Skip to content

[MLIR][XeGPU] Use operand layouts for store scatter #161447

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

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

[MLIR][XeGPU] Use operand layouts for store scatter #161447

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
24 changes: 14 additions & 10 deletions mlir/lib/Dialect/XeGPU/Transforms/XeGPUWgToSgDistribute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -824,7 +824,7 @@ struct WgToSgStoreScatterOpWithOffset
return failure();

xegpu::DistributeLayoutAttr layout =
xegpu::getDistributeLayoutAttr(op.getValue());
Copy link
Contributor

Choose a reason for hiding this comment

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

this should also work. why this change is needed?

i.e, these two calls should return the same layout.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

yes, it should work as well, but I changed it just to be consistent to call getter and setter on operand/s for store..

xegpu::getDistributeLayoutAttr(op.getOperand(0));
if (!layout || !layout.isForWorkgroup())
return failure();

Expand All @@ -844,12 +844,19 @@ struct WgToSgStoreScatterOpWithOffset
auto chunkSizeAttr = rewriter.getI64IntegerAttr(chunkSize);
for (auto [val, offs, mask] : llvm::zip(
adaptor.getValue(), adaptor.getOffsets(), adaptor.getMask())) {
xegpu::StoreScatterOp::create(rewriter, loc, val, op.getDest(), offs,
mask, chunkSizeAttr, op.getL1HintAttr(),
op.getL2HintAttr(), op.getL3HintAttr());
auto store = xegpu::StoreScatterOp::create(
rewriter, loc, val, op.getDest(), offs, mask, chunkSizeAttr,
op.getL1HintAttr(), op.getL2HintAttr(), op.getL3HintAttr());
// Update the layout attribute to drop sg_layout and sg_data.
if (auto newLayout = layout.dropSgLayoutAndData())
op->setAttr("layout", newLayout);
if (!layout.getEffectiveLaneLayoutAsInt().empty() ||
!layout.getEffectiveInstDataAsInt().empty()) {
for (OpOperand &operand : store->getOpOperands()) {
// Skip for operand one (memref)
if (operand.getOperandNumber() == 1)
continue;
xegpu::setDistributeLayoutAttr(operand, layout.dropSgLayoutAndData());
}
}
}
rewriter.eraseOp(op);
return success();
Expand Down Expand Up @@ -1247,10 +1254,7 @@ void XeGPUWgToSgDistributePass::runOnOperation() {

target.addDynamicallyLegalOp<xegpu::StoreScatterOp>(
[=](xegpu::StoreScatterOp op) -> bool {
// Check if the layout attribute is present on the result.
auto layout = op->getAttrOfType<xegpu::LayoutAttr>("layout");
if (!layout)
return true;
auto layout = xegpu::getDistributeLayoutAttr(op.getOperand(0));
return isLegal(layout);
});

Expand Down
19 changes: 12 additions & 7 deletions mlir/test/Dialect/XeGPU/xegpu-wg-to-sg-unify-ops.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -282,15 +282,20 @@ gpu.module @test_distribution {
// CHECK-LABEL: @store_scatter
// CHECK-SAME: %[[ARG0:.*]]: memref<256xf16>
gpu.func @store_scatter(%dest : memref<256xf16>) {
// CHECK: %[[VAL:.*]] = arith.constant dense<2.550000e+01> : vector<8xf16>
// CHECK: %[[CST:.*]] = arith.constant dense<0> : vector<8xindex>
// CHECK: %[[MASK:.*]] = arith.constant dense<true> : vector<8xi1>
// CHECK: %[[VAL:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<2.550000e+01> : vector<8xf16>
// CHECK: %[[CST:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<0> : vector<8xindex>
// CHECK: %[[MASK:.*]] = arith.constant {layout_result_0 = #xegpu.layout<inst_data = [8]>} dense<true> : vector<8xi1>
// CHECK: xegpu.store %[[VAL]], %[[ARG0]][%[[CST]]], %[[MASK]] <{chunk_size = 1 : i64, l1_hint = #xegpu.cache_hint<cached>}>
// CHECK-SAME: {layout_operand_0 = #xegpu.layout<inst_data = [8]>, layout_operand_2 = #xegpu.layout<inst_data = [8]>,
// CHECK-SAME: layout_operand_3 = #xegpu.layout<inst_data = [8]>}
// CHECK-SAME: : vector<8xf16>, memref<256xf16>, vector<8xindex>, vector<8xi1>
%val = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8]>} dense<25.5> : vector<256xf16>
%offset = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8]>} dense<0> : vector<256xindex>
%mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8]>} dense<1> : vector<256xi1>
xegpu.store %val, %dest[%offset], %mask {chunk_size = 1, layout = #xegpu.layout<sg_layout = [32], sg_data = [8]>, l1_hint = #xegpu.cache_hint<cached>}
%val = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<25.5> : vector<256xf16>
%offset = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<0> : vector<256xindex>
%mask = arith.constant {layout_result_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>} dense<1> : vector<256xi1>
xegpu.store %val, %dest[%offset], %mask {chunk_size = 1, layout_operand_0 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>,
layout_operand_2 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>,
layout_operand_3 = #xegpu.layout<sg_layout = [32], sg_data = [8], inst_data = [8]>,
l1_hint = #xegpu.cache_hint<cached>}
: vector<256xf16>, memref<256xf16>, vector<256xindex>, vector<256xi1>
gpu.return
}
Expand Down