[AIROCMLIR-446] Lower `linalg.matmul/batch_matmul` into `rock.gemm` by Mr-Anyone · Pull Request #2224 · ROCm/rocMLIR

Mr-Anyone · 2026-01-30T18:59:52Z

Motivation

Lower form Linalg into Rock.

Technical Details

This PR introduces a new LinalgToRock conversion pass that provides an alternative lowering path from MIGraphX through Linalg dialect to Rock.

The conversion handles linalg.matmul and linalg.batch_matmul operations, transforming them into rock.gemm operations.

Driver changes are added so that an E2E test can ran starting from migraphx -> linalg -> rock -> ... depending on the target platform.

Test Plan

Added e2e and a few lit test.

Test Result

Passed locally.

Submission Checklist

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

pabloantoniom · 2026-02-02T11:51:13Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-dot.mlir

@@ -0,0 +1,39 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg --linalg-to-rock -verify-diagnostics %s | FileCheck %s


When writing conversion tests, the recommended practice is to test only one conversion in isolation. In this case, we want to test --linalg-to-rock. As it is right now, we are testing --migraphx-to-linalg as well, which we don't want to. This is what we do in tosa-to-rock.mlir for example.

So you should take the IR after --migraphx-to-linalg and use that as input for this test, thus removing the --migraphx-to-linalg execution.

pabloantoniom · 2026-02-02T11:52:02Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-dot.mlir

@@ -0,0 +1,39 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg --linalg-to-rock -verify-diagnostics %s | FileCheck %s


Like the other PR, please format the test using the output from rocmlir-opt

pabloantoniom · 2026-02-02T11:53:27Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-dot.mlir

@@ -0,0 +1,39 @@
+// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg --linalg-to-rock -verify-diagnostics %s | FileCheck %s


For consistency with tosa-to-rock tests, please call this file "linalg-to-rock-matmul"

pabloantoniom · 2026-02-02T11:57:08Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-dot.mlir

+// CHECK-LABEL: func.func @dot_one(
+// CHECK-SAME: %[[arg0:.*]]: tensor<6xf32>, 
+// CHECK-SAME: %[[arg1:.*]]: tensor<6xf32>)
+// CHECK-NEXT:      %[[expanded:.*]] = tensor.expand_shape %[[arg1]]


The output IR is not yet rock compatible because we dont have rock.transforms.

Looking at the tosa-to-rock tests, we have tosa-to-rock-matmul which output also does not contain rock.transforms. But in the case of tosa-to-rock we have a tosa-to-rock.mlir test, which is actually generating rock.transforms with the --rock-view-to-transform pass.

Can you add a separate test called linalg-to-rock.mlir following our tosa-to-rock approach?

pabloantoniom · 2026-02-02T11:58:41Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-dot.mlir

+// CHECK-NEXT:      %[[collapsed:.*]] = tensor.collapse_shape %[[one]]
+// CHECK-NEXT:      return %[[collapsed]]
+func.func @dot_one(%arg0 : !migraphx.shaped<1x3x2xf32, 6x2x1>, %arg1: !migraphx.shaped<1x2x3xf32, 6x3x1>)
+  -> !migraphx.shaped<1x3x3xf32, 9x3x1> attributes {kernel, arch="gfx950"}{


We don't want tests to be arch-dependent. In other words, replace arch="gfx950" with arch-independent testing. Look at how we do it in https://github.com/ROCm/rocMLIR/blob/develop/mlir/test/Conversion/TosaToRock/tosa-to-rock.mlir for an example.

pabloantoniom · 2026-02-02T12:02:48Z

mlir/lib/Dialect/MIGraphX/Pipeline/Pipeline.cpp

  funcPm.addPass(migraphx::createMIGraphXTosaSimplifyPass());
 }
+
+void migraphx::addHighLevelMIGraphXToLinalg(PassManager &pm) {


Rather than having 2 separate functions (addHighLevelPipeline / addHighLevelMIGraphXToLinalg) we could have one that takes an extra argument to indicate if IR should be lowered to linalg or TOSA.

I think you should be able to make use of the lowerFromLinalg flag/option that you defined earlier.

Do you think a passing in a boolean value is fine? It seems a bit awkward to pass rock::BufferizeOptions because the disableRock member is never used in that function? Also, addHighLevelPipeline doesn't seem to be registered?

pabloantoniom · 2026-02-02T12:11:01Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+  Value a = op.getOperand(0);
+  Value b = op.getOperand(1);
+  Value cOriginal = op.getOperand(2);
+  RankedTensorType outputType = cast<RankedTensorType>(cOriginal.getType());


This would break if tensors are unranked which I think can happen? Please check it

It seems that linalg.matmul doesn't work with unranked tensor?

main.mlir:2:25: error: custom op 'linalg.matmul' Cannot build binary Linalg operation: expects allComplex, allFloatingPoint, or allInteger, got 'tensor<*xf32>' and 'tensor<*xf32>' %arg2 = linalg.matmul ins(%arg0, %arg1: tensor<*xf32>, tensor<*xf32>) outs(%arg2: tensor<*xf32>) -> tensor<*xf32>

I am not sure how to go about testing this.

Fine. I would check here if we are getting static shapes at least. If the IR comes from migraphx-to-linalg they will be, but it's better if we make this pass as solid as possible, just in case we ingest IR from another source, which might contain RankedTensor with non-static shapes.

mlir/lib/Conversion/LinalgToRock/LinalgToRockPass.cpp

justinrosner · 2026-02-02T14:52:17Z

Could you update the PR description so that it more accurately describes what the pass is doing?

justinrosner · 2026-02-02T14:54:18Z

mlir/include/mlir/Dialect/Rock/Pipelines/Pipelines.h

+
+  PassOptions::Option<bool> lowerFromLinalg{
+      *this, "lower-from-linalg",
+      desc("Disable Rock dialect targeting when bufferizing"), init(false)};


This description will need to be updated as to what this new flag is actually doing.

justinrosner · 2026-02-02T14:56:09Z

mlir/lib/Dialect/MIGraphX/Pipeline/Pipeline.cpp

  funcPm.addPass(migraphx::createMIGraphXTosaSimplifyPass());
 }
+
+void migraphx::addHighLevelMIGraphXToLinalg(PassManager &pm) {


I think you should be able to make use of the lowerFromLinalg flag/option that you defined earlier.

dhernandez0

same as the other PR, can you change the title of the commit to highlight this implements rock.gemm only.

dhernandez0 · 2026-02-03T08:42:52Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+
+  rock::StoreMethodAttr method =
+      rewriter.getAttr<rock::StoreMethodAttr>(rock::StoreMethod::Set);
+  rock::GemmOp result = rock::GemmOp::create(


are the shapes of a and b the same for migraphx.dot, linalg.matmul and rock.gemm? Please check if that's the case (for example, if they all expect B to be [batch, K, N].

Thanks for point this out, I just realized that this may not be always true. It is possible that linalg.matmul could be transposed if an indexing_map is set.

To the best of my knowledge, I think I have handled all the transpose case that comes with linalg. See matmul_transposed_* test cases.

Also, my understanding is that you cannot have transpose the c matrix:?

func.func @matmul_tranposed_B_3D(%arg0: tensor<1x3x2xf32>, %arg1: tensor<1x2x3xf32>) -> tensor<1x3x3xf32> attributes {arch = "gfx950", kernel} { %cst = arith.constant dense<0.000000e+00> : tensor<1x3x3xf32> %0 = linalg.batch_matmul indexing_maps = [affine_map<(batch, m, n, k) -> (batch, m, k)>, affine_map<(batch, m, n, k) -> (batch, k, n)>, // expected-error @+1 {{'linalg.batch_matmul' op Invalid output map result dimension.}} affine_map<(batch, m, n, k) -> (batch, n, m)>] ins(%arg0, %arg1 : tensor<1x3x2xf32>, tensor<1x2x3xf32>) outs(%cst : tensor<1x3x3xf32>) -> tensor<1x3x3xf32> return %0 : tensor<1x3x3xf32> }

Also, I think they are all [batch, K, N] and [batch, N, K]. rock.gemm seems to only support 3D or 2D.

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

Copilot

Pull request overview

This pull request introduces a new LinalgToRock conversion pass that provides an alternative lowering path from MIGraphX through the Linalg dialect to Rock, enabling end-to-end compilation flows.

Changes:

Added LinalgToRock conversion pass that converts linalg.matmul and linalg.batch_matmul operations to rock.gemm operations
Extended rocmlir-driver with new pipeline options (migraphx-linalg and highlevel-linalg) to support the Linalg-based lowering path
Added comprehensive test coverage including unit tests and end-to-end tests

Reviewed changes

Copilot reviewed 17 out of 17 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp	Core conversion logic implementing matmul pattern rewriting from Linalg to Rock
mlir/lib/Conversion/LinalgToRock/LinalgToRockPass.cpp	Pass implementation that applies the LinalgToRock conversion patterns
mlir/include/mlir/Conversion/LinalgToRock/LinalgToRock.h	Header file declaring the conversion pass and pattern population functions
mlir/include/mlir/Conversion/RocMLIRPasses.h	Added LinalgToRock header include
mlir/include/mlir/Conversion/RocMLIRPasses.td	Added LinalgToRockPass definition
mlir/lib/Conversion/LinalgToRock/CMakeLists.txt	Build configuration for the new conversion library
mlir/lib/Conversion/CMakeLists.txt	Added LinalgToRock subdirectory
mlir/lib/Dialect/Rock/Pipelines/Pipelines.cpp	Integrated LinalgToRock pass into the bufferization pipeline
mlir/lib/Dialect/Rock/Pipelines/CMakeLists.txt	Added MLIRLinalgToRock dependency
mlir/include/mlir/Dialect/Rock/Pipelines/Pipelines.h	Added lowerFromLinalg option to BufferizeOptions
mlir/lib/Dialect/MIGraphX/Pipeline/Pipeline.cpp	Modified to conditionally use LinalgToRock instead of TosaToRock
mlir/lib/Dialect/MIGraphX/Pipeline/CMakeLists.txt	Added MLIRMIGraphXToLinalg dependency
mlir/include/mlir/Dialect/MIGraphX/Pipeline/Pipeline.h	Added lowerFromLinalg parameter to addHighLevelPipeline
mlir/tools/rocmlir-driver/rocmlir-driver.cpp	Added support for new pipeline options (migraphx-linalg, highlevel-linalg)
mlir/test/Conversion/LinalgToRock/linalg-to-rock.mlir	Test verifying conversion with rock-view-to-transform
mlir/test/Conversion/LinalgToRock/linalg-to-rock-matmul.mlir	Test verifying basic LinalgToRock conversion
mlir/test/Conversion/LinalgToRock/e2e/linalg-to-rock-add-gemm.e2e.mlir	End-to-end integration test

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

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

mlir/lib/Conversion/LinalgToRock/LinalgToRockPass.cpp

mlir/include/mlir/Dialect/MIGraphX/Pipeline/Pipeline.h

mlir/include/mlir/Conversion/LinalgToRock/LinalgToRock.h

mlir/lib/Conversion/LinalgToRock/CMakeLists.txt

pabloantoniom

Don't forget to update the PR title

pabloantoniom · 2026-02-04T09:22:22Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+  Value a = op.getOperand(0);
+  Value b = op.getOperand(1);
+  Value cOriginal = op.getOperand(2);
+  RankedTensorType outputType = cast<RankedTensorType>(cOriginal.getType());


Fine. I would check here if we are getting static shapes at least. If the IR comes from migraphx-to-linalg they will be, but it's better if we make this pass as solid as possible, just in case we ingest IR from another source, which might contain RankedTensor with non-static shapes.

pabloantoniom · 2026-02-04T09:28:17Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+    LinalgMatOp op, OpAdaptor adaptor,
+    ConversionPatternRewriter &rewriter) const {
+  Location loc = op.getLoc();
+  Value a = op.getOperand(0);


nit:

Suggested change

Value a = op.getOperand(0);

Value a = op.lhs();

Value b = op.rhs();

Value c = op.getOutputs()[0];

Doesn't seem to compile?

/home/vhe/vscode/rocMLIR/mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp:41:16: error: no member named 'rhs' in 'mlir::linalg::MatmulOp' 41 | Value b = op.rhs(); | ~~ ^ /home/vhe/vscode/rocMLIR/mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp:40:16: error: no member named 'lhs' in 'mlir::linalg::BatchMatmulOp' 40 | Value a = op.lhs(); | ~~ ^

I couldn't find a method a member that has similar semantics.

mlir/lib/Conversion/LinalgToRock/CMakeLists.txt

mlir/lib/Conversion/LinalgToRock/LinalgToRockPass.cpp

mlir/include/mlir/Dialect/Rock/Pipelines/Pipelines.h

mlir/test/Conversion/LinalgToRock/e2e/linalg-to-rock-add-gemm.e2e.mlir

mlir/tools/rocmlir-driver/rocmlir-driver.cpp

umangyadav · 2026-02-09T14:57:57Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock.mlir

+// CHECK-NEXT: %[[three:.*]] = rock.gemm %[[two]] = %[[one]] * %[[zero]] storeMethod =  set
+// CHECK-NEXT: %[[four:.*]] = rock.transform %[[three]]
+// CHECK-NEXT: return %[[four]]
+func.func @matmul_3D(%arg0: tensor<6xf32>, %arg1: tensor<6xf32>) -> tensor<9xf32> attributes {arch = "##TOKEN_ARCH##", kernel} {


One more interesting test would be check if output from both TosaToRock and LinalgToRock matches or not starting from MIGraphX.

e.g.

Input
migraphx.dot

Lowering tosa -> rock

Lowering using Linalg->Rock

Check if output from both 1 and 2 matches or not.

see linalg-to-rock-vs-tosa-to-rock.mlir. Also, rocmlir-driver is used instead because it seems that tosa-to-tensor pass is not exported to romclir-opt.

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

umangyadav · 2026-02-11T20:27:38Z

mlir/test/Conversion/LinalgToRock/e2e/linalg-to-rock-add-gemm.e2e.mlir

@@ -0,0 +1,12 @@
+// RUN: rocmlir-gen -fut dot_add -arch %arch --clone-harness %s |\
+// RUN:     rocmlir-driver -host-pipeline=migraphx,highlevel -kernel-pipeline=migraphx-linalg,highlevel -targets %arch |\


not sure why you need multiple RUN lines if outputs are being piped to next command.

See other E2E tests.

rocMLIR/mlir/test/fusion/pr-e2e/mixr-gemm/mixr-gemm-add.e2e.mlir

Line 1 in 0f83e87

// RUN: rocmlir-gen -fut dot_add --arch %arch --clone-harness %s | rocmlir-driver -kernel-pipeline=migraphx,highlevel -host-pipeline=migraphx,highlevel | rocmlir-gen -ph -print-results -rand none -fut dot_add_wrapper - | rocmlir-driver -host-pipeline mhal,runner -kernel-pipeline full -targets %arch | 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 --entry-point-result=void | FileCheck %s

umangyadav · 2026-02-11T20:29:17Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+  if (auto attr = op->template getAttrOfType<StringAttr>("perf_config"))
+    result->setAttr("perf_config", attr);


add test for this

see linalg-to-rock-matmul.mlir.

umangyadav · 2026-02-11T20:29:48Z

mlir/lib/Conversion/LinalgToRock/LinalgToRock.cpp

+  if (failed(maybeAMatrixTransposed) || failed(maybeBMatrixTransposed)) {
+    return op.emitError("cannot determine if input matrix is transposed");
+  }


nit :
Would be nice to have code coverage for these lines

see linalg-to-rock-matmul.mlir

umangyadav · 2026-02-11T20:30:21Z

mlir/test/Conversion/LinalgToRock/linalg-to-rock-vs-tosa-to-rock.mlir

+// RUN: sed s/##TOKEN_ARCH##/%arch/g %s | rocmlir-driver --kernel-pipeline=migraphx-linalg,highlevel| FileCheck %s
+// RUN: sed s/##TOKEN_ARCH##/%arch/g %s | rocmlir-driver --kernel-pipeline=migraphx,highlevel | FileCheck %s


Is there a way to compare output of from both to see if they match exactly instead of relying on "CHECK" lines ?

If not, this is a good solution too.

I can't think of a good way because the diff between the tosa lowering path and the linalg lowering path is slightly different (difference in the variable name, some time ordering).

Mr-Anyone requested a review from causten as a code owner January 30, 2026 18:59

Mr-Anyone marked this pull request as draft January 30, 2026 19:00

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch 4 times, most recently from 3a1b2cc to f525a63 Compare January 30, 2026 20:15

Mr-Anyone marked this pull request as ready for review January 30, 2026 20:16

Mr-Anyone requested review from dhernandez0, justinrosner and pabloantoniom January 30, 2026 20:16

pabloantoniom reviewed Feb 2, 2026

View reviewed changes

justinrosner reviewed Feb 2, 2026

View reviewed changes

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

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch from f525a63 to 9ae19b8 Compare February 2, 2026 16:45

Mr-Anyone requested review from justinrosner and pabloantoniom February 2, 2026 19:42

dhernandez0 requested a review from Copilot February 3, 2026 08:40

Copilot started reviewing on behalf of dhernandez0 February 3, 2026 08:43 View session

dhernandez0 reviewed Feb 3, 2026

View reviewed changes

Copilot AI reviewed Feb 3, 2026

View reviewed changes

pabloantoniom approved these changes Feb 4, 2026

View reviewed changes

Mr-Anyone changed the title ~~[AIROCMLIR-446] Linalg to Rock Conversion~~ [AIROCMLIR-446] Lower linalg.matmul/batch_matmul into rock.gemm Feb 4, 2026

umangyadav reviewed Feb 9, 2026

View reviewed changes

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg branch 6 times, most recently from 42811ec to 7f4ce15 Compare February 9, 2026 22:03

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

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch from 9c0d44a to 271f8c8 Compare February 9, 2026 22:47

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 force-pushed the pr-template-migraphx-to-linalg-2 branch from 952693d to b501718 Compare February 10, 2026 20:02

Mr-Anyone requested review from dhernandez0 and umangyadav February 10, 2026 20:52

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch from 88d2296 to 80495d1 Compare February 10, 2026 20:53

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

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch from 80495d1 to 75bf455 Compare February 11, 2026 14:38

umangyadav reviewed Feb 11, 2026

View reviewed changes

Base automatically changed from pr-template-migraphx-to-linalg to develop February 11, 2026 22:46

Mr-Anyone added 8 commits February 12, 2026 15:41

Linalg to Rock Conversion

9201102

Address some comments

c89ca7a

Remove unused headers and addressing some of the comments

00a814d

Handle transpose

2cbea8e

Update testcase for better matching

260d1c5

Made some changes

d4fb651

Addressed comments

a9525f3

Set perf_conf and added testcase

d0be693

Mr-Anyone force-pushed the pr-template-migraphx-to-linalg-2 branch from 0d10133 to d0be693 Compare February 12, 2026 15:43

Mr-Anyone added 2 commits February 12, 2026 15:46

Restore unrelated changes

e53d1ba

Address comments

f1eb31b

Mr-Anyone requested a review from umangyadav February 12, 2026 17:07

use getNumResults instead

bc14b6c

		@@ -0,0 +1,39 @@
		// RUN: rocmlir-opt -split-input-file --migraphx-to-linalg --linalg-to-rock -verify-diagnostics %s \| FileCheck %s

-  Value a = op.getOperand(0);
+Value a = op.lhs();
+Value b = op.rhs();
+Value c = op.getOutputs()[0];

		@@ -0,0 +1,12 @@
		// RUN: rocmlir-gen -fut dot_add -arch %arch --clone-harness %s \|\
		// RUN: rocmlir-driver -host-pipeline=migraphx,highlevel -kernel-pipeline=migraphx-linalg,highlevel -targets %arch \|\

		if (auto attr = op->template getAttrOfType<StringAttr>("perf_config"))
		result->setAttr("perf_config", attr);

		// RUN: sed s/##TOKEN_ARCH##/%arch/g %s \| rocmlir-driver --kernel-pipeline=migraphx-linalg,highlevel\| FileCheck %s
		// RUN: sed s/##TOKEN_ARCH##/%arch/g %s \| rocmlir-driver --kernel-pipeline=migraphx,highlevel \| FileCheck %s

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!

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!

justinrosner commented Feb 2, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

dhernandez0 left a comment • 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!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

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!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Mr-Anyone commented Jan 30, 2026 •

edited

Loading

dhernandez0 left a comment •

edited

Loading

Mr-Anyone Feb 12, 2026 •

edited

Loading