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"

static cublasHandle_t cublas_handle;

Matrix *nerv_matrix_(new_)(long nrow, long ncol);

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_(tname));
    Matrix *b = luaT_checkudata(L, 3, nerv_matrix_(tname));
    MATRIX_ELEM alpha = luaL_checknumber(L, 4); /* alpha */
    MATRIX_ELEM beta = luaL_checknumber(L, 5); /* alpha */
    if (!(a->nrow == b->nrow && a->ncol == b->ncol))
        nerv_error(L, "Matrices should be of the same dimension");
    nerv_matrix_(add_)(a, b, c, alpha, beta);
    return 0;
}

static int nerv_matrix_(get_cublas_op)(char ch) {
    return (ch == 'T' || ch == 't') ? CUBLAS_OP_T : CUBLAS_OP_N;
}

static int nerv_matrix_(mul)(lua_State *L) {
#define SWAP(a, b) \
    do { int t = (a); (a) = (b); (b) = t; } while (0)

    Matrix *c = luaT_checkudata(L, 1, nerv_matrix_(tname));
    Matrix *a = luaT_checkudata(L, 2, nerv_matrix_(tname));
    Matrix *b = luaT_checkudata(L, 3, nerv_matrix_(tname));
    MATRIX_ELEM alpha = luaL_checknumber(L, 4);
    MATRIX_ELEM beta = luaL_checknumber(L, 5);
    int nargs = lua_gettop(L);
    int ta = nargs > 5 ? nerv_matrix_(get_cublas_op)(*luaL_checkstring(L, 6)) \
                            : CUBLAS_OP_N;
    int tb = nargs > 6 ? nerv_matrix_(get_cublas_op)(*luaL_checkstring(L, 7)) \
                            : CUBLAS_OP_N;
    int am = a->nrow, an = a->ncol;
    int bm = b->nrow, bn = b->ncol;
    if (ta == CUBLAS_OP_T) SWAP(am, an);
    if (tb == CUBLAS_OP_T) SWAP(bm, bn);
    if (an != bm)
        nerv_error(L, "Wrong dimension of multipliers");
/*    MATRIX_ELEM alpha = 1.0f, beta = 0.0f; */
    NERV_CUBLAS_(gemm)(cublas_handle, tb, ta,
                bn, am, bm,
                &alpha,
                MATRIX_ELEM_PTR(b), b->stride / sizeof(MATRIX_ELEM),
                MATRIX_ELEM_PTR(a), a->stride / sizeof(MATRIX_ELEM),
                &beta,
                MATRIX_ELEM_PTR(c), c->stride / sizeof(MATRIX_ELEM));
    return 0;
}

static int nerv_matrix_(create)(lua_State *L