repaired trmm bug in cgemm_kernel_4x2_bulldozer.S

diff --git a/kernel/x86_64/cgemm_kernel_4x2_bulldozer.S b/kernel/x86_64/cgemm_kernel_4x2_bulldozer.S
index 8cf8b54..1ddb66f 100644 (file)
--- a/kernel/x86_64/cgemm_kernel_4x2_bulldozer.S
+++ b/kernel/x86_64/cgemm_kernel_4x2_bulldozer.S
@@ -1255,6 +1255,9 @@
 \r
        \r
 .L2_60:\r
+#if defined(TRMMKERNEL) && !defined(LEFT)\r
+        addq    $2, KK\r
+#endif\r
 \r
        decq    J                       // j --\r
        jg      .L2_01                  // next 2 lines of N\r