Move png mmx functions into x86/png_mmx.c, remove them from DSPContext.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

7 files changed, 218 insertions, 164 deletions

diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
index 90e6440a87..d8e83f8d7b 100644
--- a/libavcodec/dsputil.c
+++ b/libavcodec/dsputil.c
@@ -38,7 +38,6 @@
 #include "config.h"
 #include "ac3dec.h"
 #include "vorbis.h"
-#include "png.h"
 
 uint8_t ff_cropTbl[256 + 2 * MAX_NEG_CROP] = {0, };
 uint32_t ff_squareTbl[512] = {0, };
@@ -1924,17 +1923,6 @@ static void add_bytes_c(uint8_t *dst, uint8_t *src, int w){
         dst[i+0] += src[i+0];
 }
 
-static void add_bytes_l2_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
-    long i;
-    for(i=0; i<=w-sizeof(long); i+=sizeof(long)){
-        long a = *(long*)(src1+i);
-        long b = *(long*)(src2+i);
-        *(long*)(dst+i) = ((a&pb_7f) + (b&pb_7f)) ^ ((a^b)&pb_80);
-    }
-    for(; i<w; i++)
-        dst[i] = src1[i]+src2[i];
-}
-
 static void diff_bytes_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
     long i;
 #if !HAVE_FAST_UNALIGNED
@@ -3087,7 +3075,6 @@ av_cold void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->ssd_int8_vs_int16 = ssd_int8_vs_int16_c;
 
     c->add_bytes= add_bytes_c;
-    c->add_bytes_l2= add_bytes_l2_c;
     c->diff_bytes= diff_bytes_c;
     c->add_hfyu_median_prediction= add_hfyu_median_prediction_c;
     c->sub_hfyu_median_prediction= sub_hfyu_median_prediction_c;
@@ -3095,9 +3082,6 @@ av_cold void dsputil_init(DSPContext* c, AVCodecContext *avctx)
     c->add_hfyu_left_prediction_bgr32 = add_hfyu_left_prediction_bgr32_c;
     c->bswap_buf= bswap_buf;
     c->bswap16_buf = bswap16_buf;
-#if CONFIG_PNG_DECODER
-    c->add_png_paeth_prediction= ff_add_png_paeth_prediction;
-#endif
 
     if (CONFIG_H263_DECODER || CONFIG_H263_ENCODER) {
         c->h263_h_loop_filter= h263_h_loop_filter_c;
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h
index 8240372270..85c2da5031 100644
--- a/libavcodec/dsputil.h
+++ b/libavcodec/dsputil.h
@@ -371,7 +371,6 @@ typedef struct DSPContext {
 
     /* 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
@@ -382,7 +381,6 @@ typedef struct DSPContext {
     int  (*add_hfyu_left_prediction)(uint8_t *dst, const uint8_t *src, int w, int left);
     void (*add_hfyu_left_prediction_bgr32)(uint8_t *dst, const uint8_t *src, int w, int *red, int *green, int *blue, int *alpha);
     /* 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 (*bswap16_buf)(uint16_t *dst, const uint16_t *src, int len);
 
diff --git a/libavcodec/png.h b/libavcodec/png.h
index bab5224851..d6fac3e673 100644
--- a/libavcodec/png.h
+++ b/libavcodec/png.h
@@ -23,6 +23,9 @@
 #define AVCODEC_PNG_H
 
 #include <stdint.h>
+#include <zlib.h>
+
+#include "avcodec.h"
 
 #define PNG_COLOR_MASK_PALETTE    1
 #define PNG_COLOR_MASK_COLOR      2
@@ -69,4 +72,41 @@ int ff_png_pass_row_size(int pass, int bits_per_pixel, int width);
 
 void ff_add_png_paeth_prediction(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
 
+typedef struct PNGDecContext {
+    const uint8_t *bytestream;
+    const uint8_t *bytestream_start;
+    const uint8_t *bytestream_end;
+    AVFrame picture1, picture2;
+    AVFrame *current_picture, *last_picture;
+
+    int state;
+    int width, height;
+    int bit_depth;
+    int color_type;
+    int compression_type;
+    int interlace_type;
+    int filter_type;
+    int channels;
+    int bits_per_pixel;
+    int bpp;
+
+    uint8_t *image_buf;
+    int image_linesize;
+    uint32_t palette[256];
+    uint8_t *crow_buf;
+    uint8_t *last_row;
+    uint8_t *tmp_row;
+    int pass;
+    int crow_size; /* compressed row size (include filter type) */
+    int row_size; /* decompressed row size */
+    int pass_row_size; /* decompress row size of the current pass */
+    int y;
+    z_stream zstream;
+
+    void (*add_bytes_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w);
+    void (*add_paeth_prediction)(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp);
+} PNGDecContext;
+
+void ff_png_init_mmx(PNGDecContext *s);
+
 #endif /* AVCODEC_PNG_H */
diff --git a/libavcodec/pngdec.c b/libavcodec/pngdec.c
index 2f9b343e5b..17fea6815b 100644
--- a/libavcodec/pngdec.c
+++ b/libavcodec/pngdec.c
@@ -18,11 +18,13 @@
  * License along with FFmpeg; if not, write to the Free Software
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
+
+//#define DEBUG
+
 #include "libavutil/imgutils.h"
 #include "avcodec.h"
 #include "bytestream.h"
 #include "png.h"
-#include "dsputil.h"
 
 /* TODO:
  * - add 2, 4 and 16 bit depth support
@@ -30,42 +32,6 @@
 
 #include <zlib.h>
 
-//#define DEBUG
-
-typedef struct PNGDecContext {
-    DSPContext dsp;
-
-    const uint8_t *bytestream;
-    const uint8_t *bytestream_start;
-    const uint8_t *bytestream_end;
-    AVFrame picture1, picture2;
-    AVFrame *current_picture, *last_picture;
-
-    int state;
-    int width, height;
-    int bit_depth;
-    int color_type;
-    int compression_type;
-    int interlace_type;
-    int filter_type;
-    int channels;
-    int bits_per_pixel;
-    int bpp;
-
-    uint8_t *image_buf;
-    int image_linesize;
-    uint32_t palette[256];
-    uint8_t *crow_buf;
-    uint8_t *last_row;
-    uint8_t *tmp_row;
-    int pass;
-    int crow_size; /* compressed row size (include filter type) */
-    int row_size; /* decompressed row size */
-    int pass_row_size; /* decompress row size of the current pass */
-    int y;
-    z_stream zstream;
-} PNGDecContext;
-
 /* Mask to determine which y pixels can be written in a pass */
 static const uint8_t png_pass_dsp_ymask[NB_PASSES] = {
     0xff, 0xff, 0x0f, 0xcc, 0x33, 0xff, 0x55,
@@ -134,7 +100,23 @@ static void png_put_interlaced_row(uint8_t *dst, int width,
     }
 }
 
-void ff_add_png_paeth_prediction(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)
+// 0x7f7f7f7f or 0x7f7f7f7f7f7f7f7f or whatever, depending on the cpu's native arithmetic size
+#define pb_7f (~0UL/255 * 0x7f)
+#define pb_80 (~0UL/255 * 0x80)
+
+static void add_bytes_l2_c(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w)
+{
+    long i;
+    for(i=0; i<=w-sizeof(long); i+=sizeof(long)){
+        long a = *(long*)(src1+i);
+        long b = *(long*)(src2+i);
+        *(long*)(dst+i) = ((a&pb_7f) + (b&pb_7f)) ^ ((a^b)&pb_80);
+    }
+    for(; i<w; i++)
+        dst[i] = src1[i]+src2[i];
+}
+
+static void add_paeth_prediction_c(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)
 {
     int i;
     for(i = 0; i < w; i++) {
@@ -191,7 +173,7 @@ void ff_add_png_paeth_prediction(uint8_t *dst, uint8_t *src, uint8_t *top, int w
     }
 
 /* NOTE: 'dst' can be equal to 'last' */
-static void png_filter_row(DSPContext *dsp, uint8_t *dst, int filter_type,
+static void png_filter_row(PNGDecContext *s, uint8_t *dst, int filter_type,
                            uint8_t *src, uint8_t *last, int size, int bpp)
 {
     int i, p, r, g, b, a;
@@ -217,7 +199,7 @@ static void png_filter_row(DSPContext *dsp, uint8_t *dst, int filter_type,
         }
         break;
     case PNG_FILTER_VALUE_UP:
-        dsp->add_bytes_l2(dst, src, last, size);
+        s->add_bytes_l2(dst, src, last, size);
         break;
     case PNG_FILTER_VALUE_AVG:
         for(i = 0; i < bpp; i++) {
@@ -235,10 +217,10 @@ static void png_filter_row(DSPContext *dsp, uint8_t *dst, int filter_type,
         if(bpp > 1 && size > 4) {
             // would write off the end of the array if we let it process the last pixel with bpp=3
             int w = bpp==4 ? size : size-3;
-            dsp->add_png_paeth_prediction(dst+i, src+i, last+i, w-i, bpp);
+            s->add_paeth_prediction(dst+i, src+i, last+i, w-i, bpp);
             i = w;
         }
-        ff_add_png_paeth_prediction(dst+i, src+i, last+i, size-i, bpp);
+        add_paeth_prediction_c(dst+i, src+i, last+i, size-i, bpp);
         break;
     }
 }
@@ -291,7 +273,7 @@ static void png_handle_row(PNGDecContext *s)
         ptr = s->image_buf + s->image_linesize * s->y;
         /* need to swap bytes correctly for RGB_ALPHA */
         if (s->color_type == PNG_COLOR_TYPE_RGB_ALPHA) {
-            png_filter_row(&s->dsp, s->tmp_row, s->crow_buf[0], s->crow_buf + 1,
+            png_filter_row(s, s->tmp_row, s->crow_buf[0], s->crow_buf + 1,
                            s->last_row, s->row_size, s->bpp);
             convert_to_rgb32(ptr, s->tmp_row, s->width, s->filter_type == PNG_FILTER_TYPE_LOCO);
             FFSWAP(uint8_t*, s->last_row, s->tmp_row);
@@ -302,7 +284,7 @@ static void png_handle_row(PNGDecContext *s)
             else
                 last_row = ptr - s->image_linesize;
 
-            png_filter_row(&s->dsp, ptr, s->crow_buf[0], s->crow_buf + 1,
+            png_filter_row(s, ptr, s->crow_buf[0], s->crow_buf + 1,
                            last_row, s->row_size, s->bpp);
         }
         /* loco lags by 1 row so that it doesn't interfere with top prediction */
@@ -325,7 +307,7 @@ static void png_handle_row(PNGDecContext *s)
                    wait for the next one */
                 if (got_line)
                     break;
-                png_filter_row(&s->dsp, s->tmp_row, s->crow_buf[0], s->crow_buf + 1,
+                png_filter_row(s, s->tmp_row, s->crow_buf[0], s->crow_buf + 1,
                                s->last_row, s->pass_row_size, s->bpp);
                 FFSWAP(uint8_t*, s->last_row, s->tmp_row);
                 got_line = 1;
@@ -633,14 +615,21 @@ static int decode_frame(AVCodecContext *avctx,
     goto the_end;
 }
 
-static av_cold int png_dec_init(AVCodecContext *avctx){
+static av_cold int png_dec_init(AVCodecContext *avctx)
+{
     PNGDecContext *s = avctx->priv_data;
 
     s->current_picture = &s->picture1;
     s->last_picture = &s->picture2;
     avcodec_get_frame_defaults(&s->picture1);
     avcodec_get_frame_defaults(&s->picture2);
-    dsputil_init(&s->dsp, avctx);
+
+    ff_png_init_mmx(s);
+
+    if (!s->add_paeth_prediction)
+        s->add_paeth_prediction = add_paeth_prediction_c;
+    if (!s->add_bytes_l2)
+        s->add_bytes_l2 = add_bytes_l2_c;
 
     return 0;
 }
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index f8d456d3ea..6416e600a3 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -25,6 +25,7 @@ MMX-OBJS-$(CONFIG_MP2FLOAT_DECODER)    += x86/mpegaudiodec_mmx.o
 MMX-OBJS-$(CONFIG_MP3FLOAT_DECODER)    += x86/mpegaudiodec_mmx.o
 MMX-OBJS-$(CONFIG_MP3ON4FLOAT_DECODER) += x86/mpegaudiodec_mmx.o
 MMX-OBJS-$(CONFIG_MP3ADUFLOAT_DECODER) += x86/mpegaudiodec_mmx.o
+MMX-OBJS-$(CONFIG_PNG_DECODER)         += x86/png_mmx.o
 MMX-OBJS-$(CONFIG_ENCODERS)            += x86/dsputilenc_mmx.o
 YASM-OBJS-$(CONFIG_ENCODERS)           += x86/dsputilenc_yasm.o
 MMX-OBJS-$(CONFIG_GPL)                 += x86/idct_mmx.o
diff --git a/libavcodec/x86/dsputil_mmx.c b/libavcodec/x86/dsputil_mmx.c
index 1b17d0ec7e..985a15d2f1 100644
--- a/libavcodec/x86/dsputil_mmx.c
+++ b/libavcodec/x86/dsputil_mmx.c
@@ -579,28 +579,6 @@ static void add_bytes_mmx(uint8_t *dst, uint8_t *src, int w){
         dst[i+0] += src[i+0];
 }
 
-static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w){
-    x86_reg i=0;
-    __asm__ volatile(
-        "jmp 2f                         \n\t"
-        "1:                             \n\t"
-        "movq   (%2, %0), %%mm0         \n\t"
-        "movq  8(%2, %0), %%mm1         \n\t"
-        "paddb  (%3, %0), %%mm0         \n\t"
-        "paddb 8(%3, %0), %%mm1         \n\t"
-        "movq %%mm0,  (%1, %0)          \n\t"
-        "movq %%mm1, 8(%1, %0)          \n\t"
-        "add $16, %0                    \n\t"
-        "2:                             \n\t"
-        "cmp %4, %0                     \n\t"
-        " js 1b                         \n\t"
-        : "+r" (i)
-        : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15)
-    );
-    for(; i<w; i++)
-        dst[i] = src1[i] + src2[i];
-}
-
 #if HAVE_7REGS && HAVE_TEN_OPERANDS
 static void add_hfyu_median_prediction_cmov(uint8_t *dst, const uint8_t *top, const uint8_t *diff, int w, int *left, int *left_top) {
     x86_reg w2 = -w;
@@ -876,80 +854,6 @@ static void draw_edges_mmx(uint8_t *buf, int wrap, int width, int height, int w,
     }
 }
 
-#define PAETH(cpu, abs3)\
-static void add_png_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\
-{\
-    x86_reg i = -bpp;\
-    x86_reg end = w-3;\
-    __asm__ volatile(\
-        "pxor      %%mm7, %%mm7 \n"\
-        "movd    (%1,%0), %%mm0 \n"\
-        "movd    (%2,%0), %%mm1 \n"\
-        "punpcklbw %%mm7, %%mm0 \n"\
-        "punpcklbw %%mm7, %%mm1 \n"\
-        "add       %4, %0 \n"\
-        "1: \n"\
-        "movq      %%mm1, %%mm2 \n"\
-        "movd    (%2,%0), %%mm1 \n"\
-        "movq      %%mm2, %%mm3 \n"\
-        "punpcklbw %%mm7, %%mm1 \n"\
-        "movq      %%mm2, %%mm4 \n"\
-        "psubw     %%mm1, %%mm3 \n"\
-        "psubw     %%mm0, %%mm4 \n"\
-        "movq      %%mm3, %%mm5 \n"\
-        "paddw     %%mm4, %%mm5 \n"\
-        abs3\
-        "movq      %%mm4, %%mm6 \n"\
-        "pminsw    %%mm5, %%mm6 \n"\
-        "pcmpgtw   %%mm6, %%mm3 \n"\
-        "pcmpgtw   %%mm5, %%mm4 \n"\
-        "movq      %%mm4, %%mm6 \n"\
-        "pand      %%mm3, %%mm4 \n"\
-        "pandn     %%mm3, %%mm6 \n"\
-        "pandn     %%mm0, %%mm3 \n"\
-        "movd    (%3,%0), %%mm0 \n"\
-        "pand      %%mm1, %%mm6 \n"\
-        "pand      %%mm4, %%mm2 \n"\
-        "punpcklbw %%mm7, %%mm0 \n"\
-        "movq      %6,    %%mm5 \n"\
-        "paddw     %%mm6, %%mm0 \n"\
-        "paddw     %%mm2, %%mm3 \n"\
-        "paddw     %%mm3, %%mm0 \n"\
-        "pand      %%mm5, %%mm0 \n"\
-        "movq      %%mm0, %%mm3 \n"\
-        "packuswb  %%mm3, %%mm3 \n"\
-        "movd      %%mm3, (%1,%0) \n"\
-        "add       %4, %0 \n"\
-        "cmp       %5, %0 \n"\
-        "jle 1b \n"\
-        :"+r"(i)\
-        :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\
-         "m"(ff_pw_255)\
-        :"memory"\
-    );\
-}
-
-#define ABS3_MMX2\
-        "psubw     %%mm5, %%mm7 \n"\
-        "pmaxsw    %%mm7, %%mm5 \n"\
-        "pxor      %%mm6, %%mm6 \n"\
-        "pxor      %%mm7, %%mm7 \n"\
-        "psubw     %%mm3, %%mm6 \n"\
-        "psubw     %%mm4, %%mm7 \n"\
-        "pmaxsw    %%mm6, %%mm3 \n"\
-        "pmaxsw    %%mm7, %%mm4 \n"\
-        "pxor      %%mm7, %%mm7 \n"
-
-#define ABS3_SSSE3\
-        "pabsw     %%mm3, %%mm3 \n"\
-        "pabsw     %%mm4, %%mm4 \n"\
-        "pabsw     %%mm5, %%mm5 \n"
-
-PAETH(mmx2, ABS3_MMX2)
-#if HAVE_SSSE3
-PAETH(ssse3, ABS3_SSSE3)
-#endif
-
 #define QPEL_V_LOW(m3,m4,m5,m6, pw_20, pw_3, rnd, in0, in1, in2, in7, out, OP)\
         "paddw " #m4 ", " #m3 "           \n\t" /* x1 */\
         "movq "MANGLE(ff_pw_20)", %%mm4   \n\t" /* 20 */\
@@ -2537,7 +2441,6 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 #endif
 
         c->add_bytes= add_bytes_mmx;
-        c->add_bytes_l2= add_bytes_l2_mmx;
 
         if (!h264_high_depth)
         c->draw_edges = draw_edges_mmx;
@@ -2658,7 +2561,6 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
                 c->add_hfyu_median_prediction = add_hfyu_median_prediction_cmov;
 #endif
 
-            c->add_png_paeth_prediction= add_png_paeth_prediction_mmx2;
         } else if (mm_flags & AV_CPU_FLAG_3DNOW) {
             c->prefetch = prefetch_3dnow;
 
@@ -2772,7 +2674,6 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
             H264_QPEL_FUNCS(3, 2, ssse3);
             H264_QPEL_FUNCS(3, 3, ssse3);
             }
-            c->add_png_paeth_prediction= add_png_paeth_prediction_ssse3;
 #if HAVE_YASM
             if (!h264_high_depth) {
             c->put_h264_chroma_pixels_tab[0]= ff_put_h264_chroma_mc8_ssse3_rnd;
diff --git a/libavcodec/x86/png_mmx.c b/libavcodec/x86/png_mmx.c
new file mode 100644
index 0000000000..cdf035962c
--- /dev/null
+++ b/libavcodec/x86/png_mmx.c
@@ -0,0 +1,141 @@
+/*
+ * MMX optimized PNG utils
+ * Copyright (c) 2008 Loren Merritt
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ *
+ */
+
+#include "libavutil/cpu.h"
+#include "libavutil/x86_cpu.h"
+#include "libavcodec/dsputil.h"
+#include "libavcodec/png.h"
+#include "dsputil_mmx.h"
+
+//#undef NDEBUG
+//#include <assert.h>
+
+static void add_bytes_l2_mmx(uint8_t *dst, uint8_t *src1, uint8_t *src2, int w)
+{
+    x86_reg i=0;
+    __asm__ volatile(
+        "jmp 2f                         \n\t"
+        "1:                             \n\t"
+        "movq   (%2, %0), %%mm0         \n\t"
+        "movq  8(%2, %0), %%mm1         \n\t"
+        "paddb  (%3, %0), %%mm0         \n\t"
+        "paddb 8(%3, %0), %%mm1         \n\t"
+        "movq %%mm0,  (%1, %0)          \n\t"
+        "movq %%mm1, 8(%1, %0)          \n\t"
+        "add $16, %0                    \n\t"
+        "2:                             \n\t"
+        "cmp %4, %0                     \n\t"
+        " js 1b                         \n\t"
+        : "+r" (i)
+        : "r"(dst), "r"(src1), "r"(src2), "r"((x86_reg)w-15)
+    );
+    for(; i<w; i++)
+        dst[i] = src1[i] + src2[i];
+}
+
+#define PAETH(cpu, abs3)\
+static void add_paeth_prediction_##cpu(uint8_t *dst, uint8_t *src, uint8_t *top, int w, int bpp)\
+{\
+    x86_reg i = -bpp;\
+    x86_reg end = w-3;\
+    __asm__ volatile(\
+        "pxor      %%mm7, %%mm7 \n"\
+        "movd    (%1,%0), %%mm0 \n"\
+        "movd    (%2,%0), %%mm1 \n"\
+        "punpcklbw %%mm7, %%mm0 \n"\
+        "punpcklbw %%mm7, %%mm1 \n"\
+        "add       %4, %0 \n"\
+        "1: \n"\
+        "movq      %%mm1, %%mm2 \n"\
+        "movd    (%2,%0), %%mm1 \n"\
+        "movq      %%mm2, %%mm3 \n"\
+        "punpcklbw %%mm7, %%mm1 \n"\
+        "movq      %%mm2, %%mm4 \n"\
+        "psubw     %%mm1, %%mm3 \n"\
+        "psubw     %%mm0, %%mm4 \n"\
+        "movq      %%mm3, %%mm5 \n"\
+        "paddw     %%mm4, %%mm5 \n"\
+        abs3\
+        "movq      %%mm4, %%mm6 \n"\
+        "pminsw    %%mm5, %%mm6 \n"\
+        "pcmpgtw   %%mm6, %%mm3 \n"\
+        "pcmpgtw   %%mm5, %%mm4 \n"\
+        "movq      %%mm4, %%mm6 \n"\
+        "pand      %%mm3, %%mm4 \n"\
+        "pandn     %%mm3, %%mm6 \n"\
+        "pandn     %%mm0, %%mm3 \n"\
+        "movd    (%3,%0), %%mm0 \n"\
+        "pand      %%mm1, %%mm6 \n"\
+        "pand      %%mm4, %%mm2 \n"\
+        "punpcklbw %%mm7, %%mm0 \n"\
+        "movq      %6,    %%mm5 \n"\
+        "paddw     %%mm6, %%mm0 \n"\
+        "paddw     %%mm2, %%mm3 \n"\
+        "paddw     %%mm3, %%mm0 \n"\
+        "pand      %%mm5, %%mm0 \n"\
+        "movq      %%mm0, %%mm3 \n"\
+        "packuswb  %%mm3, %%mm3 \n"\
+        "movd      %%mm3, (%1,%0) \n"\
+        "add       %4, %0 \n"\
+        "cmp       %5, %0 \n"\
+        "jle 1b \n"\
+        :"+r"(i)\
+        :"r"(dst), "r"(top), "r"(src), "r"((x86_reg)bpp), "g"(end),\
+         "m"(ff_pw_255)\
+        :"memory"\
+    );\
+}
+
+#define ABS3_MMX2\
+        "psubw     %%mm5, %%mm7 \n"\
+        "pmaxsw    %%mm7, %%mm5 \n"\
+        "pxor      %%mm6, %%mm6 \n"\
+        "pxor      %%mm7, %%mm7 \n"\
+        "psubw     %%mm3, %%mm6 \n"\
+        "psubw     %%mm4, %%mm7 \n"\
+        "pmaxsw    %%mm6, %%mm3 \n"\
+        "pmaxsw    %%mm7, %%mm4 \n"\
+        "pxor      %%mm7, %%mm7 \n"
+
+#define ABS3_SSSE3\
+        "pabsw     %%mm3, %%mm3 \n"\
+        "pabsw     %%mm4, %%mm4 \n"\
+        "pabsw     %%mm5, %%mm5 \n"
+
+PAETH(mmx2, ABS3_MMX2)
+#if HAVE_SSSE3
+PAETH(ssse3, ABS3_SSSE3)
+#endif
+
+void ff_png_init_mmx(PNGDecContext *s)
+{
+    int mm_flags = av_get_cpu_flags();
+
+    if (mm_flags & AV_CPU_FLAG_MMX2) {
+        s->add_bytes_l2 = add_bytes_l2_mmx;
+        s->add_paeth_prediction = add_paeth_prediction_mmx2;
+#if HAVE_SSSE3
+        if (mm_flags & AV_CPU_FLAG_SSSE3)
+            s->add_paeth_prediction = add_paeth_prediction_ssse3;
+#endif
+    }
+}