Add aten mkldnn zero_ operator #20573
Conversation
pytorchbot
added
module: mkldnn
li-roy
added
the
triaged
Summary: ### mkldnn backward ops list: - [ ] \(#20567) Add aten mkldnn conv2d backward operator 💛 - [ ] \(#20570) Add aten mkldnn backward ops: relu, linear and reshape 💛 - [ ] \(#20571) Add aten mkldnn backward ops: max_pool2d, avg_pool2d and adaptive_avg_poo2d 💛 - [ ] \(#20572) Add aten mkldnn batchnorm backward operator 💛 - [ ] \(#20573) Add aten mkldnn zero_ operator:yellow_heart: - [ ] \(#20575) Add mkldnn mul operator 💛 Pull Request resolved: #20575 Differential Revision: D15799529 Pulled By: bddppq fbshipit-source-id: 4887d8ef1a0e316ad9db199b657d9481fc13e486
Summary: ### mkldnn backward ops list: - [ ] \(pytorch/pytorch#20567) Add aten mkldnn conv2d backward operator 💛 - [ ] \(pytorch/pytorch#20570) Add aten mkldnn backward ops: relu, linear and reshape 💛 - [ ] \(pytorch/pytorch#20571) Add aten mkldnn backward ops: max_pool2d, avg_pool2d and adaptive_avg_poo2d 💛 - [ ] \(pytorch/pytorch#20572) Add aten mkldnn batchnorm backward operator 💛 - [ ] \(pytorch/pytorch#20573) Add aten mkldnn zero_ operator:yellow_heart: - [ ] \(pytorch/pytorch#20575) Add mkldnn mul operator 💛 Pull Request resolved: pytorch/pytorch#20575 Differential Revision: D15799529 Pulled By: bddppq fbshipit-source-id: 4887d8ef1a0e316ad9db199b657d9481fc13e486
XiaobingSuper
force-pushed
the
mkldnn_zero_
branch
from
683a232
to
e000509
Compare
|
@bddppq , rebased it, but the failed cases are not realsted my changes, could you help see them? thanks! |
| auto n = x.get_nelems(); | ||
| auto* x_ = static_cast<float *>(x.get_data_handle()); | ||
| parallel_for(0, n, 2048, [x_](int64_t begin, int64_t end){ | ||
| vec256::map( |
There was a problem hiding this comment.
@cpuhrsch Is it safe to unconditionally using these vectorized utilities? Or do we need to guard it by some avx macros check?
There was a problem hiding this comment.
Vec256 uses AVX and AVX2 guards to compile with the compatibility version. So calling https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/cpu/vec256/functional.h#L99 without compiling it for avx2 is correct.
But, since this isn't compiled with avx2 for OSS, it might be slow for users. Why isn't this a kernel?
There was a problem hiding this comment.
@cpuhrsch , if written as a kernel, the input should not be a TensorIterator as add_kernel , or written as SoftMaxKernel, please give me some advice, thanks!
There was a problem hiding this comment.
Inputs doesn't have to be TensorIterator. We prefer TensorIterator as input for kernels as we use same iterator for CPU and CUDA. In your case you can implement kernel that takes Tensor as input.
There was a problem hiding this comment.
Using map here is also inefficient, please consider writing something like vec256::fill which will loadu vector, fill it and store.
There was a problem hiding this comment.
@VitalyFedyunin , thanks for your suggestion, I will change the code.
There was a problem hiding this comment.
@VitalyFedyunin , I have added a fill API in ideep, perharps direct process opaque tensor is not good in pytorch, After talking with @jgong5, add soomthing like vec256::fill is only useful mkldnn tensor, perhaps this way have not big benefit. Thanks!
There was a problem hiding this comment.
@bddppq has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
There was a problem hiding this comment.
@bddppq has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Summary: ### mkldnn backward ops list: - [ ] \(pytorch/pytorch#20567) Add aten mkldnn conv2d backward operator 💛 - [ ] \(pytorch/pytorch#20570) Add aten mkldnn backward ops: relu, linear and reshape 💛 - [ ] \(pytorch/pytorch#20571) Add aten mkldnn backward ops: max_pool2d, avg_pool2d and adaptive_avg_poo2d 💛 - [ ] \(pytorch/pytorch#20572) Add aten mkldnn batchnorm backward operator 💛 - [ ] \(pytorch/pytorch#20573) Add aten mkldnn zero_ operator:yellow_heart: - [ ] \(pytorch/pytorch#20575) Add mkldnn mul operator 💚 Pull Request resolved: pytorch/pytorch#20573 Differential Revision: D15820477 Pulled By: bddppq fbshipit-source-id: 35d95f5b4e013c8db1911f52148550a2e40a2e68
mkldnn backward ops list: