Lower local dialects in GPU pipeline (#917)

Adds missing conversion after #916 to allow Perf dialect lowering used
in GPU benchmarks.

lib/TPP/GPU/GpuPipeline.cpp

@@ -114,6 +114,8 @@ struct GpuPipeline : public tpp::impl::GpuPipelineBase<GpuPipeline>,
     registry.insert<nvgpu::NVGPUDialect>();
     registry.insert<bufferization::BufferizationDialect>();
     registry.insert<spirv::SPIRVDialect>();
+    registry.insert<check::CheckDialect>();
+    registry.insert<perf::PerfDialect>();
     check::registerBufferizableOpInterfaceExternalModels(registry);
     perf::registerBufferizableOpInterfaceExternalModels(registry);
 
@@ -179,6 +181,9 @@ struct GpuPipeline : public tpp::impl::GpuPipelineBase<GpuPipeline>,
     }
     }
 
+    // Covert all local dialects like perf.
+    pm.addPass(createLocalDialectsLowering());
+
     // Clean up after the GPU pipeline.
     // Use upstream passes directly instead of the cleanup pass as the GPU
     // kernel is at the LLVM dialect level which is not compatible with the