#pragma once #ifdef USE_CUDA #define WARP_SIZE 32 #define MAX_THREADS_PER_BLOCK 1024 #define REDUCE_THREADS 256 #define HOST_AND_DEVICE __host__ __device__ #define FORCE_INLINE __forceinline__ #include #include #else #define HOST_AND_DEVICE #define FORCE_INLINE inline #endif // USE_CUDA #include #include typedef enum { INFO = 0, WARNING = 1, ERROR = 2, FATAL = 3 } level_t; const char* ToString(level_t level);