AI/rocm-systems
2
0
Tükrözés 0
Kód Hibajegyek Egyesítési kérések Actions Packages Projektek Kiadások Wiki Tevékenység

gfx12 Disable ll protocol (#1268)

Forráskód böngészése 
[ROCm/rccl commit: c246e25f8e]
This commit is contained in:
akolliasAMD
2024-07-26 08:59:55 -06:00
committed by GitHub
szülő 0aa81c3194
commit 37c44d531b
5 fájl változott, egészen pontosan 33 új sor hozzáadva és 23 régi sor törölve
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/enqueue.cc
+4 -4
Fájl megtekintése
@@ -271,7 +271,7 @@ static ncclResult_t computeCollAlignCount(struct ncclInfo* collInfo, size_t* ali
if (collInfo->protocol == NCCL_PROTO_SIMPLE) {
*alignCount = NCCL_SIMPLE_ALIGNMENT / ncclTypeSize(collInfo->datatype);
} else if (collInfo->protocol == NCCL_PROTO_LL128) {
// LL128 alignCount should be same as LL for now. NCCL_LL128_ALIGNMENT_PER_WARP needs review
// LL128 alignCount should be same as LL for now. NCCL_LL128_ALIGNMENT_PER_WARP needs review
*alignCount = NCCL_LL_ALIGNMENT_PER_THREAD / ncclTypeSize(collInfo->datatype) * collInfo->nThreads;
} else {
*alignCount = NCCL_LL_ALIGNMENT_PER_THREAD / ncclTypeSize(collInfo->datatype) * collInfo->nThreads;
@@ -591,8 +591,8 @@ static ncclResult_t addP2pToPlan(
// 1 is connIndex
struct ncclConnInfo* conn = isSendNotRecv ?
&comm->channels[channelId].peers[peer]->send[1].conn : &comm->channels[channelId].peers[peer]->recv[1].conn;
// do not use LL on gfx11
info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold()) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
// do not use LL on gfx12
info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold() && !IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx12")) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
int reg = 0;
if (info.protocol == NCCL_PROTO_SIMPLE) {
@@ -1923,7 +1923,7 @@ static ncclResult_t hostToDevRedOp(
int64_t i64;
uint64_t u64;
half f16;
float f32;
float f32;
double f64;
#if defined(RCCL_BFLOAT16)
hip_bfloat16 bf16;
projects/rccl/src/graph/tuning.cc
+4 -4
Fájl megtekintése
@@ -54,7 +54,7 @@ ncclResult_t parseList(const char* str, const char* elems[], int nelems, int* li
// Latencies in us, Bandwidths in GB/s
// Tree { LL, LL128, Simple } , Ring { LL, LL128, Simple }
static const float baseLat [NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = {
static const float baseLat [NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = {
{ 12.0, 12.0, 17.0 }, { 12.0, 12.0, 17.0 }, // Tree, Ring
{ 12.0, 12.0, 17.0 }, { 12.0, 12.0, 17.0 }, // Collnet Direct, Chain
{ 0, 0, 0 }, { 0, 0, 0 }}; // NVLS, NVLS Tree
@@ -372,7 +372,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
if (a == NCCL_ALGO_NVLS) bw = std::min(graphs[a]->bwIntra, graphs[a]->bwInter);
if (a == NCCL_ALGO_NVLS_TREE) bw = std::min(graphs[a]->bwIntra, nNodes <= 2 ? graphs[a]->bwInter : graphs[a]->bwInter/2);
// Various model refinements
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
if (nNodes <= 2)
@@ -507,8 +507,8 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
}
for (int c=0; c<NCCL_NUM_FUNCTIONS; c++) for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
// Disable LL protocol on gfx11xx
int pEnable = protoEnable[p];
// Disable LL protocol on gfx12xx
int pEnable = (p == NCCL_PROTO_LL && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx12")) ? 0 : protoEnable[p];
if (pEnable == 2 && p == NCCL_PROTO_LL128) {
#if defined(__HIP_PLATFORM_AMD__) || defined(__HCC__) || defined(__HIPCC__)
#if defined(ENABLE_LL128)