bool load_kernels(bool experimental)
{
/* verify if device was initialized */
if(!device_initialized) {
fprintf(stderr, "OpenCL: failed to initialize device.\n");
return false;
}
/* verify we have right opencl version */
if(!opencl_version_check())
return false;
/* md5 hash to detect changes */
string kernel_path = path_get("kernel");
string kernel_md5 = path_files_md5_hash(kernel_path);
string device_md5 = device_md5_hash();
/* try to use cache binary */
string clbin = string_printf("cycles_kernel_%s_%s.clbin", device_md5.c_str(), kernel_md5.c_str());;
clbin = path_user_get(path_join("cache", clbin));
if(path_exists(clbin)) {
/* if exists already, try use it */
if(!load_binary(kernel_path, clbin))
return false;
}
else {
/* compile kernel */
if(!compile_kernel(kernel_path, kernel_md5))
return false;
/* save binary for reuse */
save_binary(clbin);
}
/* find kernels */
ckPathTraceKernel = clCreateKernel(cpProgram, "kernel_ocl_path_trace", &ciErr);
if(opencl_error(ciErr))
return false;
ckFilmConvertKernel = clCreateKernel(cpProgram, "kernel_ocl_tonemap", &ciErr);
if(opencl_error(ciErr))
return false;
return true;
}
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;
}
