optimized dgemm kernel for ARMV6

diff --git a/kernel/arm/dgemm_kernel_4x2_vfp.S b/kernel/arm/dgemm_kernel_4x2_vfp.S
index 56fd815133fb0f90a4dbb3c7f83c214c1a6ac05e..55409a5ef4efdd5704ef8cfe10f0512d4e42ae91 100644 (file)
--- a/kernel/arm/dgemm_kernel_4x2_vfp.S
+++ b/kernel/arm/dgemm_kernel_4x2_vfp.S
@@ -26,10 +26,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 /**************************************************************************************
-* 2013/11/23 Saar
-*       BLASTEST               : xOK
-*       CTEST                  : xOK
-*       TEST                   : xOK
+* 2013/11/27 Saar
+*       BLASTEST               : OK
+*       CTEST                  : OK
+*       TEST                   : OK
 *
 **************************************************************************************/
 
@@ -77,7 +77,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define A_PRE  96
 #define B_PRE  96
-#define C_PRE  64
+#define C_PRE  32
 
 /**************************************************************************************
 * Macro definitions
@@ -100,26 +100,27 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro KERNEL4x2_SUB
 
+       pld     [ AO, #A_PRE ]
        fldd    d4 , [ BO ]
-       fldd    d5 , [ BO, #8 ]
 
        fldd    d0 , [ AO ]
        fldd    d1 , [ AO, #8 ]
-       fldd    d2 , [ AO, #16 ]
-       fldd    d3 , [ AO, #24 ]
 
        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
 
@@ -130,37 +131,40 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
        fldd            d0, ALPHA
 
+
        fldd    d4 , [CO1]
        fldd    d5 , [CO1, #8 ]
-       fldd    d6 , [CO1, #16 ]
-       fldd    d7 , [CO1, #24 ]
                
+       pld     [ CO1, #C_PRE ]
        fmacd   d4 , d0 , d8
+       fldd    d6 , [CO1, #16 ]
        fmacd   d5 , d0 , d9
+       fldd    d7 , [CO1, #24 ]
        fmacd   d6 , d0 , d10
+       fstd    d4 , [CO1]
        fmacd   d7 , d0 , d11
 
-       fstd    d4 , [CO1]
        fstd    d5 , [CO1, #8 ]
        fstd    d6 , [CO1, #16 ]
        fstd    d7 , [CO1, #24 ]
 
        fldd    d4 , [CO2]
        fldd    d5 , [CO2, #8 ]
-       fldd    d6 , [CO2, #16 ]
-       fldd    d7 , [CO2, #24 ]
 
+       pld     [ CO2, #C_PRE ]
        fmacd   d4 , d0 , d12
+       fldd    d6 , [CO2, #16 ]
        fmacd   d5 , d0 , d13
+       fldd    d7 , [CO2, #24 ]
        fmacd   d6 , d0 , d14
+       fstd    d4 , [CO2]
        fmacd   d7 , d0 , d15
+       add     CO1, CO1, #32
 
-       fstd    d4 , [CO2]
        fstd    d5 , [CO2, #8 ]
        fstd    d6 , [CO2, #16 ]
        fstd    d7 , [CO2, #24 ]
 
-       add     CO1, CO1, #32
 
 .endm
 
@@ -469,13 +473,18 @@ dgemm_kernel_L2_M4_20:
        .align 5
 
 dgemm_kernel_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