summaryrefslogtreecommitdiffstats
path: root/clang/lib/Headers
diff options
context:
space:
mode:
Diffstat (limited to 'clang/lib/Headers')
-rw-r--r--clang/lib/Headers/__clang_cuda_runtime_wrapper.h14
1 files changed, 14 insertions, 0 deletions
diff --git a/clang/lib/Headers/__clang_cuda_runtime_wrapper.h b/clang/lib/Headers/__clang_cuda_runtime_wrapper.h
index 47cb6a7e6a6..fb527dc9abc 100644
--- a/clang/lib/Headers/__clang_cuda_runtime_wrapper.h
+++ b/clang/lib/Headers/__clang_cuda_runtime_wrapper.h
@@ -247,5 +247,19 @@ __device__ static inline void *malloc(size_t __size) {
#include <__clang_cuda_cmath.h>
+// curand_mtgp32_kernel helpfully redeclares blockDim and threadIdx in host
+// mode, giving them their "proper" types of dim3 and uint3. This is
+// incompatible with the types we give in cuda_builtin_vars.h. As as hack,
+// force-include the header (nvcc doesn't include it by default) but redefine
+// dim3 and uint3 to our builtin types. (Thankfully dim3 and uint3 are only
+// used here for the redeclarations of blockDim and threadIdx.)
+#pragma push_macro("dim3")
+#pragma push_macro("uint3")
+#define dim3 __cuda_builtin_blockDim_t
+#define uint3 __cuda_builtin_threadIdx_t
+#include "curand_mtgp32_kernel.h"
+#pragma pop_macro("dim3")
+#pragma pop_macro("uint3")
+
#endif // __CUDA__
#endif // __CLANG_CUDA_RUNTIME_WRAPPER_H__
OpenPOWER on IntegriCloud