Skip to content

[MPS] Optimize cummin/cummax metal kernels#156794

Closed
manuelcandales wants to merge 6 commits intogh/manuelcandales/12/basefrom
gh/manuelcandales/12/head
Closed

[MPS] Optimize cummin/cummax metal kernels#156794
manuelcandales wants to merge 6 commits intogh/manuelcandales/12/basefrom
gh/manuelcandales/12/head

Conversation

@manuelcandales
Copy link
Copy Markdown
Contributor

@manuelcandales manuelcandales commented Jun 25, 2025 •
edited
Loading

Stack from ghstack (oldest at bottom):

Performance improvement (M4 Max 64GB, macOS 15.5):

                                              | Current | Previous
      cummin-dim0-32x32 (torch.float16)       |  103.4  |   102.5
      cummin-dim0-128x128 (torch.float16)     |  112.2  |   133.6
      cummin-dim0-512x512 (torch.float16)     |  146.9  |   233.1
      cummin-dim0-1024x1024 (torch.float16)   |  193.6  |   364.2
      cummin-dim1-32x32 (torch.float16)       |  102.0  |    94.4
      cummin-dim1-128x128 (torch.float16)     |  103.0  |   109.9
      cummin-dim1-512x512 (torch.float16)     |  109.1  |   227.0
      cummin-dim1-1024x1024 (torch.float16)   |  140.5  |   985.1
      cummin-1d-100 (torch.float16)           |  101.8  |   100.7
      cummin-1d-10000 (torch.float16)         |  112.8  |   805.0
      cummin-1d-1000000 (torch.float16)       | 1343.8  | 70545.6
      cummin-dim0-32x32 (torch.float32)       |  104.6  |   102.7
      cummin-dim0-128x128 (torch.float32)     |  112.3  |   137.2
      cummin-dim0-512x512 (torch.float32)     |  146.6  |   209.7
      cummin-dim0-1024x1024 (torch.float32)   |  194.0  |   340.1
      cummin-dim1-32x32 (torch.float32)       |  100.1  |    99.2
      cummin-dim1-128x128 (torch.float32)     |  101.4  |   111.9
      cummin-dim1-512x512 (torch.float32)     |  110.3  |   250.7
      cummin-dim1-1024x1024 (torch.float32)   |  141.4  |   987.9
      cummin-1d-100 (torch.float32)           |  101.0  |   100.6
      cummin-1d-10000 (torch.float32)         |  112.9  |   794.7
      cummin-1d-1000000 (torch.float32)       | 1311.7  | 71995.3
      cummin-dim0-32x32 (torch.bfloat16)      |  105.8  |   105.9
      cummin-dim0-128x128 (torch.bfloat16)    |  111.9  |   135.7
      cummin-dim0-512x512 (torch.bfloat16)    |  147.1  |   231.9
      cummin-dim0-1024x1024 (torch.bfloat16)  |  191.2  |   327.7
      cummin-dim1-32x32 (torch.bfloat16)      |  101.8  |    91.3
      cummin-dim1-128x128 (torch.bfloat16)    |  100.2  |   108.5
      cummin-dim1-512x512 (torch.bfloat16)    |  108.9  |   222.0
      cummin-dim1-1024x1024 (torch.bfloat16)  |  140.1  |   936.9
      cummin-1d-100 (torch.bfloat16)          |  103.0  |   106.6
      cummin-1d-10000 (torch.bfloat16)        |  113.1  |   795.8
      cummin-1d-1000000 (torch.bfloat16)      | 1296.8  | 68667.4

@manuelcandales
Update
a3f232a 
[ghstack-poisoned]
manuelcandales added a commit that referenced this pull request Jun 25, 2025
@manuelcandales
[MPS] Optimize cummin/cummax metal kernels
8e0f31d 
ghstack-source-id: 85e2fdf
Pull-Request: #156794
@pytorch-bot
Copy link
Copy Markdown

pytorch-bot bot commented Jun 25, 2025 •
edited
Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/156794

Note: Links to docs will display an error until the docs builds have been completed.

❗ 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

⏳ No Failures, 8 Pending

As of commit 015bfc9 with merge base 0d8e4e2 (image):
💚 Looks good so far! There are no failures yet. 💚

UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@pytorch-bot pytorch-bot bot added ciflow/mps Run MPS tests (subset of trunk) release notes: mps Release notes category labels Jun 25, 2025
@manuelcandales
Update
1927a62 
[ghstack-poisoned]
@manuelcandales manuelcandales mentioned this pull request Jun 25, 2025
Closed
@manuelcandales
Update
0ef056e 
[ghstack-poisoned]
@manuelcandales manuelcandales mentioned this pull request Jun 25, 2025
Closed
@manuelcandales
Update
29b95b9 
[ghstack-poisoned]
@manuelcandales manuelcandales mentioned this pull request Jun 25, 2025
Closed
@manuelcandales manuelcandales mentioned this pull request Jun 25, 2025
Closed
malfet
malfet reviewed Jun 26, 2025
View reviewed changes
@manuelcandales
Update
8d469be 
[ghstack-poisoned]
@malfet malfet added the topic: performance topic category label Jun 26, 2025
@manuelcandales
Update
015bfc9 
[ghstack-poisoned]
manuelcandales added a commit that referenced this pull request Jun 26, 2025
@manuelcandales
[MPS] Optimize cummin/cummax metal kernels
4b39acf 
ghstack-source-id: ac0ce95
Pull-Request: #156794
@malfet
Copy link
Copy Markdown
Contributor

malfet commented Jun 26, 2025

@pytorchbot merge -f "Let's land in trunk"

@pytorchmergebot
Copy link
Copy Markdown
Collaborator

pytorchmergebot commented Jun 26, 2025

Merge started

Your change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use -f as last resort and instead consider -i/--ignore-current to continue the merge ignoring current failures. This will allow currently pending tests to finish and report signal before the merge.

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging
Check the merge workflow status
here

@manuelcandales manuelcandales mentioned this pull request Jun 16, 2025
Open
@github-actions github-actions bot deleted the gh/manuelcandales/12/head branch July 27, 2025 02:21
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@malfet malfet malfet approved these changes

@kulinseth kulinseth Awaiting requested review from kulinseth

Assignees

No one assigned

Labels

ciflow/mps Run MPS tests (subset of trunk) Merged release notes: mps Release notes category topic: performance topic category

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@manuelcandales @malfet @pytorchmergebot