Skip to content
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

Use stride instead of order to determine block attr #2349

Merged
merged 5 commits into from
Sep 26, 2024
Merged

Use stride instead of order to determine block attr #2349

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8a4cab2
Use stride instead of order to determine block attr
alexbaden Sep 25, 2024
c09e10e
fix pitch restriction check + remove commented code
alexbaden Sep 26, 2024
9f7bfdc
++i
alexbaden Sep 26, 2024
de99c71
remove unused variable
alexbaden Sep 26, 2024
5a23de4
format
alexbaden Sep 26, 2024
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: 17 additions & 6 deletions third_party/intel/lib/TritonIntelGPUTransforms/MaterializeBlockPointer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,17 +51,27 @@ struct TritonIntelGPUMaterializeBlockPointerPass
LDBG("Found make tensor ptr op: " << makeTensorPtrOp);
auto ptrType = cast<tt::PointerType>(makeTensorPtrOp.getType());
auto tensorType = cast<RankedTensorType>(ptrType.getPointeeType());
ArrayRef<int32_t> order = makeTensorPtrOp.getOrder();
unsigned rank = order.size();
Operation::operand_range shape = makeTensorPtrOp.getShape();
unsigned rank = shape.size();
LDBG("Rank: " << rank);
if (rank == 1)
return;

unsigned fastChangeDim = order[0];
Operation::operand_range strides = makeTensorPtrOp.getStrides();
int fastChangeDim = -1;
for (size_t i = 0; i < strides.size(); ++i) {
if (mlir::triton::gpu::intel::isConstant(strides[i], 1)) {
fastChangeDim = i;
break;
}
}

LDBG("Fast change dim: " << fastChangeDim);
if (fastChangeDim >= (rank - 2)) {
Operation::operand_range strides = makeTensorPtrOp.getStrides();
if (fastChangeDim < 0) {
return;
}

if (fastChangeDim >= (rank - 2)) {
// HW 2D block read instruction only supports contiguous access.
Value fastChangeStride = strides[fastChangeDim];
LLVM_DEBUG({
Expand All @@ -77,7 +87,8 @@ struct TritonIntelGPUMaterializeBlockPointerPass
Value pitch =
strides[(fastChangeDim == rank - 1) ? rank - 2 : rank - 1];
LDBG("Pitch: " << pitch);
if (!ttgi::isDivisible(pitch, 64 / tensorType.getElementTypeBitWidth()))
if (!ttgi::isDivisible(pitch,
128 / tensorType.getElementTypeBitWidth()))
return;

loadOp->setAttr(ttgi::TritonIntelGPUDialect::getBlockIOAttrName(),
Expand Down