diff --git a/src/include/non_abi/device/pt-to-pt/utils_device.h b/src/include/non_abi/device/pt-to-pt/utils_device.h index 8342ebe06aeed8fcd5461e0fb45b36de266f7ae5..58a2a874d63e9a716544b17c9e3c55ea2b0e7e6c 100644 --- a/src/include/non_abi/device/pt-to-pt/utils_device.h +++ b/src/include/non_abi/device/pt-to-pt/utils_device.h @@ -29,6 +29,8 @@ #ifdef NVSHMEMI_COMM_DEVICE_UTILS_USE_PTX +#if defined __CUDA_ARCH__ + __device__ static inline uint64_t BSWAP64(uint64_t x) { uint64_t ret; asm volatile( @@ -83,6 +85,8 @@ __device__ static inline uint16_t BSWAP16(uint16_t x) { return ret; } +#endif /*! __CUDA_ARCH__ */ + #else /* NVSHMEMI_COMM_DEVICE_UTILS_USE_PTX */ #define BSWAP64(x) \