AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.15.1-1

Browse Source 
Add support for H100 (sm90).
Make sure NCCL kernel honor user stream priorities.
This commit is contained in:
Sylvain Jeaugey
2022-09-27 02:31:13 -07:00
parent 78313a6d21
commit da8152e57a
22 changed files with 303 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/misc/cudawrap.cc
+7 -6
View File
@@ -38,7 +38,7 @@ DECLARE_CUDA_PFN(cuGetProcAddress, 11030);
#define CUDA_DRIVER_MIN_VERSION 11030
static void *cudaLib;
static int cudaDriverVersion;
int ncclCudaDriverVersionCache = -1;
#if CUDART_VERSION >= 11030
/*
@@ -107,16 +107,17 @@ static void initOnceFunc() {
goto error;
}
res = pfn_cuDriverGetVersion(&cudaDriverVersion);
int driverVersion;
res = pfn_cuDriverGetVersion(&driverVersion);
if (res != 0) {
WARN("cuDriverGetVersion failed with %d", res);
goto error;
}
INFO(NCCL_INIT, "cudaDriverVersion %d", cudaDriverVersion);
INFO(NCCL_INIT, "cudaDriverVersion %d", driverVersion);
if (cudaDriverVersion < CUDA_DRIVER_MIN_VERSION) {
// WARN("CUDA Driver version found is %d. Minimum requirement is %d", cudaDriverVersion, CUDA_DRIVER_MIN_VERSION);
if (driverVersion < CUDA_DRIVER_MIN_VERSION) {
// WARN("CUDA Driver version found is %d. Minimum requirement is %d", driverVersion, CUDA_DRIVER_MIN_VERSION);
// Silently ignore version check mismatch for backwards compatibility
goto error;
}
@@ -148,7 +149,7 @@ error:
return;
}
ncclResult_t cudaLibraryInit() {
ncclResult_t ncclCudaLibraryInit() {
pthread_once(&initOnceControl, initOnceFunc);
return initResult;
}
src/misc/nvmlwrap.cc
+10 -1
View File
@@ -38,6 +38,7 @@ namespace {
NCCL_NVML_FN(nvmlDeviceGetNvLinkCapability, nvmlReturn_t, (nvmlDevice_t device, unsigned int link, nvmlNvLinkCapability_t capability, unsigned int *capResult))
NCCL_NVML_FN(nvmlDeviceGetCudaComputeCapability, nvmlReturn_t, (nvmlDevice_t device, int* major, int* minor))
NCCL_NVML_FN(nvmlDeviceGetP2PStatus, nvmlReturn_t, (nvmlDevice_t device1, nvmlDevice_t device2, nvmlGpuP2PCapsIndex_t p2pIndex, nvmlGpuP2PStatus_t* p2pStatus))
NCCL_NVML_FN(nvmlDeviceGetFieldValues, nvmlReturn_t, (nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values))
std::mutex lock; // NVML has had some thread safety bugs
bool initialized = false;
@@ -80,7 +81,8 @@ ncclResult_t ncclNvmlEnsureInitialized() {
{(void**)&pfn_nvmlDeviceGetNvLinkRemotePciInfo, "nvmlDeviceGetNvLinkRemotePciInfo"},
{(void**)&pfn_nvmlDeviceGetNvLinkCapability, "nvmlDeviceGetNvLinkCapability"},
{(void**)&pfn_nvmlDeviceGetCudaComputeCapability, "nvmlDeviceGetCudaComputeCapability"},
{(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"}
{(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"},
{(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"}
};
for(Symbol sym: symbols) {
*sym.ppfn = dlsym(libhandle, sym.name);
@@ -260,3 +262,10 @@ ncclResult_t ncclNvmlDeviceGetP2PStatus(
}
return ncclSuccess;
}
ncclResult_t ncclNvmlDeviceGetFieldValues(nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values) {
NCCLCHECK(ncclNvmlEnsureInitialized());
std::lock_guard<std::mutex> locked(lock);
NVMLTRY(nvmlDeviceGetFieldValues, device, valuesCount, values);
return ncclSuccess;
}
src/misc/strongstream.cc
+5 -7
View File
@@ -5,6 +5,7 @@
************************************************************************/
#include "strongstream.h"
#include "cudawrap.h"
#include "checks.h"
#include "param.h"
@@ -14,10 +15,8 @@ ncclResult_t ncclCudaGetCapturingGraph(
struct ncclCudaGraph* graph, cudaStream_t stream
) {
#if CUDART_VERSION >= 11030
thread_local int driver = -1;
if (driver == -1) {
CUDACHECK(cudaDriverGetVersion(&driver));
}
int driver;
NCCLCHECK(ncclCudaDriverVersion(&driver));
if (driver < 11030) {
cudaStreamCaptureStatus status;
unsigned long long gid;
@@ -192,11 +191,11 @@ ncclResult_t ncclStrongStreamWaitStream(
CUDACHECK(cudaEventRecord(b->event, b->stream));
}
CUDACHECK(cudaStreamWaitEvent(a->stream, b->event, 0));
a->eventIsLagging = 1;
} else {
cudaGraphNode_t pair[2] = {a->node, b->node};
CUDACHECK(cudaGraphAddEmptyNode(&a->node, graph.graph, pair, 2));
}
a->eventIsLagging = 1;
#else
CUDACHECK(cudaEventRecord(b->event, b->stream));
CUDACHECK(cudaStreamWaitEvent(a->stream, b->event, 0));
@@ -232,9 +231,8 @@ ncclResult_t ncclStrongStreamWaitStream(
}
cudaGraphNode_t pair[2] = {a->node, tie};
CUDACHECK(cudaGraphAddEmptyNode(&a->node, graph.graph, pair, 2));
a->eventIsLagging = 1;
}
// a->eventIsLagging doesn't change since we are just updating the
// dependencies of a->node.
}
#else
CUDACHECK(cudaEventRecord(a->event, b));