From 6c0671929f24f296c51568ae3d5c93a3e4d5518f Mon Sep 17 00:00:00 2001 From: Determinant Date: Fri, 22 May 2015 12:50:13 +0800 Subject: rename colsum to rowsum; add colsum implementation --- matrix/generic/cumatrix.c | 19 ++++++++++++++----- 1 file changed, 14 insertions(+), 5 deletions(-) (limited to 'matrix/generic/cumatrix.c') diff --git a/matrix/generic/cumatrix.c b/matrix/generic/cumatrix.c index d98c559..f846a73 100644 --- a/matrix/generic/cumatrix.c +++ b/matrix/generic/cumatrix.c @@ -95,25 +95,33 @@ static int nerv_matrix_(softmax)(lua_State *L) { Matrix *max = nerv_matrix_(new_)(a->nrow, 1); Matrix *dno = nerv_matrix_(new_)(a->nrow, 1); Matrix *b = nerv_matrix_(new_)(a->nrow, a->ncol); - cudak_(cuda_colmax)(a, max); + cudak_(cuda_rowmax)(a, max); cudak_(cuda_softmax_denominator)(a, max, dno); cudak_(cuda_softmax_final)(a, max, dno, b); luaT_pushudata(L, b, nerv_matrix_(tname)); return 1; } -static int nerv_matrix_(colsum)(lua_State *L) { +static int nerv_matrix_(rowsum)(lua_State *L) { Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname)); Matrix *b = nerv_matrix_(new_)(a->nrow, 1); + cudak_(cuda_rowsum)(a, b); + luaT_pushudata(L, b, nerv_matrix_(tname)); + return 1; +} + +static int nerv_matrix_(colsum)(lua_State *L) { + Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname)); + Matrix *b = nerv_matrix_(new_)(1, a->ncol); cudak_(cuda_colsum)(a, b); luaT_pushudata(L, b, nerv_matrix_(tname)); return 1; } -static int nerv_matrix_(colmax)(lua_State *L) { +static int nerv_matrix_(rowmax)(lua_State *L) { Matrix *a = luaT_checkudata(L, 1, nerv_matrix_(tname)); Matrix *b = nerv_matrix_(new_)(a->nrow, 1); - cudak_(cuda_colmax)(a, b); + cudak_(cuda_rowmax)(a, b); luaT_pushudata(L, b, nerv_matrix_(tname)); return 1; } @@ -125,7 +133,8 @@ static const luaL_Reg nerv_matrix_(extra_methods)[] = { {"sigmoid", nerv_matrix_(sigmoid)}, {"softmax", nerv_matrix_(softmax)}, {"colsum", nerv_matrix_(colsum)}, - {"colmax", nerv_matrix_(colmax)}, + {"rowsum", nerv_matrix_(rowsum)}, + {"rowmax", nerv_matrix_(rowmax)}, {NULL, NULL} }; -- cgit v1.2.3