Default to minimal 2 rings and improve LL loop
[ROCm/rccl commit: 4b77a16f3f]
Tento commit je obsažen v:
@@ -454,8 +454,8 @@ class ncclLLPrimitives {
|
||||
asm volatile ("flat_load_dwordx4 %0, %1, glc\n"
|
||||
"s_waitcnt vmcnt(0)\n"
|
||||
"buffer_wbinvl1_vol\n" : "=v"(i4) : "v"(src));
|
||||
if (checkAbort(recvConn[i]->opCountRem)) break;
|
||||
} while (i4[1] != flag || i4[3] != flag);
|
||||
if (i4[1] == flag && i4[3] == flag) break;
|
||||
} while (!checkAbort(recvConn[i]->opCountRem));
|
||||
uint64_t val64 = (uint64_t)(i4[0]) + (((uint64_t)i4[2]) << 32);
|
||||
#else
|
||||
do {
|
||||
|
||||
@@ -171,7 +171,9 @@ static ncclResult_t fillCoords(int nranks, int* matrix, int* coords, int* rankTo
|
||||
}
|
||||
}
|
||||
|
||||
#ifdef __PPC__
|
||||
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__)
|
||||
#define DEFAULT_MIN_NRINGS 2
|
||||
#elif defined(__PPC__)
|
||||
// Make the default NCCL_MIN_NRINGS=4 for IBM/Power nodes
|
||||
#define DEFAULT_MIN_NRINGS 4
|
||||
#else
|
||||
|
||||
Odkázat v novém úkolu
Zablokovat Uživatele