[AIROCMLIR-447] Lower `migraphx.conv` into Linalg by Mr-Anyone · Pull Request #2241 · ROCm/rocMLIR

Mr-Anyone · 2026-02-17T21:03:25Z

Motivation

Being able to lower migraphx.conv into linalg.

Technical Details

Forward convolution is lowered in three steps:

Apply padding to the input when the op has non-zero padding.
Expand the channel dimension into (group, channel_per_group), introducing a group dimension G. Input becomes NGC* (e.g. NGCL, NGCHW, NGCDHW) and filter becomes GFC* (e.g. GFCL, GFCHW, GFCDHW), matching the group attr.
Emit the grouped linalg convolution (1D/2D/3D), then collapse the result back to the original NFHW/NFDHW shape for the type converter.

The linalg.generic for Conv1D implements this python loop essentially

def my_conv_1d(x, filters, group, strides=1, dilation=1):
    # NGCHW 
    batch = x.shape[0]
    group = x.shape[1]
    channel = x.shape[2]
    input_height = x.shape[3]
    # GFCHW
    group = filters.shape[0]
    filter_count = filters.shape[1]
    channel = filters.shape[2]
    kernel_height = filters.shape[3]
    
    output_height = int((input_height - kernel_height -(kernel_height - 1)*(dilation - 1))/strides) + 1 
    result = np.zeros((batch, group, filter_count, output_height))
    for n in range(batch):    
        for g in range(group):
            for f in range(filter_count):
                for ho in range(output_height):
                    # starting reduction here 
                    # computing the convolution 
                    for c in range(channel):
                        for k in range(kernel_height):
                            input_width_access = ho*strides+ k*dilation
                            result[n, g, f, ho] += x[n, g,c, input_width_access] * filters[g, f, c, k]
    return result

The linalg.generic for Conv3D implements this python loop essentially

def my_conv_3d(x, filters, group, strides=(1, 1, 1), dilation=(1, 1, 1)):
    # NGCHW 
    batch = x.shape[0]
    group = x.shape[1]
    channel = x.shape[2]
    input_height = x.shape[3]
    input_width = x.shape[4]
    input_depth = x.shape[5]
    
    # GFCHW
    group = filters.shape[0]
    filter_count = filters.shape[1]
    channel = filters.shape[2]
    kernel_height = filters.shape[3]
    kernel_width = filters.shape[4]
    kernel_depth = filters.shape[5]
    
    output_height = int((input_height - kernel_height -(kernel_height - 1)*(dilation[0] - 1))/strides[0]) + 1 
    output_width = int((input_width - kernel_width - (kernel_width - 1)*(dilation[1] - 1) )/strides[1]) + 1 
    output_depth = int((input_depth - kernel_depth - (kernel_depth - 1)*(dilation[2] - 1) )/strides[2]) + 1 

    result = np.zeros((batch, group, filter_count, output_height, output_width, output_depth))
    for n in range(batch):
        for g in range(group): # paralle
            for f in range(filter_count):
                for ho in range(output_height):
                    for wo in range(output_width):
                        for do in range(output_depth):
                            # starting reduction here 
                            # computing the convolution 
                            for c in range(channel):
                                for k in range(kernel_height):
                                    for j in range(kernel_width):
                                        for z in range(kernel_depth):
                                            input_height_access = ho*strides[0] + k*dilation[0]
                                            input_width_access  = wo*strides[1] + j*dilation[1]
                                            input_depth_access  = do*strides[2] + z*dilation[2]
                                            result[n, g, f, ho, wo, do] += (
                                                x[n, g, c, input_height_access, input_width_access, input_depth_access]
                                                * filters[g, f, c, k, j, z]
                                            )
        
    return result

Test Plan

Added both e2e test and lit test. The e2e test value are verified by pytorch.

Test Result

Passed both e2e test and lit test.

Submission Checklist

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

justinrosner · 2026-02-18T22:19:59Z

mlir/test/Conversion/MIGraphXToLinalg/e2e/migraphx-to-linalg-conv-cpu.e2e.mlir

@@ -0,0 +1,35 @@
+// RUN: rocmlir-gen -fut conv -arch %arch --clone-harness %s | rocmlir-driver --host-pipeline=migraphx-linalg,highlevel | rocmlir-gen -ph -print-results -rand_type float -rand_min 0 -rand_max 0 -fut conv_wrapper --verifier clone -  | xmir-runner --shared-libs=%linalg_test_lib_dir/libmlir_rocm_runtime%shlibext,%conv_validation_wrapper_library_dir/libconv-validation-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext,%linalg_test_lib_dir/libmlir_float16_utils%shlibext,%linalg_test_lib_dir/libmlir_c_runner_utils%shlibext,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void | FileCheck %s


If you actually want to set random values use a meaningful range. E.g., something like -rand_min -1 -rand_max 1

pabloantoniom

IR looks pretty good!

pabloantoniom · 2026-02-20T13:54:37Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+    }
+  };
+
+  // Step 2: expand group dimension (NCHW -> NGCHW, FCHW -> GFCHW).


Why is this needed? The description is missing some motivation to explain why we need this.

pabloantoniom · 2026-02-20T13:57:24Z

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

@@ -0,0 +1,44 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg -verify-diagnostics %s | FileCheck %s


I think in this PR you support all the convolution flavors like dilation, padding, stride, and group. Is that correct? If so, I think we need to:

Make it clear in the description

Add proper testing for all that functionality

In particular, I think it would be beneficial to split this test into 4. 1D, 2D, 3D and unsupported cases (e.g., 4D, or other cases we don't support). And, in the 1,2,3D tests, we should add tests for dilation, padding (non-padding), stride and groups.

pabloantoniom · 2026-02-20T13:58:18Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

  patterns.add<AsUnderlyingShapeConverter, AsLogicalShapeOpConverter>(
      typeConverter, patterns.getContext());
+
+  // mhal.launch can be generated through rocmlir-gen, so we need a way to


I don't understand, can you give an example to show why we need this? Please add it to the description to motivate why we need it

During e2e testing, the rocmlir-gen seems to generate a wrapper that has mhal.launch.

Example (from test case below):

root@14298acabab2 ~/rocMLIR/build-release (pr-template-migraphx-to-linalg-conv)$ cat main.mlir func.func @conv_1d_group(%in: !migraphx.shaped<10x8x123xf32, 984x123x1>, %fil: !migraphx.shaped<12x2x7xf32, 14x7x1>) -> !migraphx.shaped<10x12x53xf32, 636x53x1> { %out = migraphx.convolution %in, %fil {dilation = [4], group = 4 : i64, padding = [3,3], padding_mode = 0 : i64, stride = [2]} : <10x8x123xf32, 984x123x1>, <12x2x7xf32, 14x7x1> -> <10x12x53xf32, 636x53x1> func.return %out : !migraphx.shaped<10x12x53xf32, 636x53x1> } root@14298acabab2 ~/rocMLIR/build-release (pr-template-migraphx-to-linalg-conv)$ ./bin/rocmlir-gen main.mlir --clone-harness --arch gfx950 -fut conv_1d_group module { func.func @conv_1d_group(%arg0: !migraphx.shaped<10x8x123xf32, 984x123x1> {mhal.read_access}, %arg1: !migraphx.shaped<12x2x7xf32, 14x7x1> {mhal.read_access}) -> (!migraphx.shaped<10x12x53xf32, 636x53x1> {mhal.write_access}) { %0 = migraphx.convolution %arg0, %arg1 {dilation = [4], group = 4 : i64, padding = [3, 3], padding_mode = 0 : i64, stride = [2]} : <10x8x123xf32, 984x123x1>, <12x2x7xf32, 14x7x1> -> <10x12x53xf32, 636x53x1> return %0 : !migraphx.shaped<10x12x53xf32, 636x53x1> } func.func @conv_1d_group_wrapper(%arg0: !migraphx.shaped<10x8x123xf32, 984x123x1>, %arg1: !migraphx.shaped<12x2x7xf32, 14x7x1>) -> !migraphx.shaped<10x12x53xf32, 636x53x1> { %token, %results = mhal.launch @conv_1d_group (%arg0, %arg1) : (!migraphx.shaped<10x8x123xf32, 984x123x1>, !migraphx.shaped<12x2x7xf32, 14x7x1>) -> !migraphx.shaped<10x12x53xf32, 636x53x1> mhal.await %token : !mhal.token return %results : !migraphx.shaped<10x12x53xf32, 636x53x1> } module @__xmodule_ attributes {mhal.arch = "gfx950", mhal.module} { func.func @conv_1d_group(%arg0: !migraphx.shaped<10x8x123xf32, 984x123x1> {mhal.read_access}, %arg1: !migraphx.shaped<12x2x7xf32, 14x7x1> {mhal.read_access}) -> (!migraphx.shaped<10x12x53xf32, 636x53x1> {mhal.write_access}) attributes {kernel, original_func = @conv_1d_group} { %0 = migraphx.convolution %arg0, %arg1 {dilation = [4], group = 4 : i64, padding = [3, 3], padding_mode = 0 : i64, stride = [2]} : <10x8x123xf32, 984x123x1>, <12x2x7xf32, 14x7x1> -> <10x12x53xf32, 636x53x1> return %0 : !migraphx.shaped<10x12x53xf32, 636x53x1> } } }

If we don't have this line above, passing through ./bin/rocmlir-driver seems to result in failure due to type not being legalized yet.

Can you post the rocmlir-driver command as well? I cannot reproduce your issue

Applying this diff:

diff --git a/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp b/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp index 742d2b29056d..61f746ae9414 100644 --- a/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp +++ b/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp @@ -799,7 +799,7 @@ void mlir::migraphx::populateMIGraphXFuncBoundaryToLinalgConversionPatterns( // mhal.launch can be generated through rocmlir-gen, so we need a way to // legalize it - populateMIGraphXToLinalgMHALLauncherConversion(patterns, typeConverter); + // populateMIGraphXToLinalgMHALLauncherConversion(patterns, typeConverter); populateAnyFunctionOpInterfaceTypeConversionPattern(patterns, typeConverter); populateReturnOpTypeConversionPattern(patterns, typeConverter); populateCallOpTypeConversionPattern(patterns, typeConverter); diff --git a/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp b/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp index 75f5588d2024..9fd00ff82ed6 100644 --- a/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp +++ b/mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp @@ -52,11 +52,11 @@ void mlir::migraphx::populateMIGraphXToLinalgBoundaryDialectConversion( target.addDynamicallyLegalOp<func::FuncOp>([&](func::FuncOp op) { return typeConverter.isSignatureLegal(op.getFunctionType()); }); - target.addDynamicallyLegalOp<mhal::LaunchOp>( - [=](mhal::LaunchOp op) -> std::optional<bool> { - return typeConverter.isLegal(op.getResultTypes()) && - typeConverter.isLegal(op.getOperandTypes()); - }); + // target.addDynamicallyLegalOp<mhal::LaunchOp>( + // [=](mhal::LaunchOp op) -> std::optional<bool> { + // return typeConverter.isLegal(op.getResultTypes()) && + // typeConverter.isLegal(op.getOperandTypes()); + // }); target.addDynamicallyLegalOp<func::ReturnOp>( [&](func::ReturnOp op) { return typeConverter.isLegal(op); }); target.addDynamicallyLegalOp<func::CallOp>(

Running:

./bin/rocmlir-gen -fut conv_3d -arch gfx950 --clone-harness <filename> | ./bin/rocmlir-driver --host-pipeline=migraphx-linalg,highlevel --kernel-pipeline=migraphx-linalg,highlevel

On:

func.func @conv_3d(%arg0: !migraphx.shaped<2x4x2x2x2xf32, 32x8x4x2x1>, %arg1: !migraphx.shaped<2x3x5x5x5xf32, 375x125x25x5x1>, %arg2: !migraphx.shaped<4x3x2x2x2xf32, 24x8x4x2x1>) -> !migraphx.shaped<2x4x2x2x2xf32, 32x8x4x2x1> { %0 = migraphx.convolution %arg1, %arg2 {dilation = [2, 2, 2], group = 1 : i64, padding = [0, 0, 0, 0, 0, 0], padding_mode = 0 : i64, stride = [2, 2, 2]} : <2x3x5x5x5xf32, 375x125x25x5x1>, <4x3x2x2x2xf32, 24x8x4x2x1> -> <2x4x2x2x2xf32, 32x8x4x2x1> %1 = migraphx.add %0, %arg0 : <2x4x2x2x2xf32, 32x8x4x2x1>, <2x4x2x2x2xf32, 32x8x4x2x1> -> <2x4x2x2x2xf32, 32x8x4x2x1> return %1 : !migraphx.shaped<2x4x2x2x2xf32, 32x8x4x2x1> }

I got this on my end?

loc("<stdin>":7:138): error: failed to legalize unresolved materialization from ('tensor<96xf32>') to ('!migraphx.shaped<4x3x2x2x2xf32, 24x8x4x2x1>') that remained live after conversion Lowering failed.

pabloantoniom · 2026-02-20T13:59:29Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  // converter.
+  Location loc = op.getLoc();
+  int64_t group = op.getGroupAttr().getInt();
+  int64_t dim = cast<RankedTensorType>(input.getType()).getRank() -


Don't we have a migraphx::ConvolutionOp method to get the rank?

pabloantoniom · 2026-02-20T14:01:05Z

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

@@ -0,0 +1,44 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg -verify-diagnostics %s | FileCheck %s


Also, does this PR support input and output with different types? For example f16 inputs with f32 outputs. If so, we need to test for that. If not, probably add it as a negative test (to the 4th test case if you follow my suggestion above)

pabloantoniom · 2026-02-20T14:02:05Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+  SmallVector<int64_t, 4> newShape;
+  int64_t n = resultType.getDimSize(0);
+  int64_t newF = resultType.getDimSize(1) / group;
+  assert(resultType.getDimSize(1) % group == 0 &&


Let's generate a proper failure error (LogicalResult) instead of using asserts

My original intention was to move this to the verifier, and make this an invariant of the operation.

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

Copilot

Pull request overview

This PR implements lowering of the migraphx.convolution operation to the Linalg dialect, enabling an alternative compilation path for MIGraphX convolutions. The implementation supports 1D, 2D, and 3D convolutions with configurable padding, stride, dilation, and grouping parameters.

Changes:

Added ConvConverter to lower migraphx.convolution to Linalg operations (generic ops for 1D/3D, named op for 2D)
Introduced new "migraphx-linalg" pipeline option in the rocmlir-driver
Added comprehensive lit tests and end-to-end tests with PyTorch validation

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
mlir/tools/rocmlir-driver/rocmlir-driver.cpp	Added "migraphx-linalg" pipeline option and updated help text
mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-not-implemented.mlir	Removed convolution test since it's now implemented
mlir/test/Conversion/MIGraphXToLinalg/migraphx-to-linalg-conv.mlir	Added lit tests for 1D, 2D, and 3D convolution lowering
mlir/test/Conversion/MIGraphXToLinalg/e2e/migraphx-to-linalg-conv-cpu.e2e.mlir	Added end-to-end test with PyTorch-validated golden values
mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp	Added `populateMIGraphXToLinalgMHALLauncherConversion` function
mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp	Added mhal.launch legalization support
mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp	Implemented ConvConverter with padding, group expansion, and linalg lowering
mlir/include/mlir/Conversion/MIGraphXToTosa/MIGraphXToTosa.h	Added function declaration for MHALLauncher conversion

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

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

Copilot · 2026-02-20T18:38:16Z

mlir/include/mlir/Conversion/MIGraphXToTosa/MIGraphXToTosa.h

+void populateMIGraphXToLinalgMHALLauncherConversion(
+    RewritePatternSet &target, TypeConverter &typeConverter);


The function populateMIGraphXToLinalgMHALLauncherConversion is declared in the wrong header file. It is declared in MIGraphXToTosa.h but should be declared in MIGraphXToLinalg.h since it's specifically used for the Linalg conversion. The function is implemented in MIGraphXToTosa.cpp and used by MIGraphXToLinalgPass.cpp, which includes both headers. While this works functionally, it creates confusing organization where Linalg-specific functionality is declared in a Tosa header file.

I have just moved it to MIGraphXToLinalg.h now.

That was somewhat intentional. MHALLauncher class is a dependency from libMLIRMIGraphXToTosa.a. I want all declaration in MIGraphXToTosa.cpp to be in MIGraphXToTosa.h originally.

mlir/test/Conversion/MIGraphXToLinalg/e2e/migraphx-to-linalg-conv.cpu.mlir

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

Copilot

Pull request overview

Copilot reviewed 11 out of 11 changed files in this pull request and generated 7 comments.

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

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

Mr-Anyone · 2026-02-23T20:31:40Z

Some changes has been made since the last revision

This PR now emits attributes for rock lowering (padding, striders, op name, etc)
conv2d is now converted into linalg.generic
1D/2D/3D convolution with non trivial parameters (!= 0 padding, >1 strides, >1group, >1 dilation) give same result as the tosa pipeline.

pabloantoniom · 2026-02-24T14:42:09Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

+
+/// Emit Conv1D expects input shape to be (batch, group, channel, height),
+/// filter to be (group, filter, channel, height)
+static Value emitGroupedConv1D(ConversionPatternRewriter &rewriter,


Looks like we now that we emit generic for all 3 cases can easily refactor emitGroupedConv1D/2D/3D into a common emitGroupedConv for 1/2/3D right?

It turns out that extending to ND is that not hard.

pabloantoniom · 2026-02-24T14:54:31Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

  patterns.add<AsUnderlyingShapeConverter, AsLogicalShapeOpConverter>(
      typeConverter, patterns.getContext());
+
+  // mhal.launch can be generated through rocmlir-gen, so we need a way to


Can you post the rocmlir-driver command as well? I cannot reproduce your issue

pabloantoniom · 2026-02-24T14:58:34Z

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

  // Convert optional attributes
  if (auto attr = (*op).template getAttrOfType<StringAttr>("perf_config"))
    newOp->setAttr("perf_config", attr);
+  newOp->setAttr("conv_op", convOpName);


I'd rather have this attribute defined in RockAttrDefs and referenced here with its name instead of hardcoding a string like this. Specially because when you work on the linalg-to-rock side of things you will need to reference the attribute again, and hardcoding strings all over the place should be avoided

Copilot

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

Mr-Anyone requested a review from causten as a code owner February 17, 2026 21:03

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch from 9704ab3 to 14d20ad Compare February 17, 2026 21:13

Mr-Anyone marked this pull request as draft February 18, 2026 02:24

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-6 branch from 3d4478e to 3d9c23f Compare February 18, 2026 15:07

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch 6 times, most recently from fe2bbc8 to 0f84c0a Compare February 18, 2026 22:07

Mr-Anyone marked this pull request as ready for review February 18, 2026 22:08

Mr-Anyone requested review from dhernandez0, justinrosner, pabloantoniom and umangyadav February 18, 2026 22:08

Mr-Anyone changed the title ~~[AIROCMLIR-447] WIP (needs testcase) Lower migraphx.conv into Linalg~~ [AIROCMLIR-447] Lower migraphx.conv into Linalg Feb 18, 2026

justinrosner reviewed Feb 18, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch 2 times, most recently from db757be to 18b0a33 Compare February 19, 2026 16:19

pabloantoniom reviewed Feb 20, 2026

View reviewed changes

umangyadav requested a review from Copilot February 20, 2026 18:29

Copilot started reviewing on behalf of umangyadav February 20, 2026 18:32 View session

Copilot AI reviewed Feb 20, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-6 branch from 3d9c23f to 2246225 Compare February 20, 2026 18:44

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch from 18b0a33 to cf79235 Compare February 20, 2026 18:45

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-6 branch from 2246225 to 56be60a Compare February 20, 2026 19:01

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch from cf79235 to 648cc4f Compare February 20, 2026 19:01

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-6 branch 2 times, most recently from 3de417c to 7ef3371 Compare February 20, 2026 19:59

umangyadav requested a review from Copilot February 23, 2026 13:21

Copilot started reviewing on behalf of umangyadav February 23, 2026 13:23 View session

Copilot AI reviewed Feb 23, 2026

View reviewed changes

umangyadav reviewed Feb 23, 2026

View reviewed changes

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

Mr-Anyone added 8 commits February 23, 2026 14:56

Lower migraphx.conv into linalg.generic and other forms

7ea87ed

smave work

6df9db2

Added test case for conv2d group

c0709d3

Added more testcase

85179af

Rename testfile to have .cpu extension

1256298

clang-format

3bd564d

Added attribute testcase

c3b6277

Emit conv op name attribute

6f67460

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-conv branch from f1a3c51 to 6f67460 Compare February 23, 2026 14:56

Mr-Anyone added 4 commits February 23, 2026 15:40

Address copilot comments

4c1a282

Fix compilation

b058424

Added type casting test

06aa357

Emits linalg.generic instead of conv2d

cbdd142

Rename op_name to have convnd prefix

17d427d

Mr-Anyone requested review from justinrosner, pabloantoniom and umangyadav February 23, 2026 20:40

pabloantoniom reviewed Feb 24, 2026

View reviewed changes

Mr-Anyone added 2 commits February 24, 2026 20:15

Code cleanup

d64aeb3

Fixed testcase after changing to LinalgConv

5ea2910

umangyadav requested a review from Copilot February 24, 2026 20:42

Copilot started reviewing on behalf of umangyadav February 24, 2026 20:45 View session

Copilot AI reviewed Feb 24, 2026

View reviewed changes

Emits rock attribute instead of string now

319e404

		@@ -0,0 +1,35 @@
		// RUN: rocmlir-gen -fut conv -arch %arch --clone-harness %s \| rocmlir-driver --host-pipeline=migraphx-linalg,highlevel \| rocmlir-gen -ph -print-results -rand_type float -rand_min 0 -rand_max 0 -fut conv_wrapper --verifier clone - \| xmir-runner --shared-libs=%linalg_test_lib_dir/libmlir_rocm_runtime%shlibext,%conv_validation_wrapper_library_dir/libconv-validation-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext,%linalg_test_lib_dir/libmlir_float16_utils%shlibext,%linalg_test_lib_dir/libmlir_c_runner_utils%shlibext,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void \| FileCheck %s

		@@ -0,0 +1,44 @@
		// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg -verify-diagnostics %s \| FileCheck %s

		void populateMIGraphXToLinalgMHALLauncherConversion(
		RewritePatternSet &target, TypeConverter &typeConverter);

Comments

Conversation

Mr-Anyone commented Feb 17, 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!

pabloantoniom 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!

Mr-Anyone Feb 23, 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!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Copilot AI Feb 20, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Mr-Anyone Feb 23, 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!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Mr-Anyone commented Feb 23, 2026

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

Mr-Anyone commented Feb 17, 2026 •

edited

Loading

Mr-Anyone Feb 23, 2026 •

edited

Loading

Mr-Anyone Feb 23, 2026 •

edited

Loading