Improving 4bit quant mat mul performance by shifting position of -8 operation. #15436
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15436
Note: Links to docs will display an error until the docs builds have been completed. ❗ 2 Active SEVsThere are 2 currently active SEVs. If your PR is affected, please view them below:
❌ 2 New Failures, 1 Unrelated FailureAs of commit f0c9c4d with merge base 3485495 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@trivedivivek has exported this pull request. If you are a Meta employee, you can view the originating Diff in D85721578. |
meta-cla
bot
added
the
CLA Signed
trivedivivek
added
the
release notes: vulkan
trivedivivek
force-pushed
the
export-D85721578
branch
from
45154bf to
65157ab
Compare
trivedivivek
added a commit
to trivedivivek/executorch
that referenced
this pull request
Oct 29, 2025
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. Differential Revision: D85721578
trivedivivek
added a commit
to trivedivivek/executorch
that referenced
this pull request
Oct 29, 2025
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. The thinking here is as follows: * The 4 bit integer weights are unsigned ranging from 0 - 15, and thus to get unsigned number 8 is subtracted from the input. * Assume WS[] is array of signed weights, M[] is matrix, S is the sum The main loop essentially performs: S += ( WS[i] - 8 ) * M[i], for i = [0, N) * This equation can rewritten as: S += WS[i] * M[i] - 8 * M[i], for i = [0, N) * 8 * M[i] need not be performed in the main loop. Also 8 * M[i], for i = [0, N) Can be substituted with A += M[i], for i = [0, N) and A *= 8 Thus, splitting parts of this equation results in a significant reduction in math ops while producing the same result. Differential Revision: D85721578
trivedivivek
force-pushed
the
export-D85721578
branch
2 times, most recently
from
1d800aa to
a54693c
Compare
trivedivivek
added a commit
to trivedivivek/executorch
that referenced
this pull request
Oct 29, 2025
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. The thinking here is as follows: * The 4 bit integer weights are unsigned ranging from 0 - 15, and thus to get unsigned number 8 is subtracted from the input. * Assume WS[] is array of signed weights, M[] is matrix, S is the sum The main loop essentially performs: S += ( WS[i] - 8 ) * M[i], for i = [0, N) * This equation can rewritten as: S += WS[i] * M[i] - 8 * M[i], for i = [0, N) * 8 * M[i] need not be performed in the main loop. Also 8 * M[i], for i = [0, N) Can be substituted with A += M[i], for i = [0, N) and A *= 8 Thus, splitting parts of this equation results in a significant reduction in math ops while producing the same result. Differential Revision: D85721578
SS-JIA
approved these changes
Oct 29, 2025
trivedivivek
added a commit
to trivedivivek/executorch
that referenced
this pull request
Oct 29, 2025
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. The thinking here is as follows: * The 4 bit integer weights are unsigned ranging from 0 - 15, and thus to get unsigned number 8 is subtracted from the input. * Assume WS[] is array of signed weights, M[] is matrix, S is the sum The main loop essentially performs: S += ( WS[i] - 8 ) * M[i], for i = [0, N) * This equation can rewritten as: S += WS[i] * M[i] - 8 * M[i], for i = [0, N) * 8 * M[i] need not be performed in the main loop. Also 8 * M[i], for i = [0, N) Can be substituted with A += M[i], for i = [0, N) and A *= 8 Thus, splitting parts of this equation results in a significant reduction in math ops while producing the same result. Reviewed By: SS-JIA Differential Revision: D85721578
trivedivivek
force-pushed
the
export-D85721578
branch
from
a54693c to
deca607
Compare
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. The thinking here is as follows: * The 4 bit integer weights are unsigned ranging from 0 - 15, and thus to get unsigned number 8 is subtracted from the input. * Assume WS[] is array of signed weights, M[] is matrix, S is the sum The main loop essentially performs: S += ( WS[i] - 8 ) * M[i], for i = [0, N) * This equation can rewritten as: S += WS[i] * M[i] - 8 * M[i], for i = [0, N) * 8 * M[i] need not be performed in the main loop. Also 8 * M[i], for i = [0, N) Can be substituted with A += M[i], for i = [0, N) and A *= 8 Thus, splitting parts of this equation results in a significant reduction in math ops while producing the same result. Reviewed By: SS-JIA Differential Revision: D85721578
trivedivivek
force-pushed
the
export-D85721578
branch
from
deca607 to
f0c9c4d
Compare
trivedivivek
added a commit
to trivedivivek/executorch
that referenced
this pull request
Oct 29, 2025
…peration. (pytorch#15436) Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime. The thinking here is as follows: * The 4 bit integer weights are unsigned ranging from 0 - 15, and thus to get unsigned number 8 is subtracted from the input. * Assume WS[] is array of signed weights, M[] is matrix, S is the sum The main loop essentially performs: S += ( WS[i] - 8 ) * M[i], for i = [0, N) * This equation can rewritten as: S += WS[i] * M[i] - 8 * M[i], for i = [0, N) * 8 * M[i] need not be performed in the main loop. Also 8 * M[i], for i = [0, N) Can be substituted with A += M[i], for i = [0, N) and A *= 8 Thus, splitting parts of this equation results in a significant reduction in math ops while producing the same result. Reviewed By: SS-JIA Differential Revision: D85721578
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.
Summary: This diff introduces performance improvements in the 4-bit quant matrix multiplication operation by adjusting the position of the -8 operation. Resulting in overall reduction in math operation performed during shader runtime.
Differential Revision: D85721578