AI/rocm-systems
2
0
Forc 0
Cód Saincheisteanna Iarratais Tarraingthe Gníomhartha Pacáistí Tionscadail Scaoileann Vicí Gníomhaíocht

[DEVICE] Enable PAT algo for RCCL 1ppn (#1756)

Brabhsáil Foinse 
* Enable PAT algo for RCCL 1ppn


[ROCm/rccl commit: e96c8473a1]
Tá an tiomantas seo le fáil i:
Bertan Dogancay
2025-07-04 13:45:18 -04:00
tiomanta ag GitHub
tuismitheoir 82a822b646
tiomantas 471fc6bff2
D'athraigh 12 comhad le 108 breiseanna agus 92 scriosta
Íoslódáil an comhad paiste Íoslódáil Comhad Diff Leathnaigh gach comhaid Laghdaigh gach comhaid
projects/rccl/src/include/collectives.h
+1 -1
Féach ar an gComhad
@@ -402,7 +402,7 @@ public:
#endif
// Need a power of two to ensure it divides by parallelFactor (which is also a power of two)
#define NCCL_PAT_NWORKERS 512
#define NCCL_PAT_NWORKERS 128
static constexpr int PatUsed = 0x1,
PatSkipped = 0x2;
projects/rccl/src/include/device.h
+30 -17
Féach ar an gComhad
@@ -693,48 +693,61 @@ extern int const ncclDevFuncRowToId[];
inline int ncclDevFuncId(int coll, int devRedOp, int type, int algo, int proto) {
int row = 0;
do {
// RING / <all_protos> / Sum / int8_t
// RING/PAT | <all_protos> | Sum | int8_t
int nAlgos = 2;
if (coll == ncclFuncAllGather) {
row += proto;
int algo1 = algo == NCCL_ALGO_RING ? 0 :
/*algo == NCCL_ALGO_PAT*/ 1;
row += algo1 * NCCL_NUM_PROTOCOLS + proto;
break;
}
row += NCCL_NUM_PROTOCOLS;
row += nAlgos * NCCL_NUM_PROTOCOLS;
// <all_algos> / <all_protos> / <all_redops> / <all_types>
// RING/TREE | <all_protos> | <all_redops> | <all_types>
nAlgos = 2;
if (coll == ncclFuncAllReduce) {
row += (((algo * NCCL_NUM_PROTOCOLS + proto) * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * (algo * NCCL_NUM_PROTOCOLS + proto);
int algo1 = algo == NCCL_ALGO_TREE ? 0 :
/*algo == NCCL_ALGO_RING*/ 1;
row += (((algo1 * NCCL_NUM_PROTOCOLS + proto) * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * (algo1 * NCCL_NUM_PROTOCOLS + proto);
break;
}
row += (NCCL_NUM_ALGORITHMS - 5) * NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);
row += nAlgos * NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);
// RING / SIMPLE / Sum / int8_t
// RING | SIMPLE | Sum | int8_t
nAlgos = 1;
if (coll == ncclFuncAllToAllPivot) break;
row += 1;
row += nAlgos * 1;
// RING / <all_protos> / Sum / int8_t
// RING | <all_protos> | Sum | int8_t
nAlgos = 1;
if (coll == ncclFuncBroadcast) {
row += proto;
break;
}
row += NCCL_NUM_PROTOCOLS;
row += nAlgos * NCCL_NUM_PROTOCOLS;
// RING / <all_protos> / <all_redops> / <all_types>
// RING | <all_protos> | <all_redops> | <all_types>
nAlgos = 1;
if (coll == ncclFuncReduce) {
row += ((proto * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * proto;
row += ((proto * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * proto;
break;
}
row += NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);
row += nAlgos * NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);
// RING / <all_protos> / <all_redops> / <all_types>
// RING/PAT | <all_protos> | <all_redops> | <all_types>
nAlgos = 2;
if (coll == ncclFuncReduceScatter) {
row += ((proto * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * proto;
int algo1 = algo == NCCL_ALGO_RING ? 0 :
/*algo == NCCL_ALGO_PAT*/ 1;
row += (((algo1 * NCCL_NUM_PROTOCOLS + proto) * ncclNumDevRedOps + devRedOp) * ncclNumTypes + type) - NCCL_NUM_FLOATS * (algo1 * NCCL_NUM_PROTOCOLS + proto);
break;
}
row += NCCL_NUM_PROTOCOLS * (ncclNumDevRedOps * ncclNumTypes - NCCL_NUM_FLOATS);
// RING / SIMPLE / Sum / int8_t
// RING | SIMPLE | Sum | int8_t
nAlgos = 1;
if (coll == ncclFuncSendRecv) break;
row += 1;
row += nAlgos * 1;
} while (false);
projects/rccl/src/include/nccl_common.h
+1 -1
Féach ar an gComhad
@@ -40,7 +40,7 @@ typedef enum {
typedef void (*ncclDebugLogger_t)(ncclDebugLogLevel level, unsigned long flags, const char *file, int line, const char *fmt, ...);
#define NCCL_NUM_ONERANK 12
#define FUNC_INDEX_TOTAL 656 + NCCL_NUM_ONERANK
#define FUNC_INDEX_TOTAL 821 + NCCL_NUM_ONERANK
#define NCCL_NUM_FUNCTIONS 5 // Send/Recv not included for now
typedef enum {