[MPS] Revamp copy_to_mps_ implementation #87475
Merged
Conversation
This file contains 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
Because why wouldn't it? Fixes #86052 Pull Request resolved: #86958 Approved by: https://github.com/kulinseth (cherry picked from commit 13cff2e)
Tensor's view in linear storage is represented by the following parameters: `.shape`, `.stride()` and `.storage_offset()`. Only tensors that are representable as 1d-views can be copied from host to device (and vice versa) using single [`copy(from:sourceOffset:to:destinationOffset:size:)`](https://developer.apple.com/documentation/metal/mtlblitcommandencoder/1400767-copyfrombuffer?language=objc) call. Modify `copy_to_mps_` function to do the following steps: - Cast `src` tensor to dst data type if needed - Expand `src` tensor to `dst` tensor shape - Clone `src` tensor if it is not stride contiguous (i.e. can not be represented by `src.view(src.numel())`) - Create an empty tensor if `dst` is not stride-contiguous or if its strides are different then potentially cloned `src` strides - Do 1d copy for `src` to (potentiall temp) `dst` - Finally do re-striding/copy on MPS if needed Add test to cover cases where stide-contiguous permuted tensor is copied to MPS, non-stride-contiguous tensor is copied to MPS and if permuted CPU tensor is copied to differently permuted MPS tensor Fixes #86954 Pull Request resolved: #86956 Approved by: https://github.com/kulinseth (cherry picked from commit ae62cf7)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/87475
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 Failures, 8 PendingAs of commit 1ad30ba: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
ciflow/mps
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Tensor's view in linear storage is represented by the following parameters:
.shape,.stride()and.storage_offset().Only tensors that are representable as 1d-views can be copied from host to device (and vice versa) using single
copy(from:sourceOffset:to:destinationOffset:size:)call.Modify
copy_to_mps_function to do the following steps:srctensor to dst data type if neededsrctensor todsttensor shapesrctensor if it is not stride contiguous (i.e. can not be represented bysrc.view(src.numel()))dstis not stride-contiguous or if its strides are different then potentially clonedsrcstridessrcto (potentiall temp)dstAdd test to cover cases where stide-contiguous permuted tensor is copied to MPS, non-stride-contiguous tensor is copied to MPS and if permuted CPU tensor is copied to differently permuted MPS tensor