AI/rocm-systems
2
0
Förgrening 0
Kod Ärenden Pull-förfrågningar Actions Packages Projekt Släpp Wiki Aktiviteter

Merge remote-tracking branch 'origin/develop' into 2.8.3

Bläddra i källkod 
[ROCm/rccl commit: ab1e7a0318]
This commit is contained in:
Wenkai Du
2021-02-04 19:46:48 -05:00
förälder 1dfb88f554 16d625ca27
incheckning ae5779702a
36 ändrade filer med 2820 tillägg och 46 borttagningar
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/tools/TransferBench/TransferBench.cpp
+26 -3
Visa fil
@@ -166,12 +166,21 @@ int main(int argc, char **argv)
links[i].blockParam = (BlockParam*)malloc(ev.numCpuPerLink * sizeof(BlockParam));
}
}
<<<<<<< HEAD
// Loop over all the different number of bytes to use per Link
for (auto N : valuesOfN)
{
if (!ev.outputToCsv) printf("Test %d: [%lu bytes]\n", testNum, N * sizeof(float));
=======
// Loop over all the different number of bytes to use per Link
for (auto N : valuesOfN)
{
if (!ev.outputToCsv) printf("Test %d: [%lu bytes]\n", testNum, N * sizeof(float));
>>>>>>> origin/develop
// Prepare links based on current N
for (int i = 0; i < numLinks; i++)
{
@@ -331,6 +340,7 @@ int main(int argc, char **argv)
DeallocateMemory(links[i].dstMemType, links[i].dstIndex, links[i].dstMem);
if (links[i].exeMemType == MEM_GPU)
<<<<<<< HEAD
{
HIP_CALL(hipEventDestroy(links[i].startEvent));
HIP_CALL(hipEventDestroy(links[i].stopEvent));
@@ -339,6 +349,16 @@ int main(int argc, char **argv)
}
else if (links[i].exeMemType == MEM_CPU)
{
=======
{
HIP_CALL(hipEventDestroy(links[i].startEvent));
HIP_CALL(hipEventDestroy(links[i].stopEvent));
HIP_CALL(hipStreamDestroy(links[i].stream));
HIP_CALL(hipFree(links[i].blockParam));
}
else if (links[i].exeMemType == MEM_CPU)
{
>>>>>>> origin/develop
free(links[i].blockParam);
}
}
@@ -540,11 +560,12 @@ void DisplayTopology()
printf(" |");
for (int j = 0; j < numGpuDevices; j++)
printf(" GPU %02d |", j);
printf("\n");
printf(" PCIe Bus ID\n");
for (int j = 0; j <= numGpuDevices; j++)
printf("--------+");
printf("\n");
printf("-------------\n");
char pciBusId[20];
for (int i = 0; i < numGpuDevices; i++)
{
printf(" GPU %02d |", i);
@@ -565,7 +586,8 @@ void DisplayTopology()
hopCount);
}
}
printf("\n");
HIP_CALL(hipDeviceGetPCIBusId(pciBusId, 20, i));
printf(" %s\n", pciBusId);
}
}
@@ -865,6 +887,7 @@ void CheckOrFill(ModeType mode, int N, bool isMemset, bool isHipCall, float* ptr
exit(1);
}
}
free(hostBuffer);
}
free(refBuffer);
projects/rccl/tools/scripts/topo_val.sh
+1 -1
Visa fil
@@ -21,7 +21,7 @@
DIR="$(cd -P "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
for i in {0..44}
for i in {0..47}
do
$DIR/../topo_expl/topo_expl -m $i > "topo_m$i.log"
$DIR/../TopoVisual/topo_visual.sh -i "topo_m$i.log"
projects/rccl/tools/topo_expl/models/topo_8p_rome_n4_1.xml
+93
Visa fil
@@ -0,0 +1,93 @@
<system version="2">
<cpu numaid="1" affinity="00000000,00000000,00000000,ffff0000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:41:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:43:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="0" sm="90" gcn="908" arch="38911" rank="0" gdr="1">
<xgmi target="0000:23:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:26:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:03:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="2" affinity="00000000,00000000,0000ffff,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:21:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:23:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="1" sm="90" gcn="908" arch="38911" rank="1" gdr="1">
<xgmi target="0000:43:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:26:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:03:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
<pci busid="0000:24:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:26:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="2" sm="90" gcn="908" arch="38911" rank="2" gdr="1">
<xgmi target="0000:43:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:23:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:03:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="3" affinity="00000000,00000000,ffff0000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:01:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:03:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="3" sm="90" gcn="908" arch="38911" rank="3" gdr="1">
<xgmi target="0000:43:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:23:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:26:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="5" affinity="00000000,ffff0000,00000000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:c1:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:c3:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="4" sm="90" gcn="908" arch="38911" rank="4" gdr="1">
<xgmi target="0000:c6:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:a3:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:83:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
<pci busid="0000:c4:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:c6:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="5" sm="90" gcn="908" arch="38911" rank="5" gdr="1">
<xgmi target="0000:c3:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:a3:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:83:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="6" affinity="0000ffff,00000000,00000000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:a1:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:a3:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="6" sm="90" gcn="908" arch="38911" rank="6" gdr="1">
<xgmi target="0000:c3:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:c6:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:83:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="7" affinity="ffff0000,00000000,00000000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:81:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:83:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="7" sm="90" gcn="908" arch="38911" rank="7" gdr="1">
<xgmi target="0000:c3:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:c6:00.0" count="1" tclass="0x038000"/>
<xgmi target="0000:a3:00.0" count="1" tclass="0x038000"/>
</gpu>
</pci>
</pci>
</cpu>
<cpu numaid="4" affinity="00000000,0000ffff,00000000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:e1:00.0" class="0x020700" link_speed="16 GT/s" link_width="16">
<nic>
<net name="mlx5_0" dev="0" speed="200000" port="1" guid="0x42fc9f00039b0398" maxconn="262144" gdr="1"/>
</nic>
</pci>
</cpu>
</system>
projects/rccl/tools/topo_expl/models/topo_8p_rome_pcie.xml
+52
Visa fil
@@ -0,0 +1,52 @@
<system version="2">
<cpu numaid="0" affinity="00000000,00000000,ffffffff,ffffffff" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:41:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:43:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="0" sm="90" gcn="906" arch="38911" rank="0" gdr="1"/>
</pci>
</pci>
<pci busid="0000:21:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:23:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="1" sm="90" gcn="906" arch="38911" rank="1" gdr="1"/>
</pci>
</pci>
<pci busid="0000:24:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:26:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="2" sm="90" gcn="906" arch="38911" rank="2" gdr="1"/>
</pci>
</pci>
<pci busid="0000:01:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:03:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="3" sm="90" gcn="906" arch="38911" rank="3" gdr="1"/>
</pci>
</pci>
</cpu>
<cpu numaid="1" affinity="ffffffff,ffffffff,00000000,00000000" arch="x86_64" vendor="AuthenticAMD" familyid="143" modelid="49">
<pci busid="0000:e1:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:e3:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="4" sm="90" gcn="906" arch="38911" rank="4" gdr="1"/>
</pci>
</pci>
<pci busid="0000:c1:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:c3:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="5" sm="90" gcn="906" arch="38911" rank="5" gdr="1"/>
</pci>
</pci>
<pci busid="0000:c4:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:c6:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="6" sm="90" gcn="906" arch="38911" rank="6" gdr="1"/>
</pci>
</pci>
<pci busid="0000:81:00.0" class="0x060400" link_speed="16 GT/s" link_width="16">
<pci busid="0000:83:00.0" class="0x038000" link_speed="16 GT/s" link_width="16">
<gpu dev="7" sm="90" gcn="906" arch="38911" rank="7" gdr="1"/>
</pci>
</pci>
<pci busid="0000:a1:00.0" class="0x020000" link_speed="8 GT/s" link_width="4">
<nic>
<net name="mlx5_0" dev="0" speed="100000" port="1" guid="0x7893510003a1420c" maxconn="262144" gdr="1"/>
<net name="mlx5_1" dev="1" speed="100000" port="2" guid="0x7893510003a1420c" maxconn="262144" gdr="1"/>
</nic>
</pci>
</cpu>
</system>
projects/rccl/tools/topo_expl/topo_expl.cpp
+3
Visa fil
@@ -114,6 +114,9 @@ NodeModelDesc model_descs[] = {
{4, "topo_4p3l_n4.xml", "4 nodes 8 gfx908 Rome NPS=4"},
{1, "topo_4p3l_n2_1.xml", "single node 8 gfx908 Rome"},
{4, "topo_4p3l_n2_1.xml", "4 nodes 8 gfx908 Rome"},
{1, "topo_8p_rome_n4_1.xml", "single node 8 gfx908 Rome NPS=4"},
{4, "topo_8p_rome_n4_1.xml", "4 nodes node 8 gfx908 Rome NPS=4"},
{2, "topo_8p_rome_pcie.xml", "2 nodes node 8 VEGA20 PCIe"},
};
int main(int argc,char* argv[])