optimized sgemm kernel for ARMV6

diff --git a/kernel/arm/sgemm_kernel_4x2_vfp.S b/kernel/arm/sgemm_kernel_4x2_vfp.S
index 3e20f86..e074e74 100644 (file)
--- a/kernel/arm/sgemm_kernel_4x2_vfp.S
+++ b/kernel/arm/sgemm_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
+* 2013/11/27 Saar
 *       BLASTEST               : OK
 *       CTEST                  : OK
 *       TEST                   : OK
@@ -101,16 +101,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro KERNEL4x2_SUB
 
        flds    s4 , [ BO ]
-       flds    s5 , [ BO, #4 ]
 
        flds    s0 , [ AO ]
        flds    s1 , [ AO, #4 ]
-       flds    s2 , [ AO, #8 ]
-       flds    s3 , [ AO, #12 ]
 
        fmacs   s8  , s0,  s4
+       flds    s2 , [ AO, #8 ]
        fmacs   s9  , s1,  s4
+       flds    s3 , [ AO, #12 ]
        fmacs   s10  , s2,  s4
+       flds    s5 , [ BO, #4 ]
        fmacs   s11  , s3,  s4
 
        fmacs   s12  , s0,  s5
@@ -469,13 +469,20 @@ sgemm_kernel_L2_M4_20:
        .align 5
 
 sgemm_kernel_L2_M4_22:
+
+       pld [ AO, #A_PRE ]
+       pld [ BO, #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
+       pld [ AO, #A_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
 
+       pld [ AO, #A_PRE ]
+       pld [ BO, #B_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB
+       pld [ AO, #A_PRE ]
        KERNEL4x2_SUB
        KERNEL4x2_SUB