From 76d5d2ccfca487fa24b69c73a8085491c4c41cae Mon Sep 17 00:00:00 2001
From: Cedric Nugteren <web@cedricnugteren.nl>
Date: Sun, 23 Oct 2016 20:49:55 +0200
Subject: Fixed a bug in the transpose-matrix function

---
 src/routines/common.hpp | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/routines/common.hpp b/src/routines/common.hpp
index 802abec4..53ca6355 100644
--- a/src/routines/common.hpp
+++ b/src/routines/common.hpp
@@ -61,8 +61,8 @@ void PadCopyTransposeMatrix(Queue &queue, const Device &device,
   if (do_transpose) {
     if (use_fast_kernel &&
         IsMultiple(src_ld, db["TRA_WPT"]) &&
-        IsMultiple(src_one, db["TRA_WPT"]*db["TRA_WPT"]) &&
-        IsMultiple(src_two, db["TRA_WPT"]*db["TRA_WPT"])) {
+        IsMultiple(src_one, db["TRA_WPT"]*db["TRA_DIM"]) &&
+        IsMultiple(src_two, db["TRA_WPT"]*db["TRA_DIM"])) {
       kernel_name = "TransposeMatrixFast";
     }
     else {
-- 
cgit v1.2.3