vulkan: Optimize contiguous copies #10254
Merged
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
Add a flops calculation for flash attention. Add one GGML_OP_CPY perf test.
github-actions
bot
added
testing
jeffbolznv
added
ggml
Add a variant of the copy shader for when the tensors are contiguous. Avoid the complex addressing calculations, and do four elements per invocation to hide some other overhead. Apply similar changes to the scale shader, since scale is always contiguous. Add a "progress bar" for shader compiles.
|
I did some more benchmarks. RTX 3090: Tesla P40: Radeon Pro VII: Looks like a good improvement all around. Edit: Similar improvements to SCALE. |
0cc4m
approved these changes
Nov 12, 2024
There was a problem hiding this comment.
The Vulkan changes look good to me, and I tested them successfully on Nvidia and AMD. From my side this can be merged.
@ggerganov Are the test-backend-ops changes fine?
slaren
approved these changes
Nov 12, 2024
arthw
pushed a commit
to arthw/llama.cpp
that referenced
this pull request
Nov 15, 2024
* tests: Fix memory bandwidth calculation for perf tests Add a flops calculation for flash attention. Add one GGML_OP_CPY perf test. * vulkan: Optimize contiguous copies Add a variant of the copy shader for when the tensors are contiguous. Avoid the complex addressing calculations, and do four elements per invocation to hide some other overhead. Apply similar changes to the scale shader, since scale is always contiguous. Add a "progress bar" for shader compiles.
arthw
pushed a commit
to arthw/llama.cpp
that referenced
this pull request
Nov 18, 2024
* tests: Fix memory bandwidth calculation for perf tests Add a flops calculation for flash attention. Add one GGML_OP_CPY perf test. * vulkan: Optimize contiguous copies Add a variant of the copy shader for when the tensors are contiguous. Avoid the complex addressing calculations, and do four elements per invocation to hide some other overhead. Apply similar changes to the scale shader, since scale is always contiguous. Add a "progress bar" for shader compiles.
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.
Split out from #10206, but the solution I went with is a bit different.
Add a variant of the copy shader for when the tensors are contiguous. Avoid the complex addressing calculations, and do four elements per invocation to hide some other overhead. In #10206, the matrix multiply is much faster if the B matrix is fp16, so there are a lot of these contiguous copies to do that conversion.
Apply similar changes to the scale shader, since scale is always contiguous.
The first commit fixes a bug in
test-backend-ops perfwhere it computed the memory footprint of one iteration but then divided by the total time for all iterations.Before/after on RTX 4070. In the after numbers, the larger copies are more or less framebuffer bandwidth-limited, and the smaller copies are hitting in L2.