Do not fail when the process is already using CUDA.
Attempt to use the selected cuda device. More strict device selection.2.0
castano
15 years ago
parent
6e9feef6f4
commit
93625c7de8
@ -1,239 +1,300 @@
|
|||||||
// Copyright NVIDIA Corporation 2007 -- Ignacio Castano <icastano@nvidia.com>
|
// Copyright NVIDIA Corporation 2007 -- Ignacio Castano <icastano@nvidia.com>
|
||||||
//
|
//
|
||||||
// Permission is hereby granted, free of charge, to any person
|
// Permission is hereby granted, free of charge, to any person
|
||||||
// obtaining a copy of this software and associated documentation
|
// obtaining a copy of this software and associated documentation
|
||||||
// files (the "Software"), to deal in the Software without
|
// files (the "Software"), to deal in the Software without
|
||||||
// restriction, including without limitation the rights to use,
|
// restriction, including without limitation the rights to use,
|
||||||
// copy, modify, merge, publish, distribute, sublicense, and/or sell
|
// copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||||
// copies of the Software, and to permit persons to whom the
|
// copies of the Software, and to permit persons to whom the
|
||||||
// Software is furnished to do so, subject to the following
|
// Software is furnished to do so, subject to the following
|
||||||
// conditions:
|
// conditions:
|
||||||
//
|
//
|
||||||
// The above copyright notice and this permission notice shall be
|
// The above copyright notice and this permission notice shall be
|
||||||
// included in all copies or substantial portions of the Software.
|
// included in all copies or substantial portions of the Software.
|
||||||
//
|
//
|
||||||
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
||||||
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
|
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
|
||||||
// OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
|
// OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
|
||||||
// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
|
// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
|
||||||
// HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
|
// HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
|
||||||
// WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
// WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
|
||||||
// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
|
// FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
|
||||||
// OTHER DEALINGS IN THE SOFTWARE.
|
// OTHER DEALINGS IN THE SOFTWARE.
|
||||||
|
|
||||||
#include <nvcore/Debug.h>
|
#include <nvcore/Debug.h>
|
||||||
#include <nvcore/Library.h>
|
#include <nvcore/Library.h>
|
||||||
#include "CudaUtils.h"
|
#include "CudaUtils.h"
|
||||||
|
|
||||||
#if defined HAVE_CUDA
|
#if defined HAVE_CUDA
|
||||||
#include <cuda.h>
|
#include <cuda.h>
|
||||||
#include <cuda_runtime_api.h>
|
#include <cuda_runtime_api.h>
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
using namespace nv;
|
using namespace nv;
|
||||||
using namespace cuda;
|
using namespace cuda;
|
||||||
|
|
||||||
/* @@ Move this to win32 utils or somewhere else.
|
/* @@ Move this to win32 utils or somewhere else.
|
||||||
#if NV_OS_WIN32
|
#if NV_OS_WIN32
|
||||||
|
|
||||||
#define WINDOWS_LEAN_AND_MEAN
|
#define WINDOWS_LEAN_AND_MEAN
|
||||||
#include <windows.h>
|
#include <windows.h>
|
||||||
|
|
||||||
static bool isWindowsVista()
|
static bool isWindowsVista()
|
||||||
{
|
{
|
||||||
OSVERSIONINFO osvi;
|
OSVERSIONINFO osvi;
|
||||||
osvi.dwOSVersionInfoSize = sizeof(OSVERSIONINFO);
|
osvi.dwOSVersionInfoSize = sizeof(OSVERSIONINFO);
|
||||||
|
|
||||||
::GetVersionEx(&osvi);
|
::GetVersionEx(&osvi);
|
||||||
return osvi.dwMajorVersion >= 6;
|
return osvi.dwMajorVersion >= 6;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
typedef BOOL (WINAPI *LPFN_ISWOW64PROCESS) (HANDLE, PBOOL);
|
typedef BOOL (WINAPI *LPFN_ISWOW64PROCESS) (HANDLE, PBOOL);
|
||||||
|
|
||||||
static bool isWow32()
|
static bool isWow32()
|
||||||
{
|
{
|
||||||
LPFN_ISWOW64PROCESS fnIsWow64Process = (LPFN_ISWOW64PROCESS)GetProcAddress(GetModuleHandle("kernel32"), "IsWow64Process");
|
LPFN_ISWOW64PROCESS fnIsWow64Process = (LPFN_ISWOW64PROCESS)GetProcAddress(GetModuleHandle("kernel32"), "IsWow64Process");
|
||||||
|
|
||||||
BOOL bIsWow64 = FALSE;
|
BOOL bIsWow64 = FALSE;
|
||||||
|
|
||||||
if (NULL != fnIsWow64Process)
|
if (NULL != fnIsWow64Process)
|
||||||
{
|
{
|
||||||
if (!fnIsWow64Process(GetCurrentProcess(), &bIsWow64))
|
if (!fnIsWow64Process(GetCurrentProcess(), &bIsWow64))
|
||||||
{
|
{
|
||||||
// Assume 32 bits.
|
// Assume 32 bits.
|
||||||
return true;
|
return true;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
return !bIsWow64;
|
return !bIsWow64;
|
||||||
}
|
}
|
||||||
|
|
||||||
#endif
|
#endif
|
||||||
*/
|
*/
|
||||||
|
|
||||||
|
|
||||||
static bool isCudaDriverAvailable(int version)
|
static bool isCudaDriverAvailable(int version)
|
||||||
{
|
{
|
||||||
#if defined HAVE_CUDA
|
#if defined HAVE_CUDA
|
||||||
#if NV_OS_WIN32
|
#if NV_OS_WIN32
|
||||||
Library nvcuda("nvcuda.dll");
|
Library nvcuda("nvcuda.dll");
|
||||||
#else
|
#else
|
||||||
Library nvcuda(NV_LIBRARY_NAME(cuda));
|
Library nvcuda(NV_LIBRARY_NAME(cuda));
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
if (!nvcuda.isValid())
|
if (!nvcuda.isValid())
|
||||||
{
|
{
|
||||||
nvDebug("*** CUDA driver not found.\n");
|
nvDebug("*** CUDA driver not found.\n");
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
if (version >= 2000)
|
if (version >= 2000)
|
||||||
{
|
{
|
||||||
void * address = nvcuda.bindSymbol("cuStreamCreate");
|
void * address = nvcuda.bindSymbol("cuStreamCreate");
|
||||||
if (address == NULL) {
|
if (address == NULL) {
|
||||||
nvDebug("*** CUDA driver version < 2.0.\n");
|
nvDebug("*** CUDA driver version < 2.0.\n");
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
if (version >= 2010)
|
if (version >= 2010)
|
||||||
{
|
{
|
||||||
void * address = nvcuda.bindSymbol("cuModuleLoadDataEx");
|
void * address = nvcuda.bindSymbol("cuModuleLoadDataEx");
|
||||||
if (address == NULL) {
|
if (address == NULL) {
|
||||||
nvDebug("*** CUDA driver version < 2.1.\n");
|
nvDebug("*** CUDA driver version < 2.1.\n");
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
if (version >= 2020)
|
if (version >= 2020)
|
||||||
{
|
{
|
||||||
typedef CUresult (CUDAAPI * PFCU_DRIVERGETVERSION)(int * version);
|
typedef CUresult (CUDAAPI * PFCU_DRIVERGETVERSION)(int * version);
|
||||||
|
|
||||||
PFCU_DRIVERGETVERSION driverGetVersion = (PFCU_DRIVERGETVERSION)nvcuda.bindSymbol("cuDriverGetVersion");
|
PFCU_DRIVERGETVERSION driverGetVersion = (PFCU_DRIVERGETVERSION)nvcuda.bindSymbol("cuDriverGetVersion");
|
||||||
if (driverGetVersion == NULL) {
|
if (driverGetVersion == NULL) {
|
||||||
nvDebug("*** CUDA driver version < 2.2.\n");
|
nvDebug("*** CUDA driver version < 2.2.\n");
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
int driverVersion;
|
int driverVersion;
|
||||||
CUresult err = driverGetVersion(&driverVersion);
|
CUresult err = driverGetVersion(&driverVersion);
|
||||||
if (err != CUDA_SUCCESS) {
|
if (err != CUDA_SUCCESS) {
|
||||||
nvDebug("*** Error querying driver version: '%s'.\n", cudaGetErrorString((cudaError_t)err));
|
nvDebug("*** Error querying driver version: '%s'.\n", cudaGetErrorString((cudaError_t)err));
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
return driverVersion >= version;
|
return driverVersion >= version;
|
||||||
}
|
}
|
||||||
#endif // HAVE_CUDA
|
#endif // HAVE_CUDA
|
||||||
|
|
||||||
return true;
|
return true;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
/// Determine if CUDA is available.
|
/// Determine if CUDA is available.
|
||||||
bool nv::cuda::isHardwarePresent()
|
bool nv::cuda::isHardwarePresent()
|
||||||
{
|
{
|
||||||
#if defined HAVE_CUDA
|
#if defined HAVE_CUDA
|
||||||
// Make sure that CUDA driver matches CUDA runtime.
|
// Make sure that CUDA driver matches CUDA runtime.
|
||||||
if (!isCudaDriverAvailable(CUDART_VERSION))
|
if (!isCudaDriverAvailable(CUDART_VERSION))
|
||||||
{
|
{
|
||||||
nvDebug("CUDA driver not available for CUDA runtime %d\n", CUDART_VERSION);
|
nvDebug("CUDA driver not available for CUDA runtime %d\n", CUDART_VERSION);
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
int count = deviceCount();
|
int count = deviceCount();
|
||||||
if (count == 1)
|
if (count == 1)
|
||||||
{
|
{
|
||||||
// Make sure it's not an emulation device.
|
// Make sure it's not an emulation device.
|
||||||
cudaDeviceProp deviceProp;
|
cudaDeviceProp deviceProp;
|
||||||
cudaGetDeviceProperties(&deviceProp, 0);
|
cudaGetDeviceProperties(&deviceProp, 0);
|
||||||
|
|
||||||
// deviceProp.name != Device Emulation (CPU)
|
// deviceProp.name != Device Emulation (CPU)
|
||||||
if (deviceProp.major == -1 || deviceProp.minor == -1)
|
if (deviceProp.major == -1 || deviceProp.minor == -1)
|
||||||
{
|
{
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
// @@ Make sure that warp size == 32
|
// @@ Make sure that warp size == 32
|
||||||
|
|
||||||
return count > 0;
|
// @@ Make sure available GPU is faster than the CPU.
|
||||||
#else
|
|
||||||
return false;
|
return count > 0;
|
||||||
#endif
|
#else
|
||||||
}
|
return false;
|
||||||
|
#endif
|
||||||
/// Get number of CUDA enabled devices.
|
}
|
||||||
int nv::cuda::deviceCount()
|
|
||||||
{
|
/// Get number of CUDA enabled devices.
|
||||||
#if defined HAVE_CUDA
|
int nv::cuda::deviceCount()
|
||||||
int gpuCount = 0;
|
{
|
||||||
|
#if defined HAVE_CUDA
|
||||||
cudaError_t result = cudaGetDeviceCount(&gpuCount);
|
int gpuCount = 0;
|
||||||
|
|
||||||
if (result == cudaSuccess)
|
cudaError_t result = cudaGetDeviceCount(&gpuCount);
|
||||||
{
|
|
||||||
return gpuCount;
|
if (result == cudaSuccess)
|
||||||
}
|
{
|
||||||
#endif
|
return gpuCount;
|
||||||
return 0;
|
}
|
||||||
}
|
#endif
|
||||||
|
return 0;
|
||||||
int nv::cuda::getFastestDevice()
|
}
|
||||||
{
|
|
||||||
int max_gflops_device = 0;
|
|
||||||
#if defined HAVE_CUDA
|
// Make sure device meets requirements:
|
||||||
int max_gflops = 0;
|
// - Not an emulation device.
|
||||||
|
// - Not an integrated device?
|
||||||
const int device_count = deviceCount();
|
// - Faster than CPU.
|
||||||
int current_device = 0;
|
bool nv::cuda::isValidDevice(int i)
|
||||||
while (current_device < device_count)
|
{
|
||||||
{
|
#if defined HAVE_CUDA
|
||||||
cudaDeviceProp device_properties;
|
|
||||||
cudaGetDeviceProperties(&device_properties, current_device);
|
cudaDeviceProp device_properties;
|
||||||
int gflops = device_properties.multiProcessorCount * device_properties.clockRate;
|
cudaGetDeviceProperties(&device_properties, i);
|
||||||
|
int gflops = device_properties.multiProcessorCount * device_properties.clockRate;
|
||||||
if (device_properties.major != -1 && device_properties.minor != -1)
|
|
||||||
{
|
if (device_properties.major == -1 || device_properties.minor == -1) {
|
||||||
if( gflops > max_gflops )
|
// Emulation device.
|
||||||
{
|
return false;
|
||||||
max_gflops = gflops;
|
}
|
||||||
max_gflops_device = current_device;
|
|
||||||
}
|
#if CUDART_VERSION >= 2030 // 2.3
|
||||||
}
|
/*if (device_properties.integrated)
|
||||||
|
{
|
||||||
current_device++;
|
// Integrated devices.
|
||||||
}
|
return false;
|
||||||
#endif
|
}*/
|
||||||
return max_gflops_device;
|
#endif
|
||||||
}
|
|
||||||
|
return true;
|
||||||
|
#else
|
||||||
/// Activate the given devices.
|
return false;
|
||||||
bool nv::cuda::setDevice(int i)
|
#endif
|
||||||
{
|
}
|
||||||
nvCheck(i < deviceCount());
|
|
||||||
#if defined HAVE_CUDA
|
int nv::cuda::getFastestDevice()
|
||||||
cudaError_t result = cudaSetDevice(i);
|
{
|
||||||
|
int max_gflops_device = -1;
|
||||||
if (result != cudaSuccess) {
|
#if defined HAVE_CUDA
|
||||||
nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
|
int max_gflops = 0;
|
||||||
}
|
|
||||||
|
const int device_count = deviceCount();
|
||||||
return result == cudaSuccess;
|
for (int i = 0; i < device_count; i++)
|
||||||
#else
|
{
|
||||||
return false;
|
if (isValidDevice(i))
|
||||||
#endif
|
{
|
||||||
}
|
cudaDeviceProp device_properties;
|
||||||
|
cudaGetDeviceProperties(&device_properties, i);
|
||||||
void nv::cuda::exit()
|
int gflops = device_properties.multiProcessorCount * device_properties.clockRate;
|
||||||
{
|
|
||||||
#if defined HAVE_CUDA
|
if (gflops > max_gflops)
|
||||||
cudaError_t result = cudaThreadExit();
|
{
|
||||||
|
max_gflops = gflops;
|
||||||
if (result != cudaSuccess) {
|
max_gflops_device = i;
|
||||||
nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
|
}
|
||||||
}
|
}
|
||||||
#endif
|
}
|
||||||
}
|
#endif
|
||||||
|
return max_gflops_device;
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
|
/// Activate the given devices.
|
||||||
|
bool nv::cuda::initDevice(int * device_ptr)
|
||||||
|
{
|
||||||
|
nvDebugCheck(device_ptr != NULL);
|
||||||
|
#if defined HAVE_CUDA
|
||||||
|
|
||||||
|
#if CUDART_VERSION >= 2030 // 2.3
|
||||||
|
|
||||||
|
// Set device flags to yield in order to play nice with other threads and to find out if CUDA was already active.
|
||||||
|
cudaError_t resul = cudaSetDeviceFlags(cudaDeviceScheduleYield);
|
||||||
|
|
||||||
|
#endif
|
||||||
|
|
||||||
|
int device = getFastestDevice();
|
||||||
|
|
||||||
|
if (device == -1)
|
||||||
|
{
|
||||||
|
// No device is fast enough.
|
||||||
|
*device_ptr = -1;
|
||||||
|
return false;
|
||||||
|
}
|
||||||
|
|
||||||
|
// Select CUDA device.
|
||||||
|
cudaError_t result = cudaSetDevice(device);
|
||||||
|
|
||||||
|
if (result == cudaErrorSetOnActiveProcess)
|
||||||
|
{
|
||||||
|
int device;
|
||||||
|
result = cudaGetDevice(&device);
|
||||||
|
|
||||||
|
*device_ptr = -1; // No device to cleanup.
|
||||||
|
return isValidDevice(device); // Return true if device is valid.
|
||||||
|
}
|
||||||
|
else if (result != cudaSuccess)
|
||||||
|
{
|
||||||
|
nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
|
||||||
|
*device_ptr = -1;
|
||||||
|
return false;
|
||||||
|
}
|
||||||
|
|
||||||
|
*device_ptr = device;
|
||||||
|
return true;
|
||||||
|
#else
|
||||||
|
return false;
|
||||||
|
#endif
|
||||||
|
}
|
||||||
|
|
||||||
|
void nv::cuda::exitDevice()
|
||||||
|
{
|
||||||
|
#if defined HAVE_CUDA
|
||||||
|
cudaError_t result = cudaThreadExit();
|
||||||
|
|
||||||
|
if (result != cudaSuccess) {
|
||||||
|
nvDebug("*** CUDA Error: %s\n", cudaGetErrorString(result));
|
||||||
|
}
|
||||||
|
#endif
|
||||||
|
}
|
||||||
|
|||||||
Loading…
Reference in New Issue