[AIROCMLIR-446] Lower `migraphx.dot` into `linalg.matmul/batch_matmul` by Mr-Anyone · Pull Request #2223 · ROCm/rocMLIR

Mr-Anyone · 2026-01-30T18:13:27Z

Motivation

Lower MIGraphX GEMM into Linalg Dialect.

Technical Details

Lower from MIGraphX to Linalg initial changes.

MIGraphX to Linalg conversion is performed in two passes:

Pass 1: Convert MIGraphX operations to their Linalg equivalents. The !migraphx.shaped type contains both shape and stride (memory layout) information. During this pass, sourceMaterialization and targetMaterialization insert temporary ops (migraphx.mlir.as_logical_shape and migraphx.mlir.as_underlying_shape) to handle the type conversions.

Pass 2: Convert the boundary/materialization operations to proper memory layout representations using tensor operations, completing the conversion from MIGraphX's shaped types to standard tensor types.

Test Plan

Lit test for diagnostics and IR output.

Test Result

One lit test for now, and passed locally.

Submission Checklist

Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

pabloantoniom · 2026-02-02T11:24:11Z

mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-dot.mlir

+// CHECK-SAME:                tensor<1x3x2xf32>, tensor<1x2x3xf32>) outs(%[[cst]] : tensor<1x3x3xf32>) -> tensor<1x3x3xf32>
+// CHECK-NEXT:      %[[collapsed:.*]] = tensor.collapse_shape %[[zero]] 
+// CHECK-NEXT:      return %[[collapsed]] : tensor<9xf32>
+func.func @dot_one(%arg0 : !migraphx.shaped<1x3x2xf32, 6x2x1>, %arg1: !migraphx.shaped<1x2x3xf32, 6x3x1>) 


nit: These tests are not formatted properly. For instance, this first line is broken down into 2 lines. Can you run rocmlir-opt with this file, it will give you the tests with proper format.

pabloantoniom · 2026-02-02T11:25:55Z

mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-dot.mlir

+// CHECK-SAME:                tensor<1x3x2xf32>, tensor<1x2x3xf32>) outs(%[[cst]] : tensor<1x3x3xf32>) -> tensor<1x3x3xf32>
+// CHECK-NEXT:      %[[collapsed:.*]] = tensor.collapse_shape %[[zero]] 
+// CHECK-NEXT:      return %[[collapsed]] : tensor<9xf32>
+func.func @dot_one(%arg0 : !migraphx.shaped<1x3x2xf32, 6x2x1>, %arg1: !migraphx.shaped<1x2x3xf32, 6x3x1>) 


I'm not sure what you mean with dot_one, dot_two, dot_three? It would make sense if it was 1D, 2D and 3D, but it's not the case. Can you rename them to dot_3D, dot_2D, and dot_4D (the one that fails)?

I have renamed them to dot_3D, dot_2D, and dot_4D.

The dot_one, and dot_two, and dot_three originally means the first dot, the second dot, and the third dot test.

pabloantoniom · 2026-02-02T11:29:52Z

mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-dot.mlir

@@ -0,0 +1,49 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg -verify-diagnostics %s | FileCheck %s


I had a look at the IR that this generates and it looks great, good job!

pabloantoniom · 2026-02-02T11:30:49Z

mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-dot.mlir

+
+func.func @dot_three(%arg0 : !migraphx.shaped<1x1x3x2xf32, 6x6x2x1>, %arg1: !migraphx.shaped<1x1x2x3xf32, 6x6x3x1>) 
+  -> !migraphx.shaped<1x1x3x3xf32, 9x9x3x1>{
+  // expected-error @+2 {{only support 2D/3D for now}}


This is fine for now, in the future we can give support to migraphx.dot in any dimension with linalg.generic

pabloantoniom · 2026-02-02T11:33:38Z

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp

  populateCallOpTypeConversionPattern(patterns, typeConverter);
 }
+
+void mlir::migraphx::populateMIGrpahXToLinalgTrivialConverter(


nit: populateMIGraphXToLinalgTrivialConverter

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp

pabloantoniom · 2026-02-02T11:35:42Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  Location loc = op.getLoc();
+  Value aIn = adaptor.getInA();
+  Value bIn = adaptor.getInB();
+  RankedTensorType aType = cast<TypedValue<RankedTensorType>>(aIn).getType();


Does migraphx::DotOp accept unranked tensors as input? I think it does. If this is the case, this code will crash.

My understanding is that the migraphx-to-tosa lowering doesn't support unranked tensor as input to my knowledge.

root@f83ce68a1182$ cat test.mlir func.func @dot_one(%arg0 : !migraphx.shaped<?x?x?xf32, ?x?x?>, %arg1: !migraphx.shaped<?x?x?xf32, ?x?x?>) -> !migraphx.shaped<?x?x?xf32, ?x?x?> { %0 = migraphx.dot %arg0, %arg1 : <?x?x?xf32, ?x?x?>, <?x?x?xf32, ?x?x?> -> <?x?x?xf32, ?x?x?> func.return %0 : !migraphx.shaped<?x?x?xf32, ?x?x?> } root@f83ce68a1182$ ./bin/rocmlir-opt test.mlir --migraphx-to-tosa <unknown>:0: error: !migraphx.shaped type with smallest stride -9223372036854775808 has no supported in-memory layout test.mlir:1:1: error: failed to legalize operation 'func.func' that was explicitly marked illegal func.func @dot_one(%arg0 : !migraphx.shaped<?x?x?xf32, ?x?x?>, %arg1: !migraphx.shaped<?x?x?xf32, ?x?x?>) -> !migraphx.shaped<?x?x?xf32, ?x?x?> { ^ test.mlir:1:1: note: see current operation: "func.func"() <{function_type = (!migraphx.shaped<?x?x?xf32, ?x?x?>, !migraphx.shaped<?x?x?xf32, ?x?x?>) -> !migraphx.shaped<?x?x?xf32, ?x?x?>, sym_name = "dot_one"}> ({ ^bb0(%arg0: !migraphx.shaped<?x?x?xf32, ?x?x?>, %arg1: !migraphx.shaped<?x?x?xf32, ?x?x?>): %0 = "migraphx.mlir.as.logical.shape"(%arg1) : (!migraphx.shaped<?x?x?xf32, ?x?x?>) -> tensor<?x?x?xf32> %1 = "migraphx.mlir.as.logical.shape"(%arg0) : (!migraphx.shaped<?x?x?xf32, ?x?x?>) -> tensor<?x?x?xf32> %2 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32> %3 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32> %4 = "tosa.matmul"(%1, %0, %2, %3) {acc_type = f32} : (tensor<?x?x?xf32>, tensor<?x?x?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?xf32> %5 = "migraphx.mlir.as.underlying.shape"(%4) : (tensor<?x?x?xf32>) -> !migraphx.shaped<?x?x?xf32, ?x?x?> "func.return"(%5) : (!migraphx.shaped<?x?x?xf32, ?x?x?>) -> () }) : () -> ()

Is this a bug?

I have also added a test case if unranked tensor type is not supported.

I can't seem to get unranked tensor type to work as a input for the migrpahx.dot, is there a different syntax?

func.func @dot_unranked(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>, %arg2: tensor<*xf32>) attributes {kernel, arch="gfx950"}{ %test = migraphx.dot %arg0, %arg1: <*xf32>, <*xf32> -> <*xf32> func.return }

Ok, it actually do not support unranked tensor, all must be static it seems. Look at this test:

func.func @mlir_reshape_dynamic_shape(%arg0: !migraphx.shaped<4096x?xf16, 0x1>) { // expected-error@+1 {{'migraphx.reshape' op Dynamic shapes are not supported}} %0 = migraphx.reshape %arg0 {dims = [4096, 4096]} : <4096x?xf16, 0x1> -> <4096x?xf16, 16536x2> return }

so we are good.

However, defensive programming does no harm. Can you add a check that the type is RankedTensorType and, if its not, throw an error? To avoid crashes. Like the "only static shape is supported for now" error below.

pabloantoniom · 2026-02-02T11:36:48Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  RankedTensorType bType = cast<TypedValue<RankedTensorType>>(bIn).getType();
+  ArrayRef<int64_t> aShape = aType.getShape();
+  ArrayRef<int64_t> bShape = bType.getShape();
+  int64_t dimension = aShape.size();


I prefer rank for this variable

Let's add more defensive programming here - add a condition that ensures that the shape of A and B has the same rank

pabloantoniom · 2026-02-02T11:40:08Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

+}
+
+void MIGraphXToLinalgPass::runOnOperation() {
+  // MIGraphX to Linalg conversion is performed in two passes:


Adding this to the PR description would be useful

I would also add this to the description of what this pass does at the top of the file.

pabloantoniom · 2026-02-02T11:42:31Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

+
+void mlir::linalg::populateMIGraphXToLinalgDialectConversion(
+    ConversionTarget &target) {
+  target.addLegalDialect<linalg::LinalgDialect, arith::ArithDialect,


Here we will be converting migraphx.dot to linalg equivalent, so I think we should add migraphx::DotOp to as IllegalOp

I have added migraphx::MIGraphXDialect as a illegal dialect except migraphx::AsLogicalShapeOp and migraphx::AsUnderlyingShapeOp

dhernandez0

Can you change the title? we only convert migraphx.dot, not all migraphx ops, right?

dhernandez0 · 2026-02-03T08:35:26Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+}
+
+//===----------------------------------------------------------------------===//
+// Base kernels (convolution, gemm)


this only converts gemms, not convolutions?

Yes, that is correct for now.

Copilot

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

dhernandez0 · 2026-02-06T11:37:38Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  int64_t rank = aShape.size();
+  if (aShape.size() != bShape.size()) {
+    return op.emitError("input a and b must have the same rank");
+  }


I think some of these checks should be in the migraphx.dot verifier if possible (if they aren't already there).

It doesn't seems like there is a migrpahx.dot verifier? I can't seem to find DotOp::verify in the codebase?

dhernandez0 · 2026-02-06T11:39:11Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+                                  rewriter.getZeroAttr(RankedTensorType::get(
+                                      outputShape, aType.getElementType())));
+    auto matMulOp =
+        linalg::MatmulOp::create(rewriter, loc, {aIn, bIn}, zero, {});


are the layouts of a and b exactly the same?

I mean migraphx and linalg expeted layouts (MxK or KxM etc)

also what about "acc_type", is there something like that in linalg.matmul?

are the layouts of a and b exactly the same?

The layouts of a and b is not the same. Usually, it is [batch, M, N] and [batch, N, M].

I don't think there is a acc_type like equivalent in linalg.batch_matmul?

dhernandez0 · 2026-02-06T11:41:13Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  }
+
+  // don't emit linalg.generic for 2D and 3D case to preserver type sugar
+  if (rank == 2) {


see MigraphxToTosa for handling of special cases such as broadcast etc

dhernandez0 · 2026-02-06T11:41:42Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+                                      outputShape, aType.getElementType())));
+    auto matMulOp =
+        linalg::MatmulOp::create(rewriter, loc, {aIn, bIn}, zero, {});
+    rewriter.replaceOp(op, matMulOp);


we should make sure to copy the perf_config attribute. See MigraphxToTosa

dhernandez0 · 2026-02-06T11:42:55Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

let's add a TODO somewhere for scaled gemms.

migraphx::DeQuantizeLinearConverter

I put it in line 107.

Copilot

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 17 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

mlir/include/mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h

mlir/lib/Conversion/MIGraphXToLinalg/CMakeLists.txt

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

mlir/include/mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h

mlir/lib/Conversion/MIGraphXToLinalg/CMakeLists.txt

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

umangyadav · 2026-02-09T14:11:15Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+    Value zero =
+        arith::ConstantOp::create(rewriter, loc,
+                                  rewriter.getZeroAttr(RankedTensorType::get(
+                                      outputShape, aType.getElementType())));


zero's type shouldn't be set as aType. it should take whatever ist the type of the output in original migrpahx.dot

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/include/mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h

Copilot

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 6 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-dot.mlir

Mr-Anyone · 2026-02-09T22:42:47Z

The logic for DotConverter now look similar to the one in MIGraphXTosa.

The important thing to note is that this PR now somewhat supports "broadcast":

func.func @dot_broadcast(%arg0: !migraphx.shaped<3x2x2x2xf32, 8x4x2x1>, %arg1: !migraphx.shaped<2x3x2x2xf32, 12x4x2x1>)
    -> !migraphx.shaped<3x2x2x2xf32, 8x4x2x1> attributes {kernel, arch="gfx950"} {
  %0 = migraphx.dot %arg0, %arg1 : <3x2x2x2xf32, 8x4x2x1>, <2x3x2x2xf32, 12x4x2x1> -> <3x2x2x2xf32, 8x4x2x1>
  func.return %0 : !migraphx.shaped<3x2x2x2xf32, 8x4x2x1>
}

now compile on both the tosa path and the linalg path. This seems to be intentional as well.

migraphx.dot is basically the following:

Input rank 2
Treat as single batch: reshape to [1, M, N] for linalg.batch_matmul.
Input rank 3
No reshape; use shapes as-is for linalg.batch_matmul.
Output rank > 3
Flatten batch dimensions: collapse leading dims to one, keep last two as matrix dims, i.e. [D0, D1, …, D_{n-2}, D_{n-1}, D_n] to [D0×D1×…×D_{n-2}, D_{n-1}, D_n], then call linalg.batch_matmul.

Mr-Anyone · 2026-02-09T23:36:13Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+    RankedTensorType newAType = RankedTensorType::get(newDimsA, elementTy);
+    RankedTensorType newBType = RankedTensorType::get(newDimsB, elementTy);
+    newOutType = RankedTensorType::get(newDimsOut, newOutElementTy);
+    inA = (rankA == 2)


clang-format seems to be doing something weird here.

umangyadav · 2026-02-10T13:19:20Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+                     .getResult(0);
+
+  // Convert optional attributes
+  if (auto attr = (*op).template getAttrOfType<StringAttr>("perf_config"))


Add a test with perf_config and make sure it is passed on when convert to linalg.

umangyadav · 2026-02-10T13:21:46Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+      !isa<RankedTensorType>(inB.getType())) {
+    return op.emitError("expected both operands to be RankedTensorType");
+  }
+  Type elementTy = cast<RankedTensorType>(inA.getType()).getElementType();


nit:
Create variables for RankedTensorType aRankedType = cast<RanekdTensorType>(inA.getType()) and use it everywhere.

umangyadav · 2026-02-10T13:22:22Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  Type outElementTy = origOutputTy.getElementType();
+  Type newOutElementTy = getTypeConverter()->convertType(outElementTy);
+
+  // check batch dimension. Tosa matmul only allow a single dimension for it,


nit:
We shouldn't mention tosa here. I get that it is to keep it aligned with tosa implementation but better not to mention it here.

umangyadav · 2026-02-10T13:26:07Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+    if (batchSizeA != batchSizeB || batchSizeC != batchSizeB) {
+      return op.emitError("cannot handle this broadcast for now");


Just looking for batchSizeA == batchSizeB could lead to mismatch.

e.g.
I see you have a test case with {3, 2, 2, 2} and {2, 3, 2, 2} shapes. where batchSize = 6 but the dimensions do not really match.

I think it wouldn't generate correct results. Can you verify ?

I think it gives correct results?

func.func @dot_broadcast(%arg0: !migraphx.shaped<3x2x2x2xf32, 8x4x2x1>, %arg1: !migraphx.shaped<6x2x2xf32, 4x2x1>) -> !migraphx.shaped<3x2x2x2xf32, 8x4x2x1> attributes {kernel, arch="gfx950"} { %0 = migraphx.dot %arg0, %arg1 : <3x2x2x2xf32, 8x4x2x1>, <6x2x2xf32, 4x2x1> -> <3x2x2x2xf32, 8x4x2x1> func.return %0 : !migraphx.shaped<3x2x2x2xf32, 8x4x2x1> }

The tosa and linalg lowering path seems to give similar results:

// ./bin/rocmlir-opt main.mlir --migraphx-to-linalg --canonicalize module { func.func @dot_broadcast(%arg0: tensor<24xf32>, %arg1: tensor<24xf32>) -> tensor<24xf32> attributes {arch = "gfx950", kernel} { %cst = arith.constant dense<0.000000e+00> : tensor<6x2x2xf32> %expanded = tensor.expand_shape %arg1 [[0, 1, 2]] output_shape [6, 2, 2] : tensor<24xf32> into tensor<6x2x2xf32> %expanded_0 = tensor.expand_shape %arg0 [[0, 1, 2]] output_shape [6, 2, 2] : tensor<24xf32> into tensor<6x2x2xf32> %0 = linalg.batch_matmul ins(%expanded_0, %expanded : tensor<6x2x2xf32>, tensor<6x2x2xf32>) outs(%cst : tensor<6x2x2xf32>) -> tensor<6x2x2xf32> %collapsed = tensor.collapse_shape %0 [[0, 1, 2]] : tensor<6x2x2xf32> into tensor<24xf32> return %collapsed : tensor<24xf32> } } // ./bin/rocmlir-opt main.mlir --migraphx-to-tosa --canonicalize module { func.func @dot_broadcast(%arg0: tensor<24xf32>, %arg1: tensor<24xf32>) -> tensor<24xf32> attributes {arch = "gfx950", kernel} { %0 = tosa.const_shape {values = dense<24> : tensor<1xindex>} : () -> !tosa.shape<1> %1 = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32> %2 = tosa.const_shape {values = dense<[6, 2, 2]> : tensor<3xindex>} : () -> !tosa.shape<3> %3 = tosa.reshape %arg1, %2 : (tensor<24xf32>, !tosa.shape<3>) -> tensor<6x2x2xf32> %4 = tosa.reshape %arg0, %2 : (tensor<24xf32>, !tosa.shape<3>) -> tensor<6x2x2xf32> %5 = tosa.matmul %4, %3, %1, %1 {acc_type = f32} : (tensor<6x2x2xf32>, tensor<6x2x2xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<6x2x2xf32> %6 = tosa.reshape %5, %0 : (tensor<6x2x2xf32>, !tosa.shape<1>) -> tensor<24xf32> return %6 : tensor<24xf32> } }

umangyadav · 2026-02-11T20:20:12Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+
+  if (rankA != rankB || rankB != outRank) {


We can support rankA != rankB in some cases.

e.g. A = {batch, m, k} and B = {k, n}. A can be folded {batch * m, k}`

umangyadav

Can you create ticket to handle cases for rankA != rankB and work on it later ?

Mr-Anyone requested a review from causten as a code owner January 30, 2026 18:13

Mr-Anyone requested review from dhernandez0, justinrosner and pabloantoniom January 30, 2026 18:13

pabloantoniom reviewed Feb 2, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch 2 times, most recently from 8759718 to 6d3702b Compare February 2, 2026 16:42

Mr-Anyone requested a review from pabloantoniom February 2, 2026 19:06

dhernandez0 reviewed Feb 3, 2026

View reviewed changes

justinrosner requested a review from Copilot February 3, 2026 15:44

Copilot AI reviewed Feb 3, 2026

View reviewed changes

Mr-Anyone changed the title ~~[AIROCMLIR-446] MIGraphX to Linalg Dialect Conversion~~ [AIROCMLIR-446] Lower migraphx.dot into linalg.matmul/batch_matmul Feb 3, 2026

pabloantoniom approved these changes Feb 4, 2026

View reviewed changes

Mr-Anyone requested a review from dhernandez0 February 4, 2026 21:36

dhernandez0 requested a review from Copilot February 6, 2026 11:36

dhernandez0 reviewed Feb 6, 2026

View reviewed changes

Copilot started reviewing on behalf of dhernandez0 February 6, 2026 11:38 View session

dhernandez0 reviewed Feb 6, 2026

View reviewed changes

Copilot AI reviewed Feb 6, 2026

View reviewed changes

umangyadav reviewed Feb 9, 2026

View reviewed changes

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp Outdated Show resolved Hide resolved

umangyadav reviewed Feb 9, 2026

View reviewed changes

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp Show resolved Hide resolved

umangyadav requested a review from Copilot February 9, 2026 14:34

umangyadav reviewed Feb 9, 2026

View reviewed changes

mlir/include/mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h Outdated Show resolved Hide resolved

Copilot AI reviewed Feb 9, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch from ccb06f3 to 98fcb90 Compare February 9, 2026 15:14

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch 2 times, most recently from 0850d14 to f12a85c Compare February 9, 2026 22:42

Mr-Anyone requested review from dhernandez0, pabloantoniom and umangyadav February 9, 2026 22:44

Mr-Anyone commented Feb 9, 2026

View reviewed changes

umangyadav reviewed Feb 10, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch 2 times, most recently from ad1c291 to ef5d903 Compare February 10, 2026 16:09

Mr-Anyone requested a review from umangyadav February 10, 2026 16:19

Mr-Anyone added 8 commits February 10, 2026 21:15

MIGraphX to Linalg Dialect Conversion

2ef2f8b

Address some comments

882ca01

Update comments

3f69da6

Added checks for RankedTensorTypes

80ac27c

Address some comments

36f167a

Updated test case

b8d381d

Added a perf_config testcase and addressed nits

428cd88

Pass CI?

405b3e8

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch from 97d65f9 to 405b3e8 Compare February 10, 2026 21:16

Mr-Anyone added 2 commits February 11, 2026 15:53

Refactor some logic to make code cleaner

32aa0d0

Use resize instead of insert

8e1ba37

umangyadav reviewed Feb 11, 2026

View reviewed changes

umangyadav approved these changes Feb 11, 2026

View reviewed changes

clang-format

b437f6a

Mr-Anyone merged commit 1aaf4c9 into develop Feb 11, 2026
7 of 14 checks passed

Mr-Anyone deleted the pr-template-migraphx-to-linalg branch February 11, 2026 22:46

		@@ -0,0 +1,49 @@
		// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg -verify-diagnostics %s \| FileCheck %s

		if (batchSizeA != batchSizeB \|\| batchSizeC != batchSizeB) {
		return op.emitError("cannot handle this broadcast for now");

Conversation

Mr-Anyone commented Jan 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pabloantoniom Feb 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dhernandez0 left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Mr-Anyone commented Jan 30, 2026 •

edited

Loading

pabloantoniom Feb 4, 2026 •

edited

Loading