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

Implement DotOperandEncodingAttr::getSizePerThread with block layout parent #5863

Merged
merged 3 commits into from
Feb 11, 2025
Merged

Implement DotOperandEncodingAttr::getSizePerThread with block layout parent #5863

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f74656b
Implement DotOperandEncodingAttr::getSizePerThread with block layout …
anmyachev Feb 8, 2025
4d2acb5
update 'getSizePerThread' impl && add dialect test
anmyachev Feb 10, 2025
16dd3ba
insert dot_op with blocked parent into 'distributedEncodings' vector;…
anmyachev Feb 10, 2025
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
5 changes: 3 additions & 2 deletions lib/Dialect/TritonGPU/IR/Dialect.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2213,10 +2213,11 @@ SmallVector<unsigned> DotOperandEncodingAttr::getSizePerThread() const {
assert(parentLayout && "DotOperandEncodingAttr must have a parent");
if (auto parentMmaLayout = mlir::dyn_cast<MmaEncodingTrait>(parentLayout)) {
return parentMmaLayout.getSizePerThreadForOperand(getKWidth(), getOpIdx());
} else if (auto blocked = mlir::dyn_cast<BlockedEncodingAttr>(parentLayout)) {
return blocked.getSizePerThread();
} else {
llvm::report_fatal_error(
"DotOperandEncodingAttr non-NvidiaMmaEncodingAttr parent not "
"supported yet");
"getSizePerThread not implemented for DotOperandEncodingAttr");
return {};
}
}
Expand Down
40 changes: 30 additions & 10 deletions unittest/Dialect/TritonGPU/DialectTest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -516,6 +516,12 @@ TEST_F(LinearEncodingTest, DistributedEncodingToLinearEncoding) {
distributedEncodings.push_back(blockedEncoding);
distributedEncodings.push_back(
triton::gpu::SliceEncodingAttr::get(&ctx, 0, blockedEncoding));
// Create an opIdx=0 and opIdx=1 encoding
for (unsigned opIdx = 0; opIdx < 2; ++opIdx) {
distributedEncodings.push_back(
triton::gpu::DotOperandEncodingAttr::get(&ctx, opIdx,
blockedEncoding, 0));
}
}
}
}
Expand All @@ -538,6 +544,12 @@ TEST_F(LinearEncodingTest, DistributedEncodingToLinearEncoding) {
}
}

auto is_dot_op_with_block_parent = [](Attribute layout) {
auto dot_layout = dyn_cast<triton::gpu::DotOperandEncodingAttr>(layout);
return dot_layout &&
isa<triton::gpu::BlockedEncodingAttr>(dot_layout.getParent());
};

for (const auto &distributedEncoding : distributedEncodings) {
for (auto shape : shapes) {
if (auto sliceEncoding =
Expand All @@ -558,29 +570,37 @@ TEST_F(LinearEncodingTest, DistributedEncodingToLinearEncoding) {
// Test that methods of DistributedEncoding return the same values
Type eltTy = Float32Type::get(&ctx);

ASSERT_EQ(getOrder(distributedEncoding), linearEncoding.getRepOrder());
if (!is_dot_op_with_block_parent(distributedEncoding)) {
ASSERT_EQ(getOrder(distributedEncoding), linearEncoding.getRepOrder());
Copy link
Contributor Author

Choose a reason for hiding this comment

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

Wrong order

}
ASSERT_EQ(distributedEncoding.getTotalElemsPerThread(shape),
linearEncoding.getTotalElemsPerThread(shape));
ASSERT_EQ(distributedEncoding.getElemsPerThread(shape),
linearEncoding.getElemsPerThread(shape));
ASSERT_EQ(distributedEncoding.getRepOrder(),
linearEncoding.getRepOrder());
ASSERT_EQ(distributedEncoding.getContigPerThread(),
linearEncoding.getContigPerThread());
if (!is_dot_op_with_block_parent(distributedEncoding)) {
ASSERT_EQ(distributedEncoding.getRepOrder(),
Copy link
Contributor Author

Choose a reason for hiding this comment

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

wrong order

linearEncoding.getRepOrder());
ASSERT_EQ(distributedEncoding.getContigPerThread(),
Copy link
Contributor Author

Choose a reason for hiding this comment

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

llvm::SmallVector<unsigned int> mlir::triton::gpu::DotOperandEncodingAttr::getContigPerThread(): Assertion kWidth !=0 && "Do not support kWidth=0"' failed.

linearEncoding.getContigPerThread());
}
// DotOperandEncodingAttr::getWarpOrder() is not defined
if (!isa<triton::gpu::DotOperandEncodingAttr>(distributedEncoding)) {
ASSERT_EQ(distributedEncoding.getWarpOrder(),
linearEncoding.getWarpOrder());
}
ASSERT_EQ(distributedEncoding.getThreadOrder(),
linearEncoding.getThreadOrder());
if (!is_dot_op_with_block_parent(distributedEncoding)) {
ASSERT_EQ(distributedEncoding.getThreadOrder(),
Copy link
Contributor Author

Choose a reason for hiding this comment

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

Wrong order: { 1, 0 } vs { 0, 1 }

linearEncoding.getThreadOrder());
}
// For slice these do not equal the total number of lines / warps
// See [Note. Divergence of methods wrt. legacy layouts]
if (!isa<triton::gpu::SliceEncodingAttr>(distributedEncoding)) {
ASSERT_EQ(distributedEncoding.getWarpsPerCTA(),
linearEncoding.getWarpsPerCTA());
ASSERT_EQ(distributedEncoding.getThreadsPerWarp(),
linearEncoding.getThreadsPerWarp());
if (!is_dot_op_with_block_parent(distributedEncoding)) {
ASSERT_EQ(distributedEncoding.getThreadsPerWarp(),
Copy link
Contributor Author

Choose a reason for hiding this comment

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

LLVM ERROR: getThreadsPerWarp not implemented for DotOperandEncodingAttr

linearEncoding.getThreadsPerWarp());
}
}
// Canonicalisation for opIdx=0 takes just a [2 x 2] subtile as it takes
// the second repetition along K as the second tile.
Expand All @@ -602,7 +622,7 @@ TEST_F(LinearEncodingTest, DistributedEncodingToLinearEncoding) {
// If we are not using CGAs, the order is meaningless
auto useCGA =
baseEncoding.getCTAsPerCGA() != SmallVector<unsigned>(rank, 1);
if (useCGA) {
if (useCGA && !is_dot_op_with_block_parent(distributedEncoding)) {
Copy link
Contributor Author

Choose a reason for hiding this comment

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

Wrong order: { 1, 0 } vs { 0, 1 }

ASSERT_EQ(baseEncoding.getCTAOrder(), linearEncoding.getCTAOrder());
}
}
Expand Down
Loading