[acc] Add acc.kernel_environment to enable compute decomposition #165455
+123 −0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters. Learn more about bidirectional Unicode characters
Introduce `acc.kernel_environment` operation to capture data mapping and asynchronous behavior from OpenACC compute constructs. This enables decomposition by separating data movement and synchronization from kernel execution parallelism, facilitating lowering to GPU dialect.
| @llvm/pr-subscribers-mlir-openacc @llvm/pr-subscribers-mlir Author: Razvan Lupusoru (razvanlupusoru) ChangesIntroduce Full diff: https://github.com/llvm/llvm-project/pull/165455.diff 2 Files Affected:
diff --git a/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td b/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td index 2f87975ebaa04..a18c18af8a753 100644 --- a/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td +++ b/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td @@ -2116,6 +2116,56 @@ def OpenACC_KernelsOp : OpenACC_Op<"kernels", let hasVerifier = 1; } +//===----------------------------------------------------------------------===// +// acc.kernel_environment +//===----------------------------------------------------------------------===// + +def OpenACC_KernelEnvironmentOp : OpenACC_Op<"kernel_environment", + [AttrSizedOperandSegments, RecursiveMemoryEffects, SingleBlock, + NoTerminator, + MemoryEffects<[MemWrite<OpenACC_ConstructResource>, + MemRead<OpenACC_CurrentDeviceIdResource>]>]> { + let summary = "Decomposition of compute constructs to capture data mapping " + "and asynchronous behavior information"; + let description = [{ + The `acc.kernel_environment` operation represents a decomposition of + any OpenACC compute construct (acc.kernels, acc.parallel, or + acc.serial) that captures data mapping and asynchronous behavior: + - data clause operands + - async clause operands + - wait clause operands + + This allows kernel execution parallelism and privatization to be + handled separately, facilitating eventual lowering to GPU dialect where + kernel launching and compute offloading are handled separately. + }]; + + let arguments = (ins + Variadic<AnyType>:$dataClauseOperands, + Variadic<IntOrIndex>:$asyncOperands, + OptionalAttr<DeviceTypeArrayAttr>:$asyncOperandsDeviceType, + OptionalAttr<DeviceTypeArrayAttr>:$asyncOnly, + Variadic<IntOrIndex>:$waitOperands, + OptionalAttr<DenseI32ArrayAttr>:$waitOperandsSegments, + OptionalAttr<DeviceTypeArrayAttr>:$waitOperandsDeviceType, + OptionalAttr<BoolArrayAttr>:$hasWaitDevnum, + OptionalAttr<DeviceTypeArrayAttr>:$waitOnly); + + let regions = (region SizedRegion<1>:$region); + + let assemblyFormat = [{ + oilist( + `dataOperands` `(` $dataClauseOperands `:` type($dataClauseOperands) `)` + | `async` `` custom<DeviceTypeOperandsWithKeywordOnly>($asyncOperands, + type($asyncOperands), $asyncOperandsDeviceType, $asyncOnly) + | `wait` `` custom<WaitClause>($waitOperands, type($waitOperands), + $waitOperandsDeviceType, $waitOperandsSegments, $hasWaitDevnum, + $waitOnly) + ) + $region attr-dict + }]; +} + //===----------------------------------------------------------------------===// // 2.6.5 data Construct //===----------------------------------------------------------------------===// diff --git a/mlir/test/Dialect/OpenACC/ops.mlir b/mlir/test/Dialect/OpenACC/ops.mlir index 77d18da49276a..042ee2503cb95 100644 --- a/mlir/test/Dialect/OpenACC/ops.mlir +++ b/mlir/test/Dialect/OpenACC/ops.mlir @@ -2243,3 +2243,76 @@ func.func @test_firstprivate_map(%arg0: memref<10xf32>) { // CHECK-NEXT: acc.yield // CHECK-NEXT: } // CHECK-NEXT: return + +// ----- + +func.func @test_kernel_environment(%arg0: memref<1024xf32>, %arg1: memref<1024xf32>) { + %c1 = arith.constant 1 : index + %c1024 = arith.constant 1024 : index + + // Create data clause operands for the kernel environment + %copyin = acc.copyin varPtr(%arg0 : memref<1024xf32>) -> memref<1024xf32> + %create = acc.create varPtr(%arg1 : memref<1024xf32>) -> memref<1024xf32> + + // Kernel environment wraps gpu.launch and captures data mapping + acc.kernel_environment dataOperands(%copyin, %create : memref<1024xf32>, memref<1024xf32>) { + gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) + threads(%tx, %ty, %tz) in (%block_x = %c1024, %block_y = %c1, %block_z = %c1) { + // Kernel body uses the mapped data + %val = memref.load %copyin[%tx] : memref<1024xf32> + %result = arith.mulf %val, %val : f32 + memref.store %result, %create[%tx] : memref<1024xf32> + gpu.terminator + } + } + + // Copy results back to host and deallocate device memory + acc.copyout accPtr(%create : memref<1024xf32>) to varPtr(%arg1 : memref<1024xf32>) + acc.delete accPtr(%copyin : memref<1024xf32>) + + return +} + +// CHECK-LABEL: func @test_kernel_environment +// CHECK: %[[COPYIN:.*]] = acc.copyin varPtr(%{{.*}} : memref<1024xf32>) -> memref<1024xf32> +// CHECK: %[[CREATE:.*]] = acc.create varPtr(%{{.*}} : memref<1024xf32>) -> memref<1024xf32> +// CHECK: acc.kernel_environment dataOperands(%[[COPYIN]], %[[CREATE]] : memref<1024xf32>, memref<1024xf32>) { +// CHECK: gpu.launch +// CHECK: memref.load %[[COPYIN]] +// CHECK: memref.store %{{.*}}, %[[CREATE]] +// CHECK: } +// CHECK: } +// CHECK: acc.copyout accPtr(%[[CREATE]] : memref<1024xf32>) to varPtr(%{{.*}} : memref<1024xf32>) +// CHECK: acc.delete accPtr(%[[COPYIN]] : memref<1024xf32>) + +// ----- + +func.func @test_kernel_environment_with_async(%arg0: memref<1024xf32>) { + %c1 = arith.constant 1 : index + %c1024 = arith.constant 1024 : index + %async_val = arith.constant 1 : i32 + + %create = acc.create varPtr(%arg0 : memref<1024xf32>) async(%async_val : i32) -> memref<1024xf32> + + // Kernel environment with async clause + acc.kernel_environment dataOperands(%create : memref<1024xf32>) async(%async_val : i32) { + gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) + threads(%tx, %ty, %tz) in (%block_x = %c1024, %block_y = %c1, %block_z = %c1) { + %f0 = arith.constant 0.0 : f32 + memref.store %f0, %create[%tx] : memref<1024xf32> + gpu.terminator + } + } + + acc.copyout accPtr(%create : memref<1024xf32>) async(%async_val : i32) to varPtr(%arg0 : memref<1024xf32>) + + return +} + +// CHECK-LABEL: func @test_kernel_environment_with_async +// CHECK: %[[ASYNC:.*]] = arith.constant 1 : i32 +// CHECK: %[[CREATE:.*]] = acc.create varPtr(%{{.*}} : memref<1024xf32>) async(%[[ASYNC]] : i32) -> memref<1024xf32> +// CHECK: acc.kernel_environment dataOperands(%[[CREATE]] : memref<1024xf32>) async(%[[ASYNC]] : i32) +// CHECK: gpu.launch +// CHECK: memref.store %{{.*}}, %[[CREATE]] +// CHECK: acc.copyout accPtr(%[[CREATE]] : memref<1024xf32>) async(%[[ASYNC]] : i32) to varPtr(%{{.*}} : memref<1024xf32>) |
| @llvm/pr-subscribers-openacc Author: Razvan Lupusoru (razvanlupusoru) ChangesIntroduce Full diff: https://github.com/llvm/llvm-project/pull/165455.diff 2 Files Affected:
diff --git a/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td b/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td index 2f87975ebaa04..a18c18af8a753 100644 --- a/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td +++ b/mlir/include/mlir/Dialect/OpenACC/OpenACCOps.td @@ -2116,6 +2116,56 @@ def OpenACC_KernelsOp : OpenACC_Op<"kernels", let hasVerifier = 1; } +//===----------------------------------------------------------------------===// +// acc.kernel_environment +//===----------------------------------------------------------------------===// + +def OpenACC_KernelEnvironmentOp : OpenACC_Op<"kernel_environment", + [AttrSizedOperandSegments, RecursiveMemoryEffects, SingleBlock, + NoTerminator, + MemoryEffects<[MemWrite<OpenACC_ConstructResource>, + MemRead<OpenACC_CurrentDeviceIdResource>]>]> { + let summary = "Decomposition of compute constructs to capture data mapping " + "and asynchronous behavior information"; + let description = [{ + The `acc.kernel_environment` operation represents a decomposition of + any OpenACC compute construct (acc.kernels, acc.parallel, or + acc.serial) that captures data mapping and asynchronous behavior: + - data clause operands + - async clause operands + - wait clause operands + + This allows kernel execution parallelism and privatization to be + handled separately, facilitating eventual lowering to GPU dialect where + kernel launching and compute offloading are handled separately. + }]; + + let arguments = (ins + Variadic<AnyType>:$dataClauseOperands, + Variadic<IntOrIndex>:$asyncOperands, + OptionalAttr<DeviceTypeArrayAttr>:$asyncOperandsDeviceType, + OptionalAttr<DeviceTypeArrayAttr>:$asyncOnly, + Variadic<IntOrIndex>:$waitOperands, + OptionalAttr<DenseI32ArrayAttr>:$waitOperandsSegments, + OptionalAttr<DeviceTypeArrayAttr>:$waitOperandsDeviceType, + OptionalAttr<BoolArrayAttr>:$hasWaitDevnum, + OptionalAttr<DeviceTypeArrayAttr>:$waitOnly); + + let regions = (region SizedRegion<1>:$region); + + let assemblyFormat = [{ + oilist( + `dataOperands` `(` $dataClauseOperands `:` type($dataClauseOperands) `)` + | `async` `` custom<DeviceTypeOperandsWithKeywordOnly>($asyncOperands, + type($asyncOperands), $asyncOperandsDeviceType, $asyncOnly) + | `wait` `` custom<WaitClause>($waitOperands, type($waitOperands), + $waitOperandsDeviceType, $waitOperandsSegments, $hasWaitDevnum, + $waitOnly) + ) + $region attr-dict + }]; +} + //===----------------------------------------------------------------------===// // 2.6.5 data Construct //===----------------------------------------------------------------------===// diff --git a/mlir/test/Dialect/OpenACC/ops.mlir b/mlir/test/Dialect/OpenACC/ops.mlir index 77d18da49276a..042ee2503cb95 100644 --- a/mlir/test/Dialect/OpenACC/ops.mlir +++ b/mlir/test/Dialect/OpenACC/ops.mlir @@ -2243,3 +2243,76 @@ func.func @test_firstprivate_map(%arg0: memref<10xf32>) { // CHECK-NEXT: acc.yield // CHECK-NEXT: } // CHECK-NEXT: return + +// ----- + +func.func @test_kernel_environment(%arg0: memref<1024xf32>, %arg1: memref<1024xf32>) { + %c1 = arith.constant 1 : index + %c1024 = arith.constant 1024 : index + + // Create data clause operands for the kernel environment + %copyin = acc.copyin varPtr(%arg0 : memref<1024xf32>) -> memref<1024xf32> + %create = acc.create varPtr(%arg1 : memref<1024xf32>) -> memref<1024xf32> + + // Kernel environment wraps gpu.launch and captures data mapping + acc.kernel_environment dataOperands(%copyin, %create : memref<1024xf32>, memref<1024xf32>) { + gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) + threads(%tx, %ty, %tz) in (%block_x = %c1024, %block_y = %c1, %block_z = %c1) { + // Kernel body uses the mapped data + %val = memref.load %copyin[%tx] : memref<1024xf32> + %result = arith.mulf %val, %val : f32 + memref.store %result, %create[%tx] : memref<1024xf32> + gpu.terminator + } + } + + // Copy results back to host and deallocate device memory + acc.copyout accPtr(%create : memref<1024xf32>) to varPtr(%arg1 : memref<1024xf32>) + acc.delete accPtr(%copyin : memref<1024xf32>) + + return +} + +// CHECK-LABEL: func @test_kernel_environment +// CHECK: %[[COPYIN:.*]] = acc.copyin varPtr(%{{.*}} : memref<1024xf32>) -> memref<1024xf32> +// CHECK: %[[CREATE:.*]] = acc.create varPtr(%{{.*}} : memref<1024xf32>) -> memref<1024xf32> +// CHECK: acc.kernel_environment dataOperands(%[[COPYIN]], %[[CREATE]] : memref<1024xf32>, memref<1024xf32>) { +// CHECK: gpu.launch +// CHECK: memref.load %[[COPYIN]] +// CHECK: memref.store %{{.*}}, %[[CREATE]] +// CHECK: } +// CHECK: } +// CHECK: acc.copyout accPtr(%[[CREATE]] : memref<1024xf32>) to varPtr(%{{.*}} : memref<1024xf32>) +// CHECK: acc.delete accPtr(%[[COPYIN]] : memref<1024xf32>) + +// ----- + +func.func @test_kernel_environment_with_async(%arg0: memref<1024xf32>) { + %c1 = arith.constant 1 : index + %c1024 = arith.constant 1024 : index + %async_val = arith.constant 1 : i32 + + %create = acc.create varPtr(%arg0 : memref<1024xf32>) async(%async_val : i32) -> memref<1024xf32> + + // Kernel environment with async clause + acc.kernel_environment dataOperands(%create : memref<1024xf32>) async(%async_val : i32) { + gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) + threads(%tx, %ty, %tz) in (%block_x = %c1024, %block_y = %c1, %block_z = %c1) { + %f0 = arith.constant 0.0 : f32 + memref.store %f0, %create[%tx] : memref<1024xf32> + gpu.terminator + } + } + + acc.copyout accPtr(%create : memref<1024xf32>) async(%async_val : i32) to varPtr(%arg0 : memref<1024xf32>) + + return +} + +// CHECK-LABEL: func @test_kernel_environment_with_async +// CHECK: %[[ASYNC:.*]] = arith.constant 1 : i32 +// CHECK: %[[CREATE:.*]] = acc.create varPtr(%{{.*}} : memref<1024xf32>) async(%[[ASYNC]] : i32) -> memref<1024xf32> +// CHECK: acc.kernel_environment dataOperands(%[[CREATE]] : memref<1024xf32>) async(%[[ASYNC]] : i32) +// CHECK: gpu.launch +// CHECK: memref.store %{{.*}}, %[[CREATE]] +// CHECK: acc.copyout accPtr(%[[CREATE]] : memref<1024xf32>) async(%[[ASYNC]] : i32) to varPtr(%{{.*}} : memref<1024xf32>) |
vzakhari approved these changes Oct 28, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit. This suggestion is invalid because no changes were made to the code. Suggestions cannot be applied while the pull request is closed. Suggestions cannot be applied while viewing a subset of changes. Only one suggestion per line can be applied in a batch. Add this suggestion to a batch that can be applied as a single commit. Applying suggestions on deleted lines is not supported. You must change the existing code in this line in order to create a valid suggestion. Outdated suggestions cannot be applied. This suggestion has been applied or marked resolved. Suggestions cannot be applied from pending reviews. Suggestions cannot be applied on multi-line comments. Suggestions cannot be applied while the pull request is queued to merge. Suggestion cannot be applied right now. Please check back later.
Introduce
acc.kernel_environmentoperation to capture data mapping and asynchronous behavior from OpenACC compute constructs. This enables decomposition by separating data movement and synchronization from kernel execution parallelism, facilitating lowering to GPU dialect.