Add meta functions for {}_embedding{}_codegen_forward_{}{}_cuda #2094
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
✅ Deploy Preview for pytorch-fbgemm-docs canceled.
|
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
ezyang
force-pushed
the
export-D50653995
branch
from
a9cd315
to
70a5189
Compare
ezyang
added a commit
to ezyang/FBGEMM
that referenced
this pull request
Nov 1, 2023
…rch#2094) Summary: For example, this hits split_embedding_codegen_lookup_rowwise_adagrad_function among others. There are a few other ways to write this and I'm open to other options. This version seemed the shortest to write, at the cost of some code duplication. The general strategy is I took embedding_forward_split_template.cu, copy pasted it into embedding_forward_split_meta_template.cpp, deleted all the CUDA specific stuff, and made it register a Meta implementation instead of a CUDA implementation. Most parts of the original were excised except the kernel itself, where all of the error checking and output tensor calculation is kept. The biggest annoyance is the code duplication for the error checking, which are copy pasted with essentially no modifications from the original CUDA kernel. I could potentially template the original logic so I can transclude it in two places, but that would also have made this diff harder to understand. I'm willing to do this if someone tells me to. Another approach would have been to implement the meta function entirely in Python (this is what I did in my original prototype), but to get it to handle all of the optimizer permutations seemed like a big pain. Reviewed By: zou3519 Differential Revision: D50653995
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
ezyang
added a commit
to ezyang/FBGEMM
that referenced
this pull request
Nov 1, 2023
…rch#2094) Summary: For example, this hits split_embedding_codegen_lookup_rowwise_adagrad_function among others. There are a few other ways to write this and I'm open to other options. This version seemed the shortest to write, at the cost of some code duplication. The general strategy is I took embedding_forward_split_template.cu, copy pasted it into embedding_forward_split_meta_template.cpp, deleted all the CUDA specific stuff, and made it register a Meta implementation instead of a CUDA implementation. Most parts of the original were excised except the kernel itself, where all of the error checking and output tensor calculation is kept. The biggest annoyance is the code duplication for the error checking, which are copy pasted with essentially no modifications from the original CUDA kernel. I could potentially template the original logic so I can transclude it in two places, but that would also have made this diff harder to understand. I'm willing to do this if someone tells me to. Another approach would have been to implement the meta function entirely in Python (this is what I did in my original prototype), but to get it to handle all of the optimizer permutations seemed like a big pain. Reviewed By: zou3519 Differential Revision: D50653995
ezyang
force-pushed
the
export-D50653995
branch
from
70a5189
to
d5a0c56
Compare
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
ezyang
added a commit
to ezyang/FBGEMM
that referenced
this pull request
Nov 1, 2023
…rch#2094) Summary: For example, this hits split_embedding_codegen_lookup_rowwise_adagrad_function among others. There are a few other ways to write this and I'm open to other options. This version seemed the shortest to write, at the cost of some code duplication. The general strategy is I took embedding_forward_split_template.cu, copy pasted it into embedding_forward_split_meta_template.cpp, deleted all the CUDA specific stuff, and made it register a Meta implementation instead of a CUDA implementation. Most parts of the original were excised except the kernel itself, where all of the error checking and output tensor calculation is kept. The biggest annoyance is the code duplication for the error checking, which are copy pasted with essentially no modifications from the original CUDA kernel. I could potentially template the original logic so I can transclude it in two places, but that would also have made this diff harder to understand. I'm willing to do this if someone tells me to. Another approach would have been to implement the meta function entirely in Python (this is what I did in my original prototype), but to get it to handle all of the optimizer permutations seemed like a big pain. Reviewed By: zou3519 Differential Revision: D50653995
ezyang
force-pushed
the
export-D50653995
branch
from
d5a0c56
to
5d31268
Compare
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
…rch#2094) Summary: For example, this hits split_embedding_codegen_lookup_rowwise_adagrad_function among others. There are a few other ways to write this and I'm open to other options. This version seemed the shortest to write, at the cost of some code duplication. The general strategy is I took embedding_forward_split_template.cu, copy pasted it into embedding_forward_split_meta_template.cpp, deleted all the CUDA specific stuff, and made it register a Meta implementation instead of a CUDA implementation. Most parts of the original were excised except the kernel itself, where all of the error checking and output tensor calculation is kept. The biggest annoyance is the code duplication for the error checking, which are copy pasted with essentially no modifications from the original CUDA kernel. I could potentially template the original logic so I can transclude it in two places, but that would also have made this diff harder to understand. I'm willing to do this if someone tells me to. Another approach would have been to implement the meta function entirely in Python (this is what I did in my original prototype), but to get it to handle all of the optimizer permutations seemed like a big pain. Reviewed By: zou3519 Differential Revision: D50653995
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
ezyang
force-pushed
the
export-D50653995
branch
from
5d31268
to
903de3c
Compare
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
1 similar comment
|
This pull request was exported from Phabricator. Differential Revision: D50653995 |
|
This pull request has been merged in 824ef10. |
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.
Summary:
For example, this hits split_embedding_codegen_lookup_rowwise_adagrad_function among others.
There are a few other ways to write this and I'm open to other options. This version seemed the shortest to write, at the cost of some code duplication. The general strategy is I took embedding_forward_split_template.cu, copy pasted it into embedding_forward_split_meta_template.cpp, deleted all the CUDA specific stuff, and made it register a Meta implementation instead of a CUDA implementation. Most parts of the original were excised except the kernel itself, where all of the error checking and output tensor calculation is kept.
The biggest annoyance is the code duplication for the error checking, which are copy pasted with essentially no modifications from the original CUDA kernel. I could potentially template the original logic so I can transclude it in two places, but that would also have made this diff harder to understand. I'm willing to do this if someone tells me to. Another approach would have been to implement the meta function entirely in Python (this is what I did in my original prototype), but to get it to handle all of the optimizer permutations seemed like a big pain.
Differential Revision: D50653995