optimized dtrmm kernel for ARMV7

diff --git a/kernel/arm/dtrmm_kernel_4x2_vfp.S b/kernel/arm/dtrmm_kernel_4x2_vfp.S
index 55a017a..762b9c5 100644 (file)
--- a/kernel/arm/dtrmm_kernel_4x2_vfp.S
+++ b/kernel/arm/dtrmm_kernel_4x2_vfp.S
@@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 /**************************************************************************************
-* 2013/11/23 Saar
+* 2013/11/28 Saar
 *       BLASTEST               : OK
 *       CTEST                  : OK
 *       TEST                   : OK
@@ -106,25 +106,26 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro KERNEL4x2_SUB
 
        fldd    d4 , [ BO ]
-       fldd    d5 , [ BO, #8 ]
-
        fldd    d0 , [ AO ]
+
        fldd    d1 , [ AO, #8 ]
-       fldd    d2 , [ AO, #16 ]
-       fldd    d3 , [ AO, #24 ]
+       pld     [ AO , #A_PRE ]
 
        fmacd   d8  , d0,  d4
+       fldd    d2 , [ AO, #16 ]
        fmacd   d9  , d1,  d4
+       fldd    d3 , [ AO, #24 ]
        fmacd   d10  , d2,  d4
+       fldd    d5 , [ BO, #8 ]
        fmacd   d11  , d3,  d4
 
        fmacd   d12  , d0,  d5
        fmacd   d13  , d1,  d5
+       add     AO , AO, #32
        fmacd   d14  , d2,  d5
+       add     BO , BO, #16
        fmacd   d15  , d3,  d5
 
-       add     AO , AO, #32
-       add     BO , BO, #16
 
 .endm
 
@@ -490,13 +491,18 @@ _L2_M4_20:
        .align 5
 
 _L2_M4_22:
+
+       pld     [ BO , #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
+       pld     [ BO , #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
 
+       pld     [ BO , #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
+       pld     [ BO , #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB