Add more constraints to enable GDR (#579)

* Add more constraints to enable GDR

* Revert deleted line

[ROCm/rccl commit: aa0d7ca882]
Dieser Commit ist enthalten in:
Wenkai Du
2022-07-08 09:52:27 -07:00
committet von GitHub
Ursprung bf15ad1d72
Commit 659cd52d5c
2 geänderte Dateien mit 11 neuen und 0 gelöschten Zeilen
+2
Datei anzeigen
@@ -56,6 +56,8 @@ ncclResult_t ncclTopoFillNet(struct ncclXml* xml, const char* pciPath, const cha
/* Remove unneeded parts */
ncclResult_t ncclTopoTrimXml(struct ncclXml* xml);
ncclResult_t ncclTopoGetStrFromSys(const char* path, const char* fileName, char* strValue);
/**************/
/* XML Struct */
/* Functions */
+9
Datei anzeigen
@@ -25,6 +25,7 @@
#include "timer.h"
#include "ibvwrap.h"
#include "graph/xml.h"
#define USE_RDMA_WRITE 1
#define MAXNAMESIZE 64
@@ -276,6 +277,14 @@ ncclResult_t ncclIbGdrSupport(int ibDev) {
if (moduleLoaded == -1) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
moduleLoaded = (access("/sys/kernel/mm/memory_peers/amdkfd/version", F_OK) == -1) ? 0 : 1;
char strValue[MAX_STR_LEN];
NCCLCHECK(ncclTopoGetStrFromSys("/sys/devices/virtual/dmi/id", "bios_version", strValue));
if (strncmp("Hyper-V UEFI Release", strValue, 20) == 0) {
int roMode = ncclParamIbPciRelaxedOrdering();
NCCLCHECK(ncclTopoGetStrFromSys("/proc/sys/kernel", "numa_balancing", strValue));
if (strcmp(strValue, "1") == 0 && roMode == 0)
moduleLoaded = 0;
}
#else
// Check for the nv_peer_mem module being loaded
moduleLoaded = ((access("/sys/kernel/mm/memory_peers/nv_mem/version", F_OK) == -1) &&