Skip to content

Use order from A matrix when determining DPAS layout #2834

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
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Use order from A matrix when determining DPAS layout #2834

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
23 changes: 19 additions & 4 deletions third_party/intel/lib/TritonIntelGPUTransforms/AccelerateMatmul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,7 +32,8 @@ namespace {

SmallVector<unsigned>
getWarpsPerTile(tt::DotOp dotOp, ttgi::DpasEncodingAttr::DPASCapability dpasCap,
const ArrayRef<int64_t> shape, unsigned numWarps) {
const ArrayRef<int64_t> shape, unsigned numWarps,
const SmallVector<unsigned> &order) {
auto filter = [&dotOp](Operation *op) {
return op->getParentRegion() == dotOp->getParentRegion();
};
Expand Down Expand Up @@ -64,7 +65,7 @@ getWarpsPerTile(tt::DotOp dotOp, ttgi::DpasEncodingAttr::DPASCapability dpasCap,
uint32_t colRowRatio =
ceil<uint32_t>(dpasCap.executionSize, dpasCap.repeatCount);

int rowDim = rank - 2, colDim = rank - 1;
int rowDim = order[rank - 2], colDim = order[rank - 1];
do {
if (ret[rowDim] * ret[colDim] >= numWarps)
break;
Expand All @@ -78,7 +79,6 @@ getWarpsPerTile(tt::DotOp dotOp, ttgi::DpasEncodingAttr::DPASCapability dpasCap,
ret[colDim] *= 2;
}
} while (true);

return ret;
}

Expand Down Expand Up @@ -115,9 +115,24 @@ class BlockedToDPAS : public OpRewritePattern<tt::DotOp> {

auto dpasCap = ttgi::DpasEncodingAttr::getDPASCapability(mod);
Type elemType = oldAType.getElementType();

unsigned opsPerChan = ttgi::DpasEncodingAttr::getOpsPerChannel(elemType);

SmallVector<unsigned> order = {0, 1};
Operation *aOp = a.getDefiningOp();
if (aOp && isa<ttg::ConvertLayoutOp>(aOp)) {
auto valueToConvert = aOp->getOperand(0);
aOp = valueToConvert.getDefiningOp();
}
if (aOp && isa<tt::LoadOp>(aOp)) {
assert(aOp->getNumResults() == 1);
Attribute layout =
cast<RankedTensorType>(aOp->getResult(0).getType()).getEncoding();
order = triton::gpu::getOrder(layout);
}

SmallVector<unsigned> warpsPerTile =
getWarpsPerTile(dotOp, dpasCap, retShape, numWarps);
getWarpsPerTile(dotOp, dpasCap, retShape, numWarps, order);
size_t rank = retShape.size();
SmallVector<unsigned> repCluster(rank, 1);

Expand Down
Loading