path: root/matrix/generic/cumatrix.c

#ifdef NERV_GENERIC_CUMATRIX
#include "matrix.h"
#include "elem_type.h"

#define MATRIX_DATA_FREE(ptr) cuda_matrix_(free)(ptr)
#define MATRIX_DATA_ALLOC(dptr, stride, width, height) \
                            cuda_matrix_(alloc)(dptr, stride, width, height)
#define MATRIX_DATA_WRITE(data, idx, val) cuda_matrix_(write)(data, idx, val)
#define MATRIX_DATA_READ(data, idx) cuda_matrix_(read)(data, idx)
#define MATRIX_INIT(L) cuda_matrix_(init)(L)
#define MATRIX_BASE_TNAME nerv_matrix_cuda_tname
#define NERV_GENERIC_MATRIX
#define NERV_GENERIC_CUKERNEL
#include "../../common.h"
#include "../cukernel.h"
#include "cuda.h"
#include "cuda_runtime.h"
#include "driver_types.h"
#include "cublas_v2.h"

#define CHECK_SAME_DIMENSION(a, b) \
    do { \
        if (!(a->nrow == b->nrow && a->ncol == b->ncol)) \
            nerv_error(L, "Matrices should be of the same dimension"); \
    } while (0)

static cublasHandle_t cublas_handle;

Matrix *nerv_matrix_(new_)(long nrow, long ncol);
void nerv_matrix_(data_free)(Matrix *self);

static void nerv_matrix_(add_)(const Matrix *a, const Matrix *b,
                                const Matrix *c,
                                MATRIX_ELEM alpha, MATRIX_ELEM beta) {
    NERV_CUBLAS_(geam)(cublas_handle, CUBLAS_OP_N, CUBLAS_OP_N,
                a->ncol, a->nrow,
                &alpha,
                MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
                &beta,
                MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM),
                MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM));
}

static int nerv_matrix_(add)(lua_State *L) {
    Matrix *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
    Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(