[Mlir-commits] [mlir] [mlir][nvgpu] Fix `transposeB` in `nvgpu.warpgroup.mma` (PR #79271)
Guray Ozen
llvmlistbot at llvm.org
Wed Jan 24 11:49:46 PST 2024
https://github.com/grypp updated https://github.com/llvm/llvm-project/pull/79271
>From 19865938501e42493ae8efbe11ef1ceba872c4b4 Mon Sep 17 00:00:00 2001
From: Guray Ozen <guray.ozen at gmail.com>
Date: Wed, 24 Jan 2024 11:44:28 +0100
Subject: [PATCH 1/3] [mlir][nvgpu] Fix `transposeB` in `nvgpu.warpgroup.mma`
The #76150 fixed meaning of `transposeB` in NVVM dialect which was initially implemented with opposite meaning.
This PR fixes the lowering of `nvgpu.warpgroup.mma` to NVVM dialect.
This will fix two integration tests:
gemm_f32_f16_f16_128x128x128.mlir
gemm_pred_f32_f16_f16_128x128x128.mlir
---
mlir/lib/Conversion/NVGPUToNVVM/NVGPUToNVVM.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mlir/lib/Conversion/NVGPUToNVVM/NVGPUToNVVM.cpp b/mlir/lib/Conversion/NVGPUToNVVM/NVGPUToNVVM.cpp
index 43d05b872a4fbc..5080956a458982 100644
--- a/mlir/lib/Conversion/NVGPUToNVVM/NVGPUToNVVM.cpp
+++ b/mlir/lib/Conversion/NVGPUToNVVM/NVGPUToNVVM.cpp
@@ -1407,7 +1407,7 @@ struct NVGPUWarpgroupMmaOpLowering
NVVM::WGMMAScaleOutAttr scaleOut = generateScaleOut();
NVVM::WGMMAScaleInAttr scaleIn = generateScaleIn();
NVVM::MMALayoutAttr layoutA = generateWgmmaLayout(op.getTransposeA());
- NVVM::MMALayoutAttr layoutB = generateWgmmaLayout(op.getTransposeB());
+ NVVM::MMALayoutAttr layoutB = generateWgmmaLayout(!op.getTransposeB());
auto overflow = NVVM::MMAIntOverflowAttr::get(
op->getContext(), NVVM::MMAIntOverflow::wrapped);
>From 7de3c17872b0db77c12b0e5f1e46f911a213e8c0 Mon Sep 17 00:00:00 2001
From: Guray Ozen <guray.ozen at gmail.com>
Date: Wed, 24 Jan 2024 12:33:56 +0100
Subject: [PATCH 2/3] fix test
---
mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
index 0ac7331e1f6987..e97ac73c6f0374 100644
--- a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
+++ b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
@@ -297,7 +297,7 @@ func.func @wgmma_f32_f16_f16(%descA : i64, %descB : i64) -> !mat64f32{
// CHECK: %[[A2:.*]] = llvm.mlir.constant(-1 : i32) : i32
// CHECK: %[[A3:.*]] = llvm.mlir.constant(-1 : i32) : i32
// CHECK: %[[A4:.*]] = llvm.mlir.constant(1 : i32) : i32
- // CHECK: %[[A5:.*]] = llvm.mlir.constant(0 : i32) : i32
+ // CHECK: %[[A5:.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[V0:.*]] = llvm.extractvalue %[[RES]][0] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V4:.*]] = llvm.extractvalue %[[RES]][4] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V11:.*]] = llvm.extractvalue %[[RES]][11] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
>From ab9f67c6e8ce1fb4bddebb0d6952419eed0eb16a Mon Sep 17 00:00:00 2001
From: Guray Ozen <guray.ozen at gmail.com>
Date: Wed, 24 Jan 2024 20:49:03 +0100
Subject: [PATCH 3/3] fix the right test
---
.../Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir | 18 +++++++++---------
.../Conversion/NVVMToLLVM/nvvm-to-llvm.mlir | 2 +-
2 files changed, 10 insertions(+), 10 deletions(-)
diff --git a/mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir b/mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir
index b25dd76bf12037..d0ff38a1d6034b 100644
--- a/mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir
@@ -880,41 +880,41 @@ func.func @warpgroup_mma_128_128_64(
// CHECK: nvvm.wgmma.fence.aligned
// CHECK: %[[UD:.+]] = llvm.mlir.undef : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
// CHECK: %[[S2:.+]] = llvm.extractvalue %[[ARG]][0] : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
-// CHECK: %[[S4:.+]] = nvvm.wgmma.mma_async %[[S0]], %[[S1]], %[[S2]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)> -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
+// CHECK: %[[S4:.+]] = nvvm.wgmma.mma_async %[[S0]], %[[S1]], %[[S2]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)> -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[S5:.+]] = llvm.mlir.constant(2 : i32) : i64
// CHECK: %[[S6:.+]] = llvm.add %[[S0]], %[[S5]] : i64
// CHECK: %[[S7:.+]] = llvm.mlir.constant(128 : i32) : i64
// CHECK: %[[S8:.+]] = llvm.add %[[S1]], %[[S7]] : i64
-// CHECK: %[[S9:.+]] = nvvm.wgmma.mma_async %[[S6]], %[[S8]], %[[S4]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S9:.+]] = nvvm.wgmma.mma_async %[[S6]], %[[S8]], %[[S4]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S10:.+]] = llvm.mlir.constant(4 : i32) : i64
// CHECK: %[[S11:.+]] = llvm.add %[[S0]], %[[S10]] : i64
// CHECK: %[[S12:.+]] = llvm.mlir.constant(256 : i32) : i64
// CHECK: %[[S13:.+]] = llvm.add %[[S1]], %[[S12]] : i64
-// CHECK: %[[S14:.+]] = nvvm.wgmma.mma_async %[[S11]], %[[S13]], %[[S9]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S14:.+]] = nvvm.wgmma.mma_async %[[S11]], %[[S13]], %[[S9]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S15:.+]] = llvm.mlir.constant(6 : i32) : i64
// CHECK: %[[S16:.+]] = llvm.add %[[S0]], %[[S15]] : i64
// CHECK: %[[S17:.+]] = llvm.mlir.constant(384 : i32) : i64
// CHECK: %[[S18:.+]] = llvm.add %[[S1]], %[[S17]] : i64
-// CHECK: %[[S19:.+]] = nvvm.wgmma.mma_async %[[S16]], %[[S18]], %[[S14]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S19:.+]] = nvvm.wgmma.mma_async %[[S16]], %[[S18]], %[[S14]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S3:.+]] = llvm.extractvalue %[[ARG]][1] : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
// CHECK: %[[S21:.+]] = llvm.mlir.constant(512 : i32) : i64
// CHECK: %[[S22:.+]] = llvm.add %[[S0]], %[[S21]] : i64
-// CHECK: %[[S23:.+]] = nvvm.wgmma.mma_async %[[S22]], %[[S1]], %[[S3]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S23:.+]] = nvvm.wgmma.mma_async %[[S22]], %[[S1]], %[[S3]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S24:.+]] = llvm.mlir.constant(514 : i32) : i64
// CHECK: %[[S25:.+]] = llvm.add %[[S0]], %[[S24]] : i64
// CHECK: %[[S26:.+]] = llvm.mlir.constant(128 : i32) : i64
// CHECK: %[[S27:.+]] = llvm.add %[[S1]], %[[S26]] : i64
-// CHECK: %[[S28:.+]] = nvvm.wgmma.mma_async %[[S25]], %[[S27]], %[[S23]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S28:.+]] = nvvm.wgmma.mma_async %[[S25]], %[[S27]], %[[S23]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S29:.+]] = llvm.mlir.constant(516 : i32) : i64
// CHECK: %[[S30:.+]] = llvm.add %[[S0]], %[[S29]] : i64
// CHECK: %[[S31:.+]] = llvm.mlir.constant(256 : i32) : i64
// CHECK: %[[S32:.+]] = llvm.add %[[S1]], %[[S31]] : i64
-// CHECK: %[[S33:.+]] = nvvm.wgmma.mma_async %[[S30]], %[[S32]], %[[S28]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S33:.+]] = nvvm.wgmma.mma_async %[[S30]], %[[S32]], %[[S28]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S34:.+]] = llvm.mlir.constant(518 : i32) : i64
// CHECK: %[[S35:.+]] = llvm.add %[[S0]], %[[S34]] : i64
// CHECK: %[[S36:.+]] = llvm.mlir.constant(384 : i32) : i64
// CHECK: %[[S37:.+]] = llvm.add %[[S1]], %[[S36]] : i64
-// CHECK: %[[S38:.+]] = nvvm.wgmma.mma_async %[[S35]], %[[S37]], %[[S33]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct
+// CHECK: %[[S38:.+]] = nvvm.wgmma.mma_async %[[S35]], %[[S37]], %[[S33]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct
// CHECK: %[[S40:.+]] = llvm.insertvalue %[[S19]], %[[UD]][0] : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
// CHECK: %[[S41:.+]] = llvm.insertvalue %[[S38]], %[[S40]][1] : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
// CHECK: nvvm.wgmma.commit.group.sync.aligned
@@ -1299,7 +1299,7 @@ func.func @warpgroup_matrix_multiply_m128n128k64(
// CHECK: nvvm.wgmma.fence.aligned
// CHECK: %[[S137:.+]] = llvm.mlir.undef : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
// CHECK: %[[S138:.+]] = llvm.extractvalue %136[0] : !llvm.struct<(struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>, struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>)>
-// CHECK: %[[S139:.+]] = nvvm.wgmma.mma_async %0, %1, %[[S138]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <col>] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)> -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
+// CHECK: %[[S139:.+]] = nvvm.wgmma.mma_async %0, %1, %[[S138]], <m = 64, n = 128, k = 16>, D[<f32>, <one>, <wrapped>], A[<f16>, <one>, <row>], B[<f16>, <one>, <row>] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)> -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: nvvm.wgmma.mma_async
// CHECK: nvvm.wgmma.mma_async
// CHECK: %[[S154:.+]] = nvvm.wgmma.mma_async
diff --git a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
index e97ac73c6f0374..0ac7331e1f6987 100644
--- a/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
+++ b/mlir/test/Conversion/NVVMToLLVM/nvvm-to-llvm.mlir
@@ -297,7 +297,7 @@ func.func @wgmma_f32_f16_f16(%descA : i64, %descB : i64) -> !mat64f32{
// CHECK: %[[A2:.*]] = llvm.mlir.constant(-1 : i32) : i32
// CHECK: %[[A3:.*]] = llvm.mlir.constant(-1 : i32) : i32
// CHECK: %[[A4:.*]] = llvm.mlir.constant(1 : i32) : i32
- // CHECK: %[[A5:.*]] = llvm.mlir.constant(1 : i32) : i32
+ // CHECK: %[[A5:.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[V0:.*]] = llvm.extractvalue %[[RES]][0] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V4:.*]] = llvm.extractvalue %[[RES]][4] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
// CHECK: %[[V11:.*]] = llvm.extractvalue %[[RES]][11] : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32, f32)>
More information about the Mlir-commits
mailing list