Skip to content

[Backend] Fix incorrect shared layout for dot operands rank==3 #4944

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

Closed
AlexAUT wants to merge 1 commit into from
Closed

[Backend] Fix incorrect shared layout for dot operands rank==3 #4944

Show file tree
Hide file tree
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
18 changes: 12 additions & 6 deletions lib/Conversion/TritonGPUToLLVM/DecomposeUnsupportedConversions.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,8 +87,8 @@ void decomposeBlockedToDotLayoutConversion(ModuleOp module) {
return;
auto srcBlocked =
dyn_cast<triton::gpu::BlockedEncodingAttr>(srcType.getEncoding());
auto dstDotOp =
dyn_cast<triton::gpu::DotOperandEncodingAttr>(dstType.getEncoding());
auto dstEncoding = dstType.getEncoding();
auto dstDotOp = dyn_cast<triton::gpu::DotOperandEncodingAttr>(dstEncoding);
if (srcBlocked && dstDotOp) {
// FIXME [Dot LL]
// We support this one via LLs, as the LocalLoad path is buggy
Expand All @@ -99,15 +99,21 @@ void decomposeBlockedToDotLayoutConversion(ModuleOp module) {
return;
}
}

auto srcOrder = triton::gpu::getOrder(srcBlocked);
auto rank = srcOrder.size();
SmallVector<unsigned> sharedOrder;
if (rank == 3) {
sharedOrder = gpu::getThreadOrder(dstEncoding);
} else {
sharedOrder = srcOrder;
}
Comment on lines +105 to +109
Copy link
Copy Markdown
Contributor

@lezcano lezcano Oct 18, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If you are going this route, you probably want to do it for all ranks, otherwise this heuristic would be incredibly counterintuitive.

Copy link
Copy Markdown
Contributor Author

@AlexAUT AlexAUT Oct 19, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

But then we might change the shared layout order for rank != 3 with this change? I can also revert all changes except for the condition in ReduceDataDuplication to make it ampere specific. Then we have the same behavior as before #4904.

Copy link
Copy Markdown
Contributor

@lezcano lezcano Oct 20, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I just find very weird the current behaviour for rank==3. @Jokeren thoughts?

Copy link
Copy Markdown
Contributor

@Jokeren Jokeren Oct 20, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actually block to dot decomposition can be deprecated soon
main...keren/dot-mma#diff-30fb59df648f6c4ec5db24c51ef728a8d56104e151bd52e8d482b6951c207aa7R95

Copy link
Copy Markdown
Contributor

@Jokeren Jokeren Oct 20, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree the special condition is weird. Can you attach a test case for us to take a look at?

Copy link
Copy Markdown
Contributor Author

@AlexAUT AlexAUT Oct 21, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When you run python/test/unit/language/test_core.py::test_dot3d[1-1-64-64-64-32-32-float16-float32] it will trigger this assert. This happens when wmma is used for the dot. The shared -> dot layout conversion for wmma will also expect that the batch dim is the slowest dimension.

Attribute sharedMemorySpace =
triton::gpu::SharedMemorySpaceAttr::get(srcType.getContext());
auto tmpType = MemDescType::get(
dstType.getShape(), dstType.getElementType(),
triton::gpu::SharedEncodingAttr::get(
module.getContext(), dstDotOp, srcType.getShape(),
srcBlocked.getOrder(), srcBlocked.getCTALayout(),
srcType.getElementType()),
module.getContext(), dstDotOp, srcType.getShape(), sharedOrder,
srcBlocked.getCTALayout(), srcType.getElementType()),
sharedMemorySpace);
auto tmp = builder.create<triton::gpu::LocalAllocOp>(
cvtOp.getLoc(), tmpType, cvtOp.getSrc());
Expand Down
19 changes: 9 additions & 10 deletions lib/Dialect/TritonGPU/Transforms/ReduceDataDuplication.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,30 +36,29 @@ class TritonGPUReduceDataDuplicationPass
auto srcType = cast<RankedTensorType>(cvtOp.getSrc().getType());
auto dstType = cast<RankedTensorType>(cvtOp.getType());
auto srcEncoding = srcType.getEncoding();
auto dstEncoding = dstType.getEncoding();
if (isa<triton::gpu::SharedEncodingAttr>(srcEncoding))
return;
auto dstDotOp =
dyn_cast<triton::gpu::DotOperandEncodingAttr>(dstType.getEncoding());
dyn_cast<triton::gpu::DotOperandEncodingAttr>(dstEncoding);
if (!dstDotOp)
return;
if (!cvtNeedsSharedMemory(srcType, dstType))
return;
// FIXME [Dot LL]
// We support this one via LLs, as the LocalLoad path is buggy
bool largeKWidth =
dstDotOp.getKWidth() * dstType.getElementTypeBitWidth() > 64;
if (largeKWidth) {
return;
if (auto mma = dyn_cast<NvidiaMmaEncodingAttr>(dstDotOp.getParent())) {
bool largeKWidth =
dstDotOp.getKWidth() * dstType.getElementTypeBitWidth() > 64;
if (mma.isAmpere() && largeKWidth) {
return;
}
}
auto srcOrder = triton::gpu::getOrder(srcEncoding);
auto rank = srcOrder.size();
SmallVector<unsigned> sharedOrder;
if (rank == 3) {
// add all elements except the element that is zero
for (unsigned i = 0; i < rank; ++i)
if (srcOrder[i] != 0)
sharedOrder.emplace_back(srcOrder[i]);
sharedOrder.emplace_back(0);
sharedOrder = gpu::getThreadOrder(dstEncoding);
} else {
sharedOrder = srcOrder;
}
Expand Down