mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-06 12:20:52 +01:00
6e1ba79b7f
This is an attempt to re-land https://github.com/pytorch/pytorch/pull/114001. The previous attempt used `std::array` in cuda kernels which wasn't compatible with Meta's internal build. Pull Request resolved: https://github.com/pytorch/pytorch/pull/116125 Approved by: https://github.com/yf225
49 lines
1.3 KiB
C++
49 lines
1.3 KiB
C++
#if !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
|
|
#include <c10/cuda/driver_api.h>
|
|
#include <c10/util/CallOnce.h>
|
|
#include <c10/util/Exception.h>
|
|
#include <dlfcn.h>
|
|
#include <iostream>
|
|
namespace c10 {
|
|
namespace cuda {
|
|
|
|
namespace {
|
|
|
|
DriverAPI create_driver_api() {
|
|
void* handle_0 = dlopen("libcuda.so.1", RTLD_LAZY | RTLD_NOLOAD);
|
|
TORCH_CHECK(handle_0, "Can't open libcuda.so.1: ", dlerror());
|
|
void* handle_1 = DriverAPI::get_nvml_handle();
|
|
DriverAPI r{};
|
|
|
|
#define LOOKUP_LIBCUDA_ENTRY(name) \
|
|
r.name##_ = ((decltype(&name))dlsym(handle_0, #name)); \
|
|
TORCH_INTERNAL_ASSERT(r.name##_, "Can't find ", #name, ": ", dlerror())
|
|
C10_LIBCUDA_DRIVER_API(LOOKUP_LIBCUDA_ENTRY)
|
|
#undef LOOKUP_LIBCUDA_ENTRY
|
|
|
|
if (handle_1) {
|
|
#define LOOKUP_NVML_ENTRY(name) \
|
|
r.name##_ = ((decltype(&name))dlsym(handle_1, #name)); \
|
|
TORCH_INTERNAL_ASSERT(r.name##_, "Can't find ", #name, ": ", dlerror())
|
|
C10_NVML_DRIVER_API(LOOKUP_NVML_ENTRY)
|
|
#undef LOOKUP_NVML_ENTRY
|
|
}
|
|
return r;
|
|
}
|
|
} // namespace
|
|
|
|
void* DriverAPI::get_nvml_handle() {
|
|
static void* nvml_hanle = dlopen("libnvidia-ml.so.1", RTLD_LAZY);
|
|
return nvml_hanle;
|
|
}
|
|
|
|
C10_EXPORT DriverAPI* DriverAPI::get() {
|
|
static DriverAPI singleton = create_driver_api();
|
|
return &singleton;
|
|
}
|
|
|
|
} // namespace cuda
|
|
} // namespace c10
|
|
|
|
#endif
|