12void CUDADeviceKernels::load(CUDADevice *device)
14 CUmodule cuModule = device->cuModule;
17 CUDADeviceKernel &kernel = kernels_[i];
23 const std::string function_name = std::string(
"kernel_gpu_") +
25 cuda_device_assert(device,
26 cuModuleGetFunction(&kernel.function, cuModule, function_name.c_str()));
28 if (kernel.function) {
29 cuda_device_assert(device, cuFuncSetCacheConfig(kernel.function, CU_FUNC_CACHE_PREFER_L1));
33 cuOccupancyMaxPotentialBlockSize(
34 &kernel.min_blocks, &kernel.num_threads_per_block, kernel.function,
nullptr, 0, 0));
37 LOG_ERROR <<
"Unable to load kernel " << function_name;
44const CUDADeviceKernel &CUDADeviceKernels::get(
DeviceKernel kernel)
const
46 return kernels_[(int)kernel];
49bool CUDADeviceKernels::available(
DeviceKernel kernel)
const
51 return kernels_[(int)kernel].function !=
nullptr;
#define CCL_NAMESPACE_END
const char * device_kernel_as_string(DeviceKernel kernel)
bool device_kernel_has_gpu_function(DeviceKernel kernel)