[mlir][gpu] Add gpu.rotate operation (#142796)

Add gpu.rotate operation and a pattern to convert gpu.rotate to SPIR-V
OpGroupNonUniformRotateKHR.

Author: Hsiangkai

mlir/include/mlir/Dialect/GPU/IR/GPUOps.td

@@ -1304,8 +1304,8 @@ def GPU_ShuffleOp : GPU_Op<
     Results<(outs AnyIntegerOrFloatOr1DVector:$shuffleResult, I1:$valid)> {
   let summary = "Shuffles values within a subgroup.";
   let description = [{
-    The "shuffle" op moves values to a across lanes (a.k.a., invocations,
-    work items) within the same subgroup. The `width` argument specifies the
+    The "shuffle" op moves values across lanes in a subgroup (a.k.a., local
+    invocation) within the same subgroup. The `width` argument specifies the
     number of lanes that participate in the shuffle, and must be uniform
     across all lanes. Further, the first `width` lanes of the subgroup must
     be active.
@@ -1366,6 +1366,54 @@ def GPU_ShuffleOp : GPU_Op<
   ];
 }
 
+def GPU_RotateOp : GPU_Op<
+    "rotate", [Pure, AllTypesMatch<["value", "rotateResult"]>]>,
+    Arguments<(ins AnyIntegerOrFloatOr1DVector:$value, I32:$offset, I32:$width)>,
+    Results<(outs AnyIntegerOrFloatOr1DVector:$rotateResult, I1:$valid)> {
+  let summary = "Rotate values within a subgroup.";
+  let description = [{
+    The "rotate" op moves values across lanes in a subgroup (a.k.a., local
+    invocations) within the same subgroup. The `width` argument specifies the
+    number of lanes that participate in the rotation, and must be uniform across
+    all participating lanes. Further, the first `width` lanes of the subgroup
+    must be active.
+
+    `width` must be a power of two, and `offset` must be in the range
+    `[0, width)`.
+
+    Return the `rotateResult` of the invocation whose id within the group is
+    calculated as follows:
+
+    ```mlir
+    Invocation ID = ((LaneId + offset) & (width - 1)) + (LaneId & ~(width - 1))
+    ```
+
+    Returns the `rotateResult` and `true` if the current lane id is smaller than
+    `width`, and poison value and `false` otherwise.
+
+    example:
+
+    ```mlir
+    %offset = arith.constant 1 : i32
+    %width = arith.constant 16 : i32
+    %1, %2 = gpu.rotate %0, %offset, %width : f32
+    ```
+
+    For lane `k`, returns the value from lane `(k + cst1) % width`.
+  }];
+
+  let assemblyFormat = [{
+    $value `,` $offset `,` $width attr-dict `:` type($value)
+  }];
+
+  let builders = [
+    // Helper function that creates a rotate with constant offset/width.
+    OpBuilder<(ins "Value":$value, "int32_t":$offset, "int32_t":$width)>
+  ];
+
+  let hasVerifier = 1;
+}
+
 def GPU_BarrierOp : GPU_Op<"barrier"> {
   let summary = "Synchronizes all work items of a workgroup.";
   let description = [{

mlir/lib/Conversion/GPUToSPIRV/GPUToSPIRV.cpp

@@ -122,6 +122,16 @@ class GPUShuffleConversion final : public OpConversionPattern<gpu::ShuffleOp> {
                   ConversionPatternRewriter &rewriter) const override;
 };
 
+/// Pattern to convert a gpu.rotate op into a spirv.GroupNonUniformRotateKHROp.
+class GPURotateConversion final : public OpConversionPattern<gpu::RotateOp> {
+public:
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(gpu::RotateOp rotateOp, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override;
+};
+
 class GPUPrintfConversion final : public OpConversionPattern<gpu::PrintfOp> {
 public:
   using OpConversionPattern::OpConversionPattern;
@@ -488,6 +498,41 @@ LogicalResult GPUShuffleConversion::matchAndRewrite(
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// Rotate
+//===----------------------------------------------------------------------===//
+
+LogicalResult GPURotateConversion::matchAndRewrite(
+    gpu::RotateOp rotateOp, OpAdaptor adaptor,
+    ConversionPatternRewriter &rewriter) const {
+  const spirv::TargetEnv &targetEnv =
+      getTypeConverter<SPIRVTypeConverter>()->getTargetEnv();
+  unsigned subgroupSize =
+      targetEnv.getAttr().getResourceLimits().getSubgroupSize();
+  IntegerAttr widthAttr;
+  if (!matchPattern(rotateOp.getWidth(), m_Constant(&widthAttr)) ||
+      widthAttr.getValue().getZExtValue() > subgroupSize)
+    return rewriter.notifyMatchFailure(
+        rotateOp,
+        "rotate width is not a constant or larger than target subgroup size");
+
+  Location loc = rotateOp.getLoc();
+  auto scope = rewriter.getAttr<spirv::ScopeAttr>(spirv::Scope::Subgroup);
+  Value rotateResult = rewriter.create<spirv::GroupNonUniformRotateKHROp>(
+      loc, scope, adaptor.getValue(), adaptor.getOffset(), adaptor.getWidth());
+  Value validVal;
+  if (widthAttr.getValue().getZExtValue() == subgroupSize) {
+    validVal = spirv::ConstantOp::getOne(rewriter.getI1Type(), loc, rewriter);
+  } else {
+    Value laneId = rewriter.create<gpu::LaneIdOp>(loc, widthAttr);
+    validVal = rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::ult,
+                                              laneId, adaptor.getWidth());
+  }
+
+  rewriter.replaceOp(rotateOp, {rotateResult, validVal});
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // Group ops
 //===----------------------------------------------------------------------===//
@@ -776,7 +821,7 @@ void mlir::populateGPUToSPIRVPatterns(const SPIRVTypeConverter &typeConverter,
                                       RewritePatternSet &patterns) {
   patterns.add<
       GPUBarrierConversion, GPUFuncOpConversion, GPUModuleConversion,
-      GPUReturnOpConversion, GPUShuffleConversion,
+      GPUReturnOpConversion, GPUShuffleConversion, GPURotateConversion,
       LaunchConfigConversion<gpu::BlockIdOp, spirv::BuiltIn::WorkgroupId>,
       LaunchConfigConversion<gpu::GridDimOp, spirv::BuiltIn::NumWorkgroups>,
       LaunchConfigConversion<gpu::BlockDimOp, spirv::BuiltIn::WorkgroupSize>,

mlir/lib/Dialect/GPU/IR/GPUDialect.cpp

@@ -1331,6 +1331,49 @@ void ShuffleOp::build(OpBuilder &builder, OperationState &result, Value value,
         mode);
 }
 
+//===----------------------------------------------------------------------===//
+// RotateOp
+//===----------------------------------------------------------------------===//
+
+void RotateOp::build(OpBuilder &builder, OperationState &result, Value value,
+                     int32_t offset, int32_t width) {
+  build(builder, result, value,
+        builder.create<arith::ConstantOp>(result.location,
+                                          builder.getI32IntegerAttr(offset)),
+        builder.create<arith::ConstantOp>(result.location,
+                                          builder.getI32IntegerAttr(width)));
+}
+
+LogicalResult RotateOp::verify() {
+  auto offsetConstOp = getOffset().getDefiningOp<arith::ConstantOp>();
+  if (!offsetConstOp)
+    return emitOpError() << "offset is not a constant value";
+
+  auto offsetIntAttr =
+      llvm::dyn_cast<mlir::IntegerAttr>(offsetConstOp.getValue());
+
+  auto widthConstOp = getWidth().getDefiningOp<arith::ConstantOp>();
+  if (!widthConstOp)
+    return emitOpError() << "width is not a constant value";
+
+  auto widthIntAttr =
+      llvm::dyn_cast<mlir::IntegerAttr>(widthConstOp.getValue());
+
+  llvm::APInt offsetValue = offsetIntAttr.getValue();
+  llvm::APInt widthValue = widthIntAttr.getValue();
+
+  if (!widthValue.isPowerOf2())
+    return emitOpError() << "width must be a power of two";
+
+  if (offsetValue.sge(widthValue) || offsetValue.slt(0)) {
+    int64_t widthValueInt = widthValue.getSExtValue();
+    return emitOpError() << "offset must be in the range [0, " << widthValueInt
+                         << ")";
+  }
+
+  return success();
+}
+
 //===----------------------------------------------------------------------===//
 // BarrierOp
 //===----------------------------------------------------------------------===//

mlir/test/Conversion/GPUToSPIRV/rotate.mlir

@@ -0,0 +1,102 @@
+// 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<v1.4, [Shader, GroupNonUniformRotateKHR], []>,
+    #spirv.resource_limits<subgroup_size = 16>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @rotate()
+  gpu.func @rotate() kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %offset = arith.constant 4 : i32
+    %width = arith.constant 16 : i32
+    %val = arith.constant 42.0 : f32
+
+    // CHECK: %[[OFFSET:.+]] = spirv.Constant 4 : i32
+    // CHECK: %[[WIDTH:.+]] = spirv.Constant 16 : i32
+    // CHECK: %[[VAL:.+]] = spirv.Constant 4.200000e+01 : f32
+    // CHECK: %{{.+}} = spirv.GroupNonUniformRotateKHR <Subgroup> %[[VAL]], %[[OFFSET]], cluster_size(%[[WIDTH]]) : f32, i32, i32 -> f32
+    // CHECK: %{{.+}} = spirv.Constant true
+    %result, %valid = gpu.rotate %val, %offset, %width : f32
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, GroupNonUniformRotateKHR], []>,
+    #spirv.resource_limits<subgroup_size = 16>>
+} {
+
+gpu.module @kernels {
+  // CHECK-LABEL:  spirv.func @rotate_width_less_than_subgroup_size()
+  gpu.func @rotate_width_less_than_subgroup_size() kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %offset = arith.constant 4 : i32
+    %width = arith.constant 8 : i32
+    %val = arith.constant 42.0 : f32
+
+    // CHECK: %[[OFFSET:.+]] = spirv.Constant 4 : i32
+    // CHECK: %[[WIDTH:.+]] = spirv.Constant 8 : i32
+    // CHECK: %[[VAL:.+]] = spirv.Constant 4.200000e+01 : f32
+    // CHECK: %{{.+}} = spirv.GroupNonUniformRotateKHR <Subgroup> %[[VAL]], %[[OFFSET]], cluster_size(%[[WIDTH]]) : f32, i32, i32 -> f32
+    // CHECK: %[[INVOCATION_ID_ADDR:.+]] = spirv.mlir.addressof @__builtin__SubgroupLocalInvocationId__
+    // CHECK: %[[INVOCATION_ID:.+]] = spirv.Load "Input" %[[INVOCATION_ID_ADDR]]
+    // CHECK: %{{.+}} = spirv.ULessThan %[[INVOCATION_ID]], %[[WIDTH]]
+    %result, %valid = gpu.rotate %val, %offset, %width : f32
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, GroupNonUniformRotateKHR], []>,
+    #spirv.resource_limits<subgroup_size = 16>>
+} {
+
+gpu.module @kernels {
+  gpu.func @rotate_with_bigger_than_subgroup_size() kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %offset = arith.constant 4 : i32
+    %width = arith.constant 32 : i32
+    %val = arith.constant 42.0 : f32
+
+    // expected-error @+1 {{failed to legalize operation 'gpu.rotate'}}
+    %result, %valid = gpu.rotate %val, %offset, %width : f32
+    gpu.return
+  }
+}
+
+}
+
+// -----
+
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, GroupNonUniformRotateKHR], []>,
+    #spirv.resource_limits<subgroup_size = 16>>
+} {
+
+gpu.module @kernels {
+  gpu.func @rotate_non_const_width(%width: i32) kernel
+    attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {
+    %offset = arith.constant 4 : i32
+    %val = arith.constant 42.0 : f32
+
+    // expected-error @+1 {{'gpu.rotate' op width is not a constant value}}
+    %result, %valid = gpu.rotate %val, %offset, %width : f32
+    gpu.return
+  }
+}
+
+}

mlir/test/Dialect/GPU/invalid.mlir

@@ -478,6 +478,84 @@ func.func @shuffle_unsupported_type_vec(%arg0 : vector<[4]xf32>, %arg1 : i32, %a
 
 // -----
 
+func.func @rotate_mismatching_type(%arg0 : f32) {
+  %offset = arith.constant 4 : i32
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op failed to verify that all of {value, rotateResult} have same type}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (i32, i1)
+  return
+}
+
+// -----
+
+func.func @rotate_unsupported_type(%arg0 : index) {
+  %offset = arith.constant 4 : i32
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op operand #0 must be Integer or Float or fixed-length vector of Integer or Float values of ranks 1, but got 'index'}}
+  %rotate, %valid = gpu.rotate %arg0, %offset, %width : index
+  return
+}
+
+// -----
+
+func.func @rotate_unsupported_type_vec(%arg0 : vector<[4]xf32>) {
+  %offset = arith.constant 4 : i32
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op operand #0 must be Integer or Float or fixed-length vector of Integer or Float values of ranks 1, but got 'vector<[4]xf32>'}}
+  %rotate, %valid = gpu.rotate %arg0, %offset, %width : vector<[4]xf32>
+  return
+}
+
+// -----
+
+func.func @rotate_unsupported_width(%arg0 : f32) {
+  %offset = arith.constant 4 : i32
+  %width = arith.constant 15 : i32
+  // expected-error@+1 {{op width must be a power of two}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (f32, i1)
+  return
+}
+
+// -----
+
+func.func @rotate_unsupported_offset(%arg0 : f32) {
+  %offset = arith.constant 16 : i32
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op offset must be in the range [0, 16)}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (f32, i1)
+  return
+}
+
+// -----
+
+func.func @rotate_unsupported_offset_minus(%arg0 : f32) {
+  %offset = arith.constant -1 : i32
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op offset must be in the range [0, 16)}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (f32, i1)
+  return
+}
+
+// -----
+
+func.func @rotate_offset_non_constant(%arg0 : f32, %offset : i32) {
+  %width = arith.constant 16 : i32
+  // expected-error@+1 {{op offset is not a constant value}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (f32, i1)
+  return
+}
+
+// -----
+
+func.func @rotate_width_non_constant(%arg0 : f32, %width : i32) {
+  %offset = arith.constant 0 : i32
+  // expected-error@+1 {{op width is not a constant value}}
+  %rotate, %valid = "gpu.rotate"(%arg0, %offset, %width) : (f32, i32, i32) -> (f32, i1)
+  return
+}
+
+// -----
+
 module {
   gpu.module @gpu_funcs {
     // expected-error @+1 {{custom op 'gpu.func' gpu.func requires named arguments}}

mlir/test/Dialect/GPU/ops.mlir

@@ -140,6 +140,10 @@ module attributes {gpu.container_module} {
       // CHECK: gpu.shuffle idx %{{.*}}, %{{.*}}, %{{.*}} : f32
       %shfl3, %pred3 = gpu.shuffle idx %arg0, %offset, %width : f32
 
+      // CHECK: gpu.rotate %{{.*}}, %{{.*}}, %{{.*}} : f32
+      %rotate_width = arith.constant 16 : i32
+      %rotate, %pred4 = gpu.rotate %arg0, %offset, %rotate_width : f32
+
       "gpu.barrier"() : () -> ()
 
       "some_op"(%bIdX, %tIdX) : (index, index) -> ()