Skip to content

#sdy Add all-to-all fusion canonicalization pattern. #516

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

#sdy Add all-to-all fusion canonicalization pattern. #516

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
50 changes: 49 additions & 1 deletion shardy/dialect/sdy/ir/canonicalization.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -314,6 +314,54 @@ class ReduceScatterFusion : public OpRewritePattern<AllSliceOp> {
}
};

class AlltoAllFusion : public OpRewritePattern<AllToAllOp> {
public:
using OpRewritePattern<AllToAllOp>::OpRewritePattern;

private:
LogicalResult matchAndRewrite(AllToAllOp allToAllOp,
PatternRewriter& rewriter) const override {
if (range_size(allToAllOp->getUsers()) != 1) {
return rewriter.notifyMatchFailure(
allToAllOp, "op has multiple users");
}
auto userAllToAllOp = dyn_cast<AllToAllOp>(*allToAllOp->user_begin());
if (!userAllToAllOp) {
return rewriter.notifyMatchFailure(
allToAllOp, "user is not all-to-all");
}
// Combine the params of the two all-to-all ops into one.
SmallVector<AllToAllParamAttr> combinedParams;
combinedParams.reserve(allToAllOp.getParams().size() +
userAllToAllOp.getParams().size());
combinedParams.append(allToAllOp.getParams().begin(),
allToAllOp.getParams().end());
combinedParams.append(userAllToAllOp.getParams().begin(),
userAllToAllOp.getParams().end());
// Check for overlap in the source and target dimensions.
BitVector seenDims(getTensorRank(allToAllOp.getResult()));
for (AllToAllParamAttr param : combinedParams) {
for (int64_t dim : {param.getSrcDim(), param.getTgtDim()}) {
if (seenDims.test(dim)) {
return rewriter.notifyMatchFailure(
allToAllOp, "overlapping dimensions in the combined parameters");
}
seenDims.set(dim);
}
}
llvm::sort(combinedParams,
[](const AllToAllParamAttr& a, const AllToAllParamAttr& b) {
return a.getSrcDim() < b.getSrcDim();
});
rewriter.replaceOpWithNewOp<AllToAllOp>(
userAllToAllOp, userAllToAllOp.getResult().getType(),
allToAllOp.getTensor(), combinedParams,
userAllToAllOp.getOutSharding());
rewriter.eraseOp(allToAllOp);
return success();
}
};

} // namespace

void ManualComputationOp::getCanonicalizationPatterns(
Expand Down Expand Up @@ -345,7 +393,7 @@ void AllReduceOp::getCanonicalizationPatterns(RewritePatternSet& results,

void AllToAllOp::getCanonicalizationPatterns(RewritePatternSet& results,
MLIRContext* context) {
// We don't have patterns for all-to-all for now.
results.add<AlltoAllFusion>(context);
}

void CollectivePermuteOp::getCanonicalizationPatterns(
Expand Down
31 changes: 31 additions & 0 deletions shardy/dialect/sdy/ir/test/collective_canonicalization.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -222,3 +222,34 @@ func.func @reduce_scatter_fusion_no_subaxis_prefix_match(%arg0 : tensor<64x16xf3
%1 = sdy.all_slice [{"r", "x", "z"}, {"y", "q"}] %0 out_sharding=<@mesh2, [{"r", "x", "z"}, {"y", "q"}]> : tensor<64x16xf32>
return %1 : tensor<64x16xf32>
}

// CHECK-LABEL: func @all_to_all_fusion_success
func.func @all_to_all_fusion_success(%arg0 : tensor<64x16x8x8xf32> {sdy.sharding=#sdy.sharding<@mesh, [{"x"}, {"y"}, {}, {}]>}) -> tensor<64x16x8x8xf32> {
// CHECK-NEXT: %0 = sdy.all_to_all [{"x"}: 0->2, {"y"}: 1->3] %arg0 out_sharding=<@mesh, [{}, {}, {"x"}, {"y"}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: return %0 : tensor<64x16x8x8xf32>
%0 = sdy.all_to_all [{"x"}: 0->2] %arg0 out_sharding=<@mesh, [{}, {"y"}, {"x"}, {}]> : tensor<64x16x8x8xf32>
%1 = sdy.all_to_all [{"y"}: 1->3] %0 out_sharding=<@mesh, [{}, {}, {"x"}, {"y"}]> : tensor<64x16x8x8xf32>
return %1 : tensor<64x16x8x8xf32>
}

// CHECK-LABEL: func @all_to_all_fusion_overlapping_dims
func.func @all_to_all_fusion_overlapping_dims(%arg0 : tensor<64x16x8x8xf32> {sdy.sharding=#sdy.sharding<@mesh, [{"x"}, {"y"}, {}, {}]>}) -> tensor<64x16x8x8xf32> {
// CHECK-NEXT: %0 = sdy.all_to_all [{"x"}: 0->2] %arg0 out_sharding=<@mesh, [{}, {"y"}, {"x"}, {}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: %1 = sdy.all_to_all [{"y"}: 1->0] %0 out_sharding=<@mesh, [{"y"}, {}, {"x"}, {}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: return %1 : tensor<64x16x8x8xf32>
%0 = sdy.all_to_all [{"x"}: 0->2] %arg0 out_sharding=<@mesh, [{}, {"y"}, {"x"}, {}]> : tensor<64x16x8x8xf32>
%1 = sdy.all_to_all [{"y"}: 1->0] %0 out_sharding=<@mesh, [{"y"}, {}, {"x"}, {}]> : tensor<64x16x8x8xf32>
return %1 : tensor<64x16x8x8xf32>
}

// CHECK-LABEL: func @all_to_all_fusion_multiple_uses
func.func @all_to_all_fusion_multiple_uses(%arg0 : tensor<64x16x8x8xf32> {sdy.sharding=#sdy.sharding<@mesh, [{"x"}, {"y"}, {}, {}]>}) -> tensor<64x16x8x8xf32> {
// CHECK-NEXT: %0 = sdy.all_to_all [{"x"}: 0->2] %arg0 out_sharding=<@mesh, [{}, {"y"}, {"x"}, {}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: %1 = sdy.all_to_all [{"y"}: 1->0] %0 out_sharding=<@mesh, [{"y"}, {}, {"x"}, {}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: %2 = sdy.all_to_all [{"x"}: 2->0] %0 out_sharding=<@mesh, [{"x"}, {"y"}, {}, {}]> : tensor<64x16x8x8xf32>
// CHECK-NEXT: return %2 : tensor<64x16x8x8xf32>
%0 = sdy.all_to_all [{"x"}: 0->2] %arg0 out_sharding=<@mesh, [{}, {"y"}, {"x"}, {}]> : tensor<64x16x8x8xf32>
%1 = sdy.all_to_all [{"y"}: 1->0] %0 out_sharding=<@mesh, [{"y"}, {}, {"x"}, {}]> : tensor<64x16x8x8xf32>
%2 = sdy.all_to_all [{"x"}: 2->0] %0 out_sharding=<@mesh, [{"x"}, {"y"}, {}, {}]> : tensor<64x16x8x8xf32>
return %2 : tensor<64x16x8x8xf32>
}
Loading