but would inadvertently reorder the expressions, which can change the semantics of the program
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17557

Differential Revision: D14250608

Pulled By: zdevito

fbshipit-source-id: d44996af4e90fe9ab9508d13cd04adbfc7bb5d1c

5 years agoCustomize the printing of namedtuple return (#17136)
Xiang Gao [Thu, 28 Feb 2019 20:59:34 +0000 (12:59 -0800)]
Customize the printing of namedtuple return (#17136)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/17112
```python
print("good", torch.randn(5,5,5).max(1))
print("terrible", torch.randn(5,5,10).max(1))
print("not as good", torch.randn(5,5,500).max(1))
print ("old behaviour = gold standard")
print(tuple(torch.randn(5,5,5).max(1)))
print(tuple(torch.randn(5,5,10).max(1)))
print(tuple(torch.randn(5,5,500).max(1)))
```
now gives
```
>>> import torch
>>> print("good", torch.randn(5,5,5).max(1))
good torch.return_types.max(
values=tensor([[ 1.2821,  1.8063,  1.8075,  1.3082, -0.1267],
        [ 0.3437,  0.7353,  1.2619,  0.7557,  1.6662],
        [ 0.8583,  1.8906,  1.0246,  1.7598,  1.1184],
        [ 1.7821,  0.0230,  0.9452,  1.0318,  1.0823],
        [ 0.4116, -0.0379, -0.1843,  1.4129,  1.8796]]),
indices=tensor([[4, 4, 3, 2, 1],
        [1, 2, 4, 1, 1],
        [2, 4, 0, 2, 1],
        [0, 2, 0, 3, 1],
        [0, 4, 4, 4, 4]]))
>>> print("terrible", torch.randn(5,5,10).max(1))
terrible torch.return_types.max(
values=tensor([[ 2.1272,  1.3664,  2.2067,  1.3974, -0.0883,  1.2505,  1.0074,  1.1217,
          0.3849,  0.6936],
        [ 0.6288, -0.4560,  1.2748,  1.5482,  1.2777,  1.6874,  0.7151,  0.6041,
          1.3572,  1.6232],
        [ 1.6703,  1.0075,  1.6480,  2.2839,  1.3390,  0.4938,  1.6449,  1.7628,
          0.8141,  2.5714],
        [ 0.7079,  1.8677,  3.2478,  1.5591,  2.4870,  0.8635, -0.1450,  1.6923,
          1.4924,  1.6298],
        [ 2.4056,  0.8002,  0.9317,  0.7455,  0.7866,  2.1191,  0.3492,  1.2095,
          1.8637,  1.7470]]),
indices=tensor([[1, 1, 0, 0, 0, 0, 3, 4, 4, 4],
        [4, 2, 2, 1, 2, 2, 3, 1, 1, 3],
        [0, 3, 3, 0, 2, 1, 4, 1, 0, 1],
        [4, 1, 3, 0, 3, 2, 0, 1, 4, 3],
        [1, 0, 3, 2, 1, 0, 0, 1, 0, 1]]))
>>> print("not as good", torch.randn(5,5,500).max(1))
not as good torch.return_types.max(
values=tensor([[ 0.3877,  0.7873,  1.8701,  ...,  0.5971,  1.6103, -0.3435],
        [ 1.1300,  2.2418,  1.4239,  ...,  1.3943,  0.3872,  1.6475],
        [ 2.0656,  1.3136,  0.9896,  ...,  2.3918,  0.8226,  1.0517],
        [ 1.1054,  0.9945,  1.0561,  ...,  2.1039,  1.1524,  3.0304],
        [ 1.5041,  2.2809,  1.0883,  ...,  0.8504,  2.4774,  1.1041]]),
indices=tensor([[4, 3, 1,  ..., 1, 4, 0],
        [4, 4, 4,  ..., 3, 0, 3],
        [3, 0, 1,  ..., 2, 2, 4],
        [0, 1, 1,  ..., 4, 2, 2],
        [1, 0, 4,  ..., 2, 0, 2]]))
>>> print ("old behaviour = gold standard")
old behaviour = gold standard
>>> print(tuple(torch.randn(5,5,5).max(1)))
(tensor([[ 1.1908,  1.1807,  1.3151,  1.7184,  0.3556],
        [ 0.3798,  0.9213,  0.3001,  1.3087,  2.2419],
        [ 1.4233,  1.4814,  1.9900,  1.7744,  1.3059],
        [ 1.0026, -0.0330,  1.3061,  1.8730,  2.0685],
        [ 1.3041,  1.6458,  1.3449,  1.8948,  3.6206]]), tensor([[0, 4, 3, 4, 0],
        [1, 1, 4, 0, 4],
        [4, 1, 0, 3, 3],
        [1, 2, 1, 4, 0],
        [3, 3, 0, 3, 3]]))
>>> print(tuple(torch.randn(5,5,10).max(1)))
(tensor([[-0.1232,  0.8275,  0.6732,  1.1223,  0.8247,  1.2851,  1.6009,  1.9979,
          1.9109,  0.7313],
        [ 0.2260,  0.5922,  1.6928,  0.6024,  2.1158,  3.0619,  0.5653,  0.7426,
          0.8316,  0.6346],
        [ 0.4319,  0.2231,  0.5255,  1.7620,  1.1657,  0.8875,  0.5782,  0.6506,
          0.5032,  1.7097],
        [ 0.4137,  1.7265,  1.4260,  2.0301,  1.2244,  0.7128,  2.6345,  0.7230,
          1.3553,  1.6508],
        [ 1.0684,  1.7195,  1.4068,  0.7076, -0.0242,  0.8474,  0.8754,  1.7108,
          0.2188,  1.1584]]), tensor([[0, 1, 3, 4, 2, 3, 4, 2, 1, 0],
        [1, 4, 0, 0, 3, 2, 0, 0, 3, 3],
        [2, 3, 1, 1, 4, 0, 1, 4, 4, 4],
        [0, 4, 1, 3, 2, 0, 2, 0, 3, 1],
        [1, 0, 0, 0, 0, 3, 3, 3, 2, 0]]))
>>> print(tuple(torch.randn(5,5,500).max(1)))
(tensor([[0.9395, 1.5572, 1.8797,  ..., 2.0494, 0.8202, 0.9623],
        [1.7937, 0.7225, 1.8836,  ..., 0.7927, 1.4976, 1.1813],
        [0.8558, 1.6943, 1.4192,  ..., 0.8327, 1.9661, 0.4197],
        [1.2993, 1.4995, 0.9357,  ..., 0.7810, 1.3030, 2.6216],
        [1.4206, 1.8315, 1.0338,  ..., 1.4312, 1.3198, 1.5233]]), tensor([[0, 4, 3,  ..., 3, 0, 2],
        [0, 1, 0,  ..., 0, 4, 3],
        [3, 4, 3,  ..., 3, 0, 0],
        [3, 2, 3,  ..., 1, 2, 1],
        [1, 2, 4,  ..., 3, 1, 3]]))
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17136

Differential Revision: D14250021

Pulled By: VitalyFedyunin

fbshipit-source-id: aae72f03b35980063b1ac1f07b8353eddb0c8b93

5 years agoRevert D14231251: [jit] alias_analysis refactor
Michael Suo [Thu, 28 Feb 2019 20:49:31 +0000 (12:49 -0800)]
Revert D14231251: [jit] alias_analysis refactor

Differential Revision:
D14231251

Original commit changeset: 6cd98ae6fced

fbshipit-source-id: 96189f47daf7cc4cf4ef5cd343022d56a2296b39

5 years agorefactor caffe2 operator constructors - 5/9 (#17086)
Sebastian Messmer [Thu, 28 Feb 2019 19:43:29 +0000 (11:43 -0800)]
refactor caffe2 operator constructors - 5/9 (#17086)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17086

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078519

fbshipit-source-id: b0ca31a52e4ab97b145a1490461d59f8fa93874a

5 years agoalias_analysis refactor (#17511)
Michael Suo [Thu, 28 Feb 2019 19:28:16 +0000 (11:28 -0800)]
alias_analysis refactor (#17511)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17511

AliasTracker was doing bookkeeping for three concepts: the points-to graph,
writes, and wildcards.

This PR makes AliasTracker's job clearer: it keeps track of the points-to
graph. Thus it has been renamed MemoryDAG. Write and wildcard information were
pulled back into AliasDb as part of this—I may decide to pull them into their
own little modules since I don't want the alias analysis stuff to get too
bloated.

This refactor is necessary because we want to start tracking information for
aliasing elements that _aren't_ first-class IR Values (e.g. the "stuff" inside
a list). So MemoryDAG can't know too much about Values

Reviewed By: houseroad

Differential Revision: D14231251

fbshipit-source-id: 6cd98ae6fced8d6c1522c2454da77c3c1b2b0504

5 years agoallow "before" and "after" alias annotations (#17480)
Michael Suo [Thu, 28 Feb 2019 19:28:16 +0000 (11:28 -0800)]
allow "before" and "after" alias annotations (#17480)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17480

This was always part of our "spec" but not implemented

Reviewed By: houseroad

Differential Revision: D14214301

fbshipit-source-id: 118db320b43ec099dc3e730c67d39487474c23ea

5 years agoONNXIFI extension & e2e tests. (#17478)
Rui Zhu [Thu, 28 Feb 2019 19:22:20 +0000 (11:22 -0800)]
ONNXIFI extension & e2e tests. (#17478)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17478

Enable onnxifi_ext in glow and build an e2e test in caffe2.

Reviewed By: yinghai

Differential Revision: D14190136

fbshipit-source-id: 26245278b487b551623109b14432f675279b17b5

5 years agoupdate slack invite instructions
Soumith Chintala [Thu, 28 Feb 2019 19:21:23 +0000 (11:21 -0800)]
update slack invite instructions

Summary: update slack invite instructions

Reviewed By: pjh5

Differential Revision: D14255348

fbshipit-source-id: 564fed0d44a6a68f80d1894fed40c3ddb360aa52

5 years agoFix errors in the description for installation on Windows (#17475)
Evgeny Mankov [Thu, 28 Feb 2019 18:31:46 +0000 (10:31 -0800)]
Fix errors in the description for installation on Windows (#17475)

Summary:
+ All quotes for ENV VARS are erroneous;
+ Toolset hasn't be specified;
+ Provide paths for all 3 Visual Studio 2017 products: Community/Professional/Enterprise.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17475

Differential Revision: D14262968

Pulled By: soumith

fbshipit-source-id: c0504e0a6be9c697ead83b06b0c5cf569b5c8625

5 years agorefactor caffe2 operator constructors - 9/9 (#17090)
Sebastian Messmer [Thu, 28 Feb 2019 17:50:19 +0000 (09:50 -0800)]
refactor caffe2 operator constructors - 9/9 (#17090)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17090

clangr codemod

Reviewed By: ezyang

Differential Revision: D14078550

fbshipit-source-id: 68e6de4298e55ce83039b7806c1a275c4d6593c8

5 years agoFix the false generated_comment (#17563)
Gemfield [Thu, 28 Feb 2019 17:37:48 +0000 (09:37 -0800)]
Fix the false generated_comment (#17563)

Summary:
The generated_comments are wrong to below generated files:
```bash
./torch/csrc/autograd/generated/VariableType_0.cpp:3:// generated from tools/autograd/templates/VariableType_0.cpp
./torch/csrc/autograd/generated/VariableType_1.cpp:3:// generated from tools/autograd/templates/VariableType_1.cpp
./torch/csrc/autograd/generated/VariableType_2.cpp:3:// generated from tools/autograd/templates/VariableType_2.cpp
./torch/csrc/autograd/generated/VariableType_3.cpp:3:// generated from tools/autograd/templates/VariableType_3.cpp
./torch/csrc/autograd/generated/VariableType_4.cpp:3:// generated from tools/autograd/templates/VariableType_4.cpp
./torch/csrc/autograd/generated/VariableTypeEverything.cpp:3:// generated from tools/autograd/templates/VariableTypeEverything.cpp

./torch/csrc/jit/generated/register_aten_ops_0.cpp:23:// generated from tools/autograd/templates/register_aten_ops_0.cpp
./torch/csrc/jit/generated/register_aten_ops_1.cpp:23:// generated from tools/autograd/templates/register_aten_ops_1.cpp
./torch/csrc/jit/generated/register_aten_ops_2.cpp:23:// generated from tools/autograd/templates/register_aten_ops_2.cpp
```

These generated files were split to speed the compile, however, the template files are not.
After this fix, the comments will look like below:
```bash
./torch/csrc/autograd/generated/VariableType_0.cpp:3:// generated from tools/autograd/templates/VariableType.cpp
./torch/csrc/autograd/generated/VariableType_1.cpp:3:// generated from tools/autograd/templates/VariableType.cpp
......
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17563

Differential Revision: D14260992

Pulled By: soumith

fbshipit-source-id: 038181367fa43bee87837e4170704ddff7f4d6f2

5 years agoRemove useless OpenCV reference
Dmytro Dzhulgakov [Thu, 28 Feb 2019 07:18:46 +0000 (23:18 -0800)]
Remove useless OpenCV reference

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17564

Differential Revision: D14255542

Pulled By: dzhulgakov

fbshipit-source-id: c129f3751ae82deedd258ee16586552b77baaca6

5 years agoconvolution/matmul/dropout (#17523)
Ailing Zhang [Thu, 28 Feb 2019 05:36:37 +0000 (21:36 -0800)]
convolution/matmul/dropout (#17523)

Summary:
* Add AD formula for _convolution & matmul & dropout
* add prim::range, fixes #17483
Example:
```
dim = 3
x = range(dim)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17523

Differential Revision: D14254002

Pulled By: ailzhang

fbshipit-source-id: ba60d77b047db347929b72beca2623fb26aec957

5 years agodisallow shape analysis with resize ops (#17518)
Elias Ellison [Thu, 28 Feb 2019 02:59:19 +0000 (18:59 -0800)]
disallow shape analysis with resize ops (#17518)

Summary:
resize_ and resize_as resize the input tensor. because our shape analysis
is flow invariant, we don't do shape analysis on any op that relies on a Tensor that can alias a resized Tensor.

E.g. in the following graph the x += 10 x may have been resized.
```
torch.jit.script
def test(x, y):
    for i in range(10):
        x += 10
        x.resize_as_([1 for i in int(range(torch.rand())))
    return x

```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17518

Differential Revision: D14249835

Pulled By: eellison

fbshipit-source-id: f281b468ccb8c29eeb0f68ca5458cc7246a166d9

5 years agoMake C10_MOBILE consistent with how feature macros are usually used (#17481)
Sebastian Messmer [Thu, 28 Feb 2019 01:54:51 +0000 (17:54 -0800)]
Make C10_MOBILE consistent with how feature macros are usually used (#17481)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17481

Usually, feature macros are either defined or undefined and checked accordingly.
C10_MOBILE was a weird special case that was always defined but either defined to 1 or to 0.

This caused a lot of confusion for me when trying to disable something from mobile build and it also disabled it
from the server build (because I was using ifdef). Also, I found a place in the existing code base that made
that wrong assumption and used the macro wrongly, see https://fburl.com/y4icohts

Reviewed By: dzhulgakov

Differential Revision: D14214825

fbshipit-source-id: f3a155b6d43d334e8839e2b2e3c40ed2c773eab6

5 years agoDisable c10 dispatcher on mobile (#17078)
Sebastian Messmer [Thu, 28 Feb 2019 01:54:50 +0000 (17:54 -0800)]
Disable c10 dispatcher on mobile (#17078)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17078

This prevents caffe2 operators from being expsoed to c10 on mobile,
which in turn causes the whole c10 dispatcher to be stripped away
and saves binary size.

We probably want to re-enable the c10 dispatcher for mobile,
but for now this is ok.

Reviewed By: ezyang

Differential Revision: D14077972

fbshipit-source-id: e4dd3e3b60cdfbde91fe0d24102c1d9708d3e5c4

5 years agoAlways synchronize src and dst streams when copying tensors (#16966)
Shen Li [Wed, 27 Feb 2019 22:54:30 +0000 (14:54 -0800)]
Always synchronize src and dst streams when copying tensors (#16966)

Summary:
fixes #15568
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16966

Differential Revision: D14213144

Pulled By: mrshenli

fbshipit-source-id: 2fcf5e07895fde80b4aee72e2736b0def876d21f

5 years agoONNX Export Adaptive Pooling
Lara Haidar [Wed, 27 Feb 2019 22:52:26 +0000 (14:52 -0800)]
ONNX Export Adaptive Pooling

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17412

Differential Revision: D14247923

Pulled By: houseroad

fbshipit-source-id: 5530cea8f80da7368bff1e29cf89c45ad53accee

5 years agoUse name for output variables instead of out in JIT (#17386)
Christian Puhrsch [Wed, 27 Feb 2019 21:48:34 +0000 (13:48 -0800)]
Use name for output variables instead of out in JIT (#17386)

Summary:
This adds 88 matches.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17386

Differential Revision: D14179139

Pulled By: cpuhrsch

fbshipit-source-id: 2c3263b8e4d084db84791e53290e8c8b1b7aecd5

5 years agoForcing UTC on Mac circleci jobs (#17516)
Jesse Hellemn [Wed, 27 Feb 2019 21:17:16 +0000 (13:17 -0800)]
Forcing UTC on Mac circleci jobs (#17516)

Summary:
And adding timestamps to linux build jobs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17516

Differential Revision: D14244533

Pulled By: pjh5

fbshipit-source-id: 26c38f59e0284c99f987d69ce6a2c2af9116c3c2

5 years agoFix math::Set for large tensor (#17539)
Xiaomeng Yang [Wed, 27 Feb 2019 20:18:52 +0000 (12:18 -0800)]
Fix math::Set for large tensor (#17539)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17539

Fix math::Set for large tensor

i-am-not-moving-c2-to-c10

Reviewed By: dzhulgakov, houseroad

Differential Revision: D14240756

fbshipit-source-id: 0ade26790be41fb26d2cc193bfa3082c7bd4e69d

5 years agoAdd sparse gradient option to `gather` operation (#17182)
Natalia Gimelshein [Wed, 27 Feb 2019 19:39:37 +0000 (11:39 -0800)]
Add sparse gradient option to `gather` operation (#17182)

Summary:
This PR allows `gather` to optionally return sparse gradients, as requested in #16329. It also allows to autograd engine to accumulate sparse gradients in place when it is safe to do so.
I've commented out size.size() check in `SparseTensor.cpp` that also caused #17152, it does not seem to me that check serves a useful purpose, but please correct me if I'm wrong and a better fix is required.
Motivating example:
For this commonly used label smoothing loss function
```
def label_smoothing_opt(x, target):
    padding_idx = 0
    smoothing = 0.1
    logprobs = torch.nn.functional.log_softmax(x, dim=-1, dtype=torch.float32)
    pad_mask = (target == padding_idx)
    ll_loss = logprobs.gather(dim=-1, index=target.unsqueeze(1), sparse = True).squeeze(1)
    smooth_loss = logprobs.mean(dim=-1)
    loss =  (smoothing - 1.0) * ll_loss - smoothing * smooth_loss
    loss.masked_fill_(pad_mask, 0)
    return loss.sum()
```
backward goes from 12.6 ms with dense gather gradients to 7.3 ms with sparse gradients, for 9K tokens x 30K vocab, which is some single percent end-to-end improvement, and also improvement in peak memory required.
Shout-out to core devs: adding python-exposed functions with keyword arguments through native_functions.yaml is very easy now!

cc gchanan apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17182

Differential Revision: D14158431

Pulled By: gchanan

fbshipit-source-id: c8b654611534198025daaf7a634482b3151fbade

5 years agoadd elastic zeus handler (#16746)
Jane Wang [Wed, 27 Feb 2019 19:26:40 +0000 (11:26 -0800)]
add elastic zeus handler (#16746)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16746

as titled. We use a special url schem elasticzeus for elastic zeus so that we dont need to change the public interface of init_process_group.

Reviewed By: aazzolini, soumith

Differential Revision: D13948151

fbshipit-source-id: 88939dcfa0ad93467dabedad6905ec32e6ec60e6

5 years agooptimize elementwise sum (#17456)
Jongsoo Park [Wed, 27 Feb 2019 18:09:53 +0000 (10:09 -0800)]
optimize elementwise sum (#17456)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17456

Using an instruction sequence similar to function in fbgemm/src/QuantUtilAvx2.cc
elementwise_sum_benchmark added

Reviewed By: protonu

Differential Revision: D14205695

fbshipit-source-id: 84939c9d3551f123deec3baf7086c8d31fbc873e

5 years agoEnable boolean_mask, adadelta, adagrad fp16 on ROCm (#17235)
rohithkrn [Wed, 27 Feb 2019 18:04:33 +0000 (10:04 -0800)]
Enable boolean_mask, adadelta, adagrad fp16 on ROCm (#17235)

Summary:
-  Fix bugs, indentation for adadelta and adagrad tests to enable fp16
- Enable boolean_mask fp16  on ROCm
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17235

Differential Revision: D14240828

Pulled By: bddppq

fbshipit-source-id: ab6e8f38aa7afb83b4b879f2f4cf2277c643198f

5 years agoEnabled HALF for fill() and zero() methods. Moved them into THTensorFill (#17536)
Iurii Zdebskyi [Wed, 27 Feb 2019 17:17:04 +0000 (09:17 -0800)]
Enabled HALF for fill() and zero() methods. Moved them into THTensorFill (#17536)

Summary:
For some additional context on this change, please, see this [PR](https://github.com/pytorch/pytorch/pull/17376)

As a part of work on Bool Tensor, we will need to add support for a bool type to _fill() and _zero() methods that are currently located in THTensorMath. As we don't need anything else and those methods are not really math related - we are moving them out into separate THTensorFill for simplicity.

Change:
-moved _fill() and _zero() from THTensorMath.h to THTensorFill
-enabled _fill() and _zero() for HALF type.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17536

Differential Revision: D14242130

Pulled By: izdeby

fbshipit-source-id: 1d8bd806f0f5510723b9299d360b70cc4ab96afb

5 years agoFix autograd with buffers requiring grad in DataParallel (#13352)
Tongzhou Wang [Wed, 27 Feb 2019 04:43:11 +0000 (20:43 -0800)]
Fix autograd with buffers requiring grad in DataParallel (#13352)

Summary:
Causing a problem with spectral norm, although SN won't use that anymore after #13350 .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13352

Differential Revision: D14209562

Pulled By: ezyang

fbshipit-source-id: f5e3183e1e7050ac5a66d203de6f8cf56e775134

5 years agoenable assymetric dilations and stride for miopen conv (#17472)
Chaitanya Sri Krishna Lolla [Wed, 27 Feb 2019 04:36:45 +0000 (20:36 -0800)]
enable assymetric dilations and stride for miopen conv (#17472)

Summary:
As of MIOpen 1.7.1 as shipped in ROCm 2.1 this works correctly and we can use MIOpen and do not need to fall back
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17472

Differential Revision: D14210323

Pulled By: ezyang

fbshipit-source-id: 4c08d0d4623e732eda304fe04cb722c835ec70e4

5 years agoEnable tests working on ROCm 2.1 dual gfx906
Johannes M Dieterich [Wed, 27 Feb 2019 04:35:28 +0000 (20:35 -0800)]
Enable tests working on ROCm 2.1 dual gfx906

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17473

Reviewed By: bddppq

Differential Revision: D14210243

Pulled By: ezyang

fbshipit-source-id: 519032a1e73c13ecb260ea93102dc8efb645e070

5 years agoFix linking errors when building dataloader test binaries on Windows (#17494)
peter [Wed, 27 Feb 2019 04:33:59 +0000 (20:33 -0800)]
Fix linking errors when building dataloader test binaries on Windows (#17494)

Summary:
Fixes #17489.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17494

Differential Revision: D14226525

Pulled By: ezyang

fbshipit-source-id: 3dfef9bc6f443d647e9f05a54bc17c5717033723

5 years agoFix typo
hysts [Wed, 27 Feb 2019 04:20:32 +0000 (20:20 -0800)]
Fix typo

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17521

Differential Revision: D14237482

Pulled By: soumith

fbshipit-source-id: 636e0fbe2c667d15fcb649136a65ae64937fa0cb

5 years agoRemove Bool/IndexTensor from schema for native functions with derivatives (#17193)
Christian Puhrsch [Wed, 27 Feb 2019 01:41:56 +0000 (17:41 -0800)]
Remove Bool/IndexTensor from schema for native functions with derivatives (#17193)

Summary:
This only deals with four functions, but is an important first step towards removing BoolTensor and IndexTensor entirely.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17193

Differential Revision: D14157829

Pulled By: cpuhrsch

fbshipit-source-id: a36f16d1d88171036c44cc7de60ac9dfed9d14f2