int cuCompilerVersion()
{
string path = cuCompilerPath();
if(path == "")
return 0;
/* get --version output */
FILE *pipe = popen((path + " --version").c_str(), "r");
if(!pipe) {
fprintf(stderr, "CUDA: failed to run compiler to retrieve version");
return 0;
}
char buf[128];
string output = "";
while(!feof(pipe))
if(fgets(buf, 128, pipe) != NULL)
output += buf;
pclose(pipe);
/* parse version number */
string marker = "Cuda compilation tools, release ";
size_t offset = output.find(marker);
if(offset == string::npos) {
fprintf(stderr, "CUDA: failed to find version number in:\n\n%s\n", output.c_str());
return 0;
}
string versionstr = output.substr(offset + marker.size(), string::npos);
int major, minor;
if(sscanf(versionstr.c_str(), "%d.%d", &major, &minor) < 2) {
fprintf(stderr, "CUDA: failed to parse version number from:\n\n%s\n", output.c_str());
return 0;
}
return 10*major + minor;
}
string compile_kernel()
{
/* compute cubin name */
int major, minor;
cuDeviceComputeCapability(&major, &minor, cuDevId);
/* attempt to use kernel provided with blender */
string cubin = path_get(string_printf("lib/kernel_sm_%d%d.cubin", major, minor));
if(path_exists(cubin))
return cubin;
/* not found, try to use locally compiled kernel */
string kernel_path = path_get("kernel");
string md5 = path_files_md5_hash(kernel_path);
cubin = string_printf("cycles_kernel_sm%d%d_%s.cubin", major, minor, md5.c_str());
cubin = path_user_get(path_join("cache", cubin));
/* if exists already, use it */
if(path_exists(cubin))
return cubin;
#ifdef _WIN32
if(cuHavePrecompiledKernels()) {
if(major < 2)
cuda_error_message(string_printf("CUDA device requires compute capability 2.0 or up, found %d.%d. Your GPU is not supported.", major, minor));
else
cuda_error_message(string_printf("CUDA binary kernel for this graphics card compute capability (%d.%d) not found.", major, minor));
return "";
}
#endif
/* if not, find CUDA compiler */
string nvcc = cuCompilerPath();
if(nvcc == "") {
cuda_error_message("CUDA nvcc compiler not found. Install CUDA toolkit in default location.");
return "";
}
/* compile */
string kernel = path_join(kernel_path, "kernel.cu");
string include = kernel_path;
const int machine = system_cpu_bits();
const int maxreg = 24;
double starttime = time_dt();
printf("Compiling CUDA kernel ...\n");
path_create_directories(cubin);
string command = string_printf("\"%s\" -arch=sm_%d%d -m%d --cubin \"%s\" "
"-o \"%s\" --ptxas-options=\"-v\" --maxrregcount=%d --opencc-options -OPT:Olimit=0 -I\"%s\" -DNVCC",
nvcc.c_str(), major, minor, machine, kernel.c_str(), cubin.c_str(), maxreg, include.c_str());
if(system(command.c_str()) == -1) {
cuda_error_message("Failed to execute compilation command, see console for details.");
return "";
}
/* verify if compilation succeeded */
if(!path_exists(cubin)) {
cuda_error_message("CUDA kernel compilation failed, see console for details.");
return "";
}
printf("Kernel compilation finished in %.2lfs.\n", time_dt() - starttime);
return cubin;
}
CCL_NAMESPACE_BEGIN
/* utility macros */
#define CUDA_LIBRARY_FIND_CHECKED(name) \
name = (t##name*)dynamic_library_find(lib, #name);
#define CUDA_LIBRARY_FIND(name) \
name = (t##name*)dynamic_library_find(lib, #name); \
assert(name);
#define CUDA_LIBRARY_FIND_V2(name) \
name = (t##name*)dynamic_library_find(lib, #name "_v2"); \
assert(name);
/* initialization function */
bool cuLibraryInit()
{
static bool initialized = false;
static bool result = false;
if(initialized)
return result;
initialized = true;
/* library paths */
#ifdef _WIN32
/* expected in c:/windows/system or similar, no path needed */
const char *path = "nvcuda.dll";
#elif defined(__APPLE__)
/* default installation path */
const char *path = "/usr/local/cuda/lib/libcuda.dylib";
#else
const char *path = "libcuda.so";
#endif
/* load library */
DynamicLibrary *lib = dynamic_library_open(path);
if(lib == NULL)
return false;
/* detect driver version */
int driver_version = 1000;
CUDA_LIBRARY_FIND_CHECKED(cuDriverGetVersion);
if(cuDriverGetVersion)
cuDriverGetVersion(&driver_version);
/* we require version 4.0 */
if(driver_version < 4000)
return false;
/* fetch all function pointers */
CUDA_LIBRARY_FIND(cuInit);
CUDA_LIBRARY_FIND(cuDeviceGet);
CUDA_LIBRARY_FIND(cuDeviceGetCount);
CUDA_LIBRARY_FIND(cuDeviceGetName);
CUDA_LIBRARY_FIND(cuDeviceComputeCapability);
CUDA_LIBRARY_FIND(cuDeviceTotalMem);
CUDA_LIBRARY_FIND(cuDeviceGetProperties);
CUDA_LIBRARY_FIND(cuDeviceGetAttribute);
CUDA_LIBRARY_FIND(cuCtxCreate);
CUDA_LIBRARY_FIND(cuCtxDestroy);
CUDA_LIBRARY_FIND(cuCtxAttach);
CUDA_LIBRARY_FIND(cuCtxDetach);
CUDA_LIBRARY_FIND(cuCtxPushCurrent);
CUDA_LIBRARY_FIND(cuCtxPopCurrent);
CUDA_LIBRARY_FIND(cuCtxGetDevice);
CUDA_LIBRARY_FIND(cuCtxSynchronize);
CUDA_LIBRARY_FIND(cuModuleLoad);
CUDA_LIBRARY_FIND(cuModuleLoadData);
CUDA_LIBRARY_FIND(cuModuleUnload);
CUDA_LIBRARY_FIND(cuModuleGetFunction);
CUDA_LIBRARY_FIND(cuModuleGetGlobal);
CUDA_LIBRARY_FIND(cuModuleGetTexRef);
CUDA_LIBRARY_FIND(cuMemGetInfo);
CUDA_LIBRARY_FIND(cuMemAlloc);
CUDA_LIBRARY_FIND(cuMemAllocPitch);
CUDA_LIBRARY_FIND(cuMemFree);
CUDA_LIBRARY_FIND(cuMemGetAddressRange);
CUDA_LIBRARY_FIND(cuMemAllocHost);
CUDA_LIBRARY_FIND(cuMemFreeHost);
CUDA_LIBRARY_FIND(cuMemHostAlloc);
CUDA_LIBRARY_FIND(cuMemHostGetDevicePointer);
CUDA_LIBRARY_FIND(cuMemcpyHtoD);
CUDA_LIBRARY_FIND(cuMemcpyDtoH);
CUDA_LIBRARY_FIND(cuMemcpyDtoD);
CUDA_LIBRARY_FIND(cuMemcpyDtoA);
CUDA_LIBRARY_FIND(cuMemcpyAtoD);
CUDA_LIBRARY_FIND(cuMemcpyHtoA);
CUDA_LIBRARY_FIND(cuMemcpyAtoH);
CUDA_LIBRARY_FIND(cuMemcpyAtoA);
CUDA_LIBRARY_FIND(cuMemcpy2D);
CUDA_LIBRARY_FIND(cuMemcpy2DUnaligned);
CUDA_LIBRARY_FIND(cuMemcpy3D);
CUDA_LIBRARY_FIND(cuMemcpyHtoDAsync);
CUDA_LIBRARY_FIND(cuMemcpyDtoHAsync);
CUDA_LIBRARY_FIND(cuMemcpyHtoAAsync);
CUDA_LIBRARY_FIND(cuMemcpyAtoHAsync);
CUDA_LIBRARY_FIND(cuMemcpy2DAsync);
CUDA_LIBRARY_FIND(cuMemcpy3DAsync);
CUDA_LIBRARY_FIND(cuMemsetD8);
CUDA_LIBRARY_FIND(cuMemsetD16);
CUDA_LIBRARY_FIND(cuMemsetD32);
CUDA_LIBRARY_FIND(cuMemsetD2D8);
CUDA_LIBRARY_FIND(cuMemsetD2D16);
CUDA_LIBRARY_FIND(cuMemsetD2D32);
CUDA_LIBRARY_FIND(cuFuncSetBlockShape);
CUDA_LIBRARY_FIND(cuFuncSetSharedSize);
CUDA_LIBRARY_FIND(cuFuncGetAttribute);
CUDA_LIBRARY_FIND(cuArrayCreate);
CUDA_LIBRARY_FIND(cuArrayGetDescriptor);
CUDA_LIBRARY_FIND(cuArrayDestroy);
CUDA_LIBRARY_FIND(cuArray3DCreate);
CUDA_LIBRARY_FIND(cuArray3DGetDescriptor);
CUDA_LIBRARY_FIND(cuTexRefCreate);
CUDA_LIBRARY_FIND(cuTexRefDestroy);
CUDA_LIBRARY_FIND(cuTexRefSetArray);
CUDA_LIBRARY_FIND(cuTexRefSetAddress);
CUDA_LIBRARY_FIND(cuTexRefSetAddress2D);
CUDA_LIBRARY_FIND(cuTexRefSetFormat);
CUDA_LIBRARY_FIND(cuTexRefSetAddressMode);
CUDA_LIBRARY_FIND(cuTexRefSetFilterMode);
CUDA_LIBRARY_FIND(cuTexRefSetFlags);
CUDA_LIBRARY_FIND(cuTexRefGetAddress);
CUDA_LIBRARY_FIND(cuTexRefGetArray);
CUDA_LIBRARY_FIND(cuTexRefGetAddressMode);
CUDA_LIBRARY_FIND(cuTexRefGetFilterMode);
CUDA_LIBRARY_FIND(cuTexRefGetFormat);
CUDA_LIBRARY_FIND(cuTexRefGetFlags);
CUDA_LIBRARY_FIND(cuParamSetSize);
CUDA_LIBRARY_FIND(cuParamSeti);
CUDA_LIBRARY_FIND(cuParamSetf);
CUDA_LIBRARY_FIND(cuParamSetv);
CUDA_LIBRARY_FIND(cuParamSetTexRef);
CUDA_LIBRARY_FIND(cuLaunch);
CUDA_LIBRARY_FIND(cuLaunchGrid);
CUDA_LIBRARY_FIND(cuLaunchGridAsync);
CUDA_LIBRARY_FIND(cuEventCreate);
CUDA_LIBRARY_FIND(cuEventRecord);
CUDA_LIBRARY_FIND(cuEventQuery);
CUDA_LIBRARY_FIND(cuEventSynchronize);
CUDA_LIBRARY_FIND(cuEventDestroy);
CUDA_LIBRARY_FIND(cuEventElapsedTime);
CUDA_LIBRARY_FIND(cuStreamCreate);
CUDA_LIBRARY_FIND(cuStreamQuery);
CUDA_LIBRARY_FIND(cuStreamSynchronize);
CUDA_LIBRARY_FIND(cuStreamDestroy);
/* cuda 2.1 */
CUDA_LIBRARY_FIND(cuModuleLoadDataEx);
CUDA_LIBRARY_FIND(cuModuleLoadFatBinary);
CUDA_LIBRARY_FIND(cuGLCtxCreate);
CUDA_LIBRARY_FIND(cuGraphicsGLRegisterBuffer);
CUDA_LIBRARY_FIND(cuGraphicsGLRegisterImage);
/* cuda 2.3 */
CUDA_LIBRARY_FIND(cuMemHostGetFlags);
CUDA_LIBRARY_FIND(cuGraphicsGLRegisterBuffer);
CUDA_LIBRARY_FIND(cuGraphicsGLRegisterImage);
/* cuda 3.0 */
CUDA_LIBRARY_FIND(cuMemcpyDtoDAsync);
CUDA_LIBRARY_FIND(cuFuncSetCacheConfig);
CUDA_LIBRARY_FIND(cuGraphicsUnregisterResource);
CUDA_LIBRARY_FIND(cuGraphicsSubResourceGetMappedArray);
CUDA_LIBRARY_FIND(cuGraphicsResourceGetMappedPointer);
CUDA_LIBRARY_FIND(cuGraphicsResourceSetMapFlags);
CUDA_LIBRARY_FIND(cuGraphicsMapResources);
CUDA_LIBRARY_FIND(cuGraphicsUnmapResources);
CUDA_LIBRARY_FIND(cuGetExportTable);
/* cuda 3.1 */
CUDA_LIBRARY_FIND(cuModuleGetSurfRef);
CUDA_LIBRARY_FIND(cuSurfRefSetArray);
CUDA_LIBRARY_FIND(cuSurfRefGetArray);
CUDA_LIBRARY_FIND(cuCtxSetLimit);
CUDA_LIBRARY_FIND(cuCtxGetLimit);
/* functions which changed 3.1 -> 3.2 for 64 bit stuff, the cuda library
* has both the old ones for compatibility and new ones with _v2 postfix,
* we load the _v2 ones here. */
CUDA_LIBRARY_FIND_V2(cuDeviceTotalMem);
CUDA_LIBRARY_FIND_V2(cuCtxCreate);
CUDA_LIBRARY_FIND_V2(cuModuleGetGlobal);
CUDA_LIBRARY_FIND_V2(cuMemGetInfo);
CUDA_LIBRARY_FIND_V2(cuMemAlloc);
CUDA_LIBRARY_FIND_V2(cuMemAllocPitch);
CUDA_LIBRARY_FIND_V2(cuMemFree);
CUDA_LIBRARY_FIND_V2(cuMemGetAddressRange);
CUDA_LIBRARY_FIND_V2(cuMemAllocHost);
CUDA_LIBRARY_FIND_V2(cuMemHostGetDevicePointer);
CUDA_LIBRARY_FIND_V2(cuMemcpyHtoD);
CUDA_LIBRARY_FIND_V2(cuMemcpyDtoH);
CUDA_LIBRARY_FIND_V2(cuMemcpyDtoD);
CUDA_LIBRARY_FIND_V2(cuMemcpyDtoA);
CUDA_LIBRARY_FIND_V2(cuMemcpyAtoD);
CUDA_LIBRARY_FIND_V2(cuMemcpyHtoA);
CUDA_LIBRARY_FIND_V2(cuMemcpyAtoH);
CUDA_LIBRARY_FIND_V2(cuMemcpyAtoA);
CUDA_LIBRARY_FIND_V2(cuMemcpyHtoAAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpyAtoHAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpy2D);
CUDA_LIBRARY_FIND_V2(cuMemcpy2DUnaligned);
CUDA_LIBRARY_FIND_V2(cuMemcpy3D);
CUDA_LIBRARY_FIND_V2(cuMemcpyHtoDAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpyDtoHAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpyDtoDAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpy2DAsync);
CUDA_LIBRARY_FIND_V2(cuMemcpy3DAsync);
CUDA_LIBRARY_FIND_V2(cuMemsetD8);
CUDA_LIBRARY_FIND_V2(cuMemsetD16);
CUDA_LIBRARY_FIND_V2(cuMemsetD32);
CUDA_LIBRARY_FIND_V2(cuMemsetD2D8);
CUDA_LIBRARY_FIND_V2(cuMemsetD2D16);
CUDA_LIBRARY_FIND_V2(cuMemsetD2D32);
CUDA_LIBRARY_FIND_V2(cuArrayCreate);
CUDA_LIBRARY_FIND_V2(cuArrayGetDescriptor);
CUDA_LIBRARY_FIND_V2(cuArray3DCreate);
CUDA_LIBRARY_FIND_V2(cuArray3DGetDescriptor);
CUDA_LIBRARY_FIND_V2(cuTexRefSetAddress);
CUDA_LIBRARY_FIND_V2(cuTexRefSetAddress2D);
CUDA_LIBRARY_FIND_V2(cuTexRefGetAddress);
CUDA_LIBRARY_FIND_V2(cuGraphicsResourceGetMappedPointer);
CUDA_LIBRARY_FIND_V2(cuGLCtxCreate);
/* cuda 4.0 */
CUDA_LIBRARY_FIND(cuCtxSetCurrent);
if(cuHavePrecompiledKernels())
result = true;
#ifndef _WIN32
else if(cuCompilerPath() != "")
result = true;
#endif
return result;
}
