[HIP] Enable specifying different default gpu arch for HIP/CUDA.

[hliao]

clang/lib/Driver/Driver.cpp

2655

Sam, could you let me know which reasonable default arch should we use here?

[tra]

clang/lib/Driver/Driver.cpp

2538

This technically depends on the CUDA version.
We do have CUDA version info in clang/lib/Driver/ToolChains/Cuda.h
The default for NVCC has been sm_30 since CUDA-9.0. In fact sm_20 is not supported at all by CUDA-9.0+ at all , so we should bump the default to sm_30 for those versions.

[hliao]

clang/lib/Driver/Driver.cpp

2538

unfortunately, when the action build is running, the CUDA is not detected yet, I probably revise the detection logic to update CUDA's default gpu arch after successful detection

[yaxunl]

clang/lib/Driver/Driver.cpp

2655

I think it should be gfx803 since as far as I know it is the lowest one that supports HIP.

[tra]

clang/lib/Driver/Driver.cpp

2538

OK. If there's no easy way to do it here, it's probably not worth doing it just to keep sm_20 as the default.
sm_20 has not been supported by the last two major CUDA releases and it makes little sense to have the default that does not work for the majority of current users.

I think we should just bump the default to sm_30 for all currently supported CUDA versions in a separate patch. I can do it once this patch lands.

clang/lib/Driver/Driver.cpp

protected:
     /// The CUDA fat binary if it was generated for the current input.
     Action *CudaFatBinary = nullptr;
 
     /// Flag that is set to true if this builder acted on the current input.
     bool IsActive = false;
 
     /// Flag for -fgpu-rdc.
     bool Relocatable = false;
+
+    /// Default GPU architecture if there's no one specified.
+    CudaArch DefaultCudaArch = CudaArch::UNKNOWN;
+
   public:
     CudaActionBuilderBase(Compilation &C, DerivedArgList &Args,
                           const Driver::InputList &Inputs,
                           Action::OffloadKind OFKind)
         : DeviceActionBuilder(C, Args, Inputs, OFKind) {}
 
     ActionBuilderReturnCode addDeviceDepences(Action *HostAction) override {
       // While generating code for CUDA, we only depend on the host input action
bool initialize() override {
       // Collect list of GPUs remaining in the set.
       for (CudaArch Arch : GpuArchs)
         GpuArchList.push_back(Arch);
 
       // Default to sm_20 which is the lowest common denominator for
       // supported GPUs.  sm_20 code should work correctly, if
       // suboptimally, on all newer GPUs.
       if (GpuArchList.empty())
-        GpuArchList.push_back(CudaArch::SM_20);
+        GpuArchList.push_back(DefaultCudaArch);
 
       return Error;
     }
   };
 
   /// \brief CUDA action builder. It injects device code in the host backend
   /// action.
   class CudaActionBuilder final : public CudaActionBuilderBase {
   public:
     CudaActionBuilder(Compilation &C, DerivedArgList &Args,
                       const Driver::InputList &Inputs)
-        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) {}
+        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_Cuda) {
+      DefaultCudaArch = CudaArch::SM_20;
+    }
 
     ActionBuilderReturnCode
     getDeviceDependences(OffloadAction::DeviceDependences &DA,
                          phases::ID CurPhase, phases::ID FinalPhase,
                          PhasesTy &Phases) override {
       if (!IsActive)
         return ABRT_Inactive;
 
class OffloadingActionBuilder final {
   /// action.
   class HIPActionBuilder final : public CudaActionBuilderBase {
     /// The linker inputs obtained for each device arch.
     SmallVector<ActionList, 8> DeviceLinkerInputs;
 
   public:
     HIPActionBuilder(Compilation &C, DerivedArgList &Args,
                      const Driver::InputList &Inputs)
-        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) {}
+        : CudaActionBuilderBase(C, Args, Inputs, Action::OFK_HIP) {
+      DefaultCudaArch = CudaArch::GFX600;
+    }
 
     bool canUseBundlerUnbundler() const override { return true; }
 
     ActionBuilderReturnCode
     getDeviceDependences(OffloadAction::DeviceDependences &DA,
                          phases::ID CurPhase, phases::ID FinalPhase,
                          PhasesTy &Phases) override {
       // amdgcn does not support linking of object files, therefore we skip

clang/test/Driver/hip-default-gpu-arch.hip

+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang -### -x hip -c %s 2>&1 | FileCheck %s
+
+// CHECK: {{.*}}clang{{.*}}"-target-cpu" "gfx600"