Remove workaround and use indirect function call (#684)

이 커밋은 다음에 포함됨:
Wenkai Du
2023-02-14 13:59:48 -08:00
커밋한 사람 GitHub
부모 9461a43168
커밋 f7a456122c
12개의 변경된 파일100개의 추가작업 그리고 8개의 파일을 삭제
+9
파일 보기
@@ -45,12 +45,21 @@ struct ncclDevRedOpFull {
nccl##func##algo##proto
/* Declare all collective operations */
#ifdef USE_INDIRECT_FUNCTION_CALL
#define DECL5(func, algo, proto, devredop, type) \
extern __device__ void NCCL_FUNC_NAME(func, algo, proto, devredop, type)(); \
extern __global__ void NCCL_KERN_NAME(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_DEBUG(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_LL128(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_LL128_DEBUG(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead);
#else
#define DECL5(func, algo, proto, devredop, type) \
extern __device__ __attribute__((noinline)) void NCCL_FUNC_NAME(func, algo, proto, devredop, type)(); \
extern __global__ void NCCL_KERN_NAME(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_DEBUG(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_LL128(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
extern __global__ void NCCL_KERN_NAME_LL128_DEBUG(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead);
#endif
#define SINGLE_ARG(...) __VA_ARGS__
#define CONCAT(a,b) a##b