kokoro: Replace K=1 conv_1d with mul_mat #64
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
danielzgtg
force-pushed
the
optimize/k1conv1d
branch
from
3ccef55 to
d379a5f
Compare
mmwillet
reviewed
May 28, 2025
Owner
mmwillet
left a comment
mmwillet
left a comment
There was a problem hiding this comment.
Mostly looks good. I'm going to check the norm to confirm it doesn't change the tensor output, but otherwise ready for merge imo.
Comment on lines
+1180
to
+1188
| n = ggml_cont(ctx, ggml_transpose(ctx, n)); | ||
| n = ggml_mul_mat(ctx, model->prosody_pred->n_proj_kernel, n); | ||
| n = squeeze_3d_2d_e0(ctx, n); | ||
| n = ggml_add(ctx, n, model->prosody_pred->n_proj_bias); |
Owner
There was a problem hiding this comment.
spacing seems weird here. Should we add a linter?
Collaborator
Author
Owner
There was a problem hiding this comment.
I do not (please forgive me I am one of those people who ignores style as long as the code runs). Dice roll for any built in style is fine by me.
danielzgtg
force-pushed
the
optimize/k1conv1d
branch
from
d379a5f to
7cea743
Compare
danielzgtg
force-pushed
the
optimize/k1conv1d
branch
from
7cea743 to
0bcb0a6
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.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.
This was while investigating #48. K=1 is not really a convolution so it reduces to a matmul in TTS.cpp without modifying
ggml-cpu.c. I originally had code for this inggml.cbut inlined it to TTS.cpp to avoid extra transposes.Unfortunately, the performance improvement was not statistically significant. It was before=1822.336528ms after=1804.631546ms with Kokoro perf_battery. Random variation reversed the trend once. The PR that improves performance nonnegligibly today on Linux is mmwillet/ggml#2 .
Nevertheless, this eliminates a bunch of transposes. There are now ~100 fewer ggml_tensors that take >1ms in
cli.