[vlc-commits] Revert "i420_yuy2/i422_yuy2/i420_rgb: purge MMX"

Rémi Denis-Courmont git at videolan.org
Tue Mar 12 18:12:02 CET 2019


vlc | branch: master | Rémi Denis-Courmont <remi at remlab.net> | Tue Mar 12 19:07:13 2019 +0200| [0620a2600ac5379b320d8b4ae738816cc44d55e4] | committer: Rémi Denis-Courmont

Revert "i420_yuy2/i422_yuy2/i420_rgb: purge MMX"

This reverts commit e04412a31bf8bc246b076843ff8eeb6d7ad1cf1b.

> http://git.videolan.org/gitweb.cgi/vlc.git/?a=commit;h=0620a2600ac5379b320d8b4ae738816cc44d55e4
---

 modules/MODULES_LIST                  |   3 +
 modules/video_chroma/Makefile.am      |  20 ++
 modules/video_chroma/i420_rgb.c       |   5 +
 modules/video_chroma/i420_rgb.h       |   2 +-
 modules/video_chroma/i420_rgb16_x86.c | 328 +++++++++++++++++++++
 modules/video_chroma/i420_rgb_mmx.h   | 540 ++++++++++++++++++++++++++++++++++
 modules/video_chroma/i420_yuy2.c      |  34 +++
 modules/video_chroma/i420_yuy2.h      | 148 +++++++++-
 modules/video_chroma/i422_yuy2.c      |  26 ++
 modules/video_chroma/i422_yuy2.h      | 111 ++++++-
 10 files changed, 1214 insertions(+), 3 deletions(-)

diff --git a/modules/MODULES_LIST b/modules/MODULES_LIST
index c6aad68600..25ea4212ed 100644
--- a/modules/MODULES_LIST
+++ b/modules/MODULES_LIST
@@ -188,14 +188,17 @@ $Id$
  * https: HTTP/TLS access module for HTTP 2.0 support
  * i420_nv12: planar YUV to semi-planar YUV conversion functions
  * i420_rgb: planar YUV to packed RGB conversion functions
+ * i420_rgb_mmx: MMX accelerated version of i420_rgb
  * i420_rgb_sse2: sse2 accelerated version of i420_rgb
  * i420_rgb_avx2: avx2 accelerated version of i420_rgb
  * i420_yuy2: planar 4:2:0 YUV to packed YUV conversion functions
  * i420_yuy2_altivec: AltiVec accelerated version of i420_yuy2
+ * i420_yuy2_mmx: MMX accelerated version of i420_yuy2
  * i420_yuy2_sse2: sse2 accelerated version of i420_yuy2
  * i420_yuy2_avx2: avx2 accelerated version of i420_yuy2
  * i422_i420: 4:2:2 to 4:2:0 conversion functions
  * i422_yuy2: planar 4:2:2 YUV to packed YUV conversion functions
+ * i422_yuy2_mmx: MMX accelerated version of i422_yuy2
  * i422_yuy2_sse2: sse2 accelerated version of i422_yuy2
  * i422_yuy2_avx2: avx2 accelerated version of i422_yuy2
  * idummy: dummy input
diff --git a/modules/video_chroma/Makefile.am b/modules/video_chroma/Makefile.am
index 0558786744..4b51140677 100644
--- a/modules/video_chroma/Makefile.am
+++ b/modules/video_chroma/Makefile.am
@@ -70,6 +70,26 @@ chroma_LTLIBRARIES += \
 	libi420_yuy2_altivec_plugin.la
 endif
 
+# MMX
+libi420_rgb_mmx_plugin_la_SOURCES = video_chroma/i420_rgb.c video_chroma/i420_rgb.h \
+	video_chroma/i420_rgb16_x86.c video_chroma/i420_rgb_mmx.h
+libi420_rgb_mmx_plugin_la_CPPFLAGS = $(AM_CPPFLAGS) -DMMX
+
+libi420_yuy2_mmx_plugin_la_SOURCES = video_chroma/i420_yuy2.c video_chroma/i420_yuy2.h
+libi420_yuy2_mmx_plugin_la_CPPFLAGS = $(AM_CPPFLAGS) \
+	-DMODULE_NAME_IS_i420_yuy2_mmx
+
+libi422_yuy2_mmx_plugin_la_SOURCES = video_chroma/i422_yuy2.c video_chroma/i422_yuy2.h
+libi422_yuy2_mmx_plugin_la_CPPFLAGS = $(AM_CPPFLAGS) \
+	-DMODULE_NAME_IS_i422_yuy2_mmx
+
+if HAVE_MMX
+chroma_LTLIBRARIES += \
+	libi420_rgb_mmx_plugin.la \
+	libi420_yuy2_mmx_plugin.la \
+	libi422_yuy2_mmx_plugin.la
+endif
+
 # SSE2
 libi420_rgb_sse2_plugin_la_SOURCES = video_chroma/i420_rgb.c video_chroma/i420_rgb.h \
 	video_chroma/i420_rgb16_x86.c video_chroma/i420_rgb_sse2.h
diff --git a/modules/video_chroma/i420_rgb.c b/modules/video_chroma/i420_rgb.c
index 7999a9dfa9..b1ed2ac57f 100644
--- a/modules/video_chroma/i420_rgb.c
+++ b/modules/video_chroma/i420_rgb.c
@@ -78,6 +78,11 @@ vlc_module_begin ()
                         "RV15,RV16,RV24,RV32 conversions") )
     set_capability( "video converter", 120 )
 # define vlc_CPU_capable() vlc_CPU_SSE2()
+#elif defined (MMX)
+    set_description( N_( "MMX I420,IYUV,YV12 to "
+                        "RV15,RV16,RV24,RV32 conversions") )
+    set_capability( "video converter", 100 )
+# define vlc_CPU_capable() vlc_CPU_MMX()
 #else
     set_description( N_("I420,IYUV,YV12 to "
                        "RGB8,RV15,RV16,RV24,RV32 conversions") )
diff --git a/modules/video_chroma/i420_rgb.h b/modules/video_chroma/i420_rgb.h
index 06c8122606..aeb9ac1c2a 100644
--- a/modules/video_chroma/i420_rgb.h
+++ b/modules/video_chroma/i420_rgb.h
@@ -21,7 +21,7 @@
  *****************************************************************************/
 #include <limits.h>
 
-#if !defined (AVX2) && !defined (SSE2)
+#if !defined (AVX2) && !defined (SSE2) && !defined (MMX)
 # define PLAIN
 #endif
 
diff --git a/modules/video_chroma/i420_rgb16_x86.c b/modules/video_chroma/i420_rgb16_x86.c
index 9e79bb819d..2bd819a58f 100644
--- a/modules/video_chroma/i420_rgb16_x86.c
+++ b/modules/video_chroma/i420_rgb16_x86.c
@@ -39,6 +39,10 @@
 # include "i420_rgb_sse2.h"
 # define VLC_TARGET VLC_SSE
 #endif
+#ifdef MMX
+# include "i420_rgb_mmx.h"
+# define VLC_TARGET VLC_MMX
+#endif
 
 /*****************************************************************************
  * SetOffset: build offset array for conversion functions
@@ -393,6 +397,60 @@ void I420_R5G5B5( filter_t *p_filter, picture_t *p_src, picture_t *p_dest )
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_16
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_15
+            );
+            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;
+
+            MMX_CALL (
+                MMX_INIT_16
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_15
+            );
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+        }
+        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 */
+    MMX_END;
 #endif
 }
 
@@ -689,6 +747,60 @@ void I420_R5G6B5( filter_t *p_filter, picture_t *p_src, picture_t *p_dest )
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_16
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_16
+            );
+            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;
+
+            MMX_CALL (
+                MMX_INIT_16
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_16
+            );
+            p_y += 8;
+            p_u += 4;
+            p_v += 4;
+        }
+        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 */
+    MMX_END;
 #endif
 }
 
@@ -985,6 +1097,60 @@ void I420_A8R8G8B8( filter_t *p_filter, picture_t *p_src,
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_ARGB
+            );
+            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;
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_ARGB
+            );
+            p_y += 8;
+            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;
+        }
+    }
+
+    /* re-enable FPU registers */
+    MMX_END;
 #endif
 }
 
@@ -1280,6 +1446,60 @@ void I420_R8G8B8A8( filter_t *p_filter, picture_t *p_src, picture_t *p_dest )
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_RGBA
+            );
+            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;
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_RGBA
+            );
+            p_y += 8;
+            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;
+        }
+    }
+
+    /* re-enable FPU registers */
+    MMX_END;
 #endif
 }
 
@@ -1575,6 +1795,60 @@ void I420_B8G8R8A8( filter_t *p_filter, picture_t *p_src, picture_t *p_dest )
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_BGRA
+            );
+            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;
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_BGRA
+            );
+            p_y += 8;
+            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;
+        }
+    }
+
+    /* re-enable FPU registers */
+    MMX_END;
 #endif
 }
 
@@ -1870,5 +2144,59 @@ void I420_A8B8G8R8( filter_t *p_filter, picture_t *p_src, picture_t *p_dest )
     /* make sure all SSE2 stores are visible thereafter */
     SSE2_END;
 
+#elif defined (MODULE_NAME_IS_i420_rgb_mmx) // MMX
+
+    i_rewind = (-(p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width)) & 7;
+
+    for( i_y = 0; i_y < (p_filter->fmt_in.video.i_y_offset + p_filter->fmt_in.video.i_visible_height); i_y++ )
+    {
+        p_pic_start = p_pic;
+        p_buffer = b_hscale ? p_buffer_start : p_pic;
+
+        for ( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x--; )
+        {
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_ABGR
+            );
+            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;
+            MMX_CALL (
+                MMX_INIT_32
+                MMX_YUV_MUL
+                MMX_YUV_ADD
+                MMX_UNPACK_32_ABGR
+            );
+            p_y += 8;
+            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;
+        }
+    }
+
+    /* re-enable FPU registers */
+    MMX_END;
 #endif
 }
diff --git a/modules/video_chroma/i420_rgb_mmx.h b/modules/video_chroma/i420_rgb_mmx.h
new file mode 100644
index 0000000000..4c2bb48ed5
--- /dev/null
+++ b/modules/video_chroma/i420_rgb_mmx.h
@@ -0,0 +1,540 @@
+/*****************************************************************************
+ * i420_rgb_mmx.h: MMX YUV transformation assembly
+ *****************************************************************************
+ * Copyright (C) 1999-2007 the VideoLAN team
+ *
+ * Authors: Olie Lho <ollie at sis.com.tw>
+ *          Gaël Hendryckx <jimmy at via.ecp.fr>
+ *          Samuel Hocevar <sam at zoy.org>
+ *          Damien Fouilleul <damienf at 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
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
+ *****************************************************************************/
+
+/* hope these constant values are cache line aligned */
+static const uint64_t mmx_80w     = 0x0080008000800080ULL; /* Will be referenced as %4 in inline asm */
+static const uint64_t mmx_10w     = 0x1010101010101010ULL; /* -- as %5 */
+static const uint64_t mmx_00ffw   = 0x00ff00ff00ff00ffULL; /* -- as %6 */
+static const uint64_t mmx_Y_coeff = 0x253f253f253f253fULL; /* -- as %7 */
+
+static const uint64_t mmx_U_green = 0xf37df37df37df37dULL; /* -- as %8 */
+static const uint64_t mmx_U_blue  = 0x4093409340934093ULL; /* -- as %9 */
+static const uint64_t mmx_V_red   = 0x3312331233123312ULL; /* -- as %10 */
+static const uint64_t mmx_V_green = 0xe5fce5fce5fce5fcULL; /* -- as %11 */
+
+static const uint64_t mmx_mask_f8 = 0xf8f8f8f8f8f8f8f8ULL; /* -- as %12 */
+static const uint64_t mmx_mask_fc = 0xfcfcfcfcfcfcfcfcULL; /* -- as %13 */
+
+#if defined(CAN_COMPILE_MMX)
+
+/* MMX assembly */
+ 
+#define MMX_CALL(MMX_INSTRUCTIONS)      \
+    do {                                \
+    __asm__ __volatile__(               \
+        ".p2align 3 \n\t"               \
+        MMX_INSTRUCTIONS                \
+        :                               \
+        : "r" (p_y), "r" (p_u),         \
+          "r" (p_v), "r" (p_buffer),    \
+          "m" (mmx_80w), "m" (mmx_10w), \
+          "m" (mmx_00ffw), "m" (mmx_Y_coeff), \
+          "m" (mmx_U_green), "m" (mmx_U_blue), \
+          "m" (mmx_V_red), "m" (mmx_V_green), \
+          "m" (mmx_mask_f8), "m" (mmx_mask_fc) \
+        : "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7" );  \
+    } while(0)
+
+#define MMX_END __asm__ __volatile__ ( "emms" )
+
+#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\
+pxor      %%mm4, %%mm4      # zero mm4                                      \n\
+movq       (%0), %%mm6      # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+"
+
+#define MMX_INIT_16_GRAY "                                                  \n\
+movq      (%0), %%mm6       # Load 8 Y        Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0       \n\
+#movl      $0, (%3)         # cache preload for image                       \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\
+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\
+"
+
+/*
+ * Do the multiply part of the conversion for even and odd pixels,
+ * register usage:
+ * mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels,
+ * mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd  pixels,
+ * mm6 -> Y even, mm7 -> Y odd
+ */
+
+#define MMX_YUV_MUL "                                                        \n\
+# convert the chroma part                                                    \n\
+punpcklbw %%mm4, %%mm0          # scatter 4 Cb     00 u3 00 u2 00 u1 00 u0   \n\
+punpcklbw %%mm4, %%mm1          # scatter 4 Cr     00 v3 00 v2 00 v1 00 v0   \n\
+psubsw    %4, %%mm0             # Cb -= 128                                  \n\
+psubsw    %4, %%mm1             # Cr -= 128                                  \n\
+psllw     $3, %%mm0             # Promote precision                          \n\
+psllw     $3, %%mm1             # Promote precision                          \n\
+movq      %%mm0, %%mm2          # Copy 4 Cb        00 u3 00 u2 00 u1 00 u0   \n\
+movq      %%mm1, %%mm3          # Copy 4 Cr        00 v3 00 v2 00 v1 00 v0   \n\
+pmulhw    %8, %%mm2             # Mul Cb with green coeff -> Cb green        \n\
+pmulhw    %11, %%mm3            # Mul Cr with green coeff -> Cr green        \n\
+pmulhw    %9, %%mm0             # Mul Cb -> Cblue  00 b3 00 b2 00 b1 00 b0   \n\
+pmulhw    %10, %%mm1            # Mul Cr -> Cred   00 r3 00 r2 00 r1 00 r0   \n\
+paddsw    %%mm3, %%mm2          # Cb green + Cr green -> Cgreen              \n\
+                                                                             \n\
+# convert the luma part                                                      \n\
+psubusb   %5, %%mm6             # Y -= 16                                    \n\
+movq      %%mm6, %%mm7          # Copy 8 Y         Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0   \n\
+pand      %6, %%mm6             # get Y even       00 Y6 00 Y4 00 Y2 00 Y0   \n\
+psrlw     $8, %%mm7             # get Y odd        00 Y7 00 Y5 00 Y3 00 Y1   \n\
+psllw     $3, %%mm6             # Promote precision                          \n\
+psllw     $3, %%mm7             # Promote precision                          \n\
+pmulhw    %7, %%mm6             # Mul 4 Y even     00 y6 00 y4 00 y2 00 y0   \n\
+pmulhw    %7, %%mm7             # Mul 4 Y odd      00 y7 00 y5 00 y3 00 y1   \n\
+"
+
+/*
+ * Do the addition part of the conversion for even and odd pixels,
+ * register usage:
+ * mm0 -> Cblue, mm1 -> Cred, mm2 -> Cgreen even pixels,
+ * mm3 -> Cblue, mm4 -> Cred, mm5 -> Cgreen odd  pixels,
+ * mm6 -> Y even, mm7 -> Y odd
+ */
+
+#define MMX_YUV_ADD "                                                        \n\
+# Do horizontal and vertical scaling                                         \n\
+movq      %%mm0, %%mm3          # Copy Cblue                                 \n\
+movq      %%mm1, %%mm4          # Copy Cred                                  \n\
+movq      %%mm2, %%mm5          # Copy Cgreen                                \n\
+paddsw    %%mm6, %%mm0          # Y even + Cblue   00 B6 00 B4 00 B2 00 B0   \n\
+paddsw    %%mm7, %%mm3          # Y odd  + Cblue   00 B7 00 B5 00 B3 00 B1   \n\
+paddsw    %%mm6, %%mm1          # Y even + Cred    00 R6 00 R4 00 R2 00 R0   \n\
+paddsw    %%mm7, %%mm4          # Y odd  + Cred    00 R7 00 R5 00 R3 00 R1   \n\
+paddsw    %%mm6, %%mm2          # Y even + Cgreen  00 G6 00 G4 00 G2 00 G0   \n\
+paddsw    %%mm7, %%mm5          # Y odd  + Cgreen  00 G7 00 G5 00 G3 00 G1   \n\
+                                                                             \n\
+# Limit RGB even to 0..255                                                   \n\
+packuswb  %%mm0, %%mm0          #                  B6 B4 B2 B0 B6 B4 B2 B0   \n\
+packuswb  %%mm1, %%mm1          #                  R6 R4 R2 R0 R6 R4 R2 R0   \n\
+packuswb  %%mm2, %%mm2          #                  G6 G4 G2 G0 G6 G4 G2 G0   \n\
+                                                                             \n\
+# Limit RGB odd to 0..255                                                    \n\
+packuswb  %%mm3, %%mm3          #                  B7 B5 B3 B1 B7 B5 B3 B1   \n\
+packuswb  %%mm4, %%mm4          #                  R7 R5 R3 R1 R7 R5 R3 R1   \n\
+packuswb  %%mm5, %%mm5          #                  G7 G5 G3 G1 G7 G5 G3 G1   \n\
+                                                                             \n\
+# Interleave RGB even and odd                                                \n\
+punpcklbw %%mm3, %%mm0          #                  B7 B6 B5 B4 B3 B2 B1 B0   \n\
+punpcklbw %%mm4, %%mm1          #                  R7 R6 R5 R4 R3 R2 R1 R0   \n\
+punpcklbw %%mm5, %%mm2          #                  G7 G6 G5 G4 G3 G2 G1 G0   \n\
+"
+
+/*
+ * Grayscale case, only use Y
+ */
+
+#define MMX_YUV_GRAY "                                                      \n\
+# convert the luma part                                                     \n\
+psubusb   %5, %%mm6                                                         \n\
+movq      %%mm6, %%mm7                                                      \n\
+pand      %6, %%mm6                                                         \n\
+psrlw     $8, %%mm7                                                         \n\
+psllw     $3, %%mm6                                                         \n\
+psllw     $3, %%mm7                                                         \n\
+pmulhw    %7, %%mm6                                                         \n\
+pmulhw    %7, %%mm7                                                         \n\
+packuswb  %%mm6, %%mm6                                                      \n\
+packuswb  %%mm7, %%mm7                                                      \n\
+punpcklbw %%mm7, %%mm6                                                      \n\
+"
+
+#define MMX_UNPACK_16_GRAY "                                                \n\
+movq      %%mm6, %%mm5                                                      \n\
+pand      %12, %%mm6                                                        \n\
+pand      %13, %%mm5                                                        \n\
+movq      %%mm6, %%mm7                                                      \n\
+psrlw     $3, %%mm7                                                         \n\
+pxor      %%mm3, %%mm3                                                      \n\
+movq      %%mm7, %%mm2                                                      \n\
+movq      %%mm5, %%mm0                                                      \n\
+punpcklbw %%mm3, %%mm5                                                      \n\
+punpcklbw %%mm6, %%mm7                                                      \n\
+psllw     $3, %%mm5                                                         \n\
+por       %%mm5, %%mm7                                                      \n\
+movq      %%mm7, (%3)                                                       \n\
+punpckhbw %%mm3, %%mm0                                                      \n\
+punpckhbw %%mm6, %%mm2                                                      \n\
+psllw     $3, %%mm0                                                         \n\
+movq      8(%0), %%mm6                                                      \n\
+por       %%mm0, %%mm2                                                      \n\
+movq      %%mm2, 8(%3)                                                      \n\
+"
+
+
+/*
+ * convert RGB plane to RGB 15 bits,
+ * mm0 -> B, mm1 -> R, mm2 -> G,
+ * mm4 -> GB, mm5 -> AR pixel 4-7,
+ * mm6 -> GB, mm7 -> AR pixel 0-3
+ */
+
+#define MMX_UNPACK_15 "         # Note, much of this shows bit patterns (of a pair of bytes) \n\
+# mask unneeded bits off                                                    \n\
+pand      %12, %%mm0            # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+psrlw     $3,%%mm0              # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pand      %12, %%mm2            # g7g6g5g4 g3______ g7g6g5g4 g3______       \n\
+pand      %12, %%mm1            # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+psrlw     $1,%%mm1              # __r7r6r5 r4r3____ __r7r6r5 r4r3____       \n\
+pxor      %%mm4, %%mm4          # zero mm4                                  \n\
+movq      %%mm0, %%mm5          # Copy B7-B0                                \n\
+movq      %%mm2, %%mm7          # Copy G7-G0                                \n\
+                                                                            \n\
+# pack the 3 separate RGB bytes into 2 for pixels 0-3                       \n\
+punpcklbw %%mm4, %%mm2          # ________ ________ g7g6g5g4 g3______       \n\
+punpcklbw %%mm1, %%mm0          # __r7r6r5 r4r3____ ______b7 b6b5b4b3       \n\
+psllw     $2,%%mm2              # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%mm2, %%mm0          # __r7r6r5 r4r3g7g6 g5g4g3b7 b6b5b4b3       \n\
+movq      %%mm0, (%3)           # store pixel 0-3                           \n\
+                                                                            \n\
+# pack the 3 separate RGB bytes into 2 for pixels 4-7                       \n\
+punpckhbw %%mm4, %%mm7          # ________ ________ g7g6g5g4 g3______       \n\
+punpckhbw %%mm1, %%mm5          # __r7r6r5 r4r3____ ______b7 b6b5b4b3       \n\
+psllw     $2,%%mm7              # ________ ____g7g6 g5g4g3__ ________       \n\
+por       %%mm7, %%mm5          # __r7r6r5 r4r3g7g6 g5g4g3b7 b6b5b4b3       \n\
+movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
+"
+
+/*
+ * convert RGB plane to RGB 16 bits,
+ * mm0 -> B, mm1 -> R, mm2 -> G,
+ * mm4 -> GB, mm5 -> AR pixel 4-7,
+ * mm6 -> GB, mm7 -> AR pixel 0-3
+ */
+
+#define MMX_UNPACK_16 "         # Note, much of this shows bit patterns (of a pair of bytes) \n\
+# mask unneeded bits off                                                    \n\
+pand      %12, %%mm0            # b7b6b5b4 b3______ b7b6b5b4 b3______       \n\
+pand      %13, %%mm2            # g7g6g5g4 g3g2____ g7g6g5g4 g3g2____       \n\
+pand      %12, %%mm1            # r7r6r5r4 r3______ r7r6r5r4 r3______       \n\
+psrlw     $3,%%mm0              # ______b7 b6b5b4b3 ______b7 b6b5b4b3       \n\
+pxor      %%mm4, %%mm4          # zero mm4                                  \n\
+movq      %%mm0, %%mm5          # Copy B7-B0                                \n\
+movq      %%mm2, %%mm7          # Copy G7-G0                                \n\
+                                                                            \n\
+# pack the 3 separate RGB bytes into 2 for pixels 0-3                       \n\
+punpcklbw %%mm4, %%mm2          # ________ ________ g7g6g5g4 g3g2____       \n\
+punpcklbw %%mm1, %%mm0          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%mm2              # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%mm2, %%mm0          # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movq      %%mm0, (%3)           # store pixel 0-3                           \n\
+                                                                            \n\
+# pack the 3 separate RGB bytes into 2 for pixels 4-7                       \n\
+punpckhbw %%mm4, %%mm7          # ________ ________ g7g6g5g4 g3g2____       \n\
+punpckhbw %%mm1, %%mm5          # r7r6r5r4 r3______ ______b7 b6b5b4b3       \n\
+psllw     $3,%%mm7              # ________ __g7g6g5 g4g3g2__ ________       \n\
+por       %%mm7, %%mm5          # r7r6r5r4 r3g7g6g5 g4g3g2b7 b6b5b4b3       \n\
+movq      %%mm5, 8(%3)          # store pixel 4-7                           \n\
+"
+
+/*
+ * convert RGB plane to RGB packed format,
+ * mm0 -> B, mm1 -> R, mm2 -> G
+ */
+
+#define MMX_UNPACK_32_ARGB "                                                \n\
+pxor      %%mm3, %%mm3  # zero mm3                                          \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 %%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, %%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 MMX_UNPACK_32_RGBA "                                                \n\
+pxor      %%mm3, %%mm3  # zero mm3                                          \n\
+movq      %%mm2, %%mm4  #                 G7 G6 G5 G4 G3 G2 G1 G0           \n\
+punpcklbw %%mm1, %%mm4  #                 R3 G3 R2 G2 R1 G1 R0 G0           \n\
+punpcklbw %%mm0, %%mm3  #                 B3 00 B2 00 B1 00 B0 00           \n\
+movq      %%mm3, %%mm5  #                 R3 00 R2 00 R1 00 R0 00           \n\
+punpcklwd %%mm4, %%mm3  #                 R1 G1 B1 00 R0 G0 B0 00           \n\
+movq      %%mm3, (%3)   # Store RGBA1 RGBA0                                 \n\
+punpckhwd %%mm4, %%mm5  #                 R3 G3 B3 00 R2 G2 B2 00           \n\
+movq      %%mm5, 8(%3)  # Store RGBA3 RGBA2                                 \n\
+pxor      %%mm6, %%mm6  # zero mm6                                          \n\
+punpckhbw %%mm1, %%mm2  #                 R7 G7 R6 G6 R5 G5 R4 G4           \n\
+punpckhbw %%mm0, %%mm6  #                 B7 00 B6 00 B5 00 B4 00           \n\
+movq      %%mm6, %%mm0  #                 B7 00 B6 00 B5 00 B4 00           \n\
+punpcklwd %%mm2, %%mm6  #                 R5 G5 B5 00 R4 G4 B4 00           \n\
+movq      %%mm6, 16(%3) # Store RGBA5 RGBA4                                 \n\
+punpckhwd %%mm2, %%mm0  #                 R7 G7 B7 00 R6 G6 B6 00           \n\
+movq      %%mm0, 24(%3) # Store RGBA7 RGBA6                                 \n\
+"
+
+#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      %%mm6, %%mm6  # zero mm6                                          \n\
+punpckhbw %%mm0, %%mm2  #                 B7 G7 B6 G6 B5 G5 B4 G4           \n\
+punpckhbw %%mm1, %%mm6  #                 R7 00 R6 00 R5 00 R4 00           \n\
+movq      %%mm6, %%mm0  #                 R7 00 R6 00 R5 00 R4 00           \n\
+punpcklwd %%mm2, %%mm6  #                 B5 G5 R5 00 B4 G4 R4 00           \n\
+movq      %%mm6, 16(%3) # Store BGRA5 BGRA4                                 \n\
+punpckhwd %%mm2, %%mm0  #                 B7 G7 R7 00 B6 G6 R6 00           \n\
+movq      %%mm0, 24(%3) # Store BGRA7 BGRA6                                 \n\
+"
+
+#define MMX_UNPACK_32_ABGR "                                                \n\
+pxor      %%mm3, %%mm3  # zero mm3                                          \n\
+movq      %%mm1, %%mm4  #                 R7 R6 R5 R4 R3 R2 R1 R0           \n\
+punpcklbw %%mm2, %%mm4  #                 G3 R3 G2 R2 G1 R1 G0 R0           \n\
+movq      %%mm0, %%mm5  #                 B7 B6 B5 B4 B3 B2 B1 B0           \n\
+punpcklbw %%mm3, %%mm5  #                 00 B3 00 B2 00 B1 00 B0           \n\
+movq      %%mm4, %%mm6  #                 G3 R3 G2 R2 G1 R1 G0 R0           \n\
+punpcklwd %%mm5, %%mm4  #                 00 B1 G1 R1 00 B0 G0 R0           \n\
+movq      %%mm4, (%3)   # Store ABGR1 ABGR0                                 \n\
+punpckhwd %%mm5, %%mm6  #                 00 B3 G3 R3 00 B2 G2 R2           \n\
+movq      %%mm6, 8(%3)  # Store ABGR3 ABGR2                                 \n\
+punpckhbw %%mm2, %%mm1  #                 G7 R7 G6 R6 G5 R5 G4 R4           \n\
+punpckhbw %%mm3, %%mm0  #                 00 B7 00 B6 00 B5 00 B4           \n\
+movq      %%mm1, %%mm2  #                 G7 R7 G6 R6 G5 R5 G4 R4           \n\
+punpcklwd %%mm0, %%mm1  #                 00 B5 G5 R5 00 B4 G4 R4           \n\
+movq      %%mm1, 16(%3) # Store ABGR5 ABGR4                                 \n\
+punpckhwd %%mm0, %%mm2  #                 B7 G7 R7 00 B6 G6 R6 00           \n\
+movq      %%mm2, 24(%3) # Store ABGR7 ABGR6                                 \n\
+"
+
+#elif defined(HAVE_MMX_INTRINSICS)
+
+/* MMX intrinsics */
+
+#include <mmintrin.h>
+
+#define MMX_CALL(MMX_INSTRUCTIONS)  \
+    do {                            \
+        __m64 mm0, mm1, mm2, mm3,   \
+              mm4, mm5, mm6, mm7;   \
+        MMX_INSTRUCTIONS            \
+    } while(0)
+
+#define MMX_END _mm_empty()
+ 
+#define MMX_INIT_16                     \
+    mm0 = _mm_cvtsi32_si64(*(int*)p_u); \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_v); \
+    mm4 = _mm_setzero_si64();           \
+    mm6 = (__m64)*(uint64_t *)p_y;
+
+#define MMX_INIT_32                     \
+    mm0 = _mm_cvtsi32_si64(*(int*)p_u); \
+    *(uint16_t *)p_buffer = 0;          \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_v); \
+    mm4 = _mm_setzero_si64();           \
+    mm6 = (__m64)*(uint64_t *)p_y;
+
+#define MMX_YUV_MUL                                 \
+    mm0 = _mm_unpacklo_pi8(mm0, mm4);               \
+    mm1 = _mm_unpacklo_pi8(mm1, mm4);               \
+    mm0 = _mm_subs_pi16(mm0, (__m64)mmx_80w);       \
+    mm1 = _mm_subs_pi16(mm1, (__m64)mmx_80w);       \
+    mm0 = _mm_slli_pi16(mm0, 3);                    \
+    mm1 = _mm_slli_pi16(mm1, 3);                    \
+    mm2 = mm0;                                      \
+    mm3 = mm1;                                      \
+    mm2 = _mm_mulhi_pi16(mm2, (__m64)mmx_U_green);  \
+    mm3 = _mm_mulhi_pi16(mm3, (__m64)mmx_V_green);  \
+    mm0 = _mm_mulhi_pi16(mm0, (__m64)mmx_U_blue);   \
+    mm1 = _mm_mulhi_pi16(mm1, (__m64)mmx_V_red);    \
+    mm2 = _mm_adds_pi16(mm2, mm3);                  \
+    \
+    mm6 = _mm_subs_pu8(mm6, (__m64)mmx_10w);        \
+    mm7 = mm6;                                      \
+    mm6 = _mm_and_si64(mm6, (__m64)mmx_00ffw);      \
+    mm7 = _mm_srli_pi16(mm7, 8);                    \
+    mm6 = _mm_slli_pi16(mm6, 3);                    \
+    mm7 = _mm_slli_pi16(mm7, 3);                    \
+    mm6 = _mm_mulhi_pi16(mm6, (__m64)mmx_Y_coeff);  \
+    mm7 = _mm_mulhi_pi16(mm7, (__m64)mmx_Y_coeff);
+
+#define MMX_YUV_ADD                     \
+    mm3 = mm0;                          \
+    mm4 = mm1;                          \
+    mm5 = mm2;                          \
+    mm0 = _mm_adds_pi16(mm0, mm6);      \
+    mm3 = _mm_adds_pi16(mm3, mm7);      \
+    mm1 = _mm_adds_pi16(mm1, mm6);      \
+    mm4 = _mm_adds_pi16(mm4, mm7);      \
+    mm2 = _mm_adds_pi16(mm2, mm6);      \
+    mm5 = _mm_adds_pi16(mm5, mm7);      \
+    \
+    mm0 = _mm_packs_pu16(mm0, mm0);     \
+    mm1 = _mm_packs_pu16(mm1, mm1);     \
+    mm2 = _mm_packs_pu16(mm2, mm2);     \
+    \
+    mm3 = _mm_packs_pu16(mm3, mm3);     \
+    mm4 = _mm_packs_pu16(mm4, mm4);     \
+    mm5 = _mm_packs_pu16(mm5, mm5);     \
+    \
+    mm0 = _mm_unpacklo_pi8(mm0, mm3);   \
+    mm1 = _mm_unpacklo_pi8(mm1, mm4);   \
+    mm2 = _mm_unpacklo_pi8(mm2, mm5);
+
+#define MMX_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 = _mm_setzero_si64();                       \
+    mm5 = mm0;                                      \
+    mm7 = mm2;                                      \
+    \
+    mm2 = _mm_unpacklo_pi8(mm2, mm4);               \
+    mm0 = _mm_unpacklo_pi8(mm0, mm1);               \
+    mm2 = _mm_slli_pi16(mm2, 2);                    \
+    mm0 = _mm_or_si64(mm0, mm2);                    \
+    *(uint64_t *)p_buffer = (uint64_t)mm0;          \
+    \
+    mm7 = _mm_unpackhi_pi8(mm7, mm4);               \
+    mm5 = _mm_unpackhi_pi8(mm5, mm1);               \
+    mm7 = _mm_slli_pi16(mm7, 2);                    \
+    mm5 = _mm_or_si64(mm5, mm7);                    \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
+
+#define MMX_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 = _mm_setzero_si64();                       \
+    mm5 = mm0;                                      \
+    mm7 = mm2;                                      \
+    \
+    mm2 = _mm_unpacklo_pi8(mm2, mm4);               \
+    mm0 = _mm_unpacklo_pi8(mm0, mm1);               \
+    mm2 = _mm_slli_pi16(mm2, 3);                    \
+    mm0 = _mm_or_si64(mm0, mm2);                    \
+    *(uint64_t *)p_buffer = (uint64_t)mm0;          \
+    \
+    mm7 = _mm_unpackhi_pi8(mm7, mm4);               \
+    mm5 = _mm_unpackhi_pi8(mm5, mm1);               \
+    mm7 = _mm_slli_pi16(mm7, 3);                    \
+    mm5 = _mm_or_si64(mm5, mm7);                    \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm5;
+
+#define MMX_UNPACK_32_ARGB                      \
+    mm3 = _mm_setzero_si64();                   \
+    mm4 = mm0;                                  \
+    mm4 = _mm_unpacklo_pi8(mm4, mm2);           \
+    mm5 = mm1;                                  \
+    mm5 = _mm_unpacklo_pi8(mm5, mm3);           \
+    mm6 = mm4;                                  \
+    mm4 = _mm_unpacklo_pi16(mm4, mm5);          \
+    *(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 MMX_UNPACK_32_RGBA                      \
+    mm3 = _mm_setzero_si64();                   \
+    mm4 = mm2;                                  \
+    mm4 = _mm_unpacklo_pi8(mm4, mm1);           \
+    mm3 = _mm_unpacklo_pi8(mm3, mm0);           \
+    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;\
+    mm6 = _mm_setzero_si64();                   \
+    mm2 = _mm_unpackhi_pi8(mm2, mm1);           \
+    mm6 = _mm_unpackhi_pi8(mm6, mm0);           \
+    mm0 = mm6;                                  \
+    mm6 = _mm_unpacklo_pi16(mm6, mm2);          \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm6;\
+    mm0 = _mm_unpackhi_pi16(mm0, mm2);          \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm0;
+
+#define MMX_UNPACK_32_BGRA                      \
+    mm3 = _mm_setzero_si64();                   \
+    mm4 = mm2;                                  \
+    mm4 = _mm_unpacklo_pi8(mm4, mm0);           \
+    mm3 = _mm_unpacklo_pi8(mm3, mm1);           \
+    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;\
+    mm6 = _mm_setzero_si64();                   \
+    mm2 = _mm_unpackhi_pi8(mm2, mm0);           \
+    mm6 = _mm_unpackhi_pi8(mm6, mm1);           \
+    mm0 = mm6;                                  \
+    mm6 = _mm_unpacklo_pi16(mm6, mm2);          \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm6;\
+    mm0 = _mm_unpackhi_pi16(mm0, mm2);          \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm0;
+
+#define MMX_UNPACK_32_ABGR                      \
+    mm3 = _mm_setzero_si64();                   \
+    mm4 = mm1;                                  \
+    mm4 = _mm_unpacklo_pi8(mm4, mm2);           \
+    mm5 = mm0;                                  \
+    mm5 = _mm_unpacklo_pi8(mm5, mm3);           \
+    mm6 = mm4;                                  \
+    mm4 = _mm_unpacklo_pi16(mm4, mm5);          \
+    *(uint64_t *)p_buffer = (uint64_t)mm4;      \
+    mm6 = _mm_unpackhi_pi16(mm6, mm5);          \
+    *(uint64_t *)(p_buffer + 2) = (uint64_t)mm6;\
+    mm1 = _mm_unpackhi_pi8(mm1, mm2);           \
+    mm0 = _mm_unpackhi_pi8(mm0, mm3);           \
+    mm2 = mm1;                                  \
+    mm1 = _mm_unpacklo_pi16(mm1, mm0);          \
+    *(uint64_t *)(p_buffer + 4) = (uint64_t)mm1;\
+    mm2 = _mm_unpackhi_pi16(mm2, mm0);          \
+    *(uint64_t *)(p_buffer + 6) = (uint64_t)mm2;
+
+#endif
+
diff --git a/modules/video_chroma/i420_yuy2.c b/modules/video_chroma/i420_yuy2.c
index bdfda6b797..1c67369c97 100644
--- a/modules/video_chroma/i420_yuy2.c
+++ b/modules/video_chroma/i420_yuy2.c
@@ -46,6 +46,9 @@
 #if defined (MODULE_NAME_IS_i420_yuy2)
 #    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422,IUYV,Y211"
 #    define VLC_TARGET
+#elif defined (MODULE_NAME_IS_i420_yuy2_mmx)
+#    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422,IUYV"
+#    define VLC_TARGET VLC_MMX
 #elif defined (MODULE_NAME_IS_i420_yuy2_sse2)
 #    define DEST_FOURCC "YUY2,YUNV,YVYU,UYVY,UYNV,Y422,IUYV"
 #    define VLC_TARGET VLC_SSE
@@ -85,6 +88,10 @@ vlc_module_begin ()
     set_description( N_("Conversions from " SRC_FOURCC " to " DEST_FOURCC) )
     set_capability( "video converter", 80 )
 # define vlc_CPU_capable() (true)
+#elif defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    set_description( N_("MMX conversions from " SRC_FOURCC " to " DEST_FOURCC) )
+    set_capability( "video converter", 160 )
+# define vlc_CPU_capable() vlc_CPU_MMX()
 #elif defined (MODULE_NAME_IS_i420_yuy2_sse2)
     set_description( N_("SSE2 conversions from " SRC_FOURCC " to " DEST_FOURCC) )
     set_capability( "video converter", 250 )
@@ -305,10 +312,14 @@ static void I420_YUY2( filter_t *p_filter, picture_t *p_source,
 
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8; i_x-- ; )
         {
+#if !defined (MODULE_NAME_IS_i420_yuy2_mmx)
             C_YUV420_YUYV( );
             C_YUV420_YUYV( );
             C_YUV420_YUYV( );
             C_YUV420_YUYV( );
+#else
+            MMX_CALL( MMX_YUV420_YUYV );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x-- ; )
         {
@@ -321,6 +332,11 @@ static void I420_YUY2( filter_t *p_filter, picture_t *p_source,
         p_line2 += i_dest_margin;
     }
 
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    /* re-enable FPU registers */
+    MMX_END;
+#endif
+
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
     }
 #endif
@@ -567,10 +583,14 @@ static void I420_YVYU( filter_t *p_filter, picture_t *p_source,
 
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8 ; i_x-- ; )
         {
+#if !defined (MODULE_NAME_IS_i420_yuy2_mmx)
             C_YUV420_YVYU( );
             C_YUV420_YVYU( );
             C_YUV420_YVYU( );
             C_YUV420_YVYU( );
+#else
+            MMX_CALL( MMX_YUV420_YVYU );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x-- ; )
         {
@@ -583,6 +603,11 @@ static void I420_YVYU( filter_t *p_filter, picture_t *p_source,
         p_line2 += i_dest_margin;
     }
 
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    /* re-enable FPU registers */
+    MMX_END;
+#endif
+
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
     }
 #endif
@@ -829,10 +854,14 @@ static void I420_UYVY( filter_t *p_filter, picture_t *p_source,
 
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8 ; i_x-- ; )
         {
+#if !defined (MODULE_NAME_IS_i420_yuy2_mmx)
             C_YUV420_UYVY( );
             C_YUV420_UYVY( );
             C_YUV420_UYVY( );
             C_YUV420_UYVY( );
+#else
+            MMX_CALL( MMX_YUV420_UYVY );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x--; )
         {
@@ -845,6 +874,11 @@ static void I420_UYVY( filter_t *p_filter, picture_t *p_source,
         p_line2 += i_dest_margin;
     }
 
+#if defined (MODULE_NAME_IS_i420_yuy2_mmx)
+    /* re-enable FPU registers */
+    MMX_END;
+#endif
+
 #if defined (MODULE_NAME_IS_i420_yuy2_altivec)
     }
 #endif
diff --git a/modules/video_chroma/i420_yuy2.h b/modules/video_chroma/i420_yuy2.h
index 7763d219d4..49eba7828f 100644
--- a/modules/video_chroma/i420_yuy2.h
+++ b/modules/video_chroma/i420_yuy2.h
@@ -22,7 +22,153 @@
  * Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  *****************************************************************************/
 
-#if defined( MODULE_NAME_IS_i420_yuy2_sse2 )
+#ifdef MODULE_NAME_IS_i420_yuy2_mmx
+
+#if defined(CAN_COMPILE_MMX)
+
+/* MMX assembly */
+ 
+#define MMX_CALL(MMX_INSTRUCTIONS)          \
+    do {                                    \
+    __asm__ __volatile__(                   \
+        ".p2align 3 \n\t                    \
+movd       (%0), %%mm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0     \n\
+movd       (%1), %%mm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0     \n\
+movq       (%2), %%mm0  # Load 8 Y1           y7 y6 y5 y4 y3 y2 y1 y0     \n\
+movq       (%3), %%mm3  # Load 8 Y2           Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
+" \
+        :                                   \
+        : "r" (p_u), "r" (p_v),             \
+          "r" (p_y1), "r" (p_y2)            \
+        : "mm0", "mm1", "mm2", "mm3");      \
+    __asm__ __volatile__(                   \
+        ".p2align 3 \n\t"                   \
+        MMX_INSTRUCTIONS                    \
+        :                                   \
+        : "r" (p_line1), "r" (p_line2)      \
+        : "mm0", "mm1", "mm2", "mm3");      \
+        p_line1 += 16; p_line2 += 16;       \
+        p_y1 += 8; p_y2 += 8;               \
+        p_u += 4; p_v += 4;                 \
+    } while(0)
+
+#define MMX_END __asm__ __volatile__ ( "emms" )
+
+#define MMX_YUV420_YUYV "                                                 \n\
+punpcklbw %%mm2, %%mm1  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+movq      %%mm0, %%mm2  #                     y7 y6 y5 y4 y3 y2 y1 y0     \n\
+punpcklbw %%mm1, %%mm2  #                     v1 y3 u1 y2 v0 y1 u0 y0     \n\
+movq      %%mm2, (%0)   # Store low YUYV                                  \n\
+punpckhbw %%mm1, %%mm0  #                     v3 y7 u3 y6 v2 y5 u2 y4     \n\
+movq      %%mm0, 8(%0)  # Store high YUYV                                 \n\
+movq      %%mm3, %%mm4  #                     Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
+punpcklbw %%mm1, %%mm4  #                     v1 Y3 u1 Y2 v0 Y1 u0 Y0     \n\
+movq      %%mm4, (%1)   # Store low YUYV                                  \n\
+punpckhbw %%mm1, %%mm3  #                     v3 Y7 u3 Y6 v2 Y5 u2 Y4     \n\
+movq      %%mm3, 8(%1)  # Store high YUYV                                 \n\
+"
+
+#define MMX_YUV420_YVYU "                                                 \n\
+punpcklbw %%mm1, %%mm2  #                     u3 v3 u2 v2 u1 v1 u0 v0     \n\
+movq      %%mm0, %%mm1  #                     y7 y6 y5 y4 y3 y2 y1 y0     \n\
+punpcklbw %%mm2, %%mm1  #                     u1 y3 v1 y2 u0 y1 v0 y0     \n\
+movq      %%mm1, (%0)   # Store low YUYV                                  \n\
+punpckhbw %%mm2, %%mm0  #                     u3 y7 v3 y6 u2 y5 v2 y4     \n\
+movq      %%mm0, 8(%0)  # Store high YUYV                                 \n\
+movq      %%mm3, %%mm4  #                     Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y0     \n\
+punpcklbw %%mm2, %%mm4  #                     u1 Y3 v1 Y2 u0 Y1 v0 Y0     \n\
+movq      %%mm4, (%1)   # Store low YUYV                                  \n\
+punpckhbw %%mm2, %%mm3  #                     u3 Y7 v3 Y6 u2 Y5 v2 Y4     \n\
+movq      %%mm3, 8(%1)  # Store high YUYV                                 \n\
+"
+
+#define MMX_YUV420_UYVY "                                                 \n\
+punpcklbw %%mm2, %%mm1  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+movq      %%mm1, %%mm2  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+punpcklbw %%mm0, %%mm2  #                     y3 v1 y2 u1 y1 v0 y0 u0     \n\
+movq      %%mm2, (%0)   # Store low UYVY                                  \n\
+movq      %%mm1, %%mm2  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+punpckhbw %%mm0, %%mm2  #                     y7 v3 y6 u3 y5 v2 y4 u2     \n\
+movq      %%mm2, 8(%0)  # Store high UYVY                                 \n\
+movq      %%mm1, %%mm4  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+punpcklbw %%mm3, %%mm4  #                     Y3 v1 Y2 u1 Y1 v0 Y0 u0     \n\
+movq      %%mm4, (%1)   # Store low UYVY                                  \n\
+punpckhbw %%mm3, %%mm1  #                     Y7 v3 Y6 u3 Y5 v2 Y4 u2     \n\
+movq      %%mm1, 8(%1)  # Store high UYVY                                 \n\
+"
+
+#elif defined(HAVE_MMX_INTRINSICS)
+
+/* MMX intrinsics */
+
+#include <mmintrin.h>
+
+#define MMX_CALL(MMX_INSTRUCTIONS)          \
+    do {                                    \
+        __m64 mm0, mm1, mm2, mm3, mm4;      \
+        MMX_INSTRUCTIONS                    \
+        p_line1 += 16; p_line2 += 16;       \
+        p_y1 += 8; p_y2 += 8;               \
+        p_u += 4; p_v += 4;                 \
+    } while(0)
+
+#define MMX_END _mm_empty()
+ 
+#define MMX_YUV420_YUYV                     \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm0 = (__m64)*(uint64_t*)p_y1;          \
+    mm3 = (__m64)*(uint64_t*)p_y2;          \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line1 = (uint64_t)mm2;    \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line1+8) = (uint64_t)mm0;\
+    mm4 = mm3;                              \
+    mm4 = _mm_unpacklo_pi8(mm4, mm1);       \
+    *(uint64_t*)p_line2 = (uint64_t)mm4;    \
+    mm3 = _mm_unpackhi_pi8(mm3, mm1);       \
+    *(uint64_t*)(p_line2+8) = (uint64_t)mm3;
+
+#define MMX_YUV420_YVYU                     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm0 = (__m64)*(uint64_t*)p_y1;          \
+    mm3 = (__m64)*(uint64_t*)p_y2;          \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line1 = (uint64_t)mm2;    \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line1+8) = (uint64_t)mm0;\
+    mm4 = mm3;                              \
+    mm4 = _mm_unpacklo_pi8(mm4, mm1);       \
+    *(uint64_t*)p_line2 = (uint64_t)mm4;    \
+    mm3 = _mm_unpackhi_pi8(mm3, mm1);       \
+    *(uint64_t*)(p_line2+8) = (uint64_t)mm3;
+
+#define MMX_YUV420_UYVY                     \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm0 = (__m64)*(uint64_t*)p_y1;          \
+    mm3 = (__m64)*(uint64_t*)p_y2;          \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm1;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm0);       \
+    *(uint64_t*)p_line1 = (uint64_t)mm2;    \
+    mm2 = mm1;                              \
+    mm2 = _mm_unpackhi_pi8(mm2, mm0);       \
+    *(uint64_t*)(p_line1+8) = (uint64_t)mm2;\
+    mm4 = mm1;                              \
+    mm4 = _mm_unpacklo_pi8(mm4, mm3);       \
+    *(uint64_t*)p_line2 = (uint64_t)mm4;    \
+    mm1 = _mm_unpackhi_pi8(mm1, mm3);       \
+    *(uint64_t*)(p_line2+8) = (uint64_t)mm1;
+
+#endif
+
+#elif defined( MODULE_NAME_IS_i420_yuy2_sse2 )
 
 #if defined(CAN_COMPILE_SSE2)
 
diff --git a/modules/video_chroma/i422_yuy2.c b/modules/video_chroma/i422_yuy2.c
index d1b7db9dbf..7ee0045f5b 100644
--- a/modules/video_chroma/i422_yuy2.c
+++ b/modules/video_chroma/i422_yuy2.c
@@ -71,6 +71,11 @@ vlc_module_begin ()
     set_capability( "video converter", 80 )
 # define vlc_CPU_capable() (true)
 # define VLC_TARGET
+#elif defined (MODULE_NAME_IS_i422_yuy2_mmx)
+    set_description( N_("MMX conversions from " SRC_FOURCC " to " DEST_FOURCC) )
+    set_capability( "video converter", 100 )
+# define vlc_CPU_capable() vlc_CPU_MMX()
+# define VLC_TARGET VLC_MMX
 #elif defined (MODULE_NAME_IS_i422_yuy2_sse2)
     set_description( N_("SSE2 conversions from " SRC_FOURCC " to " DEST_FOURCC) )
     set_capability( "video converter", 120 )
@@ -279,10 +284,14 @@ static void I422_YUY2( filter_t *p_filter, picture_t *p_source,
     {
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8 ; i_x-- ; )
         {
+#if defined (MODULE_NAME_IS_i422_yuy2)
             C_YUV422_YUYV( p_line, p_y, p_u, p_v );
             C_YUV422_YUYV( p_line, p_y, p_u, p_v );
             C_YUV422_YUYV( p_line, p_y, p_u, p_v );
             C_YUV422_YUYV( p_line, p_y, p_u, p_v );
+#elif defined (MODULE_NAME_IS_i422_yuy2_mmx)
+            MMX_CALL( MMX_YUV422_YUYV );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x-- ; )
         {
@@ -293,6 +302,9 @@ static void I422_YUY2( filter_t *p_filter, picture_t *p_source,
         p_v += i_source_margin_c;
         p_line += i_dest_margin;
     }
+#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
+    MMX_END;
+#endif
 
 #endif
 }
@@ -421,10 +433,14 @@ static void I422_YVYU( filter_t *p_filter, picture_t *p_source,
     {
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8 ; i_x-- ; )
         {
+#if defined (MODULE_NAME_IS_i422_yuy2)
             C_YUV422_YVYU( p_line, p_y, p_u, p_v );
             C_YUV422_YVYU( p_line, p_y, p_u, p_v );
             C_YUV422_YVYU( p_line, p_y, p_u, p_v );
             C_YUV422_YVYU( p_line, p_y, p_u, p_v );
+#elif defined (MODULE_NAME_IS_i422_yuy2_mmx)
+            MMX_CALL( MMX_YUV422_YVYU );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x-- ; )
         {
@@ -435,6 +451,9 @@ static void I422_YVYU( filter_t *p_filter, picture_t *p_source,
         p_v += i_source_margin_c;
         p_line += i_dest_margin;
     }
+#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
+    MMX_END;
+#endif
 
 #endif
 }
@@ -563,10 +582,14 @@ static void I422_UYVY( filter_t *p_filter, picture_t *p_source,
     {
         for( i_x = (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) / 8 ; i_x-- ; )
         {
+#if defined (MODULE_NAME_IS_i422_yuy2)
             C_YUV422_UYVY( p_line, p_y, p_u, p_v );
             C_YUV422_UYVY( p_line, p_y, p_u, p_v );
             C_YUV422_UYVY( p_line, p_y, p_u, p_v );
             C_YUV422_UYVY( p_line, p_y, p_u, p_v );
+#elif defined (MODULE_NAME_IS_i422_yuy2_mmx)
+            MMX_CALL( MMX_YUV422_UYVY );
+#endif
         }
         for( i_x = ( (p_filter->fmt_in.video.i_x_offset + p_filter->fmt_in.video.i_visible_width) % 8 ) / 2; i_x-- ; )
         {
@@ -577,6 +600,9 @@ static void I422_UYVY( filter_t *p_filter, picture_t *p_source,
         p_v += i_source_margin_c;
         p_line += i_dest_margin;
     }
+#if defined (MODULE_NAME_IS_i422_yuy2_mmx)
+    MMX_END;
+#endif
 
 #endif
 }
diff --git a/modules/video_chroma/i422_yuy2.h b/modules/video_chroma/i422_yuy2.h
index 9d05d95ad5..35b3d6f6db 100644
--- a/modules/video_chroma/i422_yuy2.h
+++ b/modules/video_chroma/i422_yuy2.h
@@ -22,7 +22,115 @@
  * Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  *****************************************************************************/
 
-#if defined( MODULE_NAME_IS_i422_yuy2_sse2 )
+#ifdef MODULE_NAME_IS_i422_yuy2_mmx
+
+#if defined(CAN_COMPILE_MMX)
+
+/* MMX assembly */
+ 
+#define MMX_CALL(MMX_INSTRUCTIONS)          \
+    do {                                    \
+    __asm__ __volatile__(                   \
+        ".p2align 3 \n\t"                   \
+        MMX_INSTRUCTIONS                    \
+        :                                   \
+        : "r" (p_line), "r" (p_y),          \
+          "r" (p_u), "r" (p_v)              \
+        : "mm0", "mm1", "mm2" );            \
+        p_line += 16; p_y += 8;             \
+        p_u += 4; p_v += 4;                 \
+    } while(0)
+
+#define MMX_END __asm__ __volatile__ ( "emms" )
+
+#define MMX_YUV422_YUYV "                                                 \n\
+movq       (%1), %%mm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0     \n\
+movd       (%2), %%mm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0     \n\
+movd       (%3), %%mm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0     \n\
+punpcklbw %%mm2, %%mm1  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+movq      %%mm0, %%mm2  #                     y7 y6 y5 y4 y3 y2 y1 y0     \n\
+punpcklbw %%mm1, %%mm2  #                     v1 y3 u1 y2 v0 y1 u0 y0     \n\
+movq      %%mm2, (%0)   # Store low YUYV                                  \n\
+punpckhbw %%mm1, %%mm0  #                     v3 y7 u3 y6 v2 y5 u2 y4     \n\
+movq      %%mm0, 8(%0)  # Store high YUYV                                 \n\
+"
+
+#define MMX_YUV422_YVYU "                                                 \n\
+movq       (%1), %%mm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0     \n\
+movd       (%2), %%mm2  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0     \n\
+movd       (%3), %%mm1  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0     \n\
+punpcklbw %%mm2, %%mm1  #                     u3 v3 u2 v2 u1 v1 u0 v0     \n\
+movq      %%mm0, %%mm2  #                     y7 y6 y5 y4 y3 y2 y1 y0     \n\
+punpcklbw %%mm1, %%mm2  #                     u1 y3 v1 y2 u0 y1 v0 y0     \n\
+movq      %%mm2, (%0)   # Store low YUYV                                  \n\
+punpckhbw %%mm1, %%mm0  #                     u3 y7 v3 y6 u2 y5 v2 y4     \n\
+movq      %%mm0, 8(%0)  # Store high YUYV                                 \n\
+"
+
+#define MMX_YUV422_UYVY "                                                 \n\
+movq       (%1), %%mm0  # Load 8 Y            y7 y6 y5 y4 y3 y2 y1 y0     \n\
+movd       (%2), %%mm1  # Load 4 Cb           00 00 00 00 u3 u2 u1 u0     \n\
+movd       (%3), %%mm2  # Load 4 Cr           00 00 00 00 v3 v2 v1 v0     \n\
+punpcklbw %%mm2, %%mm1  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+movq      %%mm1, %%mm2  #                     v3 u3 v2 u2 v1 u1 v0 u0     \n\
+punpcklbw %%mm0, %%mm2  #                     y3 v1 y2 u1 y1 v0 y0 u0     \n\
+movq      %%mm2, (%0)   # Store low UYVY                                  \n\
+punpckhbw %%mm0, %%mm1  #                     y7 v3 y6 u3 y5 v2 y4 u2     \n\
+movq      %%mm1, 8(%0)  # Store high UYVY                                 \n\
+"
+
+#elif defined(HAVE_MMX_INTRINSICS)
+
+/* MMX intrinsics */
+
+#include <mmintrin.h>
+
+#define MMX_CALL(MMX_INSTRUCTIONS)  \
+    do {                            \
+        __m64 mm0, mm1, mm2;        \
+        MMX_INSTRUCTIONS            \
+        p_line += 16; p_y += 8;     \
+        p_u += 4; p_v += 4;         \
+    } while(0)
+
+#define MMX_END _mm_empty()
+
+#define MMX_YUV422_YUYV                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm0;
+ 
+#define MMX_YUV422_YVYU                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm0;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm1);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm0 = _mm_unpackhi_pi8(mm0, mm1);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm0;
+
+#define MMX_YUV422_UYVY                     \
+    mm0 = (__m64)*(uint64_t*)p_y;           \
+    mm1 = _mm_cvtsi32_si64(*(int*)p_u);     \
+    mm2 = _mm_cvtsi32_si64(*(int*)p_v);     \
+    mm1 = _mm_unpacklo_pi8(mm1, mm2);       \
+    mm2 = mm1;                              \
+    mm2 = _mm_unpacklo_pi8(mm2, mm0);       \
+    *(uint64_t*)p_line = (uint64_t)mm2;     \
+    mm1 = _mm_unpackhi_pi8(mm1, mm0);       \
+    *(uint64_t*)(p_line+8) = (uint64_t)mm1;
+
+#endif
+ 
+#elif defined( MODULE_NAME_IS_i422_yuy2_sse2 )
 
 #if defined(CAN_COMPILE_SSE2)
 
@@ -381,3 +489,4 @@ vmovdqu    [[l]+32], ymm1    ; Store high UYVY                                 \
     *(p_line)++ = *(p_y); p_y += 2;                                         \
     *(p_line)++ = *(p_v) - 0x80; p_v += 2;                                  \
 
+



More information about the vlc-commits mailing list