bump default SM for CUDA to 75

Author: bratpiorka

clang/include/clang/Basic/OffloadArch.h

@@ -167,7 +167,7 @@ enum class OffloadArch {
   LNL_M,
   LAST,
 
-  CudaDefault = OffloadArch::SM_52,
+  CudaDefault = OffloadArch::SM_75,
   HIPDefault = OffloadArch::GFX906,
 };
 

clang/lib/Driver/Driver.cpp

@@ -5927,11 +5927,11 @@ class OffloadingActionBuilder final {
 
       // Handle defaults architectures
       for (auto &Triple : SYCLTripleList) {
-        // For NVIDIA use SM_50 as a default
+        // For NVIDIA use SM_75 as a default
         if (Triple.isNVPTX() && llvm::none_of(GpuArchList, [&](auto &P) {
               return P.first.isNVPTX();
             })) {
-          const char *DefaultArch = OffloadArchToString(OffloadArch::SM_50);
+          const char *DefaultArch = OffloadArchToString(OffloadArch::SM_75);
           GpuArchList.emplace_back(Triple, DefaultArch);
         }
 
@@ -7650,7 +7650,7 @@ Driver::getOffloadArchs(Compilation &C, const llvm::opt::DerivedArgList &Args,
       // The default arch is set for NVPTX if not provided.  For AMDGPU, emit
       // an error as the user is responsible to set the arch.
       if (TC.getTriple().isNVPTX())
-        Archs.insert(OffloadArchToString(OffloadArch::SM_50));
+        Archs.insert(OffloadArchToString(OffloadArch::SM_75));
       else if (TC.getTriple().isAMDGPU())
         C.getDriver().Diag(clang::diag::err_drv_sycl_missing_amdgpu_arch)
             << 1 << TC.getTriple().str();

libdevice/cmake/modules/SYCLLibdevice.cmake

@@ -97,7 +97,7 @@ set(imf_build_archs)
 if ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   list(APPEND full_build_archs nvptx64-nvidia-cuda)
   set(compile_opts_nvptx64-nvidia-cuda "-fsycl-targets=nvptx64-nvidia-cuda"
-  "-Xsycl-target-backend" "--cuda-gpu-arch=sm_50" "-nocudalib" "-fno-sycl-libspirv" "-Wno-unsafe-libspirv-not-linked")
+  "-Xsycl-target-backend" "--cuda-gpu-arch=sm_75" "-fno-sycl-libspirv" "-Wno-unsafe-libspirv-not-linked")
   set(opt_flags_nvptx64-nvidia-cuda "-O3" "--nvvm-reflect-enable=false")
 endif()
 if("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)

sycl-jit/jit-compiler/lib/translation/Translation.cpp

@@ -197,13 +197,14 @@ std::pair<std::string, std::string> Translator::getTargetCPUAndFeatureAttrs(
   if (CPU.empty()) {
     // Set to the lowest tested target according to the GetStartedGuide, section
     // "Build DPC++ toolchain with support for HIP AMD"
-    CPU = Format == BinaryFormat::AMDGCN ? "gfx90a" : "sm_50";
+    CPU = Format == BinaryFormat::AMDGCN ? "gfx90a" : "sm_75";
     if (KernelFunc && KernelFunc->hasFnAttribute(TARGET_CPU_ATTRIBUTE)) {
       CPU = KernelFunc->getFnAttribute(TARGET_CPU_ATTRIBUTE).getValueAsString();
     }
   }
   if (Features.empty()) {
-    Features = Format == BinaryFormat::PTX ? "+sm_50,+ptx76" : "";
+    // Turing architecture + PTX 6.3
+    Features = Format == BinaryFormat::PTX ? "+sm_75,+ptx63" : "";
     if (KernelFunc && KernelFunc->hasFnAttribute(TARGET_FEATURE_ATTRIBUTE)) {
       Features = KernelFunc->getFnAttribute(TARGET_FEATURE_ATTRIBUTE)
                      .getValueAsString();

sycl/doc/GetStartedGuide.md

@@ -198,8 +198,8 @@ extensions that require sm_80 and later architectures also require at least CUDA
 11.0.
 
 The CUDA backend should work on Windows or Linux operating systems with any GPU
-with compute capability (SM version) sm_50 or above. The default SM version for
-the NVIDIA CUDA backend is sm_50. Users of sm_3X devices can attempt to specify
+with compute capability (SM version) sm_75 or above. The default SM version for
+the NVIDIA CUDA backend is sm_75. Users of sm_3X devices can attempt to specify
 the target architecture [ahead of time](#aot-target-architectures), provided
 that they use a 11.X  or earlier CUDA toolkit version, but some features may not be
 supported. The CUDA backend has been tested with different Ubuntu Linux