Support unalign address in daxpy on loongson3a simd..

diff --git a/kernel/mips64/daxpy_loongson3a_simd.S b/kernel/mips64/daxpy_loongson3a_simd.S
index 543f1ce..9a0b8f1 100644 (file)
--- a/kernel/mips64/daxpy_loongson3a_simd.S
+++ b/kernel/mips64/daxpy_loongson3a_simd.S
@@ -72,7 +72,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "common.h"
 
                
-#define PREFETCH_DISTANCE 1864
+#define PREFETCH_DISTANCE 2016
                
 #define N      $4
 
@@ -195,11 +195,39 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        dsll    INCY, INCY, BASE_SHIFT
 
        bne     INCY, TEMP, .L20
+
+       //Dose  the address of Y algin 16 bytes?  
+       andi    TEMP,  Y, 8
+       beq     TEMP, $0, .L10  
+       //Y unalgin. Compute this unalgined element.
+       LD      a1,  0 * SIZE(X)
+       LD      b1,  0 * SIZE(Y)
+
+       daddiu  X, X, SIZE
+       daddiu  Y, Y, SIZE
+
+       MADD    t1, b1, ALPHA, a1
+       daddiu  N, N, -1
+       
+       ST      t1, -1 * SIZE(Y)
+       blez    N, .L999
+       .align 5
+       
+.L10:
+
        dsra    I, N, 4
 
        blez    I, .L15
        daddiu  I, I, -1
+       
+       //Y algin. We need test X address
+       //Dose  the address of X algin 16 bytes?  
+       andi    TEMP,  X, 8
+       bne     TEMP, $0, .L30  ///
+       .align 5
 
+.L11:
+       //X & Y algin
        gsLQC1(X_BASE,A2,A1,0*SIZE)
        gsLQC1(X_BASE,A4,A3,2*SIZE)
        gsLQC1(X_BASE,A6,A5,4*SIZE)
@@ -345,7 +373,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
        blez    I, .L999
        NOP
-       .align  3
+       .align  5
 
 .L16:
        LD      a1,  0 * SIZE(X)
@@ -382,6 +410,105 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        NOP
        .align 5
 
+.L30:
+       //Y align, X unalign, INCX==INCY==1
+       //unloop 16
+       
+       LD      a1,  0 * SIZE(X)
+       gsLQC1(X_BASE,A3,A2,1*SIZE)
+       gsLQC1(X_BASE,A5,A4,3*SIZE)
+       gsLQC1(X_BASE,A7,A6,5*SIZE)
+       gsLQC1(X_BASE,A9,A8,7*SIZE)
+
+       gsLQC1(X_BASE,A11,A10,8*SIZE)
+       gsLQC1(X_BASE,A13,A12,11*SIZE)
+       gsLQC1(X_BASE,A15,A14,13*SIZE)
+       LD      a16,  15 * SIZE(X)
+               
+       gsLQC1(Y_BASE,B2,B1,0*SIZE)
+       gsLQC1(Y_BASE,B4,B3,2*SIZE)
+       gsLQC1(Y_BASE,B6,B5,4*SIZE)
+       gsLQC1(Y_BASE,B8,B7,6*SIZE)
+               
+       blez    I, .L13
+       NOP
+       .align 5
+       
+.L31:
+       MADD    t1, b1, ALPHA, a1       
+       MADD    t2, b2, ALPHA, a2
+       gsSQC1(Y_BASE, T2, T1, 0*SIZE)          
+       gsLQC1(Y_BASE,B2,B1,8*SIZE)
+
+       MADD    t3, b3, ALPHA, a3
+       MADD    t4, b4, ALPHA, a4
+       gsSQC1(Y_BASE, T4, T3, 2*SIZE)
+       gsLQC1(Y_BASE,B4,B3,10*SIZE)
+
+       PREFETCHD(PREFETCH_DISTANCE*SIZE(Y))
+       PREFETCHD((PREFETCH_DISTANCE+4)*SIZE(Y))
+
+       MADD    t1, b5, ALPHA, a5
+       MADD    t2, b6, ALPHA, a6
+       gsSQC1(Y_BASE, T2, T1, 4*SIZE)          
+       gsLQC1(Y_BASE,B6,B5,12*SIZE)
+
+       MADD    t3, b7, ALPHA, a7
+       MADD    t4, b8, ALPHA, a8
+       gsSQC1(Y_BASE, T4, T3, 6*SIZE)
+       gsLQC1(Y_BASE,B8,B7,14*SIZE)
+
+       PREFETCHD((PREFETCH_DISTANCE+8)*SIZE(Y))
+       PREFETCHD((PREFETCH_DISTANCE+12)*SIZE(Y))
+
+       MADD    t1, b1, ALPHA, a9       
+       MADD    t2, b2, ALPHA, a10
+       gsSQC1(Y_BASE, T2, T1, 8*SIZE)          
+       gsLQC1(Y_BASE,B2,B1,16*SIZE)
+
+       MADD    t3, b3, ALPHA, a11
+       MADD    t4, b4, ALPHA, a12
+       gsSQC1(Y_BASE, T4, T3, 10*SIZE)
+       gsLQC1(Y_BASE,B4,B3,18*SIZE)
+
+       PREFETCHD(PREFETCH_DISTANCE*SIZE(X))
+       PREFETCHD((PREFETCH_DISTANCE+4)*SIZE(X))
+
+       MADD    t1, b5, ALPHA, a13      
+       MADD    t2, b6, ALPHA, a14
+       gsSQC1(Y_BASE, T2, T1, 12*SIZE)         
+       gsLQC1(Y_BASE,B6,B5,20*SIZE)
+
+       MADD    t3, b7, ALPHA, a15
+       MADD    t4, b8, ALPHA, a16
+       gsSQC1(Y_BASE, T4, T3, 14*SIZE)
+       gsLQC1(Y_BASE,B8,B7,22*SIZE)
+               
+       PREFETCHD((PREFETCH_DISTANCE+8)*SIZE(X))
+       PREFETCHD((PREFETCH_DISTANCE+12)*SIZE(X))
+
+       LD      a1,  16 * SIZE(X)
+       gsLQC1(X_BASE,A3,A2,17*SIZE)
+       gsLQC1(X_BASE,A5,A4,19*SIZE)
+       gsLQC1(X_BASE,A7,A6,21*SIZE)
+       gsLQC1(X_BASE,A9,A8,23*SIZE)
+
+       gsLQC1(X_BASE,A11,A10,25*SIZE)
+       gsLQC1(X_BASE,A13,A12,27*SIZE)
+       gsLQC1(X_BASE,A15,A14,29*SIZE)
+       LD      a16,  31 * SIZE(X)
+
+       daddiu  I, I, -1
+       daddiu  Y, Y, 16 * SIZE
+               
+       daddiu  X, X, 16 * SIZE
+       bgtz    I, .L31
+               
+       //jump back to the remain loop process.
+       b       .L13
+       .align 5
+       
+//INCX!=1 or INCY != 1 
 .L20:
        dsra    I, N, 3
        move    YY, Y
@@ -538,7 +665,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
        blez    I, .L999
        NOP
-       .align  3
+       .align  5
 
 .L26:
        LD      a1,  0 * SIZE(X)