Skip to content

Changing sgpr limits#2184

Closed
mahmoodw wants to merge 2 commits into
ROCm:developfrom
mahmoodw:sgprLimit
Closed

Changing sgpr limits#2184
mahmoodw wants to merge 2 commits into
ROCm:developfrom
mahmoodw:sgprLimit

Conversation

@mahmoodw
Copy link
Copy Markdown
Contributor

@mahmoodw mahmoodw commented Jun 4, 2025

In regards to LWPTENSILE-1696

This includes 2 changes:

  • Unrestricted the temp sgprs needed for gsu from being contiguous, avoiding overflow for certain kernels
  • Account for additional temp sgprs that will be required for code gen, up to physical limits

KKyang
KKyang previously approved these changes Jun 5, 2025
@msujon-AMD msujon-AMD added the gfx950 run CI on gfx950 label Jun 5, 2025
AlexBrownAMD
AlexBrownAMD previously approved these changes Jun 5, 2025
@mahmoodw mahmoodw dismissed stale reviews from AlexBrownAMD and KKyang via 7b85b73 June 16, 2025 20:11
@mahmoodw
Copy link
Copy Markdown
Contributor Author

/AzurePipelines run

@azure-pipelines
Copy link
Copy Markdown

Azure Pipelines successfully started running 1 pipeline(s).

# Use VGPR up to next occupancy threshold:
maxVgprs, occupancy = self.getMaxRegsForOccupancy(kernel["NumThreads"], self.vgprPool.size(), self.sgprPool.size(), \
# Account for additional temp sgprs that will be required for code gen, up to physical limits. +5 approximates upper end of required temp space for GSU sync
requiredSgprs = min(self.sgprPool.size() + 5, self.states.regCaps["MaxSgpr"])
Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

"+5" should not be required for all kernels.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

That's right, this is meant to be a temporary fix to unblock perf teams. A debt ticket will be issued to more accurately collect the high-water mark. Any suggestions or insight is appreciated.

@jayhawk-commits
Copy link
Copy Markdown
Contributor

Closing the pull request in this repo. Please refer to the migrated pull request for updates.

mahmoodw added a commit to ROCm/rocm-libraries that referenced this pull request Jul 3, 2025
In regards to LWPTENSILE-1696

This includes 2 changes:
- Unrestricted the temp sgprs needed for gsu from being contiguous,
avoiding overflow for certain kernels
- Account for additional temp sgprs that will be required for code gen,
up to physical limits

---
🔁 Imported from
[ROCm/hipBLASLt#2184](ROCm/hipBLASLt#2184)
🧑‍💻 Originally authored by @mahmoodw

---------

Co-authored-by: mahmoodw <wmahmood@amd.com>
Co-authored-by: mahmoodw <44450175+mahmoodw@users.noreply.github.com>
AlexBrownAMD pushed a commit to ROCm/rocm-libraries that referenced this pull request Jul 9, 2025
In regards to LWPTENSILE-1696

This includes 2 changes:
- Unrestricted the temp sgprs needed for gsu from being contiguous,
avoiding overflow for certain kernels
- Account for additional temp sgprs that will be required for code gen,
up to physical limits

---
🔁 Imported from
[ROCm/hipBLASLt#2184](ROCm/hipBLASLt#2184)
🧑‍💻 Originally authored by @mahmoodw

---------

Co-authored-by: mahmoodw <wmahmood@amd.com>
Co-authored-by: mahmoodw <44450175+mahmoodw@users.noreply.github.com>
mahmoodw added a commit to ROCm/rocm-libraries that referenced this pull request Jul 10, 2025
This includes 2 changes:
- Unrestricted the temp sgprs needed for gsu from being contiguous,
avoiding overflow for certain kernels
- Account for additional temp sgprs that will be required for code gen,
up to physical limits

---
🔁 Imported from
[ROCm/hipBLASLt#2184](ROCm/hipBLASLt#2184) 🧑‍💻
Originally authored by @mahmoodw

---------

Co-authored-by: assistant-librarian[bot] <210906412+assistant-librarian[bot]@users.noreply.github.com>
Co-authored-by: mahmoodw <wmahmood@amd.com>
AlexBrownAMD pushed a commit to ROCm/rocm-libraries that referenced this pull request Jul 15, 2025
This includes 2 changes:
- Unrestricted the temp sgprs needed for gsu from being contiguous,
avoiding overflow for certain kernels
- Account for additional temp sgprs that will be required for code gen,
up to physical limits

---
🔁 Imported from
[ROCm/hipBLASLt#2184](ROCm/hipBLASLt#2184) 🧑‍💻
Originally authored by @mahmoodw

---------

Co-authored-by: assistant-librarian[bot] <210906412+assistant-librarian[bot]@users.noreply.github.com>
Co-authored-by: mahmoodw <wmahmood@amd.com>
SathiyarajRam pushed a commit to ROCm/rocm-libraries that referenced this pull request Jul 15, 2025
This includes 2 changes:
- Unrestricted the temp sgprs needed for gsu from being contiguous,
avoiding overflow for certain kernels
- Account for additional temp sgprs that will be required for code gen,
up to physical limits

---
🔁 Imported from
[ROCm/hipBLASLt#2184](ROCm/hipBLASLt#2184) 🧑‍💻
Originally authored by @mahmoodw

---------

Co-authored-by: assistant-librarian[bot] <210906412+assistant-librarian[bot]@users.noreply.github.com>
Co-authored-by: mahmoodw <wmahmood@amd.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

gfx950 run CI on gfx950

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants