Do not build the CUBIN conversion pass when NVPTX Backend isn't configured

[mehdi_amini]

Summary

This pass would currently build, but fail to run when this backend isn't
linked in. On the other hand, we'd like it to initialize only the NVPTX
backend, which isn't possible if we continue to build it without the
backend available. Instead of building a broken configuration, let's
skip building the pass entirely.

[bkramer]

In the long term I'd prefer to not initialize backends inside of pass at all, but for the short term this is better than pulling in all backends.

[mehdi_amini]

ifdef out the pass registration

mlir/CMakeLists.txt

 
 # Build the CUDA conversions and run according tests if the NVPTX backend
 # is available
 if ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   set(MLIR_CUDA_CONVERSIONS_ENABLED 1)
 else()
   set(MLIR_CUDA_CONVERSIONS_ENABLED 0)
 endif()
+# TODO: we should use a config.h file like LLVM does
+add_definitions(-DMLIR_CUDA_CONVERSIONS_ENABLED=${MLIR_CUDA_CONVERSIONS_ENABLED})
 
 set(MLIR_CUDA_RUNNER_ENABLED 0 CACHE BOOL "Enable building the mlir CUDA runner")
 
 include_directories( "include")
 include_directories( ${MLIR_INCLUDE_DIR})
 
 add_subdirectory(include/mlir)
 add_subdirectory(lib)

mlir/include/mlir/InitAllPasses.h

inline void registerAllPasses() {
 
   // GPU
   createGpuKernelOutliningPass();
   createSimpleLoopsToGPUPass(0, 0);
   createLoopToGPUPass({}, {});
 
   // CUDA
   createConvertGpuLaunchFuncToCudaCallsPass();
+#if MLIR_CUDA_CONVERSIONS_ENABLED
   createConvertGPUKernelToCubinPass(
       [](const std::string &, Location, StringRef) { return nullptr; });
+#endif
   createLowerGpuOpsToNVVMOpsPass();
 
   // Linalg
   createLinalgFusionPass();
   createLinalgTilingPass();
   createLinalgTilingToParallelLoopsPass();
   createLinalgPromotionPass(0);
   createConvertLinalgToLoopsPass();

mlir/lib/Conversion/GPUToCUDA/CMakeLists.txt

-add_llvm_library(MLIRGPUtoCUDATransforms
+set(LLVM_OPTIONAL_SOURCES
   ConvertKernelFuncToCubin.cpp
+)
+
+set(SOURCES
   ConvertLaunchFuncToCudaCalls.cpp
 )
+
+if (MLIR_CUDA_CONVERSIONS_ENABLED)
+ list(APPEND SOURCES "ConvertKernelFuncToCubin.cpp")
+endif()
+
+add_llvm_library(MLIRGPUtoCUDATransforms ${SOURCES})
 target_link_libraries(MLIRGPUtoCUDATransforms
   MLIRGPU
   MLIRLLVMIR
   MLIRNVVMIR
   MLIRPass
   MLIRTargetNVVMIR
 )

mlir/lib/Conversion/GPUToCUDA/ConvertKernelFuncToCubin.cpp

public:
   GpuKernelToCubinPass(
       CubinGenerator cubinGenerator = compilePtxToCubinForTesting)
       : cubinGenerator(cubinGenerator) {}
 
   void runOnOperation() override {
     gpu::GPUModuleOp module = getOperation();
 
     // Make sure the NVPTX target is initialized.
-    llvm::InitializeAllTargets();
-    llvm::InitializeAllTargetMCs();
-    llvm::InitializeAllAsmPrinters();
+    LLVMInitializeNVPTXTarget();
+    LLVMInitializeNVPTXTargetInfo();
+    LLVMInitializeNVPTXTargetMC();
+    LLVMInitializeNVPTXAsmPrinter();
 
     auto llvmModule = translateModuleToNVVMIR(module);
     if (!llvmModule)
       return signalPassFailure();
 
     // Translate the module to CUBIN and attach the result as attribute to the
     // module.
     if (auto cubinAttr = translateGPUModuleToCubinAnnotation(