AI/rocm-systems
2
0
Forka 0
Codice Problemi Pull Request Actions Pacchetti Progetti Rilasci Wiki Attività

[DEV] Configure functions in RCCL (#986)

Sfoglia il codice sorgente 
* configure functions in rccl

[ROCm/rccl commit: 28d9b170c9]
This commit is contained in:
Bertan Dogancay
2024-01-18 15:07:16 -07:00
committato da GitHub
parent 0343d9ccac
commit 11674674fc
24 ha cambiato i file con 570 aggiunte e 749 eliminazioni
Scarica il file Patch Scarica il file Diff Expand all files Collapse all files
projects/rccl/src/init.cc
+58 -20
Vedi File
@@ -18,6 +18,7 @@
#include "enqueue.h"
#include "graph.h"
#include "argcheck.h"
#include "devcomm.h"
#if defined(ENABLE_NPKIT)
#include "npkit/npkit.h"
#endif
@@ -31,6 +32,7 @@
#include <sys/types.h>
#include <sys/stat.h>
#include <unistd.h>
#include <cstdarg>
#include "graph/topo.h"
#include "graph/xml.h"
#include "archinfo.h"
@@ -54,7 +56,7 @@
#define NCCL_GROUP_CUDA_STREAM 1 // CGMD: CUDA 9.0,9.1 Need to use an internal CUDA stream
#endif
const char* ncclFuncStr[NCCL_NUM_FUNCTIONS+2] = { "Broadcast", "Reduce", "AllGather", "ReduceScatter", "AllReduce", "SendRecv", "AllToAllPivot" };
const char* ncclFuncStr[NCCL_NUM_FUNCTIONS+2] = { "AllGather", "AllReduce", "AllToAllPivot", "Broadcast", "Reduce", "ReduceScatter", "SendRecv"};
const char* ncclAlgoStr[NCCL_NUM_ALGORITHMS] = { "Tree", "Ring", "CollNetDirect", "CollNetChain", "NVLS", "NVLSTree" };
const char* ncclProtoStr[NCCL_NUM_PROTOCOLS] = { "LL", "LL128", "Simple" };
const char* ncclDevRedOpStr[ncclNumDevRedOps] = { "Sum", "Prod", "Max", "Min", "PreMulSum", "SumPostDiv" };
@@ -177,6 +179,18 @@ void NCCL_NO_OPTIMIZE commPoison(ncclComm_t comm) {
RCCL_PARAM(KernelCollTraceEnable, "KERNEL_COLL_TRACE_ENABLE", 0);
#ifdef ENABLE_COLLTRACE
#define MAX_NAME_LENGTH 64
// Helper function to generate function names and update funcIdx
void generateFunctionName(char* func_names, int& funcIdx, const char* format, ...) {
char* line = func_names + MAX_NAME_LENGTH * funcIdx;
va_list args;
va_start(args, format);
vsnprintf(line, MAX_NAME_LENGTH, format, args);
va_end(args);
funcIdx++;
}
// Should be in sync with 'ALL_COLLS' in Generator.cmake
void *ncclCommThreadMain(void *arg) {
ncclComm_t comm = (ncclComm_t)arg;
int head[MAXCHANNELS];
@@ -184,29 +198,53 @@ void *ncclCommThreadMain(void *arg) {
memset(head, 0, sizeof(int)*MAXCHANNELS);
vega_gpu_rtc_freq = GetDeviceWallClockRateInKhz(comm->cudaDev) * 1.0E3;
#define MAX_NAME_LENGTH 64
char* func_names = (char *)malloc(MAX_NAME_LENGTH*(FUNC_INDEX_P2P+2));
for (int func = 0; func < NCCL_NUM_FUNCTIONS; func++) {
for (int al = 0; al < NCCL_NUM_ALGORITHMS; al++) {
for (int type = 0; type < ncclNumTypes; type++) {
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
for (int devredop = 0; devredop < ncclNumDevRedOps; devredop++) {
char* line = func_names+MAX_NAME_LENGTH*FUNC_INDEX(func, devredop, type, al, pr);
sprintf(line, "%s%s%s%s%s", ncclFuncStr[func], ncclAlgoStr[al], ncclProtoStr[pr],
ncclDevRedOpStr[devredop], ncclTypeStr[type]);
}
char* func_names = (char *)malloc(MAX_NAME_LENGTH*(ncclFuncId_P2p()+/*OneRankReduce*/11));
int funcIdx = 0;
// AllGather --> RING / <all_protos> / Sum / int8_t
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
generateFunctionName(func_names, funcIdx, "AllGatherRing%sSum_i8", ncclProtoStr[pr]);
}
// AllReduce --> <all_algos> / <all_protos> / <all_redops> / <all_types>
for (int al = 0; al < NCCL_NUM_ALGORITHMS - 2; al++) {
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
for (int redop = 0; redop < ncclNumDevRedOps; redop++) {
for (int ty = 0; ty < ncclNumTypes; ty++) {
if (redop == 5 && ty > 5) continue;
generateFunctionName(func_names, funcIdx, "AllReduce%s%s%s%s", ncclAlgoStr[al], ncclProtoStr[pr], ncclDevRedOpStr[redop], ncclTypeStr[ty]);
}
}
}
}
for (int type = 0; type < ncclNumTypes; type++) {
char* line = func_names+MAX_NAME_LENGTH*(FUNC_INDEX_P2P-ncclNumTypes+type);
sprintf(line, "OneRankReducePreMulSum%s", ncclTypeStr[type]);
// AllToAllPivot --> RING / SIMPLE / Sum / int8_t
generateFunctionName(func_names, funcIdx, "AllToAllPivotRingSimpleSum_i8");
// Broadcast --> RING / <all_protos> / Sum / int8_t
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
generateFunctionName(func_names, funcIdx, "BroadcastRing%sSum_i8", ncclProtoStr[pr]);
}
// Reduce --> RING / <all_protos> / <all_redops> / <all_types>
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
for (int redop = 0; redop < ncclNumDevRedOps; redop++) {
for (int ty = 0; ty < ncclNumTypes; ty++) {
if (redop == 5 && ty > 5) continue;
generateFunctionName(func_names, funcIdx, "ReduceRing%s%s%s", ncclProtoStr[pr], ncclDevRedOpStr[redop], ncclTypeStr[ty]);
}
}
}
// ReduceScatter --> RING / <all_protos> / <all_redops> / <all_types>
for (int pr = 0; pr < NCCL_NUM_PROTOCOLS; pr++) {
for (int redop = 0; redop < ncclNumDevRedOps; redop++) {
for (int ty = 0; ty < ncclNumTypes; ty++) {
if (redop == 5 && ty > 5) continue;
generateFunctionName(func_names, funcIdx, "ReduceScatterRing%s%s%s", ncclProtoStr[pr], ncclDevRedOpStr[redop], ncclTypeStr[ty]);
}
}
}
// SendRecv --> RING / SIMPLE / Sum / int8_t
generateFunctionName(func_names, funcIdx, "SendRecvRingSimpleSum_i8");
// OneRankReduce --> PreMulSum / <all_types>
for (int ty = 0; ty < ncclNumTypes; ty++) {
generateFunctionName(func_names, funcIdx, "OneRankReducePreMulSum%s", ncclTypeStr[ty]);
}
char* line = func_names+MAX_NAME_LENGTH*FUNC_INDEX_P2P;
sprintf(line, "SendRecvRingSimpleSum_i8");
line += MAX_NAME_LENGTH;
sprintf(line, "AllToAllPivotRingSimpleSum_i8");
do {
for (int channel = 0; channel < MAXCHANNELS; channel++) {
int tail = comm->collTraceTail[channel].tail%COLLTRACE_NUM_ITEMS;
@@ -232,7 +270,7 @@ void *ncclCommThreadMain(void *arg) {
(double)(td->timeStamp)/vega_gpu_rtc_freq, comm->rank, td->bid,
fIdx, td->data_0, td->opCount, td->data_1);
} else {
if (fIdx == FUNC_INDEX_P2P || type == ncclCollTraceP2pElemType)
if (fIdx == ncclFuncId_P2p() || type == ncclCollTraceP2pElemType)
sprintf(line, "## [%012.6f] [%02d:%02d] %06x-%06x", (double)(td->timeStamp)/vega_gpu_rtc_freq, comm->rank, td->bid, td->p2pOpCount[0], td->p2pOpCount[1]);
else
sprintf(line, "## [%012.6f] [%02d:%02d] %06lx", (double)(td->timeStamp)/vega_gpu_rtc_freq, comm->rank, td->bid, td->opCount);