Skip to content

[DRAFT] Move quantized dot operands in LDS in pipeliner #882

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
alefimov-amd wants to merge 1 commit into
base: pytorch/rocm7.1_internal_testing
Choose a base branch
from
Draft

[DRAFT] Move quantized dot operands in LDS in pipeliner #882

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
21 changes: 21 additions & 0 deletions third_party/amd/lib/TritonAMDGPUTransforms/StreamPipeline.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -278,6 +278,27 @@ getSharedEncIfAllUsersAreDotEnc(Value loadedValue) {
ctaLayout, opIdx, srcTy.getShape(), order, vecSize, bitWidth,
/*needTrans=*/false);
}
} else {
auto loadEncoding = dyn_cast<ttg::BlockedEncodingAttr>(
dyn_cast<RankedTensorType>(loadedValue.getType()).getEncoding());
auto sizePerThread = loadEncoding.getSizePerThread();
auto threadsShape = loadEncoding.getThreadsPerWarp();
auto order = loadEncoding.getOrder();

int vecSize = sizePerThread[order[0]];
int numBanks = 64;
int perPhase =
(32 * numBanks) / (srcTy.getShape()[order[0]] * bitWidth);
if (perPhase == 0)
perPhase = 1;
int maxPhase = threadsShape[order[1]] / perPhase;
vecSize = sizePerThread[order[0]];
tempAttr = ttg::SwizzledSharedEncodingAttr::get(
loadedValue.getContext(), vecSize, perPhase, maxPhase, order,
loadEncoding.getCTALayout());
LDBG("Deduced shared encoding candidate from blocked layout: "
<< tempAttr);
sharedEncs.push_back(tempAttr);
}
}
// Check that the shared encodings needed by the users are compatible.
Expand Down
Loading