[AIROCMLIR-555] Remove mhal.launch and mhal.await from the codebase

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

@@ -13,8 +13,6 @@
 #ifndef MLIR_CONVERSION_MIGRAPHXTOLINALG_H
 #define MLIR_CONVERSION_MIGRAPHXTOLINALG_H
 
-#include "mlir/Dialect/MIGraphX/IR/MIGraphX.h"
-#include "mlir/IR/PatternMatch.h"
 #include "mlir/Pass/Pass.h"
 #include "mlir/Transforms/DialectConversion.h"
 
@@ -40,9 +38,6 @@ void populateMIGraphXToLinalgBoundaryDialectConversion(
 void populateMIGraphXFuncBoundaryToLinalgConversionPatterns(
     RewritePatternSet &target, TypeConverter &typeConverter);
 
-/// Populates conversion patterns for function boundaries mhal.launcher
-void populateMIGraphXToLinalgMHALLauncherConversion(
-    RewritePatternSet &target, TypeConverter &typeConverter);
 } // namespace migraphx
 } // namespace mlir
 

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

@@ -16,12 +16,8 @@
 #ifndef MLIR_CONVERSION_MIGRAPHXTOTOSA_H
 #define MLIR_CONVERSION_MIGRAPHXTOTOSA_H
 
-#include "mlir/Dialect/Arith/IR/Arith.h"
-#include "mlir/Dialect/MHAL/IR/MHAL.h"
-#include "mlir/Dialect/MIGraphX/IR/MIGraphX.h"
-#include "mlir/Dialect/Tosa/IR/TosaOps.h"
-#include "mlir/IR/PatternMatch.h"
 #include "mlir/Pass/Pass.h"
+#include "mlir/Pass/PassManager.h"
 #include "mlir/Transforms/DialectConversion.h"
 
 namespace mlir {

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.cpp

@@ -10,7 +10,6 @@
 //
 //===----------------------------------------------------------------------===//
 #include "mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h"
-#include "mlir/Conversion/MIGraphXToTosa/MIGraphXToTosa.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Arith/Utils/Utils.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
@@ -1242,10 +1241,6 @@ void mlir::migraphx::populateMIGraphXFuncBoundaryToLinalgConversionPatterns(
     RewritePatternSet &patterns, TypeConverter &typeConverter) {
   patterns.add<AsUnderlyingShapeConverter, AsLogicalShapeOpConverter>(
       typeConverter, patterns.getContext());
-
-  // mhal.launch can be generated through rocmlir-gen, so we need a way to
-  // legalize it
-  populateMIGraphXToLinalgMHALLauncherConversion(patterns, typeConverter);
   populateAnyFunctionOpInterfaceTypeConversionPattern(patterns, typeConverter);
   populateReturnOpTypeConversionPattern(patterns, typeConverter);
   populateCallOpTypeConversionPattern(patterns, typeConverter);

mlir/lib/Conversion/MIGraphXToLinalg/MIGraphXToLinalgPass.cpp

@@ -16,10 +16,9 @@
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/Linalg/IR/Linalg.h"
 #include "mlir/Dialect/Math/IR/Math.h"
+#include "mlir/Dialect/MIGraphX/IR/MIGraphX.h"
 #include "mlir/Dialect/Rock/IR/Rock.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
-#include "mlir/IR/PatternMatch.h"
-#include "mlir/Pass/Pass.h"
 #include "mlir/Transforms/DialectConversion.h"
 
 using namespace mlir;
@@ -53,11 +52,6 @@ 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<func::ReturnOp>(
       [&](func::ReturnOp op) { return typeConverter.isLegal(op); });
   target.addDynamicallyLegalOp<func::CallOp>(
@@ -97,8 +91,6 @@ void MIGraphXToLinalgPass::runOnOperation() {
       boundaryConversionTarget, boundaryTypeConverter);
   migraphx::populateMIGraphXFuncBoundaryToLinalgConversionPatterns(
       boundaryPattern, boundaryTypeConverter);
-  migraphx::populateMIGraphXToLinalgMHALLauncherConversion(
-      boundaryPattern, boundaryTypeConverter);
   if (failed(applyPartialConversion(func, boundaryConversionTarget,
                                     std::move(boundaryPattern)))) {
     return signalPassFailure();

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp

@@ -11,12 +11,10 @@
 //
 //===----------------------------------------------------------------------===//
 #include "mlir/Conversion/MIGraphXToTosa/MIGraphXToTosa.h"
-#include "mlir/Conversion/MIGraphXToLinalg/MIGraphXToLinalg.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/Func/Transforms/FuncConversions.h"
 #include "mlir/Dialect/Linalg/IR/Linalg.h"
-#include "mlir/Dialect/MHAL/IR/MHAL.h"
 #include "mlir/Dialect/MIGraphX/IR/MIGraphX.h"
 #include "mlir/Dialect/MemRef/IR/MemRef.h"
 #include "mlir/Dialect/Rock/IR/Rock.h"
@@ -31,7 +29,6 @@
 #include "mlir/IR/BuiltinTypes.h"
 #include "mlir/IR/PatternMatch.h"
 #include "mlir/IR/TypeUtilities.h"
-#include "mlir/Transforms/DialectConversion.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 #include "llvm/ADT/APFloat.h"
 #include "llvm/Support/Debug.h"
@@ -1529,14 +1526,6 @@ struct AsUnderlyingShapeConverter final
                   ConversionPatternRewriter &rewriter) const final;
 };
 
-/// This mirrors the call op conversion pattern but works for mhal.launch.
-struct MHALLaunchConverter final : public OpConversionPattern<mhal::LaunchOp> {
-  using OpConversionPattern<mhal::LaunchOp>::OpConversionPattern;
-
-  LogicalResult
-  matchAndRewrite(mhal::LaunchOp op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const final;
-};
 } // namespace
 
 LogicalResult AsLogicalShapeConverter::matchAndRewrite(
@@ -1678,26 +1667,6 @@ LogicalResult AsUnderlyingShapeConverter::matchAndRewrite(
   return success();
 }
 
-LogicalResult MHALLaunchConverter::matchAndRewrite(
-    mhal::LaunchOp op, OpAdaptor adaptor,
-    ConversionPatternRewriter &rewriter) const {
-  // Convert the original function results.
-  SmallVector<Type, 2> resultTypes;
-  if (failed(typeConverter->convertTypes(op.getResultTypes(), resultTypes)))
-    return failure();
-
-  // If this isn't a one-to-one type mapping, we don't know how to aggregate
-  // the results.
-  if (op->getNumResults() != resultTypes.size())
-    return failure();
-
-  // Substitute with the new result types from the corresponding FuncType
-  // conversion.
-  rewriter.replaceOpWithNewOp<mhal::LaunchOp>(
-      op, op.getCalleeAttr(), resultTypes, adaptor.getOperands());
-  return success();
-}
-
 //===----------------------------------------------------------------------===//
 // External interface
 //===----------------------------------------------------------------------===//
@@ -1733,14 +1702,12 @@ void migraphx::populateMIGraphXToTosaConversionPatterns(
 
 void mlir::migraphx::populateMIGraphXFuncBoundaryToTosaConversionPatterns(
     RewritePatternSet &patterns, TypeConverter &typeConverter) {
+
   patterns.add<AsLogicalShapeConverter, AsUnderlyingShapeConverter,
-               TrivialConverter<func::ReturnOp, func::ReturnOp>,
-               MHALLaunchConverter>(typeConverter, patterns.getContext());
+                TrivialConverter<func::ReturnOp, func::ReturnOp>>(
+      typeConverter, patterns.getContext());
   // Add upstream patterns that take care of func.func and its friends.
   populateAnyFunctionOpInterfaceTypeConversionPattern(patterns, typeConverter);
   populateCallOpTypeConversionPattern(patterns, typeConverter);
 }
-void mlir::migraphx::populateMIGraphXToLinalgMHALLauncherConversion(
-    RewritePatternSet &patterns, TypeConverter &typeConverter) {
-  patterns.add<MHALLaunchConverter>(typeConverter, patterns.getContext());
-}
+

mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosaPass.cpp

@@ -14,15 +14,12 @@
 #include "mlir/Conversion/MIGraphXToTosa/MIGraphXToTosa.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
-#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/MHAL/IR/MHAL.h"
 #include "mlir/Dialect/MIGraphX/IR/MIGraphX.h"
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Dialect/Tosa/IR/TosaOps.h"
-#include "mlir/IR/PatternMatch.h"
 #include "mlir/Pass/PassManager.h"
 #include "mlir/Transforms/DialectConversion.h"
-#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
-#include "mlir/Transforms/Passes.h"
 
 #define DEBUG_TYPE "migraphx-to-tosa"
 
@@ -62,11 +59,6 @@ void mlir::migraphx::populateMIGraphXFuncBoundaryToTosaDialectConversion(
       [=](func::CallOp op) -> std::optional<bool> {
         return typeConverter->isSignatureLegal(op.getCalleeType());
       });
-  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) -> std::optional<bool> {
         return typeConverter->isLegal(op);

mlir/test/fusion/pr-e2e/mixr-conv-bias-clipped-relu.mlir

@@ -13,8 +13,7 @@ module {
     return %7 : !migraphx.shaped<4x4x1x1xf32, 4x1x1x1>
   }
   func.func @mlir_convolution_add_clip_wrapper(%arg0: !migraphx.shaped<1x4x1x1xf32, 4x1x1x1>, %arg1: !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>, %arg2: !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>) -> !migraphx.shaped<4x4x1x1xf32, 4x1x1x1> {
-    %token, %results = mhal.launch @mlir_convolution_add_clip (%arg0, %arg1, %arg2) : (!migraphx.shaped<1x4x1x1xf32, 4x1x1x1>, !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>, !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>) -> !migraphx.shaped<4x4x1x1xf32, 4x1x1x1>
-    mhal.await %token : !mhal.token
+    %results = func.call @mlir_convolution_add_clip(%arg0, %arg1, %arg2) : (!migraphx.shaped<1x4x1x1xf32, 4x1x1x1>, !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>, !migraphx.shaped<4x3x3x3xf32, 27x9x3x1>) -> !migraphx.shaped<4x4x1x1xf32, 4x1x1x1>
     return %results : !migraphx.shaped<4x4x1x1xf32, 4x1x1x1>
   }
   module @__xmodule_ attributes {mhal.arch = "##TOKEN_ARCH##", mhal.module} {

mlir/test/fusion/pr-e2e/mixr-dot-int4-f16.mlir

@@ -1,5 +1,5 @@
 // RUN: rocmlir-driver -kernel-pipeline=migraphx %s | rocmlir-gen -fut mlir_unpack_dequantizelinear_dot --arch %arch --clone-harness -  | FileCheck %s --check-prefix=HASINT4
-// HASINT4: mhal.launch
+// HASINT4: call @mlir_unpack_dequantizelinear_dot
 // HASINT4-SAME: tensor<64xi4>
 
 // RUN: rocmlir-driver -kernel-pipeline=migraphx %s | rocmlir-gen -fut mlir_unpack_dequantizelinear_dot --arch %arch --clone-harness - | rocmlir-driver -host-pipeline=highlevel -kernel-pipeline=highlevel | rocmlir-gen -ph -fut mlir_unpack_dequantizelinear_dot_wrapper --verifier clone - | rocmlir-driver -host-pipeline mhal,runner -kernel-pipeline full | mlir-runner -O2 --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

mlir/test/fusion/pr-e2e/mixr-expand-strides-non-multiple.mlir

@@ -1,10 +1,9 @@
-// RUN: rocmlir-gen -fut mlir_dot_sigmoid --arch %arch --clone-harness %s | rocmlir-driver -kernel-pipeline=migraphx,highlevel -host-pipeline=migraphx,highlevel -targets %arch | rocmlir-gen -ph -rand 1 -rand_type float -fut mlir_dot_sigmoid_wrapper --verifier clone - | 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,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void | FileCheck %s 
+// RUN: rocmlir-gen -fut mlir_dot_sigmoid --arch %arch --clone-harness %s | rocmlir-driver -kernel-pipeline=migraphx,highlevel -host-pipeline=migraphx,highlevel -targets %arch | rocmlir-gen -ph -rand 1 -rand_type float -fut mlir_dot_sigmoid_wrapper --verifier clone -print-verify-results=always - | 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,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void | FileCheck %s 
 
-// Only ~41% of the results will be correct since the non-contiguous strides
-// in this example is the result of concating 4x5x24 and 4x7x24 (i.e.m 4x12x24).
-// This kernel is specificall dealing with making sure that the 4x5x24 elements
-// are all in the correct place in the larger tensor.
-// CHECK: relDiff = 0 : 480/1152 (41.666667%)
+// Non-contiguous strides from a concat-like layout (4x5x24 and 4x7x24 in a
+// 4x12x24 logical tensor). Ensures dot+sigmoid + expand-strides placement;
+// with correct GPU staging, GPU matches CPU on the full output.
+// CHECK: relDiff = 0 : 1152/1152 (100.000000%)
 
 module {
   func.func @mlir_dot_sigmoid(%arg0: !migraphx.shaped<4x5x16xf16, 80x16x1>, %arg1: !migraphx.shaped<4x16x24xf16, 384x24x1>) -> !migraphx.shaped<4x5x24xf16, 288x24x1> attributes {arch = "gfx1201", kernel = "mixr", num_cu = 32 : i64} {

mlir/test/fusion/pr-e2e/mixr-gemm/mixr-gemm-tr-folding.mlir

@@ -9,8 +9,7 @@ module {
     return %3 : !migraphx.shaped<2x16x8xf32, 128x8x1>
   }
   func.func @mlir_transpose_reshape_dot_wrapper(%arg0: !migraphx.shaped<2x8x4x4xf32, 128x16x4x1>, %arg1: !migraphx.shaped<1x8x8xf32, 64x8x1> ) -> !migraphx.shaped<2x16x8xf32, 128x8x1> {
-    %token, %results = mhal.launch @mlir_transpose_reshape_dot (%arg0, %arg1) : (!migraphx.shaped<2x8x4x4xf32, 128x16x4x1>, !migraphx.shaped<1x8x8xf32, 64x8x1>) -> !migraphx.shaped<2x16x8xf32, 128x8x1>
-    mhal.await %token : !mhal.token
+    %results = func.call @mlir_transpose_reshape_dot(%arg0, %arg1) : (!migraphx.shaped<2x8x4x4xf32, 128x16x4x1>, !migraphx.shaped<1x8x8xf32, 64x8x1>) -> !migraphx.shaped<2x16x8xf32, 128x8x1>
     return %results : !migraphx.shaped<2x16x8xf32, 128x8x1>
   }
   module @__xmodule_ attributes {mhal.arch = "##TOKEN_ARCH##", mhal.module} {

mlir/test/fusion/pr-e2e/mixr-gemm/mixr-gemm-tr-folding2.mlir

@@ -8,8 +8,7 @@ module {
     return %2 : !migraphx.shaped<1x6xf32, 6x1>
   }
   func.func @mlir_transpose_reshape_dot_wrapper(%arg0: !migraphx.shaped<1x2x1x3xf32, 6x3x3x1>, %arg1: !migraphx.shaped<6x6xf32, 6x1>) -> !migraphx.shaped<1x6xf32, 6x1> {
-    %token, %results = mhal.launch @mlir_transpose_reshape_dot (%arg0, %arg1) : (!migraphx.shaped<1x2x1x3xf32, 6x3x3x1>, !migraphx.shaped<6x6xf32, 6x1>) -> !migraphx.shaped<1x6xf32, 6x1>
-    mhal.await %token : !mhal.token
+    %results = func.call @mlir_transpose_reshape_dot(%arg0, %arg1) : (!migraphx.shaped<1x2x1x3xf32, 6x3x3x1>, !migraphx.shaped<6x6xf32, 6x1>) -> !migraphx.shaped<1x6xf32, 6x1>
     return %results : !migraphx.shaped<1x6xf32, 6x1>
   }
   module @__xmodule_ attributes {mhal.arch = "##TOKEN_ARCH##", mhal.module} {

mlir/test/fusion/pr-e2e/mixr-non-contiguous-strides.mlir

@@ -1,8 +1,8 @@
-// RUN: rocmlir-gen -fut mlir_dot_sigmoid --arch %arch --clone-harness %s | rocmlir-driver -kernel-pipeline=migraphx,highlevel -host-pipeline=migraphx,highlevel -targets %arch | rocmlir-gen -ph -rand 1 -rand_type float -fut mlir_dot_sigmoid_wrapper --verifier clone - | 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,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void | FileCheck %s 
+// RUN: rocmlir-gen -fut mlir_dot_sigmoid --arch %arch --clone-harness %s | rocmlir-driver -kernel-pipeline=migraphx,highlevel -host-pipeline=migraphx,highlevel -targets %arch | rocmlir-gen -ph -rand 1 -rand_type float -fut mlir_dot_sigmoid_wrapper --verifier clone -print-verify-results=always - | 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,%linalg_test_lib_dir/libmlir_async_runtime%shlibext --entry-point-result=void | FileCheck %s 
 
-// Only half of the results will be correct since the non-contiguous strides
-// in this example means that half of the memory is uninitialized.
-// CHECK: relDiff = 0     : 2304/4608 (50.000000%)
+// Non-contiguous strides: validates expand-strides / padded layout. With
+// correct GPU host staging, CPU and GPU references match on the full tensor.
+// CHECK: relDiff = 0 : 4608/4608 (100.000000%)
 
 module {
   func.func @mlir_dot_sigmoid(%arg0: !migraphx.shaped<4x24x16xf16, 384x16x1>, %arg1: !migraphx.shaped<4x16x24xf16, 384x24x1>) -> !migraphx.shaped<4x24x24xf16, 1152x24x1> attributes {kernel = "mixr"} {

mlir/test/fusion/pr-e2e/mixr-sd-explicit-broadcasting.mlir

@@ -8,8 +8,7 @@ module {
     return %2 : !migraphx.shaped<1x1x32x32xf32, 1024x1024x32x1>
   }
   func.func @mlir_reshape_convolution_real(%arg0: !migraphx.shaped<1x1x16x1x16x1xf32, 256x256x16x16x1x1>, %arg1: !migraphx.shaped<1x1x3x3xf32, 9x9x3x1>) -> !migraphx.shaped<1x1x32x32xf32, 1024x1024x32x1> {
-    %token, %results = mhal.launch @mlir_reshape_convolution_real__part_0 (%arg0, %arg1) : (!migraphx.shaped<1x1x16x1x16x1xf32, 256x256x16x16x1x1>, !migraphx.shaped<1x1x3x3xf32, 9x9x3x1>) -> !migraphx.shaped<1x1x32x32xf32, 1024x1024x32x1>
-    mhal.await %token : !mhal.token
+    %results = func.call @mlir_reshape_convolution_real__part_0(%arg0, %arg1) : (!migraphx.shaped<1x1x16x1x16x1xf32, 256x256x16x16x1x1>, !migraphx.shaped<1x1x3x3xf32, 9x9x3x1>) -> !migraphx.shaped<1x1x32x32xf32, 1024x1024x32x1>
     return %results : !migraphx.shaped<1x1x32x32xf32, 1024x1024x32x1>
   }
   module @__xmodule_ attributes {mhal.arch = "##TOKEN_ARCH##", mhal.module} {

mlir/test/fusion/pr-e2e/tosa-to-rock-exp.e2e.mlir

@@ -1,8 +1,9 @@
-// RUN: rocmlir-gen -fut test_fusion --arch %arch --clone-harness %s | rocmlir-driver -host-pipeline highlevel -kernel-pipeline highlevel | rocmlir-gen -ph -fut test_fusion_wrapper -rand 1 -rand_type float --verifier clone - | rocmlir-driver -host-pipeline mhal,runner -kernel-pipeline full -targets %arch | mlir-runner -O2 --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 -entry-point-result=void | FileCheck %s
+// RUN: rocmlir-gen -fut test_fusion --arch %arch --clone-harness %s | rocmlir-driver -host-pipeline highlevel -kernel-pipeline highlevel | rocmlir-gen -ph -fut test_fusion_wrapper -rand 1 -rand_type float --verifier clone -print-verify-results=always - | rocmlir-driver -host-pipeline mhal,runner -kernel-pipeline full -targets %arch | mlir-runner -O2 --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 -entry-point-result=void | FileCheck %s
 
 module {
-// CHECK: RMS = {{.*}}e-09
-// CHECK: [1 0 0]
+// CHECK: Number of elements:
+// CHECK: RMS = {{.*}}
+// CHECK: [1 1 1]
   func.func @test_fusion(%arg0: tensor<128x32x32x8xf32>, %arg1: tensor<128x3x3x8xf32>) -> tensor<128x30x30x128xf32> {
 
     %zero = arith.constant dense<0.0> : tensor<128xf32>