AI/rocm-systems
2
0
Forkuj 0
Kod Zgłoszenia Oczekujące zmiany Actions Packages Projekty Wydania Wiki Aktywność

GPU allocation for CPX Unit Tests using PCI bus id (#1403)

Przeglądaj źródła 
* mapping devices wrt pci

* Gpu allocation by using pci mapping

* Passing gpuPriorityOrder in as an argument rather than making the functions non-static.

* Removing redundant testBed instance calling

[ROCm/rccl commit: 69b2b712ab]
This commit is contained in:
saurabhAMD
2024-11-04 10:51:00 -06:00
zatwierdzone przez GitHub
rodzic ad1384bea1
commit 69d976532b
9 zmienionych plików z 119 dodań i 19 usunięć
Ściągnij plik aktualizacji Ściągnij plik porównania Expand all files Collapse all files
projects/rccl/test/GroupCallTests.cpp
+10 -5
Wyświetl plik
@@ -28,7 +28,8 @@ namespace RcclUnitTesting
{
// Test either single process all GPUs, or 1 process per GPU
int const numProcesses = isMultiProcess ? totalRanks : 1;
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks), numCollPerGroup);
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
if (testBed.ev.showNames)
INFO("%s %d-ranks GroupCall Identical\n", isMultiProcess ? "MP" : "SP", totalRanks);
@@ -84,7 +85,8 @@ namespace RcclUnitTesting
{
// Test either single process all GPUs, or 1 process per GPU
int const numProcesses = isMultiProcess ? totalRanks : 1;
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks), numCollPerGroup);
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
if (testBed.ev.showNames)
INFO("%s %d-ranks GroupCall Different\n", isMultiProcess ? "MP" : "SP", totalRanks);
@@ -139,7 +141,8 @@ namespace RcclUnitTesting
{
// Test either single process all GPUs, or 1 process per GPU
int const numProcesses = isMultiProcess ? totalRanks : 1;
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks), numCollPerGroup);
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
if (testBed.ev.showNames)
INFO("%s %d-ranks GroupCall MixedDayaType\n", isMultiProcess ? "MP" : "SP", totalRanks);
@@ -194,7 +197,8 @@ namespace RcclUnitTesting
INFO("%s %d-ranks Multistream %d-Group Calls across %d streams\n",
isMultiProcess ? "MP" : "SP", totalRanks, numCollPerGroup, numStreamsPerGroup);
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks),
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder),
numCollPerGroup, numStreamsPerGroup);
// Set up each collective in group in different stream (modulo numStreamsPerGroup)
@@ -244,7 +248,8 @@ namespace RcclUnitTesting
int const numProcesses = isMultiProcess ? totalRanks : 1;
// Initialize comms by specifying the # of group calls
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks), numCollsPerGroup, numStreamsPerGroup, numGroupCalls, useBlocking);
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollsPerGroup, numStreamsPerGroup, numGroupCalls, useBlocking);
if (testBed.ev.showNames)
INFO("%s %d-ranks GroupCall MultiGroupCall\n", isMultiProcess ? "MP" : "SP", totalRanks);