[Mlir-commits] [mlir] [mlir][spirv][gpu] Add lowering for gpu.subgroup_broadcast (PR #187947)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-spirv

Author: Hank (hankluo6)

<details>
<summary>Changes</summary>

Fixes #<!-- -->157940

Add lowering for `gpu.subgroup_broadcast` and `gpu.subgroup_broadcast_first` to `spirv.GroupNonUniformBroadcast` and `spirv.GroupNonUniformBroadcastFirst`.

---
Full diff: https://github.com/llvm/llvm-project/pull/187947.diff


2 Files Affected:

- (modified) mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp (+39) 
- (added) mlir/test/Conversion/GPUToSPIRV/broadcast.mlir (+47) 


``````````diff
diff --git a/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp b/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
index c33a903d03393..d5269511d61a2 100644
--- a/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
+++ b/mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp
@@ -130,6 +130,18 @@ class GPURotateConversion final : public OpConversionPattern<gpu::RotateOp> {
                   ConversionPatternRewriter &rewriter) const override;
 };
 
+/// Pattern to convert a gpu.subgroup_broadcast op into a
+/// spirv.GroupNonUniformBroadcast op.
+class GPUSubgroupBroadcastConversion final
+    : public OpConversionPattern<gpu::SubgroupBroadcastOp> {
+public:
+  using Base::Base;
+
+  LogicalResult
+  matchAndRewrite(gpu::SubgroupBroadcastOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override;
+};
+
 class GPUPrintfConversion final : public OpConversionPattern<gpu::PrintfOp> {
 public:
   using Base::Base;
@@ -542,6 +554,32 @@ LogicalResult GPURotateConversion::matchAndRewrite(
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// Subgroup broadcast
+//===----------------------------------------------------------------------===//
+
+LogicalResult GPUSubgroupBroadcastConversion::matchAndRewrite(
+    gpu::SubgroupBroadcastOp op, OpAdaptor adaptor,
+    ConversionPatternRewriter &rewriter) const {
+  Location loc = op.getLoc();
+  auto scope = rewriter.getAttr<spirv::ScopeAttr>(spirv::Scope::Subgroup);
+  Value result;
+
+  switch (op.getBroadcastType()) {
+  case gpu::BroadcastType::specific_lane:
+    result = spirv::GroupNonUniformBroadcastOp::create(
+        rewriter, loc, scope, adaptor.getSrc(), adaptor.getLane());
+    break;
+  case gpu::BroadcastType::first_active_lane:
+    result = spirv::GroupNonUniformBroadcastFirstOp::create(
+        rewriter, loc, scope, adaptor.getSrc());
+    break;
+  }
+
+  rewriter.replaceOp(op, result);
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // Group ops
 //===----------------------------------------------------------------------===//
@@ -832,6 +870,7 @@ void mlir::populateGPUToSPIRVPatterns(const SPIRVTypeConverter &typeConverter,
   patterns.add<
       GPUBarrierConversion, GPUFuncOpConversion, GPUModuleConversion,
       GPUReturnOpConversion, GPUShuffleConversion, GPURotateConversion,
+      GPUSubgroupBroadcastConversion,
       LaunchConfigConversion<gpu::BlockIdOp, spirv::BuiltIn::WorkgroupId>,
       LaunchConfigConversion<gpu::GridDimOp, spirv::BuiltIn::NumWorkgroups>,
       LaunchConfigConversion<gpu::BlockDimOp, spirv::BuiltIn::WorkgroupSize>,
diff --git a/mlir/test/Conversion/GPUToSPIRV/broadcast.mlir b/mlir/test/Conversion/GPUToSPIRV/broadcast.mlir
new file mode 100644
index 0000000000000..41aa68b3f76f2
--- /dev/null
+++ b/mlir/test/Conversion/GPUToSPIRV/broadcast.mlir
@@ -0,0 +1,47 @@
+// RUN: mlir-opt -split-input-file -convert-gpu-to-spirv -verify-diagnostics %s -o - | FileCheck %s
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Shader, GroupNonUniformBallot], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL: spirv.func @broadcast_specific_lane()
+  gpu.func @broadcast_specific_lane() kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %lane = arith.constant 0 : i32
+    %val = arith.constant 42.0 : f32
+
+    // CHECK: %[[LANE:.+]] = spirv.Constant 0 : i32
+    // CHECK: %[[VAL:.+]] = spirv.Constant 4.200000e+01 : f32
+    // CHECK: %{{.+}} = spirv.GroupNonUniformBroadcast <Subgroup> %[[VAL]], %[[LANE]] : f32, i32
+    %result = gpu.subgroup_broadcast %val, specific_lane %lane : f32
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.3, [Shader, GroupNonUniformBallot], []>, #spirv.resource_limits<>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL: spirv.func @broadcast_first_active_lane()
+  gpu.func @broadcast_first_active_lane() kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %val = arith.constant 42.0 : f32
+
+    // CHECK: %[[VAL:.+]] = spirv.Constant 4.200000e+01 : f32
+    // CHECK: %{{.+}} = spirv.GroupNonUniformBroadcastFirst <Subgroup> %[[VAL]] : f32
+    %result = gpu.subgroup_broadcast %val, first_active_lane : f32
+    gpu.return
+  }
+}
+
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/187947