[mlir][gpu] Add GPU subgroup MMA extract and insert operations

- Introduced `gpu.subgroup_mma_extract` operation to extract values from `!gpu.mma_matrix` by invocation and indices.
- Introduced `gpu.subgroup_mma_insert` operation to insert values into `!gpu.mma_matrix` by invocation and indices.
- Updated the conversion patterns to SPIR-V for both extract and insert operations.
- Added test cases to validate the new operations in the GPU to SPIR-V conversion.

Author: Hsiangkai

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

@@ -1919,6 +1919,79 @@ def GPU_SubgroupMmaConstantMatrixOp : GPU_Op<"subgroup_mma_constant_matrix",
   }];
 }
 
+def GPU_SubgroupMmaExtractOp : GPU_Op<"subgroup_mma_extract",
+    [Pure,
+     TypesMatchWith<"value type matches element type of mma_matrix",
+                    "matrix", "res",
+                    "::llvm::cast<gpu::MMAMatrixType>($_self).getElementType()">]>{
+
+  let summary = "Extract a value from GPU warp by invocation and indices";
+
+  let description = [{
+    The `gpu.subgroup_mma_extract` operation extracts a value from `!gpu.mma_matrix`
+    by the invocation in a subgroup.
+
+    This operation takes `!gpu.mma_matrix` as its first operand. It is the source
+    matrix across a subgroup. The op returns a scalar value stored in the invocation
+    in the subgroup. If there are multiple values packed in an invocation, use
+    `indices` to specify the element to extract.
+
+    Example:
+
+    ```mlir
+    %c0 = arith.constant 0 : index
+    %val = gpu.subgroup_mma_extract %m[%c0] : !gpu.mma_matrix<16x16xf32, "AOp"> -> f32
+    ```
+  }];
+
+  let arguments = (ins GPU_MMAMatrix:$matrix, Variadic<Index>:$indices);
+
+  let results = (outs AnyIntegerOrFloat:$res);
+
+  let assemblyFormat = [{
+    $matrix`[`$indices`]` attr-dict `:` type($matrix) `->` type($res)
+  }];
+}
+
+def GPU_SubgroupMmaInsertOp : GPU_Op<"subgroup_mma_insert",
+    [Pure,
+     TypesMatchWith<"value type matches element type of mma_matrix",
+                    "matrix", "value",
+                    "::llvm::cast<gpu::MMAMatrixType>($_self).getElementType()"> ]>{
+
+  let summary = "Insert a value into GPU warp by invocation and indices";
+
+  let description = [{
+    The `gpu.subgroup_mma_insert` operation inserts a value to `!gpu.mma_matrix`
+    by the invocation in a subgroup.
+
+    This operation takes scalar value as its first operand and `!gpu.mma_matrix`
+    as its second operand. It is the matrix across a subgroup. The op inserts the
+    scalar value stored in the invocation in the subgroup to the matrix. If there
+    are multiple values packed in an invocation, use `indices` to specify the
+    location to insert in the packing.
+
+    The op returns `!gpu.mma_matrix` with the updated value.
+
+    Example:
+
+    ```mlir
+    %c0 = arith.constant 0 : index
+    %s0 = gpu.subgroup_mma_insert %val, %m[%c0] : f16, !gpu.mma_matrix<16x16xf16, "COp">
+            -> !gpu.mma_matrix<16x16xf16, "COp">
+    ```
+  }];
+
+  let arguments = (ins AnyIntegerOrFloat:$value, GPU_MMAMatrix:$matrix,
+                       Variadic<Index>:$indices);
+
+  let results = (outs GPU_MMAMatrix:$res);
+
+  let assemblyFormat = [{
+    $value`,` $matrix`[`$indices`]` attr-dict `:` type($value)`,` type($matrix) `->` type($res)
+  }];
+}
+
 def GPU_ElementwiseOpAddF  : I32EnumAttrCase<"ADDF", 0, "addf">;
 def GPU_ElementwiseOpMulF  : I32EnumAttrCase<"MULF", 1, "mulf">;
 def GPU_ElementwiseOpSUBF  : I32EnumAttrCase<"SUBF", 2, "subf">;

mlir/lib/Conversion/GPUToSPIRV/WmmaOpsToSPIRV.cpp

@@ -111,6 +111,68 @@ struct WmmaConstantOpToSPIRVLowering final
   }
 };
 
+/// Converts GPU MMA ExtractOp to CompositeExtract SPIR-V KHR/NV cooperative
+/// matrix ops.
+struct WmmaExtractOpToSPIRVLowering final
+    : OpConversionPattern<gpu::SubgroupMmaExtractOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(gpu::SubgroupMmaExtractOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Value matrix = adaptor.getMatrix();
+    auto coopType =
+        getTypeConverter()->convertType<spirv::CooperativeMatrixType>(
+            matrix.getType());
+    if (!coopType)
+      return rewriter.notifyMatchFailure(op, "type conversion failed");
+
+    SmallVector<int32_t> intValues;
+    for (Value val : op.getIndices()) {
+      if (auto constOp = val.getDefiningOp<arith::ConstantIndexOp>()) {
+        intValues.push_back(static_cast<int32_t>(constOp.value()));
+      } else {
+        return rewriter.notifyMatchFailure(op, "indices must be constants");
+      }
+    }
+
+    Type elementType = coopType.getElementType();
+    rewriter.replaceOpWithNewOp<spirv::CompositeExtractOp>(
+        op, elementType, matrix, rewriter.getI32ArrayAttr(intValues));
+    return success();
+  }
+};
+
+/// Converts GPU MMA InsertOp to CompositeInsert SPIR-V KHR/NV cooperative
+/// matrix ops.
+struct WmmaInsertOpToSPIRVLowering final
+    : OpConversionPattern<gpu::SubgroupMmaInsertOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(gpu::SubgroupMmaInsertOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const override {
+    Value value = adaptor.getValue();
+    Value matrix = adaptor.getMatrix();
+    auto coopType = getTypeConverter()->convertType(matrix.getType());
+    if (!coopType)
+      return rewriter.notifyMatchFailure(op, "type conversion failed");
+
+    SmallVector<int32_t> intValues;
+    for (Value val : op.getIndices()) {
+      if (auto constOp = val.getDefiningOp<arith::ConstantIndexOp>()) {
+        intValues.push_back(static_cast<int32_t>(constOp.value()));
+      } else {
+        return rewriter.notifyMatchFailure(op, "indices must be constants");
+      }
+    }
+
+    rewriter.replaceOpWithNewOp<spirv::CompositeInsertOp>(
+        op, coopType, value, matrix, rewriter.getI32ArrayAttr(intValues));
+    return success();
+  }
+};
+
 /// Converts elementwise ops to SPIR-V cooperative matrix elementwise ops for
 /// the default case.
 struct WmmaElementwiseOpToSPIRVDefaultLowering final
@@ -296,6 +358,7 @@ void mlir::populateGpuWMMAToSPIRVCoopMatrixKHRConversionPatterns(
   MLIRContext *context = patterns.getContext();
   patterns.add<khr::WmmaLoadOpToSPIRVLowering, khr::WmmaMmaOpToSPIRVLowering,
                khr::WmmaStoreOpToSPIRVLowering, WmmaConstantOpToSPIRVLowering,
+               WmmaExtractOpToSPIRVLowering, WmmaInsertOpToSPIRVLowering,
                WmmaElementwiseOpToSPIRVDefaultLowering>(converter, context);
   // Give the following patterns higher benefit to prevail over the default one.
   patterns.add<WmmaElementwiseOpToSPIRVScalarMulLowering>(converter, context,

mlir/test/Conversion/GPUToSPIRV/wmma-ops-to-spirv-khr-coop-matrix.mlir

@@ -93,6 +93,33 @@ module attributes {
       gpu.return
     }
 
+    // CHECK-LABEL: spirv.func @gpu_wmma_extract_op
+    // CHECK-SAME: %[[ARG0:.+]]: !spirv.coopmatrix<16x16xf32, Subgroup, MatrixA>
+    gpu.func @gpu_wmma_extract_op(%m: !gpu.mma_matrix<16x16xf32, "AOp">,
+                                  %ptr: memref<16x16xf32, #spirv.storage_class<StorageBuffer>>) kernel
+      attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [32, 4, 1]>} {
+      // CHECK: spirv.CompositeExtract %[[ARG0]][0 : i32] : !spirv.coopmatrix<16x16xf32, Subgroup, MatrixA>
+      %c0 = arith.constant 0 : index
+      %val = gpu.subgroup_mma_extract %m[%c0] : !gpu.mma_matrix<16x16xf32, "AOp"> -> f32
+      memref.store %val, %ptr[%c0, %c0] : memref<16x16xf32, #spirv.storage_class<StorageBuffer>>
+      gpu.return
+    }
+
+    // CHECK-LABEL: spirv.func @gpu_wmma_insert_op
+    // CHECK-SAME: %[[ARG0:.+]]: f16
+    // CHECK-SAME: %[[ARG1:.+]]: !spirv.coopmatrix<16x16xf16, Subgroup, MatrixAcc>
+    gpu.func @gpu_wmma_insert_op(%val: f16,
+                                 %m: !gpu.mma_matrix<16x16xf16, "COp">,
+                                 %ptr: memref<16x16xf16, #spirv.storage_class<StorageBuffer>>) kernel
+      attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [32, 4, 1]>} {
+      // CHECK: spirv.CompositeInsert %[[ARG0]], %[[ARG1]][0 : i32] : f16 into !spirv.coopmatrix<16x16xf16, Subgroup, MatrixAcc>
+      %c0 = arith.constant 0 : index
+      %s0 = gpu.subgroup_mma_insert %val, %m[%c0] : f16, !gpu.mma_matrix<16x16xf16, "COp"> -> !gpu.mma_matrix<16x16xf16, "COp">
+      gpu.subgroup_mma_store_matrix %s0, %ptr[%c0,%c0] {leadDimension = 16 : index} :
+        !gpu.mma_matrix<16x16xf16, "COp">, memref<16x16xf16, #spirv.storage_class<StorageBuffer>>
+      gpu.return
+    }
+
     // CHECK-LABEL: spirv.func @gpu_wmma_elementwise_op_default
     // CHECK-SAME:    !spirv.coopmatrix<16x16xf16, Subgroup, MatrixAcc>
     // CHECK-SAME:    !spirv.coopmatrix<16x16xf16, Subgroup, MatrixAcc>