#include "common.h"
-#define PREFETCH_DISTANCE 48
+#define PREFETCH_DISTANCE 48
#define N $4
blez I, .L13
NOP
- .align 3
+ .align 5
.L12:
PREFETCHD(PREFETCH_DISTANCE*SIZE(X))
ST t3, 2 * SIZE(Y)
ST t4, 3 * SIZE(Y)
+ PREFETCHD((PREFETCH_DISTANCE+4)*SIZE(X))
+ PREFETCHD((PREFETCH_DISTANCE+4)*SIZE(Y))
+
MADD t1, b5, ALPHA, a5
LD a5, 12 * SIZE(X)
LD b5, 12 * SIZE(Y)
bgtz I, .L12
daddiu X, X, 8 * SIZE
- .align 3
+ .align 5
.L13:
MADD t1, b1, ALPHA, a1
daddiu X, X, 8 * SIZE
daddiu Y, Y, 8 * SIZE
- .align 3
+ .align 5
.L15:
andi I, N, 7
j $31
NOP
- .align 3
+ .align 5
.L20:
dsra I, N, 3
blez I, .L23
NOP
- .align 3
+ .align 5
.L22:
MADD t1, b1, ALPHA, a1
bgtz I, .L22
daddu YY, YY, INCY
- .align 3
+ .align 5
.L23:
MADD t1, b1, ALPHA, a1
daddu YY, YY, INCY
ST t4, 0 * SIZE(YY)
daddu YY, YY, INCY
- .align 3
+ .align 5
.L25:
andi I, N, 7
bgtz I, .L26
daddu Y, Y, INCY
- .align 3
+ .align 5
.L999:
#ifndef __64BIT__