Complex support on GPU for dynamic casting #29612
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
c10/util/TypeCast.h
Outdated
Comment on lines
106
to
126
| template<typename dest_t, typename complex_src_t> | ||
| struct MaybeReal { | ||
| static C10_HOST_DEVICE inline typename complex_src_t::value_type maybe_real(complex_src_t src) { | ||
| return src.real(); | ||
| } | ||
| }; | ||
|
|
||
| template<typename complex_src_t> | ||
| struct MaybeReal<std::complex<float>, complex_src_t> { | ||
| static C10_HOST_DEVICE inline complex_src_t maybe_real(complex_src_t src) { | ||
| return src; | ||
| } | ||
| }; | ||
|
|
||
| template<typename complex_src_t> | ||
| struct MaybeReal<std::complex<double>, complex_src_t> { | ||
| static C10_HOST_DEVICE inline complex_src_t maybe_real(complex_src_t src) { | ||
| return src; | ||
| } | ||
| }; | ||
|
|
Contributor
There was a problem hiding this comment.
Thanks for adding MaybeReal. I'm surprised that other scalar_t types like Bool, Half, BFFloat16 can call the scalar_t::real() method, but I guess it worked for std::real(scalar_t)?
Collaborator
Author
There was a problem hiding this comment.
@dylanbespalko No, the way it is used makes sure that complex_src_t could only be std::complex<something_t>.
Contributor
There was a problem hiding this comment.
Also, you could really help me out by adding this change to c10/util/TypeCast.h::line48::52. I think it should be:
template <typename dest_t, typename src_t>
C10_HOST_DEVICE inline dest_t static_cast_with_inter_type(src_t src) {
return static_cast<dest_t>(
static_cast<inter_copy_type_t<dest_t>>(MaybeReal<dest_t, src_t>(src)));
}
Contributor
There was a problem hiding this comment.
Oh, I see. Clever...
Collaborator
Author
There was a problem hiding this comment.
@dylanbespalko Thanks for the review. That makes great sense. And I will add that.
Contributor
|
I would prefer if you accept these changes before mine. I will pull the update from master. |
3 tasks
Collaborator
Author
|
@dylanbespalko I have moved logics into |
Contributor
Looks good to me. |
ezyang
approved these changes
Nov 12, 2019
Contributor
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@ezyang is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
Contributor
facebook-github-bot
pushed a commit
that referenced
this pull request
Nov 13, 2019
Summary: In-tree changes to pytorch to support complex numbers are being submitted here. Out-of-tree support for complex numbers is here: [pytorch-cpu-strided-complex extension](https://gitlab.com/pytorch-complex/pytorch-cpu-strided-complex) - [x] Replaced std:real(a) with a.real() in kernel level code. - [x] Fixed Vec256_base implementation of complex ops so that it works correctly on Non-AVX devices. - [ ] Clean up CopyKernel after #29612 is approved. zasdfgbnm is fixing this issue in #29612. This should be added first. cc: iotamudelta, ezyang, bddppq Pull Request resolved: #29607 Differential Revision: D18451046 Pulled By: ezyang fbshipit-source-id: b9dcd8e25e91cab13bd131b070d027b090cdedc9
zdevito
pushed a commit
to zdevito/ATen
that referenced
this pull request
Nov 13, 2019
Summary: In-tree changes to pytorch to support complex numbers are being submitted here. Out-of-tree support for complex numbers is here: [pytorch-cpu-strided-complex extension](https://gitlab.com/pytorch-complex/pytorch-cpu-strided-complex) - [x] Replaced std:real(a) with a.real() in kernel level code. - [x] Fixed Vec256_base implementation of complex ops so that it works correctly on Non-AVX devices. - [ ] Clean up CopyKernel after pytorch/pytorch#29612 is approved. zasdfgbnm is fixing this issue in pytorch/pytorch#29612. This should be added first. cc: iotamudelta, ezyang, bddppq Pull Request resolved: pytorch/pytorch#29607 Differential Revision: D18451046 Pulled By: ezyang fbshipit-source-id: b9dcd8e25e91cab13bd131b070d027b090cdedc9
facebook-github-bot
pushed a commit
that referenced
this pull request
Nov 13, 2019
Summary: After #29612 get merged, `static_cast_with_inter_type` can now automatically convert complex types to its real values, therefore there is no need to do it inside copy kernel. This should wait until #29612 get merged, otherwise it won't pass CI. Pull Request resolved: #29631 Differential Revision: D18485676 Pulled By: ezyang fbshipit-source-id: 0bbfd551e3d3010f87eef0fce23a1f8a094b7d31
zdevito
pushed a commit
to zdevito/ATen
that referenced
this pull request
Nov 13, 2019
Summary: After pytorch/pytorch#29612 get merged, `static_cast_with_inter_type` can now automatically convert complex types to its real values, therefore there is no need to do it inside copy kernel. This should wait until pytorch/pytorch#29612 get merged, otherwise it won't pass CI. Pull Request resolved: pytorch/pytorch#29631 Differential Revision: D18485676 Pulled By: ezyang fbshipit-source-id: 0bbfd551e3d3010f87eef0fce23a1f8a094b7d31
This was referenced Nov 14, 2019
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Currently, the dynamic casting mechanism is implemented assuming no support of complex on GPU. This will no longer be true in the soon future.
#29547 could clear some clang warning but the complex support on GPU is still not complete:
This PR is what should be done for type promotion in order to add support to complex dtype on GPU, as suggested in #755 (comment)
Note that what is newly added here in this PR is not tested due to the lack of basic support of complex dtypes (I can not construct a complex tensor). But his PR shouldn't break any existing part of PyTorch.
For the merge this PR, consider two options: