[Mlir-commits] [mlir] [mlir][AMDGPU] Add PermlaneOp (PR #154345)
Tim Gymnich
llvmlistbot at llvm.org
Tue Aug 19 08:34:34 PDT 2025
https://github.com/tgymnich updated https://github.com/llvm/llvm-project/pull/154345
>From a061dbaf32aaafe1ad302217382f9cfbb8cfac89 Mon Sep 17 00:00:00 2001
From: Tim Gymnich <tim at gymni.ch>
Date: Mon, 18 Aug 2025 13:18:00 +0000
Subject: [PATCH 1/2] add AMDGPU PermlaneOp
---
mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td | 29 ++++
.../AMDGPUToROCDL/AMDGPUToROCDL.cpp | 52 +++++-
.../Conversion/AMDGPUToROCDL/permlane.mlir | 153 ++++++++++++++++++
mlir/test/Dialect/AMDGPU/ops.mlir | 14 ++
4 files changed, 247 insertions(+), 1 deletion(-)
create mode 100644 mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
diff --git a/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td b/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
index 2c646934c11c2..17386a2888788 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
@@ -656,6 +656,35 @@ def AMDGPU_SwizzleBitModeOp : AMDGPU_Op<"swizzle_bitmode",
}];
}
+def AMDGPU_PermlanePerm : I32EnumAttr<"PermlanePerm",
+ "The possible permutations for a permlane operation",
+ [
+ I32EnumAttrCase<"swap_16", 0>,
+ I32EnumAttrCase<"swap_32", 1>,
+ ]> {
+ let genSpecializedAttr = 0;
+ let cppNamespace = "::mlir::amdgpu";
+}
+
+def AMDGPU_PermlanePermAttr : EnumAttr<AMDGPU_Dialect, AMDGPU_PermlanePerm,
+ "permlane_perm">;
+
+def AMDGPU_PermlaneOp : AMDGPU_Op<"permlane", [Pure, AllTypesMatch<["result", "src"]>]>,
+Arguments<(ins AnyIntegerOrFloatOr1DVector:$src,
+ AMDGPU_PermlanePermAttr:$kind)> {
+ let summary = "AMDGPU permlane op";
+ let description = [{
+ High-level wrapper on `rocdl.permlane` variants.
+
+ Supports arbitrary int/float/vector types, which will be repacked to i32 and
+ one or more `rocdl.permlane` ops during lowering.
+ }];
+ let results = (outs AnyIntegerOrFloatOr1DVector:$result);
+ let assemblyFormat = [{
+ $src $kind attr-dict `:` type($result)
+ }];
+}
+
def AMDGPU_LDSBarrierOp : AMDGPU_Op<"lds_barrier"> {
let summary = "Barrier that includes a wait for LDS memory operations.";
let description = [{
diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index 64720bfe6cf50..4b0f8bcb40811 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -14,6 +14,7 @@
#include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h"
#include "mlir/Dialect/AMDGPU/Utils/Chipset.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/Dialect/LLVMIR/LLVMTypes.h"
#include "mlir/Dialect/LLVMIR/ROCDLDialect.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/TypeUtilities.h"
@@ -1876,6 +1877,55 @@ struct AMDGPUSwizzleBitModeLowering
}
};
+struct AMDGPUPermlaneLowering : public ConvertOpToLLVMPattern<PermlaneOp> {
+ using ConvertOpToLLVMPattern::ConvertOpToLLVMPattern;
+
+ AMDGPUPermlaneLowering(const LLVMTypeConverter &converter, Chipset chipset)
+ : ConvertOpToLLVMPattern<PermlaneOp>(converter), chipset(chipset) {}
+ Chipset chipset;
+
+ LogicalResult
+ matchAndRewrite(PermlaneOp op, OpAdaptor adaptor,
+ ConversionPatternRewriter &rewriter) const override {
+ Location loc = op.getLoc();
+ Type i32 = rewriter.getI32Type();
+ Value src = adaptor.getSrc();
+ auto kind = op.getKind();
+ auto fi = rewriter.getBoolAttr(false);
+ auto boundctrl = rewriter.getBoolAttr(false);
+
+ if (chipset < kGfx950)
+ return op->emitOpError("permlane_swap is only supported on gfx950+");
+
+ SmallVector<Value> decomposed =
+ LLVM::decomposeValue(rewriter, loc, src, i32);
+
+ SmallVector<Value> permuted;
+ for (Value v : decomposed) {
+ Value res;
+ Type i32pair = LLVM::LLVMStructType::getLiteral(
+ rewriter.getContext(), {v.getType(), v.getType()});
+ switch (kind) {
+ case PermlanePerm::swap_16:
+ res = ROCDL::Permlane16SwapOp::create(rewriter, loc, i32pair, v, v, fi,
+ boundctrl);
+ break;
+ case PermlanePerm::swap_32:
+ res = ROCDL::Permlane32SwapOp::create(rewriter, loc, i32pair, v, v, fi,
+ boundctrl);
+ break;
+ }
+
+ Value vdstNew = LLVM::ExtractValueOp::create(rewriter, loc, res, {0});
+ permuted.emplace_back(vdstNew);
+ }
+
+ Value result = LLVM::composeValue(rewriter, loc, permuted, src.getType());
+ rewriter.replaceOp(op, result);
+ return success();
+ }
+};
+
struct ConvertAMDGPUToROCDLPass
: public impl::ConvertAMDGPUToROCDLPassBase<ConvertAMDGPUToROCDLPass> {
using Base::Base;
@@ -1944,6 +1994,6 @@ void mlir::populateAMDGPUToROCDLConversionPatterns(LLVMTypeConverter &converter,
WMMAOpLowering, ExtPackedFp8OpLowering, ScaledExtPackedOpLowering,
PackedScaledTruncOpLowering, PackedTrunc2xFp8OpLowering,
PackedStochRoundFp8OpLowering, GatherToLDSOpLowering,
- TransposeLoadOpLowering>(converter, chipset);
+ TransposeLoadOpLowering, AMDGPUPermlaneLowering>(converter, chipset);
patterns.add<AMDGPUSwizzleBitModeLowering>(converter);
}
diff --git a/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir b/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
new file mode 100644
index 0000000000000..5b216e393043f
--- /dev/null
+++ b/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
@@ -0,0 +1,153 @@
+// RUN: mlir-opt -convert-amdgpu-to-rocdl=chipset=gfx950 --canonicalize %s | FileCheck %s
+
+// CHECK-LABEL: func @test_permlane16_i32
+// CHECK-SAME: (%[[ARG0:.*]]: i32)
+func.func @test_permlane16_i32(%arg0 : i32) -> i32 {
+// CHECK: %[[PERM:.*]] = rocdl.permlane16.swap %[[ARG0]], %[[ARG0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: return %[[RES]] : i32
+ %0 = amdgpu.permlane %arg0 swap_16 : i32
+ return %0 : i32
+}
+
+// CHECK-LABEL: func @test_permlane32_i32
+// CHECK-SAME: (%[[ARG0:.*]]: i32)
+func.func @test_permlane32_i32(%arg0 : i32) -> i32 {
+// CHECK: %[[PERM:.*]] = rocdl.permlane32.swap %[[ARG0]], %[[ARG0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: return %[[RES]] : i32
+ %0 = amdgpu.permlane %arg0 swap_32 : i32
+ return %0 : i32
+}
+
+// CHECK-LABEL: func @test_permlane16_f32
+// CHECK-SAME: (%[[ARG0:.*]]: f32)
+func.func @test_permlane16_f32(%arg0 : f32) -> f32 {
+// CHECK: %[[CAST:.*]] = llvm.bitcast %[[ARG0]] : f32 to i32
+// CHECK: %[[PERM:.*]] = rocdl.permlane16.swap %[[CAST]], %[[CAST]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[RES_CAST:.*]] = llvm.bitcast %[[RES]] : i32 to f32
+// CHECK: return %[[RES_CAST]] : f32
+ %0 = amdgpu.permlane %arg0 swap_16 : f32
+ return %0 : f32
+}
+
+// CHECK-LABEL: func @test_permlane32_f32
+// CHECK-SAME: (%[[ARG0:.*]]: f32)
+func.func @test_permlane32_f32(%arg0 : f32) -> f32 {
+// CHECK: %[[CAST:.*]] = llvm.bitcast %[[ARG0]] : f32 to i32
+// CHECK: %[[PERM:.*]] = rocdl.permlane32.swap %[[CAST]], %[[CAST]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[RES_CAST:.*]] = llvm.bitcast %[[RES]] : i32 to f32
+// CHECK: return %[[RES_CAST]] : f32
+ %0 = amdgpu.permlane %arg0 swap_32 : f32
+ return %0 : f32
+}
+
+// CHECK-LABEL: func @test_permlane16_f16
+// CHECK-SAME: (%[[ARG0:.*]]: f16)
+func.func @test_permlane16_f16(%arg0 : f16) -> f16 {
+// CHECK: %[[CAST:.*]] = llvm.bitcast %[[ARG0]] : f16 to i16
+// CHECK: %[[ZEXT:.*]] = llvm.zext %[[CAST]] : i16 to i32
+// CHECK: %[[PERM:.*]] = rocdl.permlane16.swap %[[ZEXT]], %[[ZEXT]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[TRUNC:.*]] = llvm.trunc %[[RES]] : i32 to i16
+// CHECK: %[[RES_CAST:.*]] = llvm.bitcast %[[TRUNC]] : i16 to f16
+// CHECK: return %[[RES_CAST]] : f16
+ %0 = amdgpu.permlane %arg0 swap_16 : f16
+ return %0 : f16
+}
+
+// CHECK-LABEL: func @test_permlane32_f16
+// CHECK-SAME: (%[[ARG0:.*]]: f16)
+func.func @test_permlane32_f16(%arg0 : f16) -> f16 {
+// CHECK: %[[CAST:.*]] = llvm.bitcast %[[ARG0]] : f16 to i16
+// CHECK: %[[ZEXT:.*]] = llvm.zext %[[CAST]] : i16 to i32
+// CHECK: %[[PERM:.*]] = rocdl.permlane32.swap %[[ZEXT]], %[[ZEXT]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[TRUNC:.*]] = llvm.trunc %[[RES]] : i32 to i16
+// CHECK: %[[RES_CAST:.*]] = llvm.bitcast %[[TRUNC]] : i16 to f16
+// CHECK: return %[[RES_CAST]] : f16
+ %0 = amdgpu.permlane %arg0 swap_32 : f16
+ return %0 : f16
+}
+
+// CHECK-LABEL: func @test_permlane16_2xi32
+// CHECK-SAME: (%[[ARG0:.*]]: vector<2xi32>)
+func.func @test_permlane16_2xi32(%arg0 : vector<2xi32>) -> vector<2xi32> {
+// CHECK-DAG: %[[POISON:.*]] = llvm.mlir.poison : vector<2xi32>
+// CHECK-DAG: %[[C1:.*]] = llvm.mlir.constant(1 : i32) : i32
+// CHECK-DAG: %[[C0:.*]] = llvm.mlir.constant(0 : i32) : i32
+// CHECK: %[[ELEM0:.*]] = llvm.extractelement %[[ARG0]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[ELEM1:.*]] = llvm.extractelement %[[ARG0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[PERM0_TUPLE:.*]] = rocdl.permlane16.swap %[[ELEM0]], %[[ELEM0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM0:.*]] = llvm.extractvalue %[[PERM0_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[PERM1_TUPLE:.*]] = rocdl.permlane16.swap %[[ELEM1]], %[[ELEM1]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM1:.*]] = llvm.extractvalue %[[PERM1_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[VEC_INSERT0:.*]] = llvm.insertelement %[[PERM0]], %[[POISON]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[VEC_INSERT1:.*]] = llvm.insertelement %[[PERM1]], %[[VEC_INSERT0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: return %[[VEC_INSERT1]] : vector<2xi32>
+ %0 = amdgpu.permlane %arg0 swap_16 : vector<2xi32>
+ return %0 : vector<2xi32>
+}
+
+// CHECK-LABEL: func @test_permlane32_2xi32
+// CHECK-SAME: (%[[ARG0:.*]]: vector<2xi32>)
+func.func @test_permlane32_2xi32(%arg0 : vector<2xi32>) -> vector<2xi32> {
+// CHECK-DAG: %[[POISON:.*]] = llvm.mlir.poison : vector<2xi32>
+// CHECK-DAG: %[[C1:.*]] = llvm.mlir.constant(1 : i32) : i32
+// CHECK-DAG: %[[C0:.*]] = llvm.mlir.constant(0 : i32) : i32
+// CHECK: %[[ELEM0:.*]] = llvm.extractelement %[[ARG0]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[ELEM1:.*]] = llvm.extractelement %[[ARG0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[PERM0_TUPLE:.*]] = rocdl.permlane32.swap %[[ELEM0]], %[[ELEM0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM0:.*]] = llvm.extractvalue %[[PERM0_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[PERM1_TUPLE:.*]] = rocdl.permlane32.swap %[[ELEM1]], %[[ELEM1]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM1:.*]] = llvm.extractvalue %[[PERM1_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[VEC_INSERT0:.*]] = llvm.insertelement %[[PERM0]], %[[POISON]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[VEC_INSERT1:.*]] = llvm.insertelement %[[PERM1]], %[[VEC_INSERT0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: return %[[VEC_INSERT1]] : vector<2xi32>
+ %0 = amdgpu.permlane %arg0 swap_32 : vector<2xi32>
+ return %0 : vector<2xi32>
+}
+
+// CHECK-LABEL: func @test_permlane16_4xf16
+// CHECK-SAME: (%[[ARG0:.*]]: vector<4xf16>)
+func.func @test_permlane16_4xf16(%arg0 : vector<4xf16>) -> vector<4xf16> {
+// CHECK-DAG: %[[POISON:.*]] = llvm.mlir.poison : vector<2xi32>
+// CHECK-DAG: %[[C1:.*]] = llvm.mlir.constant(1 : i32) : i32
+// CHECK-DAG: %[[C0:.*]] = llvm.mlir.constant(0 : i32) : i32
+// CHECK: %[[CAST1:.*]] = llvm.bitcast %[[ARG0]] : vector<4xf16> to vector<2xi32>
+// CHECK: %[[ELEM0:.*]] = llvm.extractelement %[[CAST1]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[ELEM1:.*]] = llvm.extractelement %[[CAST1]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[PERM0_TUPLE:.*]] = rocdl.permlane16.swap %[[ELEM0]], %[[ELEM0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM0:.*]] = llvm.extractvalue %[[PERM0_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[PERM1_TUPLE:.*]] = rocdl.permlane16.swap %[[ELEM1]], %[[ELEM1]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM1:.*]] = llvm.extractvalue %[[PERM1_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[VEC_INSERT0:.*]] = llvm.insertelement %[[PERM0]], %[[POISON]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[VEC_INSERT1:.*]] = llvm.insertelement %[[PERM1]], %[[VEC_INSERT0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[CAST2:.*]] = llvm.bitcast %[[VEC_INSERT1]] : vector<2xi32> to vector<4xf16>
+// CHECK: return %[[CAST2]] : vector<4xf16>
+ %0 = amdgpu.permlane %arg0 swap_16 : vector<4xf16>
+ return %0 : vector<4xf16>
+}
+
+// CHECK-LABEL: func @test_permlane32_4xf16
+// CHECK-SAME: (%[[ARG0:.*]]: vector<4xf16>)
+func.func @test_permlane32_4xf16(%arg0 : vector<4xf16>) -> vector<4xf16> {
+// CHECK-DAG: %[[POISON:.*]] = llvm.mlir.poison : vector<2xi32>
+// CHECK-DAG: %[[C1:.*]] = llvm.mlir.constant(1 : i32) : i32
+// CHECK-DAG: %[[C0:.*]] = llvm.mlir.constant(0 : i32) : i32
+// CHECK: %[[CAST1:.*]] = llvm.bitcast %[[ARG0]] : vector<4xf16> to vector<2xi32>
+// CHECK: %[[ELEM0:.*]] = llvm.extractelement %[[CAST1]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[ELEM1:.*]] = llvm.extractelement %[[CAST1]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[PERM0_TUPLE:.*]] = rocdl.permlane32.swap %[[ELEM0]], %[[ELEM0]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM0:.*]] = llvm.extractvalue %[[PERM0_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[PERM1_TUPLE:.*]] = rocdl.permlane32.swap %[[ELEM1]], %[[ELEM1]], false, false : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[PERM1:.*]] = llvm.extractvalue %[[PERM1_TUPLE]][0] : !llvm.struct<(i32, i32)>
+// CHECK: %[[VEC_INSERT0:.*]] = llvm.insertelement %[[PERM0]], %[[POISON]][%[[C0]] : i32] : vector<2xi32>
+// CHECK: %[[VEC_INSERT1:.*]] = llvm.insertelement %[[PERM1]], %[[VEC_INSERT0]][%[[C1]] : i32] : vector<2xi32>
+// CHECK: %[[CAST2:.*]] = llvm.bitcast %[[VEC_INSERT1]] : vector<2xi32> to vector<4xf16>
+// CHECK: return %[[CAST2]] : vector<4xf16>
+ %0 = amdgpu.permlane %arg0 swap_32 : vector<4xf16>
+ return %0 : vector<4xf16>
+}
diff --git a/mlir/test/Dialect/AMDGPU/ops.mlir b/mlir/test/Dialect/AMDGPU/ops.mlir
index 87e11c028c62a..5eb6d53d4d9fe 100644
--- a/mlir/test/Dialect/AMDGPU/ops.mlir
+++ b/mlir/test/Dialect/AMDGPU/ops.mlir
@@ -524,6 +524,20 @@ func.func @swizzle_bitmode(%arg0 : f32) -> f32 {
func.return %0 : f32
}
+// CHECK-LABEL: func @permlane16_swap
+func.func @permlane16_swap(%arg0 : f32) -> f32 {
+ // CHECK: amdgpu.permlane
+ %0 = amdgpu.permlane %arg0 swap_16 : f32
+ func.return %0 : f32
+}
+
+// CHECK-LABEL: func @permlane32_swap
+func.func @permlane32_swap(%arg0 : f32) -> f32 {
+ // CHECK: amdgpu.permlane
+ %0 = amdgpu.permlane %arg0 swap_32 : f32
+ func.return %0 : f32
+}
+
// CHECK-LABEL: func @scaled_mfma
func.func @scaled_mfma(%arg0 : f8E8M0FNU, %arg1 : vector<32xf6E2M3FN>, %arg2 : vector<16xf32>) -> vector<16xf32> {
// CHECK: amdgpu.scaled_mfma
>From fa28436f47c1528fbe223d160885b6b1a014dc90 Mon Sep 17 00:00:00 2001
From: Tim Gymnich <tim at gymni.ch>
Date: Tue, 19 Aug 2025 10:34:19 -0500
Subject: [PATCH 2/2] expose boundCtrl and fi
---
mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td | 4 +++-
mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp | 4 ++--
mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir | 10 ++++++++++
3 files changed, 15 insertions(+), 3 deletions(-)
diff --git a/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td b/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
index 17386a2888788..019da46246841 100644
--- a/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
+++ b/mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
@@ -671,7 +671,9 @@ def AMDGPU_PermlanePermAttr : EnumAttr<AMDGPU_Dialect, AMDGPU_PermlanePerm,
def AMDGPU_PermlaneOp : AMDGPU_Op<"permlane", [Pure, AllTypesMatch<["result", "src"]>]>,
Arguments<(ins AnyIntegerOrFloatOr1DVector:$src,
- AMDGPU_PermlanePermAttr:$kind)> {
+ AMDGPU_PermlanePermAttr:$kind,
+ DefaultValuedAttr<BoolAttr, "false">:$fi,
+ DefaultValuedAttr<BoolAttr, "false">:$bound_ctrl)> {
let summary = "AMDGPU permlane op";
let description = [{
High-level wrapper on `rocdl.permlane` variants.
diff --git a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
index 4b0f8bcb40811..1f5c6bec92c42 100644
--- a/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
+++ b/mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
@@ -1891,8 +1891,8 @@ struct AMDGPUPermlaneLowering : public ConvertOpToLLVMPattern<PermlaneOp> {
Type i32 = rewriter.getI32Type();
Value src = adaptor.getSrc();
auto kind = op.getKind();
- auto fi = rewriter.getBoolAttr(false);
- auto boundctrl = rewriter.getBoolAttr(false);
+ auto fi = op.getFi();
+ auto boundctrl = op.getBoundCtrl();
if (chipset < kGfx950)
return op->emitOpError("permlane_swap is only supported on gfx950+");
diff --git a/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir b/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
index 5b216e393043f..81b98c4e6bc3e 100644
--- a/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
+++ b/mlir/test/Conversion/AMDGPUToROCDL/permlane.mlir
@@ -10,6 +10,16 @@ func.func @test_permlane16_i32(%arg0 : i32) -> i32 {
return %0 : i32
}
+// CHECK-LABEL: func @test_permlane16_i32_optional_attr
+// CHECK-SAME: (%[[ARG0:.*]]: i32)
+func.func @test_permlane16_i32_optional_attr(%arg0 : i32) -> i32 {
+// CHECK: %[[PERM:.*]] = rocdl.permlane16.swap %[[ARG0]], %[[ARG0]], true, true : (i32, i32) -> <(i32, i32)>
+// CHECK: %[[RES:.*]] = llvm.extractvalue %[[PERM]][0] : !llvm.struct<(i32, i32)>
+// CHECK: return %[[RES]] : i32
+ %0 = amdgpu.permlane %arg0 swap_16 { fi = true, bound_ctrl = true } : i32
+ return %0 : i32
+}
+
// CHECK-LABEL: func @test_permlane32_i32
// CHECK-SAME: (%[[ARG0:.*]]: i32)
func.func @test_permlane32_i32(%arg0 : i32) -> i32 {
More information about the Mlir-commits
mailing list