changeset 11112:ffe1ea72ea29 libavcodec

ARMv6 optimised pix_abs8
author mru
date Tue, 09 Feb 2010 16:13:34 +0000
parents 95b1fd6057d3
children b529129c4563
files arm/dsputil_armv6.S arm/dsputil_init_armv6.c
diffstat 2 files changed, 36 insertions(+), 0 deletions(-) [+]
line wrap: on
line diff
--- a/arm/dsputil_armv6.S	Tue Feb 09 16:13:31 2010 +0000
+++ b/arm/dsputil_armv6.S	Tue Feb 09 16:13:34 2010 +0000
@@ -428,3 +428,33 @@
 
         pop             {r4-r11, pc}
 .endfunc
+
+function ff_pix_abs8_armv6, export=1
+        pld             [r2, r3]
+        ldr             r12, [sp]
+        push            {r4-r9, lr}
+        mov             r0,  #0
+        mov             lr,  #0
+        ldrd            r4,  r5,  [r1], r3
+1:
+        subs            r12, r12, #2
+        ldr             r7,  [r2, #4]
+        ldr             r6,  [r2], r3
+        ldrd            r8,  r9,  [r1], r3
+        usada8          r0,  r4,  r6,  r0
+        pld             [r2, r3]
+        usada8          lr,  r5,  r7,  lr
+        ldr             r7,  [r2, #4]
+        ldr             r6,  [r2], r3
+        beq             2f
+        ldrd            r4,  r5,  [r1], r3
+        usada8          r0,  r8,  r6,  r0
+        pld             [r2, r3]
+        usada8          lr,  r9,  r7,  lr
+        b               1b
+2:
+        usada8          r0,  r8,  r6,  r0
+        usada8          lr,  r9,  r7,  lr
+        add             r0,  r0,  lr
+        pop             {r4-r9, pc}
+.endfunc
--- a/arm/dsputil_init_armv6.c	Tue Feb 09 16:13:31 2010 +0000
+++ b/arm/dsputil_init_armv6.c	Tue Feb 09 16:13:34 2010 +0000
@@ -57,6 +57,9 @@
 int ff_pix_abs16_y2_armv6(void *s, uint8_t *blk1, uint8_t *blk2,
                           int line_size, int h);
 
+int ff_pix_abs8_armv6(void *s, uint8_t *blk1, uint8_t *blk2,
+                       int line_size, int h);
+
 void av_cold ff_dsputil_init_armv6(DSPContext* c, AVCodecContext *avctx)
 {
     if (!avctx->lowres && (avctx->idct_algo == FF_IDCT_AUTO ||
@@ -94,5 +97,8 @@
     c->pix_abs[0][1] = ff_pix_abs16_x2_armv6;
     c->pix_abs[0][2] = ff_pix_abs16_y2_armv6;
 
+    c->pix_abs[1][0] = ff_pix_abs8_armv6;
+
     c->sad[0] = ff_pix_abs16_armv6;
+    c->sad[1] = ff_pix_abs8_armv6;
 }