Skip to content

[mlir][amdgpu] Update mfma assembly format with intrinsic shape #165037

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 25, 2025
Merged

[mlir][amdgpu] Update mfma assembly format with intrinsic shape #165037

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
56f0578
[mlir][amdgpu] Update mfma assembly format with intrinsic shape
kuhar Oct 24, 2025
d195dc0
Clean up example
kuhar Oct 24, 2025
37f4385
Clean up example
kuhar Oct 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 10 additions & 8 deletions mlir/include/mlir/Dialect/AMDGPU/IR/AMDGPU.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -923,10 +923,10 @@ def AMDGPU_MFMAOp :
AMDGPU_Op<"mfma", [AllTypesMatch<["destC", "destD"]>,
Pure]>,
Arguments<(ins
I32Attr:$m,
I32Attr:$n,
I32Attr:$k,
I32Attr:$blocks,
ConfinedAttr<I32Attr, [IntIsOneOf<[4, 16, 32]>]>:$m,
ConfinedAttr<I32Attr, [IntIsOneOf<[4, 16, 32]>]>:$n,
ConfinedAttr<I32Attr, [IntIsOneOf<[1, 2, 4, 8, 16, 32, 64, 128]>]>:$k,
DefaultValuedAttr<ConfinedAttr<I32Attr, [IntIsOneOf<[1, 2, 4, 16]>]>, "1">:$blocks,
MFMAInTypes:$sourceA,
MFMAInTypes:$sourceB,
MFMAOutTypes:$destC,
Expand Down Expand Up @@ -969,14 +969,16 @@ def AMDGPU_MFMAOp :

Example:
```mlir
%0 = amdgpu.mfma %matA * %matB + %matC
{ abid = 1 : i32, cbsz = 1 : i32,
m = 32 : i32, n = 32 : i32, k = 1 : i32, blocks = 2 : i32 }
%0 = amdgpu.mfma 16x16x16 %matA * %matB + %matC
: vector<4xf16>, vector<4xf16>, vector<4xf32>

%1 = amdgpu.mfma 32x32x1 %matD * %matE + %matF
{ abid = 1 : i32, cbsz = 1 : i32, blocks = 2 : i32 }
blgp = bcast_second_32 : f32, f32, vector<32xf32>
```
}];
let assemblyFormat = [{
$sourceA `*` $sourceB `+` $destC
custom<MNKDimensionList>($m, $n, $k) $sourceA `*` $sourceB `+` $destC
attr-dict
`blgp` `=` $blgp
`:` type($sourceA) `,` type($sourceB) `,` type($destC)
Expand Down
6 changes: 3 additions & 3 deletions mlir/lib/Dialect/AMDGPU/IR/AMDGPUDialect.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -422,11 +422,11 @@ LogicalResult MFMAOp::verify() {

Type sourceElem = sourceType, destElem = destType;
uint32_t sourceLen = 1, destLen = 1;
if (auto sourceVector = llvm::dyn_cast<VectorType>(sourceType)) {
if (auto sourceVector = dyn_cast<VectorType>(sourceType)) {
sourceLen = sourceVector.getNumElements();
sourceElem = sourceVector.getElementType();
}
if (auto destVector = llvm::dyn_cast<VectorType>(destType)) {
if (auto destVector = dyn_cast<VectorType>(destType)) {
destLen = destVector.getNumElements();
destElem = destVector.getElementType();
}
Expand All @@ -451,7 +451,7 @@ LogicalResult MFMAOp::verify() {
return emitOpError("expected both non-small-float source operand types "
"to match exactly");
}
// Normalize the wider integer types the compiler expects to i8
// Normalize the wider integer types the compiler expects to i8.
if (sourceElem.isInteger(32)) {
sourceLen *= 4;
sourceElem = b.getI8Type();
Expand Down
54 changes: 27 additions & 27 deletions mlir/test/Conversion/AMDGPUToROCDL/mfma-gfx950.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,46 +8,46 @@ func.func @mfma_to_rocdl(%arg0 : vector<8xf16>, %arg1 : vector<16xf32>,
// CHECK: %[[c0:.+]] = llvm.mlir.constant(0 : i32) : i32

// CHECK: rocdl.mfma.f32.32x32x16.f16{{.*}}: (vector<8xf16>, vector<8xf16>, vector<16xf32>, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg0 * %arg0 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 16 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<8xf16>, vector<8xf16>, vector<16xf32>
amdgpu.mfma 32x32x16 %arg0 * %arg0 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<8xf16>, vector<8xf16>, vector<16xf32>
// CHECK: rocdl.mfma.f32.16x16x32.f16{{.*}}: (vector<8xf16>, vector<8xf16>, vector<4xf32>, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg0 * %arg0 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 32 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<8xf16>, vector<8xf16>, vector<4xf32>
amdgpu.mfma 16x16x32 %arg0 * %arg0 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<8xf16>, vector<8xf16>, vector<4xf32>
// CHECK: rocdl.mfma.f32.32x32x16.bf16{{.*}}: (vector<8xbf16>, vector<8xbf16>, vector<16xf32>, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg3 * %arg3 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 16 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<8xbf16>, vector<8xbf16>, vector<16xf32>
amdgpu.mfma 32x32x16 %arg3 * %arg3 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<8xbf16>, vector<8xbf16>, vector<16xf32>
// CHECK: rocdl.mfma.f32.16x16x32.bf16{{.*}}: (vector<8xbf16>, vector<8xbf16>, vector<4xf32>, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg3 * %arg3 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 32 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<8xbf16>, vector<8xbf16>, vector<4xf32>
amdgpu.mfma 16x16x32 %arg3 * %arg3 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<8xbf16>, vector<8xbf16>, vector<4xf32>
// CHECK: rocdl.mfma.i32.32x32x32.i8{{.*}}: (vector<4xi32>, vector<4xi32>, vector<16xi32>, i32, i32, i32) -> vector<16xi32>
amdgpu.mfma %arg4 * %arg4 + %arg5 { abid = 0 : i32, cbsz = 0 : i32, k = 32 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<16xi8>, vector<16xi8>, vector<16xi32>
amdgpu.mfma 32x32x32 %arg4 * %arg4 + %arg5 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<16xi8>, vector<16xi8>, vector<16xi32>
// CHECK: rocdl.mfma.i32.16x16x64.i8{{.*}}: (vector<4xi32>, vector<4xi32>, vector<4xi32>, i32, i32, i32) -> vector<4xi32>
amdgpu.mfma %arg4 * %arg4 + %arg6 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<16xi8>, vector<16xi8>, vector<4xi32>
amdgpu.mfma 16x16x64 %arg4 * %arg4 + %arg6 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<16xi8>, vector<16xi8>, vector<4xi32>
// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c0]], %[[c0]], %[[c0]], %[[c0]]{{.*}}: (vector<8xi32>, vector<8xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg7 * %arg7 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf8E4M3FN>, vector<32xf8E4M3FN>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg7 * %arg7 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf8E4M3FN>, vector<32xf8E4M3FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[c0]], %[[c0]], %[[c0]]{{.*}}: (vector<8xi32>, vector<8xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg7 * %arg7 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf8E4M3FN>, vector<32xf8E4M3FN>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg7 * %arg7 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf8E4M3FN>, vector<32xf8E4M3FN>, vector<4xf32>
// CHECK: %[[c1:.+]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c1]], %[[c1]], %[[c0]], %[[c0]]{{.*}}: (vector<8xi32>, vector<8xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg8 * %arg8 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf8E5M2>, vector<32xf8E5M2>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg8 * %arg8 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf8E5M2>, vector<32xf8E5M2>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c1]], %[[c1]], %[[c0]], %[[c0]]{{.*}}: (vector<8xi32>, vector<8xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg8 * %arg8 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf8E5M2>, vector<32xf8E5M2>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg8 * %arg8 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf8E5M2>, vector<32xf8E5M2>, vector<4xf32>
// CHECK: %[[c2:.+]] = llvm.mlir.constant(2 : i32) : i32
// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c2]], %[[c2]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<6xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg9 * %arg9 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf6E2M3FN>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg9 * %arg9 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf6E2M3FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c2]], %[[c2]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<6xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg9 * %arg9 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf6E2M3FN>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg9 * %arg9 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf6E2M3FN>, vector<4xf32>
// CHECK: %[[c3:.+]] = llvm.mlir.constant(3 : i32) : i32
// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c3]], %[[c3]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<6xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg10 * %arg10 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E3M2FN>, vector<32xf6E3M2FN>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg10 * %arg10 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E3M2FN>, vector<32xf6E3M2FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c3]], %[[c3]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<6xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg10 * %arg10 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E3M2FN>, vector<32xf6E3M2FN>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg10 * %arg10 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E3M2FN>, vector<32xf6E3M2FN>, vector<4xf32>
// CHECK-DAG: %[[c4:.+]] = llvm.mlir.constant(4 : i32) : i32
// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c4]], %[[c4]], %[[c0]], %[[c0]]{{.*}}: (vector<4xi32>, vector<4xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg11 * %arg11 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf4E2M1FN>, vector<32xf4E2M1FN>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg11 * %arg11 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf4E2M1FN>, vector<32xf4E2M1FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c4]], %[[c4]], %[[c0]], %[[c0]]{{.*}}: (vector<4xi32>, vector<4xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg11 * %arg11 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf4E2M1FN>, vector<32xf4E2M1FN>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg11 * %arg11 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf4E2M1FN>, vector<32xf4E2M1FN>, vector<4xf32>

// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c2]], %[[c4]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<4xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.mfma %arg9 * %arg11 + %arg1 { abid = 0 : i32, cbsz = 0 : i32, k = 64 : i32, m = 32 : i32, n = 32 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf4E2M1FN>, vector<16xf32>
amdgpu.mfma 32x32x64 %arg9 * %arg11 + %arg1 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf4E2M1FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c2]], %[[c4]], %[[c0]], %[[c0]]{{.*}}: (vector<6xi32>, vector<4xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.mfma %arg9 * %arg11 + %arg2 { abid = 0 : i32, cbsz = 0 : i32, k = 128 : i32, m = 16 : i32, n = 16 : i32, blocks = 1 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf4E2M1FN>, vector<4xf32>
amdgpu.mfma 16x16x128 %arg9 * %arg11 + %arg2 { abid = 0 : i32, cbsz = 0 : i32 } blgp = none : vector<32xf6E2M3FN>, vector<32xf4E2M1FN>, vector<4xf32>

func.return
}
Expand All @@ -57,9 +57,9 @@ func.func @mfma_to_rocdl(%arg0 : vector<8xf16>, %arg1 : vector<16xf32>,
func.func @scaled_mfma_to_rocdl(%arg0 : vector<16xf32>,
%arg1 : vector<4xf32>, %arg2 : vector<32xf8E4M3FN>,
%arg3 : vector<32xf8E5M2>, %arg4 : vector<32xf6E2M3FN>,
%arg5 : vector<32xf6E3M2FN>, %arg6 : vector<32xf4E2M1FN>,
%arg5 : vector<32xf6E3M2FN>, %arg6 : vector<32xf4E2M1FN>,
%arg7 : vector<4xf8E8M0FNU>, %arg8 : f8E8M0FNU) {

// CHECK: %[[c0:.+]] = llvm.mlir.constant(0 : i32) : i32
// CHECK: %[[c1:.+]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: %[[b0:.+]] = llvm.bitcast {{.*}} : vector<4xi8> to i32
Expand All @@ -69,32 +69,32 @@ func.func @scaled_mfma_to_rocdl(%arg0 : vector<16xf32>,
amdgpu.scaled_mfma(%arg7[0] * %arg2) * (%arg8[1] * %arg2) + %arg0 { k = 64 : i32, m = 32 : i32, n = 32 : i32 } : vector<4xf8E8M0FNU>, vector<32xf8E4M3FN>, f8E8M0FNU, vector<32xf8E4M3FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<8xi32>, vector<8xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg2) * (%arg8[1] * %arg2) + %arg1 { k = 128 : i32, m = 16 : i32, n = 16 : i32 } : vector<4xf8E8M0FNU>, vector<32xf8E4M3FN>, f8E8M0FNU, vector<32xf8E4M3FN>, vector<4xf32>

// CHECK: llvm.bitcast

// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<8xi32>, vector<8xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg3) * (%arg8[1] * %arg3) + %arg0 { k = 64 : i32, m = 32 : i32, n = 32 : i32 } : vector<4xf8E8M0FNU>, vector<32xf8E5M2>, f8E8M0FNU, vector<32xf8E5M2>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<8xi32>, vector<8xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg3) * (%arg8[1] * %arg3) + %arg1 { k = 128 : i32, m = 16 : i32, n = 16 : i32 } : vector<4xf8E8M0FNU>, vector<32xf8E5M2>, f8E8M0FNU, vector<32xf8E5M2>, vector<4xf32>

// CHECK: llvm.bitcast

// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<6xi32>, vector<6xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg4) * (%arg8[1] * %arg4) + %arg0 { k = 64 : i32, m = 32 : i32, n = 32 : i32 } : vector<4xf8E8M0FNU>, vector<32xf6E2M3FN>, f8E8M0FNU, vector<32xf6E2M3FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<6xi32>, vector<6xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg4) * (%arg8[1] * %arg4) + %arg1 { k = 128 : i32, m = 16 : i32, n = 16 : i32 } : vector<4xf8E8M0FNU>, vector<32xf6E2M3FN>, f8E8M0FNU, vector<32xf6E2M3FN>, vector<4xf32>

// CHECK: llvm.bitcast
// CHECK: llvm.mlir.constant(3 : i32) : i32

// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<6xi32>, vector<6xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg5) * (%arg8[1] * %arg5) + %arg0 { k = 64 : i32, m = 32 : i32, n = 32 : i32 } : vector<4xf8E8M0FNU>, vector<32xf6E3M2FN>, f8E8M0FNU, vector<32xf6E3M2FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<6xi32>, vector<6xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg5) * (%arg8[1] * %arg5) + %arg1 { k = 128 : i32, m = 16 : i32, n = 16 : i32 } : vector<4xf8E8M0FNU>, vector<32xf6E3M2FN>, f8E8M0FNU, vector<32xf6E3M2FN>, vector<4xf32>

// CHECK: llvm.bitcast
// CHECK: llvm.mlir.constant(4 : i32) : i32

// CHECK: rocdl.mfma.scale.f32.32x32x64.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<4xi32>, vector<4xi32>, vector<16xf32>, i32, i32, i32, i32, i32, i32) -> vector<16xf32>
amdgpu.scaled_mfma(%arg7[0] * %arg6) * (%arg8[1] * %arg6) + %arg0 { k = 64 : i32, m = 32 : i32, n = 32 : i32 } : vector<4xf8E8M0FNU>, vector<32xf4E2M1FN>, f8E8M0FNU, vector<32xf4E2M1FN>, vector<16xf32>
// CHECK: rocdl.mfma.scale.f32.16x16x128.f8f6f4{{.*}}, %[[c0]], %[[b0]], %[[c1]], %[[z0]] : (vector<4xi32>, vector<4xi32>, vector<4xf32>, i32, i32, i32, i32, i32, i32) -> vector<4xf32>
Expand Down
Loading