Skip to content

Rock: cap K-copy vectorization at kPerBlock (fix int4 M=1 GEMV on WMMA) #2463

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
rlegithub wants to merge 1 commit into
base: develop
Choose a base branch
from
+7 −1
Draft

Rock: cap K-copy vectorization at kPerBlock (fix int4 M=1 GEMV on WMMA) #2463

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 7 additions & 1 deletion mlir/lib/Dialect/Rock/utility/loweringUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1142,7 +1142,13 @@ computeCopyPerThread(Type elementType, int64_t copyPerThread, int64_t kPerBlock,
copyKPerThread = copyPerThread / copyDPerThread;
dim = GemmDimension::MorN;
} else {
copyKPerThread = math_util::gcd(maxVlen, copyPerThread);
// Cap the K-vectorization at kPerBlock: with narrow types (e.g. int4,
// maxVlen=32) the greedy gcd(maxVlen, copyPerThread) can exceed kPerBlock
// (e.g. 16), which makes a thread copy more K elements than the block tile
// holds -> "incoherent tuning parameters". Capping keeps the copy coherent
// (still divides copyPerThread and maxVlen; thread tiling is unchanged since
// copyKPerThread*copyDPerThread == copyPerThread).
copyKPerThread = math_util::gcd(math_util::gcd(maxVlen, copyPerThread), kPerBlock);
copyDPerThread = copyPerThread / copyKPerThread;
dim = GemmDimension::K;
}
Expand Down
Loading