optimized gemv_n_vfp.S

diff --git a/kernel/arm/gemv_n_vfp.S b/kernel/arm/gemv_n_vfp.S
index 4726599..f1cf9a0 100644 (file)
--- a/kernel/arm/gemv_n_vfp.S
+++ b/kernel/arm/gemv_n_vfp.S
@@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 
 /**************************************************************************************
-* 2013/11/24 Saar
+* 2013/11/28 Saar
 *       BLASTEST               : OK
 *       CTEST                  : OK
 *       TEST                   : OK
@@ -74,44 +74,75 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #if    defined(DOUBLE)
 
-.macro INIT_F4
+.macro INIT_F8
 
        pld     [ YO , #Y_PRE ]
+       pld     [ YO , #Y_PRE+32 ]
 
-       vsub.f64        d12 , d12 , d12
-       vmov.f64        d13 , d12
-       vmov.f64        d14 , d12
-       vmov.f64        d15 , d12
+       vsub.f64        d8 , d8 , d8
+       vmov.f64        d9  , d8
+       vmov.f64        d10 , d8
+       vmov.f64        d11 , d8
+       vmov.f64        d12 , d8
+       vmov.f64        d13 , d8
+       vmov.f64        d14 , d8
+       vmov.f64        d15 , d8
 
 .endm
 
-.macro KERNEL_F4X4
+.macro KERNEL_F8X8
 
        pld     [ XO , #X_PRE ]
-       KERNEL_F4X1
-       KERNEL_F4X1
-       KERNEL_F4X1
-       KERNEL_F4X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+
+       pld     [ XO , #X_PRE ]
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
 
 .endm
 
 
-.macro KERNEL_F4X1
+.macro KERNEL_F8X1
 
+       pld     [ AO2 , #A_PRE ]
        fldmiad XO! ,  { d2 }
-       fldmiad AO1 ,  { d8 - d11 }
+       fldmiad AO1 ,  { d4 - d7 }
 
-       vmla.f64        d12 , d2 , d8
-       pld     [ AO2 , #A_PRE ]
-       vmla.f64        d13 , d2 , d9
+       vmla.f64        d8  , d2 , d4
+       pld     [ AO2 , #4*SIZE ]
+       vmla.f64        d9  , d2 , d5
+       add     r3, AO1, #4*SIZE
+       vmla.f64        d10 , d2 , d6
+       vmla.f64        d11 , d2 , d7
+
+
+       fldmiad r3 ,  { d4 - d7 }
+
+       vmla.f64        d12 , d2 , d4
+       vmla.f64        d13 , d2 , d5
        add             AO1, AO1, LDA
-       vmla.f64        d14 , d2 , d10
-       vmla.f64        d15 , d2 , d11
+       vmla.f64        d14 , d2 , d6
        add             AO2, AO2, LDA
+       vmla.f64        d15 , d2 , d7
+
 
 .endm
 
-.macro SAVE_F4
+.macro SAVE_F8
+
+       fldmiad YO,  { d4 - d7 }
+
+       vmla.f64        d4 , d0, d8
+       vmla.f64        d5 , d0, d9
+       vmla.f64        d6 , d0, d10
+       vmla.f64        d7 , d0, d11
+
+       fstmiad YO!, { d4 - d7 }
 
        fldmiad YO,  { d4 - d7 }
 
@@ -244,43 +275,73 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #else  /************************* SINGLE PRECISION *****************************************/
 
-.macro INIT_F4
+.macro INIT_F8
 
        pld     [ YO , #Y_PRE ]
 
-       vsub.f32        s12 , s12 , s12
-       vmov.f32        s13 , s12
-       vmov.f32        s14 , s12
-       vmov.f32        s15 , s12
+       vsub.f32        s8 , s8 , s8
+       vmov.f32        s9  , s8
+       vmov.f32        s10 , s8
+       vmov.f32        s11 , s8
+       vmov.f32        s12 , s8
+       vmov.f32        s13 , s8
+       vmov.f32        s14 , s8
+       vmov.f32        s15 , s8
 
 .endm
 
-.macro KERNEL_F4X4
+.macro KERNEL_F8X8
 
        pld     [ XO , #X_PRE ]
-       KERNEL_F4X1
-       KERNEL_F4X1
-       KERNEL_F4X1
-       KERNEL_F4X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
+       KERNEL_F8X1
 
 .endm
 
 
-.macro KERNEL_F4X1
+.macro KERNEL_F8X1
 
+       pld     [ AO2, #A_PRE ]
        fldmias XO! ,  { s2 }
-       fldmias AO1 ,  { s8 - s11 }
+       fldmias AO1 ,  { s4 - s7 }
+
+       vmla.f32        s8  , s2 , s4
+       vmla.f32        s9  , s2 , s5
+       vmla.f32        s10 , s2 , s6
+       vmla.f32        s11 , s2 , s7
+
+       add     r3, AO1, #4*SIZE
+
+       fldmias r3 ,  { s4 - s7 }
+
+       vmla.f32        s12 , s2 , s4
+       vmla.f32        s13 , s2 , s5
+       vmla.f32        s14 , s2 , s6
+       vmla.f32        s15 , s2 , s7
 
-       vmla.f32        s12 , s2 , s8
-       vmla.f32        s13 , s2 , s9
-       vmla.f32        s14 , s2 , s10
-       vmla.f32        s15 , s2 , s11
        add             AO1, AO1, LDA
        add             AO2, AO2, LDA
 
 .endm
 
-.macro SAVE_F4
+.macro SAVE_F8
+
+       fldmias YO,  { s4 - s7 }
+
+       vmla.f32        s4 , s0, s8
+       vmla.f32        s5 , s0, s9
+       vmla.f32        s6 , s0, s10
+       vmla.f32        s7 , s0, s11
+
+       fstmias YO!, { s4 - s7 }
+
 
        fldmias YO,  { s4 - s7 }
 
@@ -332,8 +393,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro KERNEL_S4X4
 
+       pld     [ AO2 , #A_PRE ]
        KERNEL_S4X1
        KERNEL_S4X1
+       pld     [ AO2 , #A_PRE ]
        KERNEL_S4X1
        KERNEL_S4X1
 
@@ -342,7 +405,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro KERNEL_S4X1
 
-       pld     [ AO2 , #A_PRE ]
        fldmias XO  ,  { s2 }
        fldmias AO1 ,  { s8 - s11 }
 
@@ -471,27 +533,30 @@ gemvn_kernel_F4_BEGIN:
        ldr     YO , Y
 
        ldr     I, M
-       asrs    I, I, #2                                        // I = M / 4
+       asrs    I, I, #3                                        // I = M / 8
        ble     gemvn_kernel_F1_BEGIN
 
 gemvn_kernel_F4X4:
 
        ldr     AO1, A
        add     AO2, AO1, LDA
-       add     r3 , AO1, #4*SIZE
+       add     r3 , AO1, #8*SIZE
        str     r3 , A
 
+       add     AO2, AO2, LDA
+       add     AO2, AO2, LDA
+
        ldr     XO , X
 
-       INIT_F4
+       INIT_F8
 
-       asrs    J, N, #2                                        // J = N / 4
+       asrs    J, N, #3                                        // J = N / 8
        ble     gemvn_kernel_F4X1
 
 
 gemvn_kernel_F4X4_10:
 
-       KERNEL_F4X4
+       KERNEL_F8X8
 
        subs    J, J, #1
        bne     gemvn_kernel_F4X4_10
@@ -499,12 +564,12 @@ gemvn_kernel_F4X4_10:
 
 gemvn_kernel_F4X1:
 
-       ands    J, N , #3
+       ands    J, N , #7
        ble     gemvn_kernel_F4_END
 
 gemvn_kernel_F4X1_10:
 
-       KERNEL_F4X1
+       KERNEL_F8X1
 
        subs    J, J, #1
        bne     gemvn_kernel_F4X1_10
@@ -512,7 +577,7 @@ gemvn_kernel_F4X1_10:
 
 gemvn_kernel_F4_END:
 
-       SAVE_F4
+       SAVE_F8
 
        subs    I , I , #1
        bne     gemvn_kernel_F4X4
@@ -521,7 +586,7 @@ gemvn_kernel_F4_END:
 gemvn_kernel_F1_BEGIN:
 
        ldr     I, M
-       ands    I,  I , #3
+       ands    I,  I , #7
        ble     gemvn_kernel_L999
 
 gemvn_kernel_F1X1: