1 files changed, 52 insertions, 8 deletions

diff --git a/nerv/matrix/generic/mmatrix.c b/nerv/matrix/generic/mmatrix.c
index a4e8489..93562d0 100644
--- a/nerv/matrix/generic/mmatrix.c
+++ b/nerv/matrix/generic/mmatrix.c
@@ -9,9 +9,15 @@
 #include "../../lib/common.h"
 #include "../../lib/matrix/generic/mmatrix.h"
 #include "../../io/chunk_file.h"
-#include "string.h"
+#include <string.h>
+#include <cblas.h>
 
-int nerv_matrix_(lua_get_elem)(lua_State *L) {
+#define BLAS_OP_N CblasNoTrans
+static int nerv_matrix_(lua_get_blas_op)(char ch) {
+    return (ch == 'T' || ch == 't') ? CblasTrans : CblasNoTrans;
+}
+
+static int nerv_matrix_(lua_get_elem)(lua_State *L) {
     Matrix *self = luaT_checkudata(L, 1, nerv_matrix_(tname));
     int idx = luaL_checkinteger(L, 2);
     if (idx < 0 || idx >= self->nmax)
@@ -20,7 +26,7 @@ int nerv_matrix_(lua_get_elem)(lua_State *L) {
     return 1;
 }
 
-int nerv_matrix_(lua_set_elem)(lua_State *L) {
+static int nerv_matrix_(lua_set_elem)(lua_State *L) {
     Matrix *self = luaT_checkudata(L, 1, nerv_matrix_(tname));
     int idx = luaL_checkinteger(L, 2);
     MATRIX_ELEM v = luaL_checknumber(L, 3);
@@ -40,7 +46,7 @@ static void host_matrix_(init)(lua_State *L) {
 
 #include "matrix.c"
 
-int nerv_matrix_(lua_load)(lua_State *L) {
+static int nerv_matrix_(lua_load)(lua_State *L) {
     Status status;
     ChunkData *cdp = luaT_checkudata(L, 1, nerv_chunk_data_tname);
     Matrix *self = nerv_matrix_(load)(cdp, &status);
@@ -49,7 +55,7 @@ int nerv_matrix_(lua_load)(lua_State *L) {
     return 1;
 }
 
-int nerv_matrix_(lua_save)(lua_State *L) {
+static int nerv_matrix_(lua_save)(lua_State *L) {
     Status status;
     ChunkFile *cfp = luaT_checkudata(L, 2,
                             nerv_chunk_file_handle_tname);
@@ -59,7 +65,7 @@ int nerv_matrix_(lua_save)(lua_State *L) {
     return 0;
 }
 
-int nerv_matrix_(lua_copy_from)(lua_State *L) {
+static int nerv_matrix_(lua_copy_fromh)(lua_State *L) {
     Status status;
     Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname));
     const Matrix *b = luaT_checkudata(L, 2, nerv_matrix_(tname));
@@ -67,15 +73,53 @@ int nerv_matrix_(lua_copy_from)(lua_State *L) {
     int b_begin = nargs > 2 ? luaL_checkinteger(L, 3) : 0;
     int b_end = nargs > 3 ? luaL_checkinteger(L, 4) : b->nrow;
     int a_begin = nargs > 4 ? luaL_checkinteger(L, 5) : 0;
-    nerv_matrix_(copy_from)(a, b, a_begin, b_begin, b_end, &status);
+    nerv_matrix_(copy_fromh)(a, b, a_begin, b_begin, b_end, &status);
     NERV_LUA_CHECK_STATUS(L, status);
     return 0;
 }
 
+static int nerv_matrix_(lua_copy_rows_fromh_by_idx)(lua_State *L)
+{
+    Status status;
+    Matrix *a=luaT_checkudata(L,1,nerv_matrix_(tname));
+    const Matrix *b=luaT_checkudata(L,2,nerv_matrix_(tname));
+    const Matrix *idx=luaT_checkudata(L,3,nerv_matrix_(tname));
+    int b_begin=lua_gettop(L)>3?luaL_checkinteger(L,4):0;
+    nerv_matrix_(copy_rows_fromh_by_idx)(a,b,idx,b_begin,&status);
+    NERV_LUA_CHECK_STATUS(L,status);
+    return 0;
+}
+
 static const luaL_Reg nerv_matrix_(extra_methods)[] = {
+    {"colsum", nerv_matrix_(lua_colsum)},
+    {"colsame", nerv_matrix_(lua_colsame)},
+    {"rowsum", nerv_matrix_(lua_rowsum)},
+    {"rowmax", nerv_matrix_(lua_rowmax)},
+    {"rowmax_idx", nerv_matrix_(lua_rowmax_idx)},
+    {"trans", nerv_matrix_(lua_trans)},
+    {"decompress", nerv_matrix_(lua_decompress)},
+    /* in-place calc */
+    {"copy_fromh", nerv_matrix_(lua_copy_fromh)},
+    /* alias for copy_from */
+    {"copy_from", nerv_matrix_(lua_copy_fromh)},
+    {"add", nerv_matrix_(lua_add)},
+    {"mul", nerv_matrix_(lua_mul)},
+    {"add_row", nerv_matrix_(lua_add_row)},
+    {"clip", nerv_matrix_(lua_clip)},
+    {"fill", nerv_matrix_(lua_fill)},
+    {"sigmoid", nerv_matrix_(lua_sigmoid)},
+    {"sigmoid_grad", nerv_matrix_(lua_sigmoid_grad)},
+    {"softmax", nerv_matrix_(lua_softmax)},
+    {"mul_elem", nerv_matrix_(lua_mul_elem)},
+    {"log_elem", nerv_matrix_(lua_log_elem)},
+    {"copy_rows_fromh_by_idx", nerv_matrix_(lua_copy_rows_fromh_by_idx)},
+    {"copy_rows_from_by_idx", nerv_matrix_(lua_copy_rows_fromh_by_idx)},
+    {"expand_frm", nerv_matrix_(lua_expand_frm)},
+    {"rearrange_frm", nerv_matrix_(lua_rearrange_frm)},
+    {"scale_rows_by_row", nerv_matrix_(lua_scale_rows_by_row)},
+    {"scale_rows_by_col", nerv_matrix_(lua_scale_rows_by_col)},
     {"load", nerv_matrix_(lua_load)},
     {"save", nerv_matrix_(lua_save)},
-    {"copy_from", nerv_matrix_(lua_copy_from)},
     {NULL, NULL}
 };