Skip to content

[Operators] Improving fp32 matrix multiplication on x86 CPUs#378

Merged
yaoyaoding merged 141 commits intohidet-org:mainfrom
BolinSNLHM:main
Dec 14, 2023
Merged

[Operators] Improving fp32 matrix multiplication on x86 CPUs#378
yaoyaoding merged 141 commits intohidet-org:mainfrom
BolinSNLHM:main

Conversation

@BolinSNLHM
Copy link
Copy Markdown
Contributor

No description provided.

@yaoyaoding
Copy link
Copy Markdown
Member

Hi @BolinSNLHM,

Thanks a lot for the contribution! I will review it soon. Before that, could you spend some time on benchmark on the new implementation vs. baselines (e.g., pytorch and tvm).

@BolinSNLHM
Copy link
Copy Markdown
Contributor Author

BolinSNLHM commented Nov 27, 2023

The preliminary performance results of the improved version of the matmul operator:
matmul-performance

(The results for the 1x1024x1024, as the bars are too low to be visible:
Hidet: 0.22ms; Ansor: 0.04ms; Torch: 0.08ms)

Copy link
Copy Markdown
Member

@yaoyaoding yaoyaoding left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks @BolinSNLHM !

It looks good to me.

@@ -0,0 +1,148 @@
import numpy as np
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Remove this file.


super().__init__(
name='matmul_f32_x86',
name='matmul_f32_x86_v2',
Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Maybe still use matmul_f32_x86?

@yaoyaoding
Copy link
Copy Markdown
Member

Thanks @BolinSNLHM !

@yaoyaoding yaoyaoding merged commit 264beec into hidet-org:main Dec 14, 2023
vadiklyutiy pushed a commit that referenced this pull request Dec 26, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants