[mlir][spirv] Implement lowering `gpu.subgroup_reduce` with cluster size for SPIRV #141402

fairywreath · 2025-05-25T08:56:51Z

Implement lowering of gpu.subgroup_reduce with a cluster size attribute to SPIRV by using the ClusteredReduce group operation.

…r size for SPIRV

github-actions · 2025-05-25T08:57:11Z

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

llvmbot · 2025-05-25T08:57:44Z

@llvm/pr-subscribers-mlir-spirv
@llvm/pr-subscribers-mlir-gpu

@llvm/pr-subscribers-mlir

Author: Darren Wihandi (fairywreath)

Changes

Implement lowering of gpu.subgroup_reduce with a cluster size attribute to SPIRV by using the ClusteredReduce group operation.

Full diff: https://github.com/llvm/llvm-project/pull/141402.diff

2 Files Affected:

(modified) mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp (+29-16)
(modified) mlir/test/Conversion/GPUToSPIRV/reductions.mlir (+41)

diff --git a/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp b/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
index 3cc64b82950b5..f42605a6e8ce1 100644
--- a/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
+++ b/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
@@ -464,27 +464,39 @@ LogicalResult GPUShuffleConversion::matchAndRewrite(
 
 template <typename UniformOp, typename NonUniformOp>
 static Value createGroupReduceOpImpl(OpBuilder &builder, Location loc,
-                                     Value arg, bool isGroup, bool isUniform) {
+                                     Value arg, bool isGroup, bool isUniform,
+                                     std::optional<uint32_t> clusterSize) {
   Type type = arg.getType();
   auto scope = mlir::spirv::ScopeAttr::get(builder.getContext(),
                                            isGroup ? spirv::Scope::Workgroup
                                                    : spirv::Scope::Subgroup);
-  auto groupOp = spirv::GroupOperationAttr::get(builder.getContext(),
-                                                spirv::GroupOperation::Reduce);
+  auto groupOp = spirv::GroupOperationAttr::get(
+      builder.getContext(), clusterSize.has_value()
+                                ? spirv::GroupOperation::ClusteredReduce
+                                : spirv::GroupOperation::Reduce);
   if (isUniform) {
     return builder.create<UniformOp>(loc, type, scope, groupOp, arg)
         .getResult();
   }
-  return builder.create<NonUniformOp>(loc, type, scope, groupOp, arg, Value{})
+
+  Value clusterSizeValue =
+      clusterSize.has_value()
+          ? builder.create<spirv::ConstantOp>(
+                loc, builder.getI32Type(),
+                builder.getIntegerAttr(builder.getI32Type(), *clusterSize))
+          : Value{};
+  return builder
+      .create<NonUniformOp>(loc, type, scope, groupOp, arg, clusterSizeValue)
       .getResult();
 }
 
-static std::optional<Value> createGroupReduceOp(OpBuilder &builder,
-                                                Location loc, Value arg,
-                                                gpu::AllReduceOperation opType,
-                                                bool isGroup, bool isUniform) {
+static std::optional<Value>
+createGroupReduceOp(OpBuilder &builder, Location loc, Value arg,
+                    gpu::AllReduceOperation opType, bool isGroup,
+                    bool isUniform, std::optional<uint32_t> clusterSize) {
   enum class ElemType { Float, Boolean, Integer };
-  using FuncT = Value (*)(OpBuilder &, Location, Value, bool, bool);
+  using FuncT = Value (*)(OpBuilder &, Location, Value, bool, bool,
+                          std::optional<uint32_t>);
   struct OpHandler {
     gpu::AllReduceOperation kind;
     ElemType elemType;
@@ -548,7 +560,7 @@ static std::optional<Value> createGroupReduceOp(OpBuilder &builder,
 
   for (const OpHandler &handler : handlers)
     if (handler.kind == opType && elementType == handler.elemType)
-      return handler.func(builder, loc, arg, isGroup, isUniform);
+      return handler.func(builder, loc, arg, isGroup, isUniform, clusterSize);
 
   return std::nullopt;
 }
@@ -571,7 +583,7 @@ class GPUAllReduceConversion final
 
     auto result =
         createGroupReduceOp(rewriter, op.getLoc(), adaptor.getValue(), *opType,
-                            /*isGroup*/ true, op.getUniform());
+                            /*isGroup*/ true, op.getUniform(), std::nullopt);
     if (!result)
       return failure();
 
@@ -589,16 +601,17 @@ class GPUSubgroupReduceConversion final
   LogicalResult
   matchAndRewrite(gpu::SubgroupReduceOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const override {
-    if (op.getClusterSize())
+    if (op.getClusterStride() > 1) {
       return rewriter.notifyMatchFailure(
-          op, "lowering for clustered reduce not implemented");
+          op, "lowering for cluster stride > 1 is not implemented");
+    }
 
     if (!isa<spirv::ScalarType>(adaptor.getValue().getType()))
       return rewriter.notifyMatchFailure(op, "reduction type is not a scalar");
 
-    auto result = createGroupReduceOp(rewriter, op.getLoc(), adaptor.getValue(),
-                                      adaptor.getOp(),
-                                      /*isGroup=*/false, adaptor.getUniform());
+    auto result = createGroupReduceOp(
+        rewriter, op.getLoc(), adaptor.getValue(), adaptor.getOp(),
+        /*isGroup=*/false, adaptor.getUniform(), op.getClusterSize());
     if (!result)
       return failure();
 
diff --git a/mlir/test/Conversion/GPUToSPIRV/reductions.mlir b/mlir/test/Conversion/GPUToSPIRV/reductions.mlir
index ae834b9915d50..08d9b094a5303 100644
--- a/mlir/test/Conversion/GPUToSPIRV/reductions.mlir
+++ b/mlir/test/Conversion/GPUToSPIRV/reductions.mlir
@@ -789,3 +789,44 @@ gpu.module @kernels {
   }
 }
 }
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupUniformArithmeticKHR, GroupNonUniformClustered], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @test
+  //  CHECK-SAME: (%[[ARG:.*]]: f32)
+  //  CHECK: %[[CLUSTER_SIZE:.*]] = spirv.Constant 8 : i32
+  gpu.func @test22(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // CHECK: %{{.*}} = spirv.GroupNonUniformFAdd <Subgroup> <ClusteredReduce> %[[ARG]] cluster_size(%[[CLUSTER_SIZE]]) : f32, i32 -> f32
+    %reduced = gpu.subgroup_reduce add %arg cluster(size = 8) : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+// Subgrop reduce with cluster stride > 1 is not yet supported.
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Kernel, Addresses, Groups, GroupUniformArithmeticKHR, GroupNonUniformClustered], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  gpu.func @test22(%arg : f32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    // expected-error @+1 {{failed to legalize operation 'gpu.subgroup_reduce'}}
+    %reduced = gpu.subgroup_reduce add %arg cluster(size = 8, stride = 2) : (f32) -> (f32)
+    gpu.return
+  }
+}
+
+}

mlir/test/Conversion/GPUToSPIRV/reductions.mlir

kuhar

Nice, thanks for adding this

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

fairywreath · 2025-06-03T21:41:25Z

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

-  return builder.create<NonUniformOp>(loc, type, scope, groupOp, arg, Value{})
+
+  Value clusterSizeValue = {};
+  if (clusterSize.has_value())


Is there any preference on whether to use has_value() or using the variable directly as a bool for std::optional?

either way is fine IMO

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

kuhar · 2025-06-03T22:05:47Z

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

-  return builder.create<NonUniformOp>(loc, type, scope, groupOp, arg, Value{})
+
+  Value clusterSizeValue = {};
+  if (clusterSize.has_value())


either way is fine IMO

github-actions · 2025-06-06T16:50:39Z

@fairywreath Congratulations on having your first Pull Request (PR) merged into the LLVM Project!

Your changes will be combined with recent changes from other authors, then tested by our build bots. If there is a problem with a build, you may receive a report in an email or a comment on this PR.

Please check whether problems have been caused by your change specifically, as the builds can include changes from many authors. It is not uncommon for your change to be included in a build that fails due to someone else's changes, or infrastructure issues.

How to do this, and the rest of the post-merge process, is covered in detail here.

If your change does cause a problem, it may be reverted, or you can revert it yourself. This is a normal part of LLVM development. You can fix your changes and open a new PR to merge them again.

If you don't get any reports, no action is required from you. Your changes are working as expected, well done!

…ize for SPIRV (llvm#141402) Implement lowering of `gpu.subgroup_reduce` with a cluster size attribute to SPIRV by using the `ClusteredReduce` group operation.

[mlir][spirv] Implement lowering of gpu.subgroup_reduce with cluste…

377841b

…r size for SPIRV

fairywreath requested review from antiagainst and kuhar as code owners May 25, 2025 08:56

llvmbot added mlir:gpu mlir:spirv mlir labels May 25, 2025

fairywreath mentioned this pull request May 25, 2025

[CodeGen][SPIRV] Lowering for clustered reduce not implemented iree-org/iree#20872

Open

antiagainst requested changes May 25, 2025

View reviewed changes

mlir/test/Conversion/GPUToSPIRV/reductions.mlir Outdated Show resolved Hide resolved

Use better function names in test

9dbd3d2

kuhar approved these changes Jun 2, 2025

View reviewed changes

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp Outdated Show resolved Hide resolved

Use if statement instead of ternary

a865235

fairywreath commented Jun 3, 2025

View reviewed changes

kuhar reviewed Jun 3, 2025

View reviewed changes

Address review comment

bc26255

Muzammiluddin-Syed-ECE mentioned this pull request Jun 4, 2025

[Codegen][GPU] Lower gpu.subgroup_reduce to DPP intrinsics on AMD GPUs iree-org/iree#20468

Merged

Merge branch 'main' into spirv-subgroup-reduce-cluster-size

e0de60b

kuhar merged commit c9c6017 into llvm:main Jun 6, 2025
7 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[mlir][spirv] Implement lowering `gpu.subgroup_reduce` with cluster size for SPIRV #141402

[mlir][spirv] Implement lowering `gpu.subgroup_reduce` with cluster size for SPIRV #141402

Uh oh!

fairywreath commented May 25, 2025

Uh oh!

github-actions bot commented May 25, 2025

Uh oh!

llvmbot commented May 25, 2025 •

edited

Loading

Uh oh!

Uh oh!

kuhar left a comment

Uh oh!

Uh oh!

fairywreath Jun 3, 2025

Uh oh!

kuhar Jun 3, 2025

Uh oh!

Uh oh!

kuhar Jun 3, 2025

Uh oh!

Uh oh!

github-actions bot commented Jun 6, 2025

Uh oh!

Uh oh!

[mlir][spirv] Implement lowering gpu.subgroup_reduce with cluster size for SPIRV #141402

[mlir][spirv] Implement lowering gpu.subgroup_reduce with cluster size for SPIRV #141402

Uh oh!

Conversation

fairywreath commented May 25, 2025

Uh oh!

github-actions bot commented May 25, 2025

Uh oh!

llvmbot commented May 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

kuhar left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

fairywreath Jun 3, 2025

Choose a reason for hiding this comment

Uh oh!

kuhar Jun 3, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

kuhar Jun 3, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions bot commented Jun 6, 2025

Uh oh!

Uh oh!

[mlir][spirv] Implement lowering `gpu.subgroup_reduce` with cluster size for SPIRV #141402

[mlir][spirv] Implement lowering `gpu.subgroup_reduce` with cluster size for SPIRV #141402

llvmbot commented May 25, 2025 •

edited

Loading