Use correct WARP_SIZE for gfx1030 (#348)
[ROCm/rccl commit: 9c718ce6d6]
Este commit está contenido en:
Wenkai Du
GitHub
@@ -61,7 +61,11 @@ union ncclLLFifoLine {
|
||||
int4 i4;
|
||||
};
|
||||
|
||||
#if defined(__gfx1030__)
|
||||
#define WARP_SIZE 32
|
||||
#else
|
||||
#define WARP_SIZE 64
|
||||
#endif
|
||||
#define MAXCHANNELS 32
|
||||
#define NCCL_MAX_NTHREADS 256
|
||||
#define NCCL_SIMPLE_MAX_NTHREADS NCCL_MAX_NTHREADS
|
||||
|
||||
Referencia en una nueva incidencia
Block a user