AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove workaround and use indirect function call (#684)

Browse Source 
[ROCm/rccl commit: f7a456122c]
This commit is contained in:
Wenkai Du
2023-02-14 13:59:48 -08:00
committed by GitHub
parent cb7e2e8eeb
commit 4fb1ebcf4b
12 changed files with 100 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/collectives.h
+9
View File
@@ -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