NVIDIA / NVIDIA/cutlass

Why not [RSC] but [C/64, R, S, 64] in kloop of conv implicit gemm?

Open
#1,797 4 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

? - Needs Triage inactive-30d inactive-90d question
Dominant language
C++
Stars
10.5k
Forks
2.1k
Avg merge
3d 11h
Merged PRs (30d)
7

Description

suppose R=2, S=2, C=128, ThreadBlockShape=<128,128,64>.
When main loop in implicit GEMM_K, the memory access sequence in cutlass will be r=0,s=0,c=0-63, r=0,s=1,c=0-63, r=1,s=0,c=0-63, r=1,s=1,c=0-63, r=0,s=0,c=63-127, r=0,s=1,c=63-127, r=1,s=0,c=63-127, r=1,s=1,c=63-127.
But why not r=0,s=0,c=0-63, r=0,s=0,c=63-127, r=0,s=1,c=0-63, r=0,s=1,c=63-127 r=1,s=0,c=0-63, r=1,s=0,c=63-127, r=1,s=1,c=0-63, r=1,s=1,c=63-127, Isn't it a better memory access strategy to access C first?

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start at the implicit GEMM_K main loop and inspect how ThreadBlockShape<128,128,64> determines the r, s, and c traversal order. Compare the reported [RSC] sequence with the proposed [C/64, R, S, 64] sequence, then use the relevant CUTLASS performance measurements to determine whether the alternative improves memory access; done means a measured conclusion or a clearly scoped change.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
performance
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.