Improve collective trace
Tá an tiomantas seo le fáil i:
@@ -9,6 +9,9 @@
|
||||
#define NCCL_COLLECTIVES_H_
|
||||
|
||||
#define FUNC_INDEX_P2P (NCCL_NUM_FUNCTIONS*NCCL_NUM_ALGORITHMS*NCCL_NUM_PROTOCOLS*ncclNumTypes*ncclNumOps)
|
||||
#define FUNC_INDEX_A2A (FUNC_INDEX_P2P+1)
|
||||
#define FUNC_INDEX_A2AV (FUNC_INDEX_P2P+2)
|
||||
|
||||
#define FUNC_INDEX(func, redop, ncclType, al, pr) ((func >= NCCL_NUM_FUNCTIONS) \
|
||||
? (func-NCCL_NUM_FUNCTIONS+NCCL_NUM_FUNCTIONS*NCCL_NUM_ALGORITHMS*NCCL_NUM_PROTOCOLS*ncclNumTypes*ncclNumOps) \
|
||||
: ((((((func)*ncclNumOps + (redop))*ncclNumTypes) + (ncclType))*NCCL_NUM_ALGORITHMS+(al))*NCCL_NUM_PROTOCOLS+(pr)))
|
||||
|
||||
@@ -286,7 +286,18 @@ struct ncclCollTrace {
|
||||
uint32_t data_0;
|
||||
uint64_t timeStamp;
|
||||
uint64_t opCount;
|
||||
uint64_t data_1;
|
||||
union {
|
||||
uint64_t data_1;
|
||||
struct {
|
||||
uint16_t nThreads;
|
||||
uint8_t bid;
|
||||
uint8_t nChannels;
|
||||
} coll;
|
||||
struct {
|
||||
uint16_t nThreads;
|
||||
uint16_t delta;
|
||||
} p2p;
|
||||
};
|
||||
};
|
||||
static_assert(sizeof(struct ncclCollTrace) == 8*sizeof(int), "ncclCollTrace must have a pow2 size");
|
||||
|
||||
|
||||
Tagairt in Eagrán Nua
Cuir bac ar úsáideoir