diff dsputil.h @ 6384:0a403ade8c81 libavcodec

simd and unroll png_filter_row cycles per 1000 pixels on core2: left: 9211->5170 top: 9283->2138 avg: 12215->7611 paeth: 64024->17360 overall rgb png decoding speed: +45% overall greyscale png decoding speed: +6%

--- a/dsputil.h	Thu Feb 21 00:06:07 2008 +0000
+++ b/dsputil.h	Thu Feb 21 07:10:46 2008 +0000
@@ -304,12 +304,15 @@
 
     /* huffyuv specific */
     void (*add_bytes)(uint8_t *dst/*align 16*/, uint8_t *src/*align 16*/, int w);
+    void (*add_bytes_l2)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 16*/, int w);
     void (*diff_bytes)(uint8_t *dst/*align 16*/, uint8_t *src1/*align 16*/, uint8_t *src2/*align 1*/,int w);
     /**
      * subtract huffyuv's variant of median prediction
      * note, this might read from src1[-1], src2[-1]
      */
     void (*sub_hfyu_median_prediction)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w, int *left, int *left_top);
+    /* this might write to dst[w] */
+    void (*add_png_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
     void (*bswap_buf)(uint32_t *dst, const uint32_t *src, int w);
 
     void (*h264_v_loop_filter_luma)(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0);