gpu: a GEMM with 128×128 tiles to measure the Ampere strategy

Open
#3,442 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
25/100
Issue type
Feature
Clarity
Mostly clear
Activity status
Active
Tech stack
rust
Domain
performance

Research direction

Start from the existing v5 GEMM linked in #3400 and the copy pipeline in #3412; the issue names no files, so locate the 64x64/4-warp kernel and its benchmark first. Implement a v8 block tile of 128x128 with 8 warps, each warp taking 64x32, then benchmark on the RTX A6000 with and without the #3412 pipeline, comparing TFLOP/s and cuBLAS percentage. Done means v8 compiles and both measurements are reported; if register pressure or a compiler pass blocks the larger tile, file that as a separate issue.

Written by the indexing model from the issue text.

Description

The Vx GEMMs in #3400 compute a 64×64 tile of C in each block, with 4 warps. On an RTX A6000, v5 reaches 50 TFLOP/s, 44% of cuBLAS (https://github.com/vx-lang/Vx/issues/3400#issuecomment-6883701143). cuBLAS uses tiles of 128×128 or larger. A 64×64 tile does too little math for each byte it loads into shared memory.

Write a v8 that computes a 128×128 tile in each block, with 8 warps, each warp taking 64×32. Then measure it with and without the copy pipeline of #3412. With larger tiles, fewer blocks fit on each SM, so other blocks can no longer use the tensor cores while one block waits for its copies. That is the case where the pipeline should pay off.

If v8 hits a compiler limit, such as register use or a pass that does not fit the larger tile, file it as its own issue.

Dominant language
Rust
Stars
35
Forks
5
Avg merge
12h 26m
Merged PRs (30d)
799

Getting set up

We have not checked this project's setup files yet. Start from its README, and see our first-contribution guide for the general steps.

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.

More from vx-lang/Vx

All issues in vx-lang/Vx

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.