Class DeviceDnVecFloat

XM/include/Utils/memory.h:964–1013

dense vector wrapper on device: float type

// dense vector wrapper on device: float type
class DeviceDnVecFloat {
    public:
        int gpu_id;
        int size;
        float* vals;
        cusparseDnVecDescr_t cusparse_descr;

        DeviceDnVecFloat(): gpu_id(0), size(0), vals(nullptr), cusparse_descr(NULL) {}
        DeviceDnVecFloat(const int gpu_id, const int size, bool as_byte = false):
            gpu_id(gpu_id), size(size), vals(nullptr), cusparse_descr(NULL) {
            this->allocate(this->gpu_id, this->size, as_byte);
        }

        inline void allocate(const int gpu_id, const int size, bool as_byte = false) {
            if (this->vals == nullptr) {
                this->gpu_id = gpu_id;
                CHECK_CUDA( cudaSetDevice(this->gpu_id) );
                // as_byte is used to allocate buffer size, which is usually given in terms of bytes
                if (as_byte) {
                    this->size = (size + sizeof(float) - 1) / sizeof(float);
                } else {
                    this->size = size;
                }
                CHECK_CUDA( cudaMalloc((void**) &this->vals, sizeof(float) * this->size) );
                CHECK_CUSPARSE( cusparseCreateDnVec(&this->cusparse_descr, this->size, this->vals, CUDA_R_32F) );
            }
            return;
        }
        inline float get_norm(const DeviceBlasHandle& cublas_H) {
            float norm;
            CHECK_CUDA( cudaSetDevice(this->gpu_id) );
            CHECK_CUBLAS( cublasSnrm2_v2(
                cublas_H.cublas_handle, this->size, this->vals, 1, &norm
            ) );
            return norm;
        }

        ~DeviceDnVecFloat() {
            CHECK_CUDA( cudaSetDevice(this->gpu_id) );
            if (this->vals != nullptr) {
                CHECK_CUDA( cudaFree(this->vals) );
                this->vals = nullptr;
            }
            if (this->cusparse_descr != NULL) {
                CHECK_CUSPARSE( cusparseDestroyDnVec(this->cusparse_descr) );
                this->cusparse_descr = NULL;
            }
            // std::cout << "DeviceDnVecFloat destructor called!" << std::endl;
        }
};

// dense vector wrapper on device: int type
class DeviceDnVecInt {