#define NERV_GENERIC_CUMATRIX
#include "../lib/common.h"
#include "../lib/matrix/cumatrix.h"
#include "../lib/matrix/cuda_helper.h"
#include <string.h>

const char *nerv_cuda_context_tname = "nerv.CuContext";

int nerv_cuda_context_lua_select_gpu(lua_State *L) {
    Status status;
    nerv_cuda_context_select_gpu(luaT_checkudata(L, 1, nerv_cuda_context_tname),
                            luaL_checkinteger(L, 2), &status);
    NERV_LUA_CHECK_STATUS(L, status);
    return 0;
}

int nerv_cuda_context_lua_print_profile(lua_State *L) {
    nerv_cuda_context_print_profile(luaT_checkudata(L, 1, nerv_cuda_context_tname));
    return 0;
}

int nerv_cuda_context_lua_clear_profile(lua_State *L) {
    nerv_cuda_context_clear_profile(luaT_checkudata(L, 1, nerv_cuda_context_tname));
    return 0;
}

int nerv_cuda_context_lua_new(lua_State *L) {
    Status status;
    int dev = lua_gettop(L) > 0 ? luaL_checkinteger(L, 1) : -1;
    CuContext *self = nerv_cuda_context_create(dev, &status);
    NERV_LUA_CHECK_STATUS(L, status);
    luaT_pushudata(L, self, nerv_cuda_context_tname);
    return 1;
}

int nerv_cuda_context_lua_destroy(lua_State *L) {
    Status status;
    CuContext *self = luaT_checkudata(L, 1, nerv_cuda_context_tname);
    nerv_cuda_context_destroy(self, &status);
    NERV_LUA_CHECK_STATUS(L, status);
    return 1;
}

static const luaL_Reg nerv_cuda_context_methods[] = {
    {"print_profile", nerv_cuda_context_lua_print_profile},
    {"clear_profile", nerv_cuda_context_lua_clear_profile},
    {"select_gpu", nerv_cuda_context_lua_select_gpu},
    {NULL, NULL}
};

void nerv_cuda_context_lua_init(lua_State *L) {
    luaT_newmetatable(L, nerv_cuda_context_tname, NULL,
                        nerv_cuda_context_lua_new,
                        nerv_cuda_context_lua_destroy, NULL);
    luaL_register(L, NULL, nerv_cuda_context_methods);
}

extern void nerv_matrix_cuda_float_lua_init(lua_State *L);
extern void nerv_matrix_cuda_double_lua_init(lua_State *L);

static const luaL_Reg cumatrix_methods[] = {
    {NULL, NULL}
};

void nerv_lua_cumatrix_init(lua_State *L) {
    luaL_register(L, NULL, cumatrix_methods);
    nerv_cuda_context_lua_init(L);
    nerv_matrix_cuda_float_lua_init(L);
    nerv_matrix_cuda_double_lua_init(L);
}

#define MATRIX_CONTEXT CuContext
#define MATRIX_CONTEXT_TNAME nerv_cuda_context_tname

#define MATRIX_USE_FLOAT
#define cuda_matrix_(NAME) cuda_matrix_float_##NAME
#define nerv_matrix_(NAME) nerv_matrix_cuda_float_##NAME
#define cudak_(NAME) cudak_float_ ## NAME
#define NERV_CUBLAS_(NAME) cublasS##NAME
#define MATRIX_CUMATRIX_HOST_TNAME nerv_matrix_host_float_tname
const char *nerv_matrix_(tname) = "nerv.CuMatrixFloat";
#include "generic/cumatrix.c"
#undef NERV_CUBLAS_
#undef cudak_
#undef nerv_matrix_
#undef cuda_matrix_
#undef MATRIX_USE_FLOAT
#undef MATRIX_ELEM
#undef MATRIX_ELEM_PTR
#undef MATRIX_ELEM_PTR_BASE
#undef MATRIX_ELEM_FMT
#undef MATRIX_ELEM_WRITE_FMT
#undef MATRIX_CUMATRIX_HOST_TNAME

#define MATRIX_USE_DOUBLE
#define cuda_matrix_(NAME) cuda_matrix_double_##NAME
#define nerv_matrix_(NAME) nerv_matrix_cuda_double_##NAME
#define cudak_(NAME) cudak_double_ ## NAME
#define NERV_CUBLAS_(NAME) cublasD##NAME
#define MATRIX_CUMATRIX_HOST_TNAME nerv_matrix_host_double_tname
const char *nerv_matrix_(tname) = "nerv.CuMatrixDouble";
#include "generic/cumatrix.c"