CUDA: Add support for CUDA_ARCHITECTURES=native

CUDA 11.6 added the `nvcc -arch=native` flag to automatically compile
for the host GPUs' architectures.  Add support for specifying this
special `native` value in `CMAKE_CUDA_ARCHITECTURES` and
`CUDA_ARCHITECTURES`.  During the compiler ABI detection step,
detect the native architectures so we can pass them explicitly
when using Clang or older versions of nvcc.

Fixes: #22375

1 files changed, 31 insertions, 2 deletions

diff --git a/Modules/CMakeCUDACompilerABI.cu b/Modules/CMakeCUDACompilerABI.cu
index 449a079..8463e86 100644
--- a/Modules/CMakeCUDACompilerABI.cu
+++ b/Modules/CMakeCUDACompilerABI.cu
@@ -2,6 +2,10 @@
 #  error "A C or C++ compiler has been selected for CUDA"
 #endif
 
+#include <cstdio>
+
+#include <cuda_runtime.h>
+
 #include "CMakeCompilerABI.h"
 
 int main(int argc, char* argv[])
@@ -13,6 +17,31 @@ int main(int argc, char* argv[])
 #if defined(ABI_ID)
   require += info_abi[argc];
 #endif
-  (void)argv;
-  return require;
+  static_cast<void>(argv);
+
+  int count = 0;
+  if (cudaGetDeviceCount(&count) != cudaSuccess || count == 0) {
+    std::fprintf(stderr, "No CUDA devices found.\n");
+    return -1;
+  }
+
+  int found = 0;
+  const char* sep = "";
+  for (int device = 0; device < count; ++device) {
+    cudaDeviceProp prop;
+    if (cudaGetDeviceProperties(&prop, device) == cudaSuccess) {
+      std::printf("%s%d%d", sep, prop.major, prop.minor);
+      sep = ";";
+      found = 1;
+    }
+  }
+
+  if (!found) {
+    std::fprintf(stderr, "No CUDA architecture detected from any devices.\n");
+    // Convince the compiler that the non-zero return value depends
+    // on the info strings so they are not optimized out.
+    return require ? -1 : 1;
+  }
+
+  return 0;
 }