[x264-devel] NOTE: Altivec/PPC port is broken [PATCH2]

[Noboru Asai]

Hi,

>  > Here's a patch that fixes hpel_filter_altivec, although I don't know
>  > whether this is a correct way to fix it.
>
>
> Me neither. I've looked at the patch a bit and can't figure out for
>  the moment why this change is needed. As you said previously, in
>  practice the original code produces the right output. It must break on
>  some extreme cases...

Sorry, I missed x range ( x < width ) of vertical hpel filter. x <
width + 3 is right.

Attached patch is based on Manuel's patch and remove unneeded bits.


Noboru Asai

diff --git a/common/ppc/mc.c b/common/ppc/mc.c
index 3c69b26..c99cd31 100644
--- a/common/ppc/mc.c
+++ b/common/ppc/mc.c
@@ -611,11 +611,30 @@ void x264_hpel_filter_altivec( uint8_t *dsth,
uint8_t *dstv, uint8_t *dstc, uint
             HPEL_FILTER_CENTRAL();
         }

+        /* Partial vertical filter */
+        VEC_LOAD( &src[x+i_stride*(y-2)], src1v, 16, vec_u8_t );
+        VEC_LOAD( &src[x+i_stride*(y-1)], src2v, 16, vec_u8_t );
+        VEC_LOAD( &src[x+i_stride*(y-0)], src3v, 16, vec_u8_t );
+        VEC_LOAD( &src[x+i_stride*(y+1)], src4v, 16, vec_u8_t );
+        VEC_LOAD( &src[x+i_stride*(y+2)], src5v, 16, vec_u8_t );
+        VEC_LOAD( &src[x+i_stride*(y+3)], src6v, 16, vec_u8_t );
+
+        temp1v = vec_u8_to_s16_h( src1v );
+        temp2v = vec_u8_to_s16_h( src2v );
+        temp3v = vec_u8_to_s16_h( src3v );
+        temp4v = vec_u8_to_s16_h( src4v );
+        temp5v = vec_u8_to_s16_h( src5v );
+        temp6v = vec_u8_to_s16_h( src6v );
+
+        HPEL_FILTER_1( temp1v, temp2v, temp3v,
+                      temp4v, temp5v, temp6v );
+
         /* central_filter */
         tempav = tempcv;
         tempbv = tempdv;
         tempcv = tempev;
-        tempdv = vec_splat( tempcv, 7 ); /* last only */
+        tempdv = temp1v;
+        /* tempev is not used */

         HPEL_FILTER_CENTRAL();
     }