- video_chromas: more SSE2 and MMX support and optimization, added SSE2 i420 -> RGB acceleration

configure.ac

@@ -1274,7 +1274,7 @@ MMXEXT_MODULES="memcpymmxext"
 #MMXEXT_MODULES="${MMXEXT_MODULES} idctmmxext motionmmxext"
 THREEDNOW_MODULES="memcpy3dn"
 SSE_MODULES=""
-SSE2_MODULES="i420_yuy2_sse2"
+SSE2_MODULES="i420_rgb_sse2 i420_yuy2_sse2"
 ALTIVEC_MODULES="memcpyaltivec i420_yuy2_altivec"
 #ALTIVEC_MODULES="${ALTIVEC_MODULES} idctaltivec motionaltivec"
 
@@ -1325,7 +1325,7 @@ AC_CACHE_CHECK([if \$CC groks SSE2 intrinsics],
                     [ac_cv_c_sse2_intrinsics=no])])
 if test "${ac_cv_c_sse2_intrinsics}" != "no"; then
   AC_DEFINE(HAVE_SSE2_INTRINSICS, 1, Define if SSE2 intrinsics are available.)
-  dnl VLC_ADD_CFLAGS([i420_rgb_sse2],[-msse2])
+  VLC_ADD_CFLAGS([i420_rgb_sse2],[-msse2])
 fi
 
 AC_CACHE_CHECK([if \$CC groks MMX inline assembly],

modules/video_chroma/Modules.am

@@ -13,6 +13,13 @@ SOURCES_i420_rgb_mmx = \
 	i420_rgb_mmx.h \
 	$(NULL)
 
+SOURCES_i420_rgb_sse2 = \
+	i420_rgb.c \
+	i420_rgb.h \
+	i420_rgb16.c \
+	i420_rgb_mmx.h \
+	$(NULL)
+
 SOURCES_i420_yuy2 = \
 	i420_yuy2.c \
 	i420_yuy2.h \

modules/video_chroma/i420_rgb.c

@@ -4,7 +4,8 @@
  * Copyright (C) 2000, 2001, 2004 the VideoLAN team
  * $Id$
  *
- * Author: Sam Hocevar <sam@zoy.org>
+ * Authors: Sam Hocevar <sam@zoy.org>
+ *          Damien Fouilleul <damienf@videolan.org>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -72,6 +73,11 @@ vlc_module_begin();
                         "RV15,RV16,RV24,RV32 conversions") );
     set_capability( "chroma", 100 );
     add_requirement( MMX );
+#elif defined (MODULE_NAME_IS_i420_rgb_sse2)
+    set_description( _( "SSE2 I420,IYUV,YV12 to "
+                        "RV15,RV16,RV24,RV32 conversions") );
+    set_capability( "chroma", 120 );
+    add_requirement( SSE2 );
 #endif
     set_callbacks( Activate, Deactivate );
 vlc_module_end();
@@ -107,19 +113,30 @@ static int Activate( vlc_object_t *p_this )
 #endif
                 case VLC_FOURCC('R','V','1','5'):
                 case VLC_FOURCC('R','V','1','6'):
-#if defined (MODULE_NAME_IS_i420_rgb_mmx)
+#if ! defined (MODULE_NAME_IS_i420_rgb)
                     /* If we don't have support for the bitmasks, bail out */
-                    if( ( p_vout->output.i_rmask != 0x7c00
-                           || p_vout->output.i_gmask != 0x03e0
-                           || p_vout->output.i_bmask != 0x001f )
-                     && ( p_vout->output.i_rmask != 0xf800
-                           || p_vout->output.i_gmask != 0x07e0
-                           || p_vout->output.i_bmask != 0x001f ) )
+                    if( ( p_vout->output.i_rmask == 0x7c00
+                       && p_vout->output.i_gmask == 0x03e0
+                       && p_vout->output.i_bmask == 0x001f ) )
                     {
-                        return -1;
+                        /* R5G5B6 pixel format */
+                        msg_Dbg(p_this, "RGB pixel format is R5G5B5");
+                        p_vout->chroma.pf_convert = E_(I420_R5G5B5);
                     }
-#endif
+                    else if( ( p_vout->output.i_rmask == 0xf800
+                            && p_vout->output.i_gmask == 0x07e0
+                            && p_vout->output.i_bmask == 0x001f ) )
+                    {
+                        /* R5G6B5 pixel format */
+                        msg_Dbg(p_this, "RGB pixel format is R5G6B5");
+                        p_vout->chroma.pf_convert = E_(I420_R5G6B5);
+                    }
+                    else
+                        return -1;
+#else
+                    // generic C chroma converter */
                     p_vout->chroma.pf_convert = E_(I420_RGB16);
+#endif
                     break;
 
 #if 0
@@ -128,16 +145,30 @@ static int Activate( vlc_object_t *p_this )
 #endif
 
                 case VLC_FOURCC('R','V','3','2'):
-#if defined (MODULE_NAME_IS_i420_rgb_mmx)
+#if ! defined (MODULE_NAME_IS_i420_rgb)
                     /* If we don't have support for the bitmasks, bail out */
-                    if( p_vout->output.i_rmask != 0x00ff0000
-                         || p_vout->output.i_gmask != 0x0000ff00
-                         || p_vout->output.i_bmask != 0x000000ff )
+                    if( p_vout->output.i_rmask == 0x00ff0000
+                     && p_vout->output.i_gmask == 0x0000ff00
+                     && p_vout->output.i_bmask == 0x000000ff )
                     {
-                        return -1;
+                        /* A8R8G8B8 pixel format */
+                        msg_Dbg(p_this, "RGB pixel format is A8R8G8B8");
+                        p_vout->chroma.pf_convert = E_(I420_A8R8G8B8);
                     }
-#endif
+                    else if( p_vout->output.i_rmask == 0x0000ff00
+                          && p_vout->output.i_gmask == 0x00ff0000
+                          && p_vout->output.i_bmask == 0xff000000 )
+                    {
+                        /* B8G8R8A8 pixel format */
+                        msg_Dbg(p_this, "RGB pixel format is B8G8R8A8");
+                        p_vout->chroma.pf_convert = E_(I420_B8G8R8A8);
+                    }
+                    else
+                        return -1;
+#else
+                    // generic C chroma converter */
                     p_vout->chroma.pf_convert = E_(I420_RGB32);
+#endif
                     break;
 
                 default:

modules/video_chroma/i420_rgb.h

@@ -58,9 +58,14 @@ struct chroma_sys_t
 #ifdef MODULE_NAME_IS_i420_rgb
 void E_(I420_RGB8)         ( vout_thread_t *, picture_t *, picture_t * );
 void E_(I420_RGB16_dither) ( vout_thread_t *, picture_t *, picture_t * );
-#endif
 void E_(I420_RGB16)        ( vout_thread_t *, picture_t *, picture_t * );
 void E_(I420_RGB32)        ( vout_thread_t *, picture_t *, picture_t * );
+#else // if defined(MODULE_NAME_IS_i420_rgb_mmx)
+void E_(I420_R5G5B5)       ( vout_thread_t *, picture_t *, picture_t * );
+void E_(I420_R5G6B5)       ( vout_thread_t *, picture_t *, picture_t * );
+void E_(I420_A8R8G8B8)     ( vout_thread_t *, picture_t *, picture_t * );
+void E_(I420_B8G8R8A8)     ( vout_thread_t *, picture_t *, picture_t * );
+#endif
 
 /*****************************************************************************
  * CONVERT_*_PIXEL: pixel conversion macros

modules/video_chroma/i420_rgb16.c

@@ -5,6 +5,7 @@
  * $Id$
  *
  * Authors: Samuel Hocevar <sam@zoy.org>
+ *          Damien Fouilleul <damienf@videolan.org>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -38,6 +39,11 @@
 #       include <mmintrin.h>
 #   endif
 #   include "i420_rgb_mmx.h"
+#elif defined (MODULE_NAME_IS_i420_rgb_sse2)
+#   if defined(HAVE_SSE2_INTRINSICS)
+#       include <emmintrin.h>
+#   endif
+#   include "i420_rgb_mmx.h"
 #endif
 
 static void SetOffset( int, int, int, int, vlc_bool_t *,
@@ -202,6 +208,9 @@ void E_(I420_RGB16_dither)( vout_thread_t *p_vout, picture_t *p_src,
  *  - input: 2 lines (2 Y lines, 1 U/V line)
  *  - output: 1 line
  *****************************************************************************/
+
+#if defined (MODULE_NAME_IS_i420_rgb)
+
 void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
                                             picture_t *p_dest )
 {
@@ -220,12 +229,10 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
     int         i_scale_count;                       /* scale modulo counter */
     int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
     uint16_t *  p_pic_start;       /* beginning of the current line for copy */
-#if defined (MODULE_NAME_IS_i420_rgb)
     int         i_uval, i_vval;                           /* U and V samples */
     int         i_red, i_green, i_blue;          /* U and V modified samples */
     uint16_t *  p_yuv = p_vout->chroma.p_sys->p_rgb16;
     uint16_t *  p_ybase;                     /* Y dependant conversion table */
-#endif
 
     /* Conversion buffer pointer */
     uint16_t *  p_buffer_start = (uint16_t*)p_vout->chroma.p_sys->p_buffer;
@@ -268,7 +275,6 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
         p_pic_start = p_pic;
         p_buffer = b_hscale ? p_buffer_start : p_pic;
 
-#if defined (MODULE_NAME_IS_i420_rgb)
         for ( i_x = p_vout->render.i_width / 8; i_x--; )
         {
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
@@ -276,131 +282,20 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
         }
-#elif defined (MODULE_NAME_IS_i420_rgb_mmx)
-        if( p_vout->output.i_rmask == 0x7c00 )
-        {
-            /* 15bpp 5/5/5 */
-            for ( i_x = p_vout->render.i_width / 8; i_x--; )
-            {
-#   if defined (HAVE_MMX_INTRINSICS)
-                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
-                uint64_t tmp64;
-                INTRINSICS_INIT_16
-                INTRINSICS_YUV_MUL
-                INTRINSICS_YUV_ADD
-                INTRINSICS_UNPACK_15
-#   else
-                __asm__( MMX_INIT_16
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-
-                __asm__( ".p2align 3"
-                         MMX_YUV_MUL
-                         MMX_YUV_ADD
-                         MMX_UNPACK_15
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
-
-                p_y += 8;
-                p_u += 4;
-                p_v += 4;
-                p_buffer += 8;
-            }
-        }
-        else
-        {
-            /* 16bpp 5/6/5 */
-            for ( i_x = p_vout->render.i_width / 8; i_x--; )
-            {
-#   if defined (HAVE_MMX_INTRINSICS)
-                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
-                uint64_t tmp64;
-                INTRINSICS_INIT_16
-                INTRINSICS_YUV_MUL
-                INTRINSICS_YUV_ADD
-                INTRINSICS_UNPACK_16
-#   else
-                __asm__( MMX_INIT_16
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-
-                __asm__( ".p2align 3"
-                         MMX_YUV_MUL
-                         MMX_YUV_ADD
-                         MMX_UNPACK_16
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
-
-                p_y += 8;
-                p_u += 4;
-                p_v += 4;
-                p_buffer += 8;
-            }
-        }
-#endif
 
         /* Here we do some unaligned reads and duplicate conversions, but
          * at least we have all the pixels */
         if( i_rewind )
         {
-#if defined (MODULE_NAME_IS_i420_rgb_mmx)
-#   if defined (HAVE_MMX_INTRINSICS)
-            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
-            uint64_t tmp64;
-#   endif
-#endif
             p_y -= i_rewind;
             p_u -= i_rewind >> 1;
             p_v -= i_rewind >> 1;
             p_buffer -= i_rewind;
-#if defined (MODULE_NAME_IS_i420_rgb)
+
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
             CONVERT_YUV_PIXEL(2);  CONVERT_Y_PIXEL(2);
-#elif defined (MODULE_NAME_IS_i420_rgb_mmx)
-
-#   if defined (HAVE_MMX_INTRINSICS)
-            INTRINSICS_INIT_16
-#   else
-            __asm__( MMX_INIT_16
-                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
-
-            if( p_vout->output.i_rmask == 0x7c00 )
-            {
-                /* 15bpp 5/5/5 */
-#   if defined (HAVE_MMX_INTRINSICS)
-                INTRINSICS_YUV_MUL
-                INTRINSICS_YUV_ADD
-                INTRINSICS_UNPACK_15
-#   else
-                __asm__( ".p2align 3"
-                         MMX_YUV_MUL
-                         MMX_YUV_ADD
-                         MMX_UNPACK_15
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
-            }
-            else
-            {
-#   if defined (HAVE_MMX_INTRINSICS)
-                INTRINSICS_YUV_MUL
-                INTRINSICS_YUV_ADD
-                INTRINSICS_UNPACK_16
-#   else
-                /* 16bpp 5/6/5 */
-                __asm__( ".p2align 3"
-                         MMX_YUV_MUL
-                         MMX_YUV_ADD
-                         MMX_UNPACK_16
-                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
-            }
-
-            p_y += 8;
-            p_u += 4;
-            p_v += 4;
-            p_buffer += 8;
-#endif
         }
         SCALE_WIDTH;
         SCALE_HEIGHT( 420, 2 );
@@ -414,21 +309,13 @@ void E_(I420_RGB16)( vout_thread_t *p_vout, picture_t *p_src,
     }
 }
 
-/*****************************************************************************
- * I420_RGB32: color YUV 4:2:0 to RGB 32 bpp
- *****************************************************************************
- * Horizontal alignment needed:
- *  - input: 8 pixels (8 Y bytes, 4 U/V bytes), margins not allowed
- *  - output: 1 pixel (2 bytes), margins allowed
- * Vertical alignment needed:
- *  - input: 2 lines (2 Y lines, 1 U/V line)
- *  - output: 1 line
- *****************************************************************************/
-void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
+#else // defined (MODULE_NAME_IS_i420_rgb_mmx)
+
+void E_(I420_R5G5B5)( vout_thread_t *p_vout, picture_t *p_src,
                                             picture_t *p_dest )
 {
     /* We got this one from the old arguments */
-    uint32_t *p_pic = (uint32_t*)p_dest->p->p_pixels;
+    uint16_t *p_pic = (uint16_t*)p_dest->p->p_pixels;
     uint8_t  *p_y   = p_src->Y_PIXELS;
     uint8_t  *p_u   = p_src->U_PIXELS;
     uint8_t  *p_v   = p_src->V_PIXELS;
@@ -441,17 +328,11 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
     int         i_rewind;
     int         i_scale_count;                       /* scale modulo counter */
     int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
-    uint32_t *  p_pic_start;       /* beginning of the current line for copy */
-#if defined (MODULE_NAME_IS_i420_rgb)
-    int         i_uval, i_vval;                           /* U and V samples */
-    int         i_red, i_green, i_blue;          /* U and V modified samples */
-    uint32_t *  p_yuv = p_vout->chroma.p_sys->p_rgb32;
-    uint32_t *  p_ybase;                     /* Y dependant conversion table */
-#endif
+    uint16_t *  p_pic_start;       /* beginning of the current line for copy */
 
     /* Conversion buffer pointer */
-    uint32_t *  p_buffer_start = (uint32_t*)p_vout->chroma.p_sys->p_buffer;
-    uint32_t *  p_buffer;
+    uint16_t *  p_buffer_start = (uint16_t*)p_vout->chroma.p_sys->p_buffer;
+    uint16_t *  p_buffer;
 
     /* Offset array pointer */
     int *       p_offset_start = p_vout->chroma.p_sys->p_offset;
@@ -464,15 +345,6 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
 
     i_right_margin = p_dest->p->i_pitch - p_dest->p->i_visible_pitch;
 
-    if( p_vout->render.i_width & 7 )
-    {
-        i_rewind = 8 - ( p_vout->render.i_width & 7 );
-    }
-    else
-    {
-        i_rewind = 0;
-    }
-
     /* Rule: when a picture of size (x1,y1) with aspect ratio r1 is rendered
      * on a picture of size (x2,y2) with aspect ratio r2, if x1 grows to x1'
      * then y1 grows to y1' = x1' * y2/x2 * r2/r1 */
@@ -480,11 +352,191 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
                p_vout->output.i_width, p_vout->output.i_height,
                &b_hscale, &i_vscale, p_offset_start );
 
+
     /*
      * Perform conversion
      */
     i_scale_count = ( i_vscale == 1 ) ?
                     p_vout->output.i_height : p_vout->render.i_height;
+
+#if defined (MODULE_NAME_IS_i420_rgb_sse2)
+
+    if( p_vout->render.i_width & 15 )
+    {
+        i_rewind = 16 - ( p_vout->render.i_width & 15 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    /*
+    ** SSE2 128 bits fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+
+    p_buffer = b_hscale ? p_buffer_start : p_pic;
+    if( 0 == (15 & (p_src->p[Y_PLANE].i_pitch|
+                    p_dest->p->i_pitch|
+                    ((int)p_y)|
+                    ((int)p_buffer))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+
+            for ( i_x = p_vout->render.i_width/16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_ALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_15_ALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+                SSE2_INTRINSICS_INIT_16_ALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_15_ALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_15_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_16_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_15_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 2 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+        /* make sure all SSE2 stores are visible thereafter */
+#if defined (CAN_COMPILE_SSE2)
+        __asm__ __volatile__ ( "sfence" );
+#else
+        _mm_sfence();
+#endif
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+            for ( i_x = p_vout->render.i_width/16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_15_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+                SSE2_INTRINSICS_INIT_16_UNALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_15_UNALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_15_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_16_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_15_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 2 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+    }
+#else // defined (MODULE_NAME_IS_i420_rgb_mmx)
+
+    if( p_vout->render.i_width & 7 )
+    {
+        i_rewind = 8 - ( p_vout->render.i_width & 7 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
     for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
     {
         p_pic_start = p_pic;
@@ -492,81 +544,62 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
 
         for ( i_x = p_vout->render.i_width / 8; i_x--; )
         {
-#if defined (MODULE_NAME_IS_i420_rgb)
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-#elif defined (MODULE_NAME_IS_i420_rgb_mmx)
-#   if defined (HAVE_MMX_INTRINSICS)
-            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
-            uint64_t tmp64;
-            INTRINSICS_INIT_32
-            INTRINSICS_YUV_MUL
-            INTRINSICS_YUV_ADD
-            INTRINSICS_UNPACK_32
-#   else
-            __asm__( MMX_INIT_32
-                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-
+#if defined (CAN_COMPILE_MMX)
             __asm__( ".p2align 3"
+                     MMX_INIT_16
                      MMX_YUV_MUL
                      MMX_YUV_ADD
-                     MMX_UNPACK_32
+                     MMX_UNPACK_15
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
+#else
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+            uint64_t tmp64;
+            MMX_INTRINSICS_INIT_16
+            MMX_INTRINSICS_YUV_MUL
+            MMX_INTRINSICS_YUV_ADD
+            MMX_INTRINSICS_UNPACK_15
+#endif
 
             p_y += 8;
             p_u += 4;
             p_v += 4;
             p_buffer += 8;
-#endif
         }
 
         /* Here we do some unaligned reads and duplicate conversions, but
          * at least we have all the pixels */
         if( i_rewind )
         {
-#if defined (MODULE_NAME_IS_i420_rgb_mmx)
-#   if defined (HAVE_MMX_INTRINSICS)
-            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
-            uint64_t tmp64;
-#   endif
-#endif
             p_y -= i_rewind;
             p_u -= i_rewind >> 1;
             p_v -= i_rewind >> 1;
             p_buffer -= i_rewind;
-#if defined (MODULE_NAME_IS_i420_rgb)
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
-#elif defined (MODULE_NAME_IS_i420_rgb_mmx)
-#   if defined (HAVE_MMX_INTRINSICS)
-            INTRINSICS_INIT_32
-            INTRINSICS_YUV_MUL
-            INTRINSICS_YUV_ADD
-            INTRINSICS_UNPACK_32
-#   else
-            __asm__( MMX_INIT_32
-                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
 
+#if defined (CAN_COMPILE_MMX)
             __asm__( ".p2align 3"
+                     MMX_INIT_16
                      MMX_YUV_MUL
                      MMX_YUV_ADD
-                     MMX_UNPACK_32
+                     MMX_UNPACK_15
                      : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
-#   endif
+#else
+            {
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                uint64_t tmp64;
 
+                MMX_INTRINSICS_INIT_16
+                MMX_INTRINSICS_YUV_MUL
+                MMX_INTRINSICS_YUV_ADD
+                MMX_INTRINSICS_UNPACK_15
+            }
+#endif
             p_y += 8;
             p_u += 4;
             p_v += 4;
             p_buffer += 8;
-#endif
         }
         SCALE_WIDTH;
-        SCALE_HEIGHT( 420, 4 );
+        SCALE_HEIGHT( 420, 2 );
 
         p_y += i_source_margin;
         if( i_y % 2 )
@@ -575,8 +608,1081 @@ void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
             p_v += i_source_margin_c;
         }
     }
+    /* re-enable FPU registers */
+#if defined (CAN_COMPILE_MMX)
+    __asm__ __volatile__ ( "emms" );
+#else
+    _mm_empty();
+#endif
+
+#endif
 }
 
+void E_(I420_R5G6B5)( vout_thread_t *p_vout, picture_t *p_src,
+                                            picture_t *p_dest )
+{
+    /* We got this one from the old arguments */
+    uint16_t *p_pic = (uint16_t*)p_dest->p->p_pixels;
+    uint8_t  *p_y   = p_src->Y_PIXELS;
+    uint8_t  *p_u   = p_src->U_PIXELS;
+    uint8_t  *p_v   = p_src->V_PIXELS;
+
+    vlc_bool_t  b_hscale;                         /* horizontal scaling type */
+    unsigned int i_vscale;                          /* vertical scaling type */
+    unsigned int i_x, i_y;                /* horizontal and vertical indexes */
+
+    int         i_right_margin;
+    int         i_rewind;
+    int         i_scale_count;                       /* scale modulo counter */
+    int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
+    uint16_t *  p_pic_start;       /* beginning of the current line for copy */
+
+    /* Conversion buffer pointer */
+    uint16_t *  p_buffer_start = (uint16_t*)p_vout->chroma.p_sys->p_buffer;
+    uint16_t *  p_buffer;
+
+    /* Offset array pointer */
+    int *       p_offset_start = p_vout->chroma.p_sys->p_offset;
+    int *       p_offset;
+
+    const int i_source_margin = p_src->p[0].i_pitch
+                                 - p_src->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_src->p[1].i_pitch
+                                 - p_src->p[1].i_visible_pitch;
+
+    i_right_margin = p_dest->p->i_pitch - p_dest->p->i_visible_pitch;
+
+    /* Rule: when a picture of size (x1,y1) with aspect ratio r1 is rendered
+     * on a picture of size (x2,y2) with aspect ratio r2, if x1 grows to x1'
+     * then y1 grows to y1' = x1' * y2/x2 * r2/r1 */
+    SetOffset( p_vout->render.i_width, p_vout->render.i_height,
+               p_vout->output.i_width, p_vout->output.i_height,
+               &b_hscale, &i_vscale, p_offset_start );
+
+
+    /*
+     * Perform conversion
+     */
+    i_scale_count = ( i_vscale == 1 ) ?
+                    p_vout->output.i_height : p_vout->render.i_height;
+
+#if defined (MODULE_NAME_IS_i420_rgb_sse2)
+
+    if( p_vout->render.i_width & 15 )
+    {
+        i_rewind = 16 - ( p_vout->render.i_width & 15 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    /*
+    ** SSE2 128 bits fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+
+    p_buffer = b_hscale ? p_buffer_start : p_pic;
+    if( 0 == (15 & (p_src->p[Y_PLANE].i_pitch|
+                    p_dest->p->i_pitch|
+                    ((int)p_y)|
+                    ((int)p_buffer))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+
+            for ( i_x = p_vout->render.i_width/16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_ALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_16_ALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+                SSE2_INTRINSICS_INIT_16_ALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_16_ALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_16_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_16_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_16_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 2 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+        /* make sure all SSE2 stores are visible thereafter */
+#if defined (CAN_COMPILE_SSE2)
+        __asm__ __volatile__ ( "sfence" );
+#else
+        _mm_sfence();
+#endif
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+            for ( i_x = p_vout->render.i_width/16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_16_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+                SSE2_INTRINSICS_INIT_16_UNALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_16_UNALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+
+#if defined (CAN_COMPILE_SSE2)
+                __asm__( ".p2align 3"
+                         SSE2_INIT_16_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_16_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_16_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_16_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 2 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+    }
+#else // defined (MODULE_NAME_IS_i420_rgb_mmx)
+
+    if( p_vout->render.i_width & 7 )
+    {
+        i_rewind = 8 - ( p_vout->render.i_width & 7 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = p_vout->render.i_width / 8; i_x--; )
+        {
+#if defined (CAN_COMPILE_MMX)
+            __asm__( ".p2align 3"
+                     MMX_INIT_16
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_16
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+            uint64_t tmp64;
+            MMX_INTRINSICS_INIT_16
+            MMX_INTRINSICS_YUV_MUL
+            MMX_INTRINSICS_YUV_ADD
+            MMX_INTRINSICS_UNPACK_16
+#endif
+
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+
+        /* Here we do some unaligned reads and duplicate conversions, but
+         * at least we have all the pixels */
+        if( i_rewind )
+        {
+            p_y -= i_rewind;
+            p_u -= i_rewind >> 1;
+            p_v -= i_rewind >> 1;
+            p_buffer -= i_rewind;
+
+#if defined (CAN_COMPILE_MMX)
+            __asm__( ".p2align 3"
+                     MMX_INIT_16
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_16
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            {
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                uint64_t tmp64;
+
+                MMX_INTRINSICS_INIT_16
+                MMX_INTRINSICS_YUV_MUL
+                MMX_INTRINSICS_YUV_ADD
+                MMX_INTRINSICS_UNPACK_16
+            }
+#endif
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+        SCALE_WIDTH;
+        SCALE_HEIGHT( 420, 2 );
+
+        p_y += i_source_margin;
+        if( i_y % 2 )
+        {
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+        }
+    }
+    /* re-enable FPU registers */
+#if defined (CAN_COMPILE_MMX)
+    __asm__ __volatile__ ( "emms" );
+#else
+    _mm_empty();
+#endif
+
+#endif
+}
+
+#endif
+
+/*****************************************************************************
+ * I420_RGB32: color YUV 4:2:0 to RGB 32 bpp
+ *****************************************************************************
+ * Horizontal alignment needed:
+ *  - input: 8 pixels (8 Y bytes, 4 U/V bytes), margins not allowed
+ *  - output: 1 pixel (2 bytes), margins allowed
+ * Vertical alignment needed:
+ *  - input: 2 lines (2 Y lines, 1 U/V line)
+ *  - output: 1 line
+ *****************************************************************************/
+
+#if defined (MODULE_NAME_IS_i420_rgb)
+
+void E_(I420_RGB32)( vout_thread_t *p_vout, picture_t *p_src,
+                                            picture_t *p_dest )
+{
+    /* We got this one from the old arguments */
+    uint32_t *p_pic = (uint32_t*)p_dest->p->p_pixels;
+    uint8_t  *p_y   = p_src->Y_PIXELS;
+    uint8_t  *p_u   = p_src->U_PIXELS;
+    uint8_t  *p_v   = p_src->V_PIXELS;
+
+    vlc_bool_t  b_hscale;                         /* horizontal scaling type */
+    unsigned int i_vscale;                          /* vertical scaling type */
+    unsigned int i_x, i_y;                /* horizontal and vertical indexes */
+
+    int         i_right_margin;
+    int         i_rewind;
+    int         i_scale_count;                       /* scale modulo counter */
+    int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
+    uint32_t *  p_pic_start;       /* beginning of the current line for copy */
+    int         i_uval, i_vval;                           /* U and V samples */
+    int         i_red, i_green, i_blue;          /* U and V modified samples */
+    uint32_t *  p_yuv = p_vout->chroma.p_sys->p_rgb32;
+    uint32_t *  p_ybase;                     /* Y dependant conversion table */
+
+    /* Conversion buffer pointer */
+    uint32_t *  p_buffer_start = (uint32_t*)p_vout->chroma.p_sys->p_buffer;
+    uint32_t *  p_buffer;
+
+    /* Offset array pointer */
+    int *       p_offset_start = p_vout->chroma.p_sys->p_offset;
+    int *       p_offset;
+
+    const int i_source_margin = p_src->p[0].i_pitch
+                                 - p_src->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_src->p[1].i_pitch
+                                 - p_src->p[1].i_visible_pitch;
+
+    i_right_margin = p_dest->p->i_pitch - p_dest->p->i_visible_pitch;
+
+    if( p_vout->render.i_width & 7 )
+    {
+        i_rewind = 8 - ( p_vout->render.i_width & 7 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    /* Rule: when a picture of size (x1,y1) with aspect ratio r1 is rendered
+     * on a picture of size (x2,y2) with aspect ratio r2, if x1 grows to x1'
+     * then y1 grows to y1' = x1' * y2/x2 * r2/r1 */
+    SetOffset( p_vout->render.i_width, p_vout->render.i_height,
+               p_vout->output.i_width, p_vout->output.i_height,
+               &b_hscale, &i_vscale, p_offset_start );
+
+    /*
+     * Perform conversion
+     */
+    i_scale_count = ( i_vscale == 1 ) ?
+                    p_vout->output.i_height : p_vout->render.i_height;
+    for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = p_vout->render.i_width / 8; i_x--; )
+        {
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+        }
+
+        /* Here we do some unaligned reads and duplicate conversions, but
+         * at least we have all the pixels */
+        if( i_rewind )
+        {
+            p_y -= i_rewind;
+            p_u -= i_rewind >> 1;
+            p_v -= i_rewind >> 1;
+            p_buffer -= i_rewind;
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+            CONVERT_YUV_PIXEL(4);  CONVERT_Y_PIXEL(4);
+        }
+        SCALE_WIDTH;
+        SCALE_HEIGHT( 420, 4 );
+
+        p_y += i_source_margin;
+        if( i_y % 2 )
+        {
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+        }
+    }
+}
+
+#else // defined (MODULE_NAME_IS_i420_rgb_mmx) || defined (MODULE_NAME_IS_i420_rgb_sse2)
+
+void E_(I420_A8R8G8B8)( vout_thread_t *p_vout, picture_t *p_src,
+                                            picture_t *p_dest )
+{
+    /* We got this one from the old arguments */
+    uint32_t *p_pic = (uint32_t*)p_dest->p->p_pixels;
+    uint8_t  *p_y   = p_src->Y_PIXELS;
+    uint8_t  *p_u   = p_src->U_PIXELS;
+    uint8_t  *p_v   = p_src->V_PIXELS;
+
+    vlc_bool_t  b_hscale;                         /* horizontal scaling type */
+    unsigned int i_vscale;                          /* vertical scaling type */
+    unsigned int i_x, i_y;                /* horizontal and vertical indexes */
+
+    int         i_right_margin;
+    int         i_rewind;
+    int         i_scale_count;                       /* scale modulo counter */
+    int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
+    uint32_t *  p_pic_start;       /* beginning of the current line for copy */
+    /* Conversion buffer pointer */
+    uint32_t *  p_buffer_start = (uint32_t*)p_vout->chroma.p_sys->p_buffer;
+    uint32_t *  p_buffer;
+
+    /* Offset array pointer */
+    int *       p_offset_start = p_vout->chroma.p_sys->p_offset;
+    int *       p_offset;
+
+    const int i_source_margin = p_src->p[0].i_pitch
+                                 - p_src->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_src->p[1].i_pitch
+                                 - p_src->p[1].i_visible_pitch;
+
+    i_right_margin = p_dest->p->i_pitch - p_dest->p->i_visible_pitch;
+
+    /* Rule: when a picture of size (x1,y1) with aspect ratio r1 is rendered
+     * on a picture of size (x2,y2) with aspect ratio r2, if x1 grows to x1'
+     * then y1 grows to y1' = x1' * y2/x2 * r2/r1 */
+    SetOffset( p_vout->render.i_width, p_vout->render.i_height,
+               p_vout->output.i_width, p_vout->output.i_height,
+               &b_hscale, &i_vscale, p_offset_start );
+
+    /*
+     * Perform conversion
+     */
+    i_scale_count = ( i_vscale == 1 ) ?
+                    p_vout->output.i_height : p_vout->render.i_height;
+
+#if defined (MODULE_NAME_IS_i420_rgb_sse2)
+
+    if( p_vout->render.i_width & 15 )
+    {
+        i_rewind = 16 - ( p_vout->render.i_width & 15 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    /*
+    ** SSE2 128 bits fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+
+    p_buffer = b_hscale ? p_buffer_start : p_pic;
+    if( 0 == (15 & (p_src->p[Y_PLANE].i_pitch|
+                    p_dest->p->i_pitch|
+                    ((int)p_y)|
+                    ((int)p_buffer))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+
+            for ( i_x = p_vout->render.i_width / 16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_ALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_ARGB_ALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 C intrinsics wrappers */
+                __m128i  xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                SSE2_INTRINSICS_INIT_32_ALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_32_ARGB_ALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_ARGB_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 intrinsics wrappers */
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_32_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_32_ARGB_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 4;
+                p_v += 4;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 4 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+        /* make sure all SSE2 stores are visible thereafter */
+#if defined (CAN_COMPILE_SSE2)
+        __asm__ __volatile__ ( "sfence" );
+#else
+        _mm_sfence();
+#endif
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+            for ( i_x = p_vout->render.i_width / 16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_ARGB_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 C intrinsics wrappers */
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                SSE2_INTRINSICS_INIT_32_UNALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_32_ARGB_UNALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_ARGB_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 intrinsics wrappers */
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_32_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_32_ARGB_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 4 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+    }
+
+#else
+
+    if( p_vout->render.i_width & 7 )
+    {
+        i_rewind = 8 - ( p_vout->render.i_width & 7 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = p_vout->render.i_width / 8; i_x--; )
+        {
+#if defined (CAN_COMPILE_MMX)
+            /* use inline MMX assembly */
+            __asm__( MMX_INIT_32
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+
+            __asm__( ".p2align 3"
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_32_ARGB
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            /* otherwise use MMX C intrinsics wrappers */
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+            uint64_t tmp64;
+
+            MMX_INTRINSICS_INIT_32
+            MMX_INTRINSICS_YUV_MUL
+            MMX_INTRINSICS_YUV_ADD
+            MMX_INTRINSICS_UNPACK_32_ARGB
+#endif
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+
+        /* Here we do some unaligned reads and duplicate conversions, but
+         * at least we have all the pixels */
+        if( i_rewind )
+        {
+            p_y -= i_rewind;
+            p_u -= i_rewind >> 1;
+            p_v -= i_rewind >> 1;
+            p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_MMX)
+            /* use inline MMX assembly */
+            __asm__( ".p2align 3"
+                     MMX_INIT_32
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_32_ARGB
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            /* otherwise use MMX intrinsics wrappers */
+            {
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                uint64_t tmp64;
+
+                MMX_INTRINSICS_INIT_32
+                MMX_INTRINSICS_YUV_MUL
+                MMX_INTRINSICS_YUV_ADD
+                MMX_INTRINSICS_UNPACK_32_ARGB
+            }
+#endif
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+        SCALE_WIDTH;
+        SCALE_HEIGHT( 420, 4 );
+
+        p_y += i_source_margin;
+        if( i_y % 2 )
+        {
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+        }
+    }
+    /* re-enable FPU registers */
+#if defined (CAN_COMPILE_MMX)
+    __asm__ __volatile__ ( "emms" );
+#else
+    _mm_empty();
+#endif
+
+#endif
+}
+
+void E_(I420_B8G8R8A8)( vout_thread_t *p_vout, picture_t *p_src,
+                                            picture_t *p_dest )
+{
+    /* We got this one from the old arguments */
+    uint32_t *p_pic = (uint32_t*)p_dest->p->p_pixels;
+    uint8_t  *p_y   = p_src->Y_PIXELS;
+    uint8_t  *p_u   = p_src->U_PIXELS;
+    uint8_t  *p_v   = p_src->V_PIXELS;
+
+    vlc_bool_t  b_hscale;                         /* horizontal scaling type */
+    unsigned int i_vscale;                          /* vertical scaling type */
+    unsigned int i_x, i_y;                /* horizontal and vertical indexes */
+
+    int         i_right_margin;
+    int         i_rewind;
+    int         i_scale_count;                       /* scale modulo counter */
+    int         i_chroma_width = p_vout->render.i_width / 2; /* chroma width */
+    uint32_t *  p_pic_start;       /* beginning of the current line for copy */
+    /* Conversion buffer pointer */
+    uint32_t *  p_buffer_start = (uint32_t*)p_vout->chroma.p_sys->p_buffer;
+    uint32_t *  p_buffer;
+
+    /* Offset array pointer */
+    int *       p_offset_start = p_vout->chroma.p_sys->p_offset;
+    int *       p_offset;
+
+    const int i_source_margin = p_src->p[0].i_pitch
+                                 - p_src->p[0].i_visible_pitch;
+    const int i_source_margin_c = p_src->p[1].i_pitch
+                                 - p_src->p[1].i_visible_pitch;
+
+    i_right_margin = p_dest->p->i_pitch - p_dest->p->i_visible_pitch;
+
+    /* Rule: when a picture of size (x1,y1) with aspect ratio r1 is rendered
+     * on a picture of size (x2,y2) with aspect ratio r2, if x1 grows to x1'
+     * then y1 grows to y1' = x1' * y2/x2 * r2/r1 */
+    SetOffset( p_vout->render.i_width, p_vout->render.i_height,
+               p_vout->output.i_width, p_vout->output.i_height,
+               &b_hscale, &i_vscale, p_offset_start );
+
+    /*
+     * Perform conversion
+     */
+    i_scale_count = ( i_vscale == 1 ) ?
+                    p_vout->output.i_height : p_vout->render.i_height;
+
+#if defined (MODULE_NAME_IS_i420_rgb_sse2)
+
+    if( p_vout->render.i_width & 15 )
+    {
+        i_rewind = 16 - ( p_vout->render.i_width & 15 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    /*
+    ** SSE2 128 bits fetch/store instructions are faster 
+    ** if memory access is 16 bytes aligned
+    */
+
+    p_buffer = b_hscale ? p_buffer_start : p_pic;
+    if( 0 == (15 & (p_src->p[Y_PLANE].i_pitch|
+                    p_dest->p->i_pitch|
+                    ((int)p_y)|
+                    ((int)p_buffer))) )
+    {
+        /* use faster SSE2 aligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+
+            for ( i_x = p_vout->render.i_width / 16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_ALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_BGRA_ALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 C intrinsics wrappers */
+                __m128i  xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                SSE2_INTRINSICS_INIT_32_ALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_32_BGRA_ALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_BGRA_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 intrinsics wrappers */
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_32_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_32_BGRA_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 4;
+                p_v += 4;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 4 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+        /* make sure all SSE2 stores are visible thereafter */
+#if defined (CAN_COMPILE_SSE2)
+        __asm__ __volatile__ ( "sfence" );
+#else
+        _mm_sfence();
+#endif
+    }
+    else
+    {
+        /* use slower SSE2 unaligned fetch and store */
+        for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+        {
+            p_pic_start = p_pic;
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+            for ( i_x = p_vout->render.i_width / 16; i_x--; )
+            {
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_BGRA_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 C intrinsics wrappers */
+                __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                SSE2_INTRINSICS_INIT_32_UNALIGNED
+                SSE2_INTRINSICS_YUV_MUL
+                SSE2_INTRINSICS_YUV_ADD
+                SSE2_INTRINSICS_UNPACK_32_BGRA_UNALIGNED
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+                p_buffer += 16;
+            }
+
+            /* Here we do some unaligned reads and duplicate conversions, but
+             * at least we have all the pixels */
+            if( i_rewind )
+            {
+                p_y -= i_rewind;
+                p_u -= i_rewind >> 1;
+                p_v -= i_rewind >> 1;
+                p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_SSE2)
+                /* use inline SSE2 assembly */
+                __asm__( ".p2align 3"
+                         SSE2_INIT_32_UNALIGNED
+                         SSE2_YUV_MUL
+                         SSE2_YUV_ADD
+                         SSE2_UNPACK_32_BGRA_UNALIGNED
+                         : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) : "eax" );
+#else
+                /* otherwise use SSE2 intrinsics wrappers */
+                {
+                    __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
+
+                    SSE2_INTRINSICS_INIT_32_UNALIGNED
+                    SSE2_INTRINSICS_YUV_MUL
+                    SSE2_INTRINSICS_YUV_ADD
+                    SSE2_INTRINSICS_UNPACK_32_BGRA_UNALIGNED
+                }
+#endif
+                p_y += 16;
+                p_u += 8;
+                p_v += 8;
+            }
+            SCALE_WIDTH;
+            SCALE_HEIGHT( 420, 4 );
+
+            p_y += i_source_margin;
+            if( i_y % 2 )
+            {
+                p_u += i_source_margin_c;
+                p_v += i_source_margin_c;
+            }
+            p_buffer = b_hscale ? p_buffer_start : p_pic;
+        }
+    }
+
+#else
+
+    if( p_vout->render.i_width & 7 )
+    {
+        i_rewind = 8 - ( p_vout->render.i_width & 7 );
+    }
+    else
+    {
+        i_rewind = 0;
+    }
+
+    for( i_y = 0; i_y < p_vout->render.i_height; i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = p_vout->render.i_width / 8; i_x--; )
+        {
+#if defined (CAN_COMPILE_MMX)
+            /* use inline MMX assembly */
+            __asm__( MMX_INIT_32
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+
+            __asm__( ".p2align 3"
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_32_ARGB
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            /* otherwise use MMX C intrinsics wrappers */
+            __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+            uint64_t tmp64;
+
+            MMX_INTRINSICS_INIT_32
+            MMX_INTRINSICS_YUV_MUL
+            MMX_INTRINSICS_YUV_ADD
+            MMX_INTRINSICS_UNPACK_32_BGRA
+#endif
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+
+        /* Here we do some unaligned reads and duplicate conversions, but
+         * at least we have all the pixels */
+        if( i_rewind )
+        {
+            p_y -= i_rewind;
+            p_u -= i_rewind >> 1;
+            p_v -= i_rewind >> 1;
+            p_buffer -= i_rewind;
+#if defined (CAN_COMPILE_MMX)
+            /* use inline MMX assembly */
+            __asm__( ".p2align 3"
+                     MMX_INIT_32
+                     MMX_YUV_MUL
+                     MMX_YUV_ADD
+                     MMX_UNPACK_32_BGRA
+                     : : "r" (p_y), "r" (p_u), "r" (p_v), "r" (p_buffer) );
+#else
+            /* otherwise use MMX intrinsics wrappers */
+            {
+                __m64 mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
+                uint64_t tmp64;
+
+                MMX_INTRINSICS_INIT_32
+                MMX_INTRINSICS_YUV_MUL
+                MMX_INTRINSICS_YUV_ADD
+                MMX_INTRINSICS_UNPACK_32_BGRA
+            }
+#endif
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+            p_buffer += 8;
+        }
+        SCALE_WIDTH;
+        SCALE_HEIGHT( 420, 4 );
+
+        p_y += i_source_margin;
+        if( i_y % 2 )
+        {
+            p_u += i_source_margin_c;
+            p_v += i_source_margin_c;
+        }
+    }
+    /* re-enable FPU registers */
+#if defined (CAN_COMPILE_MMX)
+    __asm__ __volatile__ ( "emms" );
+#else
+    _mm_empty();
+#endif
+
+#endif
+}
+
+#endif
+
 /* Following functions are local */
 
 /*****************************************************************************

modules/video_chroma/i420_rgb_mmx.h

 /*****************************************************************************
  * transforms_yuvmmx.h: MMX YUV transformation assembly
  *****************************************************************************
- * Copyright (C) 1999-2004 the VideoLAN team
+ * Copyright (C) 1999-2007 the VideoLAN team
  * $Id$
  *
  * Authors: Olie Lho <ollie@sis.com.tw>
  *          Gaël Hendryckx <jimmy@via.ecp.fr>
  *          Samuel Hocevar <sam@zoy.org>
+ *          Damien Fouilleul <damienf@videolan.org>
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
@@ -53,21 +54,48 @@ USED_U64(mmx_mask_fc) = 0xfcfcfcfcfcfcfcfcULL;
 #endif
 
 #define MMX_INIT_16 "                                                       \n\
-movd      (%1), %%mm0       # Load 4 Cb       00 00 00 00 u3 u2 u1 u0       \n\
-movd      (%2), %%mm1       # Load 4 Cr       00 00 00 00 v3 v2 v1 v0       \n\
+movd       (%1), %%mm0      # Load 4 Cb       00 00 00 00 u3 u2 u1 u0       \n\
+movd       (%2), %%mm1      # Load 4 Cr       00 00 00 00 v3 v2 v1 v0       \n\
 pxor      %%mm4, %%mm4      # zero mm4                                      \n\
-movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
-#movl      $0, (%3)         # cache preload for image                       \n\
+movq       (%0), %%mm6      # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 "
 
-#define INTRINSICS_INIT_16 \
-    tmp64 = *(uint32_t *)p_u; \
-    mm0 = (__m64)tmp64; \
-    tmp64 = *(uint32_t *)p_v; \
-    mm1 = (__m64)tmp64; \
-    mm4 = (__m64)(uint64_t)0; \
-    mm6 = (__m64)*(uint64_t *)p_y; \
-    /* *(uint16_t *)p_buffer = 0; */
+#define SSE2_INIT_16_ALIGNED "                                              \n\
+prefetcht1  (%3)            # cache preload for image                       \n\
+movq        (%1), %%xmm0    # Load 8 Cb       00 00 00 00 u3 u2 u1 u0       \n\
+movq        (%2), %%xmm1    # Load 8 Cr       00 00 00 00 v3 v2 v1 v0       \n\
+pxor      %%xmm4, %%xmm4    # zero mm4                                      \n\
+movdqa      (%0), %%xmm6    # Load 16 Y       Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+"
+
+#define SSE2_INIT_16_UNALIGNED "                                            \n\
+prefetcht1  (%3)            # cache preload for image                       \n\
+movq        (%1), %%xmm0    # Load 8 Cb       00 00 00 00 u3 u2 u1 u0       \n\
+movq        (%2), %%xmm1    # Load 8 Cr       00 00 00 00 v3 v2 v1 v0       \n\
+pxor      %%xmm4, %%xmm4    # zero mm4                                      \n\
+movdqu      (%0), %%xmm6    # Load 16 Y       Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+"
+
+#define MMX_INTRINSICS_INIT_16      \
+    tmp64 = *(uint32_t *)p_u;       \
+    mm0 = (__m64)tmp64;             \
+    tmp64 = *(uint32_t *)p_v;       \
+    mm1 = (__m64)tmp64;             \
+    mm4 = _mm_setzero_si64();       \
+    mm6 = (__m64)*(uint64_t *)p_y;  \
+
+#define SSE2_INTRINSICS_INIT_16_ALIGNED     \
+    xmm0 = _mm_loadl_epi64((__m128i *)p_u); \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v); \
+    xmm4 = _mm_setzero_si128();             \
+    xmm6 = _mm_load_si128((__m128i *)p_y);  \
+
+#define SSE2_INTRINSICS_INIT_16_UNALIGNED   \
+    _mm_prefetch(p_buffer, _MM_HINT_T1);    \
+    xmm0 = _mm_loadl_epi64((__m128i *)p_u); \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v); \
+    xmm4 = _mm_setzero_si128();             \
+    xmm6 = _mm_loadu_si128((__m128i *)p_y); \
 
 #define MMX_INIT_16_GRAY "                                                  \n\
 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
@@ -76,21 +104,49 @@ movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 
 #define MMX_INIT_32 "                                                       \n\
 movd      (%1), %%mm0       # Load 4 Cb       00 00 00 00 u3 u2 u1 u0       \n\
-movl      $0, (%3)          # cache preload for image                       \n\
+movl        $0, (%3)        # cache preload for image                       \n\
 movd      (%2), %%mm1       # Load 4 Cr       00 00 00 00 v3 v2 v1 v0       \n\
-pxor      %%mm4, %%mm4      # zero mm4                                      \n\
+pxor     %%mm4, %%mm4       # zero mm4                                      \n\
 movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
 "
 
-#define INTRINSICS_INIT_32 \
-    tmp64 = *(uint32_t *)p_u; \
-    mm0 = (__m64)tmp64; \
-    *(uint16_t *)p_buffer = 0; \
-    tmp64 = *(uint32_t *)p_v; \
-    mm1 = (__m64)tmp64; \
-    mm4 = (__m64)(uint64_t)0; \
+#define SSE2_INIT_32_ALIGNED "                                              \n\
+movq        (%1), %%xmm0    # Load 8 Cb       00 00 00 00 u3 u2 u1 u0       \n\
+movq        (%2), %%xmm1    # Load 8 Cr       00 00 00 00 v3 v2 v1 v0       \n\
+pxor      %%xmm4, %%xmm4    # zero mm4                                      \n\
+movdqa      (%0), %%xmm6    # Load 16 Y       Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+"
+
+#define SSE2_INIT_32_UNALIGNED "                                            \n\
+prefetcht1  (%3)            # cache preload for image                       \n\
+movq        (%1), %%xmm0    # Load 8 Cb       00 00 00 00 u3 u2 u1 u0       \n\
+movq        (%2), %%xmm1    # Load 8 Cr       00 00 00 00 v3 v2 v1 v0       \n\
+pxor      %%xmm4, %%xmm4    # zero mm4                                      \n\
+movdqu      (%0), %%xmm6    # Load 16 Y       Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+"
+
+#define MMX_INTRINSICS_INIT_32      \
+    tmp64 = *(uint32_t *)p_u;       \
+    mm0 = (__m64)tmp64;             \
+    *(uint16_t *)p_buffer = 0;      \
+    tmp64 = *(uint32_t *)p_v;       \
+    mm1 = (__m64)tmp64;             \
+    mm4 = _mm_setzero_si64();       \
     mm6 = (__m64)*(uint64_t *)p_y;
 
+#define SSE2_INTRINSICS_INIT_32_ALIGNED      \
+    xmm0 = _mm_loadl_epi64((__m128i *)p_u); \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v); \
+    xmm4 = _mm_setzero_si128();             \
+    xmm6 = _mm_load_si128((__m128i *)p_y);  \
+
+#define SSE2_INTRINSICS_INIT_32_UNALIGNED    \
+    _mm_prefetch(p_buffer, _MM_HINT_T1);     \
+    xmm0 = _mm_loadl_epi64((__m128i *)p_u); \
+    xmm1 = _mm_loadl_epi64((__m128i *)p_v); \
+    xmm4 = _mm_setzero_si128();             \
+    xmm6 = _mm_loadu_si128((__m128i *)p_y); \
+
 /*
  * Do the multiply part of the conversion for even and odd pixels,
  * register usage:
@@ -126,7 +182,58 @@ pmulhw    mmx_Y_coeff"G", %%mm6 # Mul 4 Y even    00 y6 00 y4 00 y2 00 y0   \n\
 pmulhw    mmx_Y_coeff"G", %%mm7 # Mul 4 Y odd     00 y7 00 y5 00 y3 00 y1   \n\
 "
 
-#define INTRINSICS_YUV_MUL \
+#define SSE2_YUV_MUL "                                                      \n\
+# convert the chroma part                                                   \n\
+punpcklbw %%xmm4, %%xmm0        # scatter 8 Cb    00 u3 00 u2 00 u1 00 u0   \n\
+punpcklbw %%xmm4, %%xmm1        # scatter 8 Cr    00 v3 00 v2 00 v1 00 v0   \n\
+movl      $0x00800080, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to     0080 0080 ... 0080 0080   \n\
+psubsw    %%xmm5, %%xmm0        # Cb -= 128                                 \n\
+psubsw    %%xmm5, %%xmm1        # Cr -= 128                                 \n\
+psllw     $3, %%xmm0            # Promote precision                         \n\
+psllw     $3, %%xmm1            # Promote precision                         \n\
+movdqa    %%xmm0, %%xmm2        # Copy 8 Cb       00 u3 00 u2 00 u1 00 u0   \n\
+movdqa    %%xmm1, %%xmm3        # Copy 8 Cr       00 v3 00 v2 00 v1 00 v0   \n\
+movl      $0xf37df37d, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to     f37d f37d ... f37d f37d   \n\
+pmulhw    %%xmm5, %%xmm2        # Mul Cb with green coeff -> Cb green       \n\
+movl      $0xe5fce5fc, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to     e5fc e5fc ... e5fc e5fc   \n\
+pmulhw    %%xmm5, %%xmm3        # Mul Cr with green coeff -> Cr green       \n\
+movl      $0x40934093, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to     4093 4093 ... 4093 4093   \n\
+pmulhw    %%xmm5, %%xmm0        # Mul Cb -> Cblue 00 b3 00 b2 00 b1 00 b0   \n\
+movl      $0x33123312, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to     3312 3312 ... 3312 3312   \n\
+pmulhw    %%xmm5, %%xmm1        # Mul Cr -> Cred  00 r3 00 r2 00 r1 00 r0   \n\
+paddsw    %%xmm3, %%xmm2        # Cb green + Cr green -> Cgreen             \n\
+                                                                            \n\
+# convert the luma part                                                     \n\
+movl      $0x10101010, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # Set xmm5 to   1010 1010 ... 1010 1010     \n\
+psubusb   %%xmm5, %%xmm6        # Y -= 16                                   \n\
+movdqa    %%xmm6, %%xmm7        # Copy 16 Y       Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0   \n\
+movl      $0x00ff00ff, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     00ff 00ff ... 00ff 00ff   \n\
+pand      %%xmm5, %%xmm6        # get Y even      00 Y6 00 Y4 00 Y2 00 Y0   \n\
+psrlw     $8, %%xmm7            # get Y odd       00 Y7 00 Y5 00 Y3 00 Y1   \n\
+psllw     $3, %%xmm6            # Promote precision                         \n\
+psllw     $3, %%xmm7            # Promote precision                         \n\
+movl      $0x253f253f, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     253f 253f ... 253f 253f   \n\
+pmulhw    %%xmm5, %%xmm6        # Mul 8 Y even    00 y6 00 y4 00 y2 00 y0   \n\
+pmulhw    %%xmm5, %%xmm7        # Mul 8 Y odd     00 y7 00 y5 00 y3 00 y1   \n\
+"
+
+#define MMX_INTRINSICS_YUV_MUL \
     mm0 = _mm_unpacklo_pi8(mm0, mm4); \
     mm1 = _mm_unpacklo_pi8(mm1, mm4); \
     mm0 = _mm_subs_pi16(mm0, (__m64)mmx_80w); \
@@ -150,6 +257,38 @@ pmulhw    mmx_Y_coeff"G", %%mm7 # Mul 4 Y odd     00 y7 00 y5 00 y3 00 y1   \n\
     mm6 = _mm_mulhi_pi16(mm6, (__m64)mmx_Y_coeff); \
     mm7 = _mm_mulhi_pi16(mm7, (__m64)mmx_Y_coeff);
 
+#define SSE2_INTRINSICS_YUV_MUL \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm4); \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm4); \
+    xmm5 = _mm_set1_epi32(0x80808080UL); \
+    xmm0 = _mm_subs_epi16(xmm0, xmm5); \
+    xmm1 = _mm_subs_epi16(xmm1, xmm5); \
+    xmm0 = _mm_slli_epi16(xmm0, 3); \
+    xmm1 = _mm_slli_epi16(xmm1, 3); \
+    xmm2 = xmm0; \
+    xmm3 = xmm1; \
+    xmm5 = _mm_set1_epi32(0xf37df37dUL); \
+    xmm2 = _mm_mulhi_epi16(xmm2, xmm5); \
+    xmm5 = _mm_set1_epi32(0xe5fce5fcUL); \
+    xmm3 = _mm_mulhi_epi16(xmm3, xmm5); \
+    xmm5 = _mm_set1_epi32(0x40934093UL); \
+    xmm0 = _mm_mulhi_epi16(xmm0, xmm5); \
+    xmm5 = _mm_set1_epi32(0x33123312UL); \
+    xmm1 = _mm_mulhi_epi16(xmm1, xmm5); \
+    xmm2 = _mm_adds_epi16(xmm2, xmm3); \
+    \
+    xmm5 = _mm_set1_epi32(0x10101010UL); \
+    xmm6 = _mm_subs_epu8(xmm6, xmm5); \
+    xmm7 = xmm6; \
+    xmm5 = _mm_set1_epi32(0x00ff00ffUL); \
+    xmm6 = _mm_and_si128(xmm6, xmm5); \
+    xmm7 = _mm_srli_epi16(xmm7, 8); \
+    xmm6 = _mm_slli_epi16(xmm6, 3); \
+    xmm7 = _mm_slli_epi16(xmm7, 3); \
+    xmm5 = _mm_set1_epi32(0x253f253fUL); \
+    xmm6 = _mm_mulhi_epi16(xmm6, xmm5); \
+    xmm7 = _mm_mulhi_epi16(xmm7, xmm5);
+
 /*
  * Do the addition part of the conversion for even and odd pixels,
  * register usage:
@@ -186,7 +325,35 @@ punpcklbw %%mm4, %%mm1          #                 R7 R6 R5 R4 R3 R2 R1 R0   \n\
 punpcklbw %%mm5, %%mm2          #                 G7 G6 G5 G4 G3 G2 G1 G0   \n\
 "
 
-#define INTRINSICS_YUV_ADD \
+#define SSE2_YUV_ADD "                                                      \n\
+# Do horizontal and vertical scaling                                        \n\
+movdqa    %%xmm0, %%xmm3        # Copy Cblue                                \n\
+movdqa    %%xmm1, %%xmm4        # Copy Cred                                 \n\
+movdqa    %%xmm2, %%xmm5        # Copy Cgreen                               \n\
+paddsw    %%xmm6, %%xmm0        # Y even + Cblue  00 B6 00 B4 00 B2 00 B0   \n\
+paddsw    %%xmm7, %%xmm3        # Y odd  + Cblue  00 B7 00 B5 00 B3 00 B1   \n\
+paddsw    %%xmm6, %%xmm1        # Y even + Cred   00 R6 00 R4 00 R2 00 R0   \n\
+paddsw    %%xmm7, %%xmm4        # Y odd  + Cred   00 R7 00 R5 00 R3 00 R1   \n\
+paddsw    %%xmm6, %%xmm2        # Y even + Cgreen 00 G6 00 G4 00 G2 00 G0   \n\
+paddsw    %%xmm7, %%xmm5        # Y odd  + Cgreen 00 G7 00 G5 00 G3 00 G1   \n\
+                                                                            \n\
+# Limit RGB even to 0..255                                                  \n\
+packuswb  %%xmm0, %%xmm0        # B6 B4 B2 B0 / B6 B4 B2 B0                 \n\
+packuswb  %%xmm1, %%xmm1        # R6 R4 R2 R0 / R6 R4 R2 R0                 \n\
+packuswb  %%xmm2, %%xmm2        # G6 G4 G2 G0 / G6 G4 G2 G0                 \n\
+                                                                            \n\
+# Limit RGB odd to 0..255                                                   \n\
+packuswb  %%xmm3, %%xmm3        # B7 B5 B3 B1 / B7 B5 B3 B1                 \n\
+packuswb  %%xmm4, %%xmm4        # R7 R5 R3 R1 / R7 R5 R3 R1                 \n\
+packuswb  %%xmm5, %%xmm5        # G7 G5 G3 G1 / G7 G5 G3 G1                 \n\
+                                                                            \n\
+# Interleave RGB even and odd                                               \n\
+punpcklbw %%xmm3, %%xmm0        #                 B7 B6 B5 B4 B3 B2 B1 B0   \n\
+punpcklbw %%xmm4, %%xmm1        #                 R7 R6 R5 R4 R3 R2 R1 R0   \n\
+punpcklbw %%xmm5, %%xmm2        #                 G7 G6 G5 G4 G3 G2 G1 G0   \n\
+"
+
+#define MMX_INTRINSICS_YUV_ADD \
     mm3 = mm0; \
     mm4 = mm1; \
     mm5 = mm2; \
@@ -209,6 +376,29 @@ punpcklbw %%mm5, %%mm2          #                 G7 G6 G5 G4 G3 G2 G1 G0   \n\
     mm1 = _mm_unpacklo_pi8(mm1, mm4); \
     mm2 = _mm_unpacklo_pi8(mm2, mm5);
 
+#define SSE2_INTRINSICS_YUV_ADD \
+    xmm3 = xmm0; \
+    xmm4 = xmm1; \
+    xmm5 = xmm2; \
+    xmm0 = _mm_adds_epi16(xmm0, xmm6); \
+    xmm3 = _mm_adds_epi16(xmm3, xmm7); \
+    xmm1 = _mm_adds_epi16(xmm1, xmm6); \
+    xmm4 = _mm_adds_epi16(xmm4, xmm7); \
+    xmm2 = _mm_adds_epi16(xmm2, xmm6); \
+    xmm5 = _mm_adds_epi16(xmm5, xmm7); \
+    \
+    xmm0 = _mm_packus_epi16(xmm0, xmm0); \
+    xmm1 = _mm_packus_epi16(xmm1, xmm1); \
+    xmm2 = _mm_packus_epi16(xmm2, xmm2); \
+    \
+    xmm3 = _mm_packus_epi16(xmm3, xmm3); \
+    xmm4 = _mm_packus_epi16(xmm4, xmm4); \
+    xmm5 = _mm_packus_epi16(xmm5, xmm5); \
+    \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm3); \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm4); \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
+
 /*
  * Grayscale case, only use Y
  */
@@ -287,13 +477,71 @@ movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 "
 
-#define INTRINSICS_UNPACK_15 \
+#define SSE2_UNPACK_15_ALIGNED "                                            \n\
+# mask unneeded bits off                                                    \n\
+movl      $0xf8f8f8f8, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm0        # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+psrlw     $3,%%xmm0             # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pand      %%xmm5, %%xmm2        # g7g6g5g4 g3______ g7g6g5g4 g3______       \n\
+pand      %%xmm5, %%xmm1        # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+psrlw     $1,%%xmm1             # __r7r6r5 r4r3____ __r7r6r5 r4r3____       \n\
+pxor      %%xmm4, %%xmm4        # zero mm4                                  \n\
+movdqa    %%xmm0, %%xmm5        # Copy B15-B0                               \n\
+movdqa    %%xmm2, %%xmm7        # Copy G15-G0                               \n\
+                                                                            \n\
+# convert rgb24 plane to rgb15 pack for pixel 0-7                           \n\
+punpcklbw %%xmm4, %%xmm2        # ________ ________ g7g6g5g4 g3______       \n\
+punpcklbw %%xmm1, %%xmm0        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $2,%%xmm2             # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%xmm2, %%xmm0        # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
+movntdq   %%xmm0, (%3)          # store pixel 0-7                           \n\
+                                                                            \n\
+# convert rgb24 plane to rgb15 pack for pixel 8-15                          \n\
+punpckhbw %%xmm4, %%xmm7        # ________ ________ g7g6g5g4 g3______       \n\
+punpckhbw %%xmm1, %%xmm5        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $2,%%xmm7             # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%xmm7, %%xmm5        # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
+movntdq   %%xmm5, 16(%3)        # store pixel 4-7                           \n\
+"
+
+#define SSE2_UNPACK_15_UNALIGNED "                                          \n\
+# mask unneeded bits off                                                    \n\
+movl      $0xf8f8f8f8, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm0        # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+psrlw     $3,%%xmm0             # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pand      %%xmm5, %%xmm2        # g7g6g5g4 g3______ g7g6g5g4 g3______       \n\
+pand      %%xmm5, %%xmm1        # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+psrlw     $1,%%xmm1             # __r7r6r5 r4r3____ __r7r6r5 r4r3____       \n\
+pxor      %%xmm4, %%xmm4        # zero mm4                                  \n\
+movdqa    %%xmm0, %%xmm5        # Copy B15-B0                               \n\
+movdqa    %%xmm2, %%xmm7        # Copy G15-G0                               \n\
+                                                                            \n\
+# convert rgb24 plane to rgb15 pack for pixel 0-7                           \n\
+punpcklbw %%xmm4, %%xmm2        # ________ ________ g7g6g5g4 g3______       \n\
+punpcklbw %%xmm1, %%xmm0        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $2,%%xmm2             # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%xmm2, %%xmm0        # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
+movdqu    %%xmm0, (%3)          # store pixel 0-7                           \n\
+                                                                            \n\
+# convert rgb24 plane to rgb15 pack for pixel 8-15                          \n\
+punpckhbw %%xmm4, %%xmm7        # ________ ________ g7g6g5g4 g3______       \n\
+punpckhbw %%xmm1, %%xmm5        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $2,%%xmm7             # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%xmm7, %%xmm5        # r7r6r5r4 r3__g7g6 g5g4g3b7 b6b5b4b3       \n\
+movdqu    %%xmm5, 16(%3)        # store pixel 4-7                           \n\
+"
+
+#define MMX_INTRINSICS_UNPACK_15 \
     mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
     mm0 = _mm_srli_pi16(mm0, 3); \
     mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_f8); \
     mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
     mm1 = _mm_srli_pi16(mm1, 1); \
-    mm4 = (__m64)(uint64_t)0; \
+    mm4 = _mm_setzero_si64(); \
     mm5 = mm0; \
     mm7 = mm2; \
     \
@@ -315,6 +563,52 @@ movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
     mm1 = (__m64)tmp64; \
     *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
 
+#define SSE2_INTRINSICS_UNPACK_15_ALIGNED \
+    xmm5 = _mm_set1_epi32(0xf8f8f8f8UL); \
+    xmm0 = _mm_and_si128(xmm0, xmm5); \
+    xmm0 = _mm_srli_epi16(xmm0, 3); \
+    xmm2 = _mm_and_si128(xmm2, xmm5); \
+    xmm1 = _mm_and_si128(xmm1, xmm5); \
+    xmm1 = _mm_srli_epi16(xmm1, 1); \
+    xmm4 = _mm_setzero_si128();          \
+    xmm5 = xmm0; \
+    xmm7 = xmm2; \
+    \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm4); \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
+    xmm2 = _mm_slli_epi16(xmm2, 2); \
+    xmm0 = _mm_or_si128(xmm0, xmm2); \
+    _mm_stream_si128((__m128i*)p_buffer, xmm0); \
+    \
+    xmm7 = _mm_unpackhi_epi8(xmm7, xmm4); \
+    xmm5 = _mm_unpackhi_epi8(xmm5, xmm1); \
+    xmm7 = _mm_slli_epi16(xmm7, 2); \
+    xmm5 = _mm_or_si128(xmm5, xmm7); \
+    _mm_stream_si128((__m128i*)(p_buffer+8), xmm5);
+
+#define SSE2_INTRINSICS_UNPACK_15_UNALIGNED \
+    xmm5 = _mm_set1_epi32(0xf8f8f8f8UL); \
+    xmm0 = _mm_and_si128(xmm0, xmm5); \
+    xmm0 = _mm_srli_epi16(xmm0, 3); \
+    xmm2 = _mm_and_si128(xmm2, xmm5); \
+    xmm1 = _mm_and_si128(xmm1, xmm5); \
+    xmm1 = _mm_srli_epi16(xmm1, 1); \
+    xmm4 = _mm_setzero_si128();          \
+    xmm5 = xmm0; \
+    xmm7 = xmm2; \
+    \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm4); \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
+    xmm2 = _mm_slli_epi16(xmm2, 2); \
+    xmm0 = _mm_or_si128(xmm0, xmm2); \
+    _mm_storeu_si128((__m128i*)p_buffer, xmm0); \
+    \
+    xmm7 = _mm_unpackhi_epi8(xmm7, xmm4); \
+    xmm5 = _mm_unpackhi_epi8(xmm5, xmm1); \
+    xmm7 = _mm_slli_epi16(xmm7, 2); \
+    xmm5 = _mm_or_si128(xmm5, xmm7); \
+    _mm_storeu_si128((__m128i*)(p_buffer+16), xmm5);
+
 /*
  * convert RGB plane to RGB 16 bits,
  * mm0 -> B, mm1 -> R, mm2 -> G,
@@ -350,12 +644,74 @@ movd      4(%2), %%mm1          # Load 4 Cr       __ __ __ __ v3 v2 v1 v0   \n\
 movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
 "
 
-#define INTRINSICS_UNPACK_16 \
+#define SSE2_UNPACK_16_ALIGNED "                                            \n\
+# mask unneeded bits off                                                    \n\
+movl      $0xf8f8f8f8, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm0        # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+pand      %%xmm5, %%xmm1        # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+movl      $0xfcfcfcfc, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm2        # g7g6g5g4 g3g2____ g7g6g5g4 g3g2____       \n\
+psrlw     $3,%%xmm0             # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pxor      %%xmm4, %%xmm4        # zero mm4                                  \n\
+movdqa    %%xmm0, %%xmm5        # Copy B15-B0                               \n\
+movdqa    %%xmm2, %%xmm7        # Copy G15-G0                               \n\
+                                                                            \n\
+# convert rgb24 plane to rgb16 pack for pixel 0-7                           \n\
+punpcklbw %%xmm4, %%xmm2        # ________ ________ g7g6g5g4 g3g2____       \n\
+punpcklbw %%xmm1, %%xmm0        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%xmm2             # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%xmm2, %%xmm0        # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movntdq   %%xmm0, (%3)          # store pixel 0-7                           \n\
+                                                                            \n\
+# convert rgb24 plane to rgb16 pack for pixel 8-15                          \n\
+punpckhbw %%xmm4, %%xmm7        # ________ ________ g7g6g5g4 g3g2____       \n\
+punpckhbw %%xmm1, %%xmm5        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%xmm7             # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%xmm7, %%xmm5        # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movntdq   %%xmm5, 16(%3)        # store pixel 4-7                           \n\
+"
+
+#define SSE2_UNPACK_16_UNALIGNED "                                          \n\
+# mask unneeded bits off                                                    \n\
+movl      $0xf8f8f8f8, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm0        # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+pand      %%xmm5, %%xmm1        # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+movl      $0xfcfcfcfc, %%eax    #                                           \n\
+movd      %%eax, %%xmm5         #                                           \n\
+pshufd    $0, %%xmm5, %%xmm5    # set xmm5 to     f8f8 f8f8 ... f8f8 f8f8   \n\
+pand      %%xmm5, %%xmm2        # g7g6g5g4 g3g2____ g7g6g5g4 g3g2____       \n\
+psrlw     $3,%%xmm0             # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pxor      %%xmm4, %%xmm4        # zero mm4                                  \n\
+movdqa    %%xmm0, %%xmm5        # Copy B15-B0                               \n\
+movdqa    %%xmm2, %%xmm7        # Copy G15-G0                               \n\
+                                                                            \n\
+# convert rgb24 plane to rgb16 pack for pixel 0-7                           \n\
+punpcklbw %%xmm4, %%xmm2        # ________ ________ g7g6g5g4 g3g2____       \n\
+punpcklbw %%xmm1, %%xmm0        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%xmm2             # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%xmm2, %%xmm0        # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movdqu    %%xmm0, (%3)          # store pixel 0-7                           \n\
+                                                                            \n\
+# convert rgb24 plane to rgb16 pack for pixel 8-15                          \n\
+punpckhbw %%xmm4, %%xmm7        # ________ ________ g7g6g5g4 g3g2____       \n\
+punpckhbw %%xmm1, %%xmm5        # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%xmm7             # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%xmm7, %%xmm5        # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movdqu    %%xmm5, 16(%3)        # store pixel 4-7                           \n\
+"
+
+#define MMX_INTRINSICS_UNPACK_16 \
     mm0 = _mm_and_si64(mm0, (__m64)mmx_mask_f8); \
     mm2 = _mm_and_si64(mm2, (__m64)mmx_mask_fc); \
     mm1 = _mm_and_si64(mm1, (__m64)mmx_mask_f8); \
     mm0 = _mm_srli_pi16(mm0, 3); \
-    mm4 = (__m64)(uint64_t)0; \
+    mm4 = _mm_setzero_si64(); \
     mm5 = mm0; \
     mm7 = mm2; \
     \
@@ -377,62 +733,294 @@ movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
     mm1 = (__m64)tmp64; \
     *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
 
+#define SSE2_INTRINSICS_UNPACK_16_ALIGNED \
+    xmm5 = _mm_set1_epi32(0xf8f8f8f8UL); \
+    xmm0 = _mm_and_si128(xmm0, xmm5); \
+    xmm1 = _mm_and_si128(xmm1, xmm5); \
+    xmm5 = _mm_set1_epi32(0xfcfcfcfcUL); \
+    xmm2 = _mm_and_si128(xmm2, xmm5); \
+    xmm0 = _mm_srli_epi16(xmm0, 3); \
+    xmm4 = _mm_setzero_si128();          \
+    xmm5 = xmm0; \
+    xmm7 = xmm2; \
+    \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm4); \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
+    xmm2 = _mm_slli_epi16(xmm2, 3); \
+    xmm0 = _mm_or_si128(xmm0, xmm2); \
+    _mm_stream_si128((__m128i*)p_buffer, xmm0); \
+    \
+    xmm7 = _mm_unpackhi_epi8(xmm7, xmm4); \
+    xmm5 = _mm_unpackhi_epi8(xmm5, xmm1); \
+    xmm7 = _mm_slli_epi16(xmm7, 3); \
+    xmm5 = _mm_or_si128(xmm5, xmm7); \
+    _mm_stream_si128((__m128i*)(p_buffer+8), xmm5);
+
+#define SSE2_INTRINSICS_UNPACK_16_UNALIGNED \
+    xmm5 = _mm_set1_epi32(0xf8f8f8f8UL); \
+    xmm0 = _mm_and_si128(xmm0, xmm5); \
+    xmm1 = _mm_and_si128(xmm1, xmm5); \
+    xmm5 = _mm_set1_epi32(0xfcfcfcfcUL); \
+    xmm2 = _mm_and_si128(xmm2, xmm5); \
+    xmm0 = _mm_srli_epi16(xmm0, 3); \
+    xmm4 = _mm_setzero_si128();          \
+    xmm5 = xmm0; \
+    xmm7 = xmm2; \
+    \
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm4); \
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1); \
+    xmm2 = _mm_slli_epi16(xmm2, 3); \
+    xmm0 = _mm_or_si128(xmm0, xmm2); \
+    _mm_storeu_si128((__m128i*)p_buffer, xmm0); \
+    \
+    xmm7 = _mm_unpackhi_epi8(xmm7, xmm4); \
+    xmm5 = _mm_unpackhi_epi8(xmm5, xmm1); \
+    xmm7 = _mm_slli_epi16(xmm7, 3); \
+    xmm5 = _mm_or_si128(xmm5, xmm7); \
+    _mm_storeu_si128((__m128i*)(p_buffer+8), xmm5);
+
 /*
  * convert RGB plane to RGB packed format,
- * mm0 -> B, mm1 -> R, mm2 -> G, mm3 -> 0,
- * mm4 -> GB, mm5 -> AR pixel 4-7,
- * mm6 -> GB, mm7 -> AR pixel 0-3
+ * mm0 -> B, mm1 -> R, mm2 -> G
  */
 
-#define MMX_UNPACK_32 "                                                     \n\
+#define MMX_UNPACK_32_ARGB "                                                \n\
 pxor      %%mm3, %%mm3  # zero mm3                                          \n\
-movq      %%mm0, %%mm6  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
-movq      %%mm1, %%mm7  #                 R7 R6 R5 R4 R3 R2 R1 R0           \n\
 movq      %%mm0, %%mm4  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
+punpcklbw %%mm2, %%mm4  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
 movq      %%mm1, %%mm5  #                 R7 R6 R5 R4 R3 R2 R1 R0           \n\
-punpcklbw %%mm2, %%mm6  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
-punpcklbw %%mm3, %%mm7  #                 00 R3 00 R2 00 R1 00 R0           \n\
-punpcklwd %%mm7, %%mm6  #                 00 R1 B1 G1 00 R0 B0 G0           \n\
-movq      %%mm6, (%3)   # Store ARGB1 ARGB0                                 \n\
-movq      %%mm0, %%mm6  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
-punpcklbw %%mm2, %%mm6  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
-punpckhwd %%mm7, %%mm6  #                 00 R3 G3 B3 00 R2 B3 G2           \n\
+punpcklbw %%mm3, %%mm5  #                 00 R3 00 R2 00 R1 00 R0           \n\
+movq      %%mm4, %%mm6  #                 G3 B3 G2 B2 G1 B1 G0 B0           \n\
+punpcklwd %%mm5, %%mm4  #                 00 R1 B1 G1 00 R0 B0 G0           \n\
+movq      %%mm4, (%3)   # Store ARGB1 ARGB0                                 \n\
+punpckhwd %%mm5, %%mm6  #                 00 R3 B3 G3 00 R2 B2 G2           \n\
 movq      %%mm6, 8(%3)  # Store ARGB3 ARGB2                                 \n\
-punpckhbw %%mm2, %%mm4  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
-punpckhbw %%mm3, %%mm5  #                 00 R7 00 R6 00 R5 00 R4           \n\
-punpcklwd %%mm5, %%mm4  #                 00 R5 B5 G5 00 R4 B4 G4           \n\
-movq      %%mm4, 16(%3) # Store ARGB5 ARGB4                                 \n\
-movq      %%mm0, %%mm4  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
-punpckhbw %%mm2, %%mm4  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
-punpckhwd %%mm5, %%mm4  #                 00 R7 G7 B7 00 R6 B6 G6           \n\
-movq      %%mm4, 24(%3) # Store ARGB7 ARGB6                                 \n\
-                                                                            \n\
-#movd      4(%1), %%mm0  # Load 4 Cb       00 00 00 00 u3 u2 u1 u0           \n\
-#movd      4(%2), %%mm1  # Load 4 Cr       00 00 00 00 v3 v2 v1 v0           \n\
-#pxor      %%mm4, %%mm4  # zero mm4                                          \n\
-#movq      8(%0), %%mm6  # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0           \n\
+punpckhbw %%mm2, %%mm0  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpckhbw %%mm3, %%mm1  #                 00 R7 00 R6 00 R5 00 R4           \n\
+movq      %%mm0, %%mm5  #                 G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpcklwd %%mm1, %%mm5  #                 00 R5 B5 G5 00 R4 B4 G4           \n\
+movq      %%mm5, 16(%3) # Store ARGB5 ARGB4                                 \n\
+punpckhwd %%mm1, %%mm0  #                 00 R7 B7 G7 00 R6 B6 G6           \n\
+movq      %%mm0, 24(%3) # Store ARGB7 ARGB6                                 \n\
+"
+
+#define SSE2_UNPACK_32_ARGB_ALIGNED "                                       \n\
+pxor      %%xmm3, %%xmm3  # zero xmm3                                       \n\
+movdqa    %%xmm0, %%xmm4  #               B7 B6 B5 B4 B3 B2 B1 B0           \n\
+punpcklbw %%xmm2, %%xmm4  #               G3 B3 G2 B2 G1 B1 G0 B0           \n\
+movdqa    %%xmm1, %%xmm5  #               R7 R6 R5 R4 R3 R2 R1 R0           \n\
+punpcklbw %%xmm3, %%xmm5  #               00 R3 00 R2 00 R1 00 R0           \n\
+movdqa    %%xmm4, %%xmm6  #               G3 B3 G2 B2 G1 B1 G0 B0           \n\
+punpcklwd %%xmm5, %%xmm4  #               00 R1 B1 G1 00 R0 B0 G0           \n\
+movntdq   %%xmm4, (%3)    # Store ARGB3 ARGB2 ARGB1 ARGB0                   \n\
+punpckhwd %%xmm5, %%xmm6  #               00 R3 B3 G3 00 R2 B2 G2           \n\
+movntdq   %%xmm6, 16(%3)  # Store ARGB7 ARGB6 ARGB5 ARGB4                   \n\
+punpckhbw %%xmm2, %%xmm0  #               G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpckhbw %%xmm3, %%xmm1  #               00 R7 00 R6 00 R5 00 R4           \n\
+movdqa    %%xmm0, %%xmm5  #               G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpcklwd %%xmm1, %%xmm5  #               00 R5 B5 G5 00 R4 B4 G4           \n\
+movntdq   %%xmm5, 32(%3)  # Store ARGB11 ARGB10 ARGB9 ARGB8                 \n\
+punpckhwd %%xmm1, %%xmm0  #               00 R7 B7 G7 00 R6 B6 G6           \n\
+movntdq   %%xmm0, 48(%3)  # Store ARGB15 ARGB14 ARGB13 ARGB12               \n\
+"
+
+#define SSE2_UNPACK_32_ARGB_UNALIGNED "                                     \n\
+pxor      %%xmm3, %%xmm3  # zero xmm3                                       \n\
+movdqa    %%xmm0, %%xmm4  #               B7 B6 B5 B4 B3 B2 B1 B0           \n\
+punpcklbw %%xmm2, %%xmm4  #               G3 B3 G2 B2 G1 B1 G0 B0           \n\
+movdqa    %%xmm1, %%xmm5  #               R7 R6 R5 R4 R3 R2 R1 R0           \n\
+punpcklbw %%xmm3, %%xmm5  #               00 R3 00 R2 00 R1 00 R0           \n\
+movdqa    %%xmm4, %%xmm6  #               G3 B3 G2 B2 G1 B1 G0 B0           \n\
+punpcklwd %%xmm5, %%xmm4  #               00 R1 B1 G1 00 R0 B0 G0           \n\
+movdqu    %%xmm4, (%3)    # Store ARGB3 ARGB2 ARGB1 ARGB0                   \n\
+punpckhwd %%xmm5, %%xmm6  #               00 R3 B3 G3 00 R2 B2 G2           \n\
+movdqu    %%xmm6, 16(%3)  # Store ARGB7 ARGB6 ARGB5 ARGB4                   \n\
+punpckhbw %%xmm2, %%xmm0  #               G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpckhbw %%xmm3, %%xmm1  #               00 R7 00 R6 00 R5 00 R4           \n\
+movdqa    %%xmm0, %%xmm5  #               G7 B7 G6 B6 G5 B5 G4 B4           \n\
+punpcklwd %%xmm1, %%xmm5  #               00 R5 B5 G5 00 R4 B4 G4           \n\
+movdqu    %%xmm5, 32(%3)  # Store ARGB11 ARGB10 ARGB9 ARGB8                 \n\
+punpckhwd %%xmm1, %%xmm0  #               00 R7 B7 G7 00 R6 B6 G6           \n\
+movdqu    %%xmm0, 48(%3)  # Store ARGB15 ARGB14 ARGB13 ARGB12               \n\
 "
 
-#define INTRINSICS_UNPACK_32 \
-    mm3 = (__m64)(uint64_t)0; \
-    mm6 = mm0; \
-    mm7 = mm1; \
+#define MMX_INTRINSICS_UNPACK_32_ARGB \
+    mm3 = _mm_setzero_si64(); \
     mm4 = mm0; \
+    mm4 = _mm_unpacklo_pi8(mm4, mm2); \
     mm5 = mm1; \
-    mm6 = _mm_unpacklo_pi8(mm6, mm2); \
-    mm7 = _mm_unpacklo_pi8(mm7, mm3); \
-    mm6 = _mm_unpacklo_pi16(mm6, mm7); \
-    *(uint64_t *)p_buffer = (uint64_t)mm6; \
-    mm6 = mm0; \
-    mm6 = _mm_unpacklo_pi8(mm6, mm2); \
-    mm6 = _mm_unpackhi_pi16(mm6, mm7); \
-    *(uint64_t *)(p_buffer + 2) = (uint64_t)mm6; \
-    mm4 = _mm_unpackhi_pi8(mm4, mm2); \
-    mm5 = _mm_unpackhi_pi8(mm5, mm3); \
+    mm5 = _mm_unpacklo_pi8(mm5, mm3); \
+    mm6 = mm4; \
     mm4 = _mm_unpacklo_pi16(mm4, mm5); \
-    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm4; \
-    mm4 = mm0; \
-    mm4 = _mm_unpackhi_pi8(mm4, mm2); \
-    mm4 = _mm_unpackhi_pi16(mm4, mm5); \
-    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm4; \
+    *(uint64_t *)p_buffer = (uint64_t)mm4; \
+    mm6 = _mm_unpackhi_pi16(mm6, mm5); \
+    *(uint64_t *)(p_buffer + 2) = (uint64_t)mm6; \
+    mm0 = _mm_unpackhi_pi8(mm0, mm2); \
+    mm1 = _mm_unpackhi_pi8(mm1, mm3); \
+    mm5 = mm0; \
+    mm5 = _mm_unpacklo_pi16(mm5, mm1); \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5; \
+    mm0 = _mm_unpackhi_pi16(mm0, mm1); \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm0;
+
+#define SSE2_INTRINSICS_UNPACK_32_ARGB_ALIGNED \
+    xmm3 = _mm_setzero_si128();          \
+    xmm4 = xmm0; \
+    xmm4 = _mm_unpacklo_epi8(xmm4, xmm2); \
+    xmm5 = xmm1; \
+    xmm5 = _mm_unpacklo_epi8(xmm5, xmm3); \
+    xmm6 = xmm4; \
+    xmm4 = _mm_unpacklo_epi16(xmm4, xmm5); \
+    _mm_stream_si128((__m128i*)(p_buffer), xmm4); \
+    xmm6 = _mm_unpackhi_epi16(xmm6, xmm5); \
+    _mm_stream_si128((__m128i*)(p_buffer+4), xmm6); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm2); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm3); \
+    xmm5 = xmm0; \
+    xmm5 = _mm_unpacklo_epi16(xmm5, xmm1); \
+    _mm_stream_si128((__m128i*)(p_buffer+8), xmm5); \
+    xmm0 = _mm_unpackhi_epi16(xmm0, xmm1); \
+    _mm_stream_si128((__m128i*)(p_buffer+12), xmm0);
+
+#define SSE2_INTRINSICS_UNPACK_32_ARGB_UNALIGNED \
+    xmm3 = _mm_setzero_si128();          \
+    xmm4 = xmm0; \
+    xmm4 = _mm_unpacklo_epi8(xmm4, xmm2); \
+    xmm5 = xmm1; \
+    xmm5 = _mm_unpacklo_epi8(xmm5, xmm3); \
+    xmm6 = xmm4; \
+    xmm4 = _mm_unpacklo_epi16(xmm4, xmm5); \
+    _mm_storeu_si128((__m128i*)(p_buffer), xmm4); \
+    xmm6 = _mm_unpackhi_epi16(xmm6, xmm5); \
+    _mm_storeu_si128((__m128i*)(p_buffer+4), xmm6); \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm2); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm3); \
+    xmm5 = xmm0; \
+    xmm5 = _mm_unpacklo_epi16(xmm5, xmm1); \
+    _mm_storeu_si128((__m128i*)(p_buffer+8), xmm5); \
+    xmm0 = _mm_unpackhi_epi16(xmm0, xmm1); \
+    _mm_storeu_si128((__m128i*)(p_buffer+12), xmm0);
+
+#define MMX_UNPACK_32_BGRA "                                                \n\
+pxor      %%mm3, %%mm3  # zero mm3                                          \n\
+movq      %%mm2, %%mm4  #                 G7 G6 G5 G4 G3 G2 G1 G0           \n\
+punpcklbw %%mm0, %%mm4  #                 B3 G3 B2 G2 B1 G1 B0 G0           \n\
+punpcklbw %%mm1, %%mm3  #                 R3 00 R2 00 R1 00 R0 00           \n\
+movq      %%mm3, %%mm5  #                 R3 00 R2 00 R1 00 R0 00           \n\
+punpcklwd %%mm4, %%mm3  #                 B1 G1 R1 00 B0 G0 R0 00           \n\
+movq      %%mm3, (%3)   # Store BGRA1 BGRA0                                 \n\
+punpckhwd %%mm4, %%mm5  #                 B3 G3 R3 00 B2 G2 R2 00           \n\
+movq      %%mm5, 8(%3)  # Store BGRA3 BGRA2                                 \n\
+pxor      %%mm3, %%mm3  # zero mm3                                          \n\
+movq      %%mm2, %%mm4  #                 G7 G6 G5 G4 G3 G2 G1 G0           \n\
+punpckhbw %%mm0, %%mm4  #                 B7 G7 B6 G6 B5 G5 B4 G4           \n\
+punpckhbw %%mm1, %%mm3  #                 R7 00 R6 00 R5 00 R4 00           \n\
+movq      %%mm3, %%mm5  #                 R7 00 R6 00 R5 00 R4 00           \n\
+punpcklwd %%mm1, %%mm3  #                 B5 G5 R5 00 B4 G4 R4 00           \n\
+movq      %%mm3, 16(%3) # Store BGRA5 BGRA4                                 \n\
+punpckhwd %%mm4, %%mm5  #                 B7 G7 R7 00 B6 G6 R6 00           \n\
+movq      %%mm5, 24(%3) # Store BGRA7 BGRA6                                 \n\
+"
+
+#define SSE2_UNPACK_32_BGRA_ALIGNED "                                       \n\
+pxor      %%xmm3, %%xmm3  # zero mm3                                        \n\
+movdqa    %%xmm2, %%xmm4  #                 G7 G6 G5 G4 G3 G2 G1 G0         \n\
+punpcklbw %%xmm0, %%xmm4  #                 B3 G3 B2 G2 B1 G1 B0 G0         \n\
+punpcklbw %%xmm1, %%xmm3  #                 R3 00 R2 00 R1 00 R0 00         \n\
+movdqa    %%xmm3, %%xmm5  #                 R3 00 R2 00 R1 00 R0 00         \n\
+punpcklwd %%xmm4, %%xmm3  #                 B1 G1 R1 00 B0 G0 R0 00         \n\
+movntdq   %%xmm3, (%3)    # Store BGRA3 BGRA2 BGRA1 BGRA0                   \n\
+punpckhwd %%xmm4, %%xmm5  #                 B3 G3 R3 00 B2 G2 R2 00         \n\
+movntdq   %%xmm5, 8(%3)   # Store BGRA7 BGRA6 BGRA5 BGRA4                   \n\
+pxor      %%xmm3, %%xmm3  # zero mm3                                        \n\
+movdqa    %%xmm2, %%xmm4  #                 G7 G6 G5 G4 G3 G2 G1 G0         \n\
+punpckhbw %%xmm0, %%xmm4  #                 B7 G7 B6 G6 B5 G5 B4 G4         \n\
+punpckhbw %%xmm1, %%xmm3  #                 R7 00 R6 00 R5 00 R4 00         \n\
+movdqa    %%xmm3, %%xmm5  #                 R7 00 R6 00 R5 00 R4 00         \n\
+punpcklwd %%xmm1, %%xmm3  #                 B5 G5 R5 00 B4 G4 R4 00         \n\
+movntdq   %%xmm3, 16(%3)  # Store BGRA11 BGRA10 BGRA9 BGRA8                 \n\
+punpckhwd %%xmm4, %%xmm5  #                 B7 G7 R7 00 B6 G6 R6 00         \n\
+movntdq   %%xmm5, 24(%3)  # Store BGRA15 BGRA14 BGRA13 BGRA12               \n\
+"
+
+#define SSE2_UNPACK_32_BGRA_UNALIGNED "                                     \n\
+pxor      %%xmm3, %%xmm3  # zero mm3                                        \n\
+movdqa    %%xmm2, %%xmm4  #                 G7 G6 G5 G4 G3 G2 G1 G0         \n\
+punpcklbw %%xmm0, %%xmm4  #                 B3 G3 B2 G2 B1 G1 B0 G0         \n\
+punpcklbw %%xmm1, %%xmm3  #                 R3 00 R2 00 R1 00 R0 00         \n\
+movdqa    %%xmm3, %%xmm5  #                 R3 00 R2 00 R1 00 R0 00         \n\
+punpcklwd %%xmm4, %%xmm3  #                 B1 G1 R1 00 B0 G0 R0 00         \n\
+movdqu    %%xmm3, (%3)    # Store BGRA3 BGRA2 BGRA1 BGRA0                   \n\
+punpckhwd %%xmm4, %%xmm5  #                 B3 G3 R3 00 B2 G2 R2 00         \n\
+movdqu    %%xmm5, 8(%3)   # Store BGRA7 BGRA6 BGRA5 BGRA4                   \n\
+pxor      %%xmm3, %%xmm3  # zero mm3                                        \n\
+movdqa    %%xmm2, %%xmm4  #                 G7 G6 G5 G4 G3 G2 G1 G0         \n\
+punpckhbw %%xmm0, %%xmm4  #                 B7 G7 B6 G6 B5 G5 B4 G4         \n\
+punpckhbw %%xmm1, %%xmm3  #                 R7 00 R6 00 R5 00 R4 00         \n\
+movdqa    %%xmm3, %%xmm5  #                 R7 00 R6 00 R5 00 R4 00         \n\
+punpcklwd %%xmm1, %%xmm3  #                 B5 G5 R5 00 B4 G4 R4 00         \n\
+movdqu    %%xmm3, 16(%3)  # Store BGRA11 BGRA10 BGRA9 BGRA8                 \n\
+punpckhwd %%xmm4, %%xmm5  #                 B7 G7 R7 00 B6 G6 R6 00         \n\
+movdqu    %%xmm5, 24(%3)  # Store BGRA15 BGRA14 BGRA13 BGRA12               \n\
+"
+
+#define MMX_INTRINSICS_UNPACK_32_BGRA \
+    mm3 = _mm_setzero_si64(); \
+    mm4 = mm2; \
+    mm4 = _mm_unpacklo_pi8(mm4, mm0); \
+    mm1 = _mm_unpacklo_pi8(mm1, mm3); \
+    mm5 = mm3; \
+    mm3 = _mm_unpacklo_pi16(mm3, mm4); \
+    *(uint64_t *)p_buffer = (uint64_t)mm3; \
+    mm5 = _mm_unpackhi_pi16(mm5, mm4); \
+    *(uint64_t *)(p_buffer + 2) = (uint64_t)mm5; \
+    mm3 = _mm_setzero_si64(); \
+    mm4 = mm2; \
+    mm0 = _mm_unpackhi_pi8(mm0, mm4); \
+    mm1 = _mm_unpackhi_pi8(mm1, mm3); \
+    mm5 = mm3; \
+    mm3 = _mm_unpacklo_pi16(mm3, mm1); \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm3; \
+    mm5 = _mm_unpackhi_pi16(mm5, mm4); \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm5; \
+
+#define SSE2_INTRINSICS_UNPACK_32_BGRA_ALIGNED \
+    xmm3 = _mm_setzero_si128(); \
+    xmm4 = xmm2; \
+    xmm4 = _mm_unpacklo_epi8(xmm4, xmm0); \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm3); \
+    xmm5 = xmm3; \
+    xmm3 = _mm_unpacklo_epi16(xmm3, xmm4); \
+    _mm_stream_si128((__m128i*)(p_buffer), xmm3); \
+    xmm5 = _mm_unpackhi_epi16(xmm5, xmm4); \
+    _mm_stream_si128((__m128i*)(p_buffer+4), xmm5); \
+    xmm3 = _mm_setzero_si128(); \
+    xmm4 = xmm2; \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm4); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm3); \
+    xmm5 = xmm3; \
+    xmm3 = _mm_unpacklo_epi16(xmm3, xmm1); \
+    _mm_stream_si128((__m128i*)(p_buffer+8), xmm3); \
+    xmm5 = _xmm_unpackhi_pi16(xmm5, xmm4); \
+    _mm_stream_si128((__m128i*)(p_buffer+12), xmm5); \
+
+#define SSE2_INTRINSICS_UNPACK_32_BGRA_UNALIGNED \
+    xmm3 = _mm_setzero_si128(); \
+    xmm4 = xmm2; \
+    xmm4 = _mm_unpacklo_epi8(xmm4, xmm0); \
+    xmm1 = _mm_unpacklo_epi8(xmm1, xmm3); \
+    xmm5 = xmm3; \
+    xmm3 = _mm_unpacklo_epi16(xmm3, xmm4); \
+    _mm_storeu_si128((__m128i*)(p_buffer), xmm3); \
+    xmm5 = _mm_unpackhi_epi16(xmm5, xmm4); \
+    _mm_storeu_si128((__m128i*)(p_buffer+4), xmm5); \
+    xmm3 = _mm_setzero_si128(); \
+    xmm4 = xmm2; \
+    xmm0 = _mm_unpackhi_epi8(xmm0, xmm4); \
+    xmm1 = _mm_unpackhi_epi8(xmm1, xmm3); \
+    xmm5 = xmm3; \
+    xmm3 = _mm_unpacklo_epi16(xmm3, xmm1); \
+    _mm_storeu_si128((__m128i*)(p_buffer+8), xmm3); \
+    xmm5 = _xmm_unpackhi_pi16(xmm5, xmm4); \
+    _mm_storeu_si128((__m128i*)(p_buffer+12), xmm5); \
 

modules/video_chroma/i420_yuy2.c

@@ -306,7 +306,8 @@ static void I420_YUY2( vout_thread_t *p_vout, picture_t *p_source,
     }
 
 #if defined (MODULE_NAME_IS_i420_yuy2_mmx)
-    __asm__ __volatile__("emms" :: );
+    /* re-enable FPU registers */
+    __asm__ __volatile__ ( "emms" );
 #endif
 
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
@@ -347,6 +348,8 @@ static void I420_YUY2( vout_thread_t *p_vout, picture_t *p_source,
             p_line1 += i_dest_margin;
             p_line2 += i_dest_margin;
         }
+        /* make sure all SSE2 stores are visible thereafter */
+        __asm__ __volatile__ ( "sfence" );
     }
     else
     {
@@ -514,7 +517,8 @@ static void I420_YVYU( vout_thread_t *p_vout, picture_t *p_source,
     }
 
 #if defined (MODULE_NAME_IS_i420_yuy2_mmx)
-    __asm__ __volatile__("emms" :: );
+    /* re-enable FPU registers */
+    __asm__ __volatile__ ( "emms" );
 #endif
 
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
@@ -554,6 +558,8 @@ static void I420_YVYU( vout_thread_t *p_vout, picture_t *p_source,
             p_line1 += i_dest_margin;
             p_line2 += i_dest_margin;
         }
+        /* make sure all SSE2 stores are visible thereafter */
+        __asm__ __volatile__ ( "sfence" );
     }
     else
     {
@@ -720,7 +726,8 @@ static void I420_UYVY( vout_thread_t *p_vout, picture_t *p_source,
     }
 
 #if defined (MODULE_NAME_IS_i420_yuy2_mmx)
-    __asm__ __volatile__("emms" :: );
+    /* re-enable FPU registers */
+    __asm__ __volatile__ ( "emms" );
 #endif
 
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
@@ -760,6 +767,8 @@ static void I420_UYVY( vout_thread_t *p_vout, picture_t *p_source,
             p_line1 += i_dest_margin;
             p_line2 += i_dest_margin;
         }
+        /* make sure all SSE2 stores are visible thereafter */
+        __asm__ __volatile__ ( "sfence" );
     }
     else
     {
@@ -861,7 +870,8 @@ static void I420_cyuv( vout_thread_t *p_vout, picture_t *p_source,
     }
 
 #if defined (MODULE_NAME_IS_i420_yuy2_mmx)
-    __asm__ __volatile__("emms" :: );
+    /* re-enable FPU registers */
+    __asm__ __volatile__ ( "emms" );
 #endif
 
 #else // defined(MODULE_NAME_IS_i420_yuy2_sse2)
@@ -897,6 +907,8 @@ static void I420_cyuv( vout_thread_t *p_vout, picture_t *p_source,
             p_line1 += i_dest_margin;
             p_line2 += i_dest_margin;
         }
+        /* make sure all SSE2 stores are visible thereafter */
+        __asm__ __volatile__ ( "sfence" );
     }
     else
     {

modules/video_chroma/i420_yuy2.h

@@ -136,14 +136,14 @@ movdqa      (%3), %%xmm3  # Load 16 Y         Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
 punpcklbw %%xmm2, %%xmm1  #                   v7 u7 v6 u6 .. u1 v0 u0     \n\
 movdqa    %%xmm0, %%xmm2  #                   y15 y14 y13 .. y2 y1 y0     \n\
 punpcklbw %%xmm1, %%xmm2  #                   v3 y7 u3 .. v0 y1 u0 y0     \n\
-movdqa    %%xmm2, (%0)    # Store low YUYV                                \n\
+movntdq   %%xmm2, (%0)    # Store low YUYV                                \n\
 punpckhbw %%xmm1, %%xmm0  #                   v3 y7 u3 y6 v2 y5 u2 y4     \n\
-movdqa    %%xmm0, 16(%0)  # Store high YUYV                               \n\
+movntdq   %%xmm0, 16(%0)  # Store high YUYV                               \n\
 movdqa    %%xmm3, %%xmm4  #                   Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
 punpcklbw %%xmm1, %%xmm4  #                   v1 Y3 u1 Y2 v0 Y1 u0 Y0     \n\
-movdqa    %%xmm4, (%1)    # Store low YUYV                                \n\
+movntdq   %%xmm4, (%1)    # Store low YUYV                                \n\
 punpckhbw %%xmm1, %%xmm3  #                   v3 Y7 u3 Y6 v2 Y5 u2 Y4     \n\
-movdqa    %%xmm3, 16(%1)  # Store high YUYV                               \n\
+movntdq   %%xmm3, 16(%1)  # Store high YUYV                               \n\
 "
 
 #define SSE2_YUV420_YUYV_UNALIGNED "                                      \n\
@@ -172,14 +172,14 @@ movdqa      (%3), %%xmm3  # Load 16 Y           Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
 punpcklbw %%xmm2, %%xmm1  #                     u3 v3 u2 v2 u1 v1 u0 v0     \n\
 movdqa    %%xmm0, %%xmm2  #                     y7 y6 y5 y4 y3 y2 y1 y0     \n\
 punpcklbw %%xmm1, %%xmm2  #                     u1 y3 v1 y2 u0 y1 v0 y0     \n\
-movdqa    %%xmm2, (%0)    # Store low YUYV                                  \n\
+movntdq   %%xmm2, (%0)    # Store low YUYV                                  \n\
 punpckhbw %%xmm1, %%xmm0  #                     u3 y7 v3 y6 u2 y5 v2 y4     \n\
-movdqa    %%xmm0, 16(%0)  # Store high YUYV                                 \n\
+movntdq   %%xmm0, 16(%0)  # Store high YUYV                                 \n\
 movdqa    %%xmm3, %%xmm4  #                     Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
 punpcklbw %%xmm1, %%xmm4  #                     u1 Y3 v1 Y2 u0 Y1 v0 Y0     \n\
-movdqa    %%xmm4, (%1)    # Store low YUYV                                  \n\
+movntdq   %%xmm4, (%1)    # Store low YUYV                                  \n\
 punpckhbw %%xmm1, %%xmm3  #                     u3 Y7 v3 Y6 u2 Y5 v2 Y4     \n\
-movdqa    %%xmm3, 16(%1)  # Store high YUYV                                 \n\
+movntdq   %%xmm3, 16(%1)  # Store high YUYV                                 \n\
 "
 
 #define SSE2_YUV420_YVYU_UNALIGNED "                                        \n\
@@ -208,15 +208,15 @@ movdqa      (%3), %%xmm3  # Load 16 Y           Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
 punpcklbw %%xmm2, %%xmm1  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
 movdqa    %%xmm1, %%xmm2  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
 punpcklbw %%xmm0, %%xmm2  #                     y3 v1 y2 u1 y1 v0 y0 u0     \n\
-movdqa    %%xmm2, (%0)    # Store low UYVY                                  \n\
+movntdq   %%xmm2, (%0)    # Store low UYVY                                  \n\
 movdqa    %%xmm1, %%xmm2  #                     u3 v3 u2 v2 u1 v1 u0 v0     \n\
 punpckhbw %%xmm0, %%xmm2  #                     y3 v1 y2 u1 y1 v0 y0 u0     \n\
-movdqa    %%xmm2, 16(%0)  # Store high UYVY                                 \n\
+movntdq   %%xmm2, 16(%0)  # Store high UYVY                                 \n\
 movdqa    %%xmm1, %%xmm4  #                     u3 v3 u2 v2 u1 v1 u0 v0     \n\
 punpcklbw %%xmm3, %%xmm4  #                     Y3 v1 Y2 u1 Y1 v0 Y0 u0     \n\
-movdqa    %%xmm4, (%1)    # Store low UYVY                                  \n\
+movntdq   %%xmm4, (%1)    # Store low UYVY                                  \n\
 punpckhbw %%xmm3, %%xmm1  #                     Y7 v3 Y6 u3 Y5 v2 Y4 u2     \n\
-movdqa    %%xmm1, 16(%1)  # Store high UYVY                                 \n\
+movntdq   %%xmm1, 16(%1)  # Store high UYVY                                 \n\
 "
 
 #define SSE2_YUV420_UYVY_UNALIGNED "                                        \n\