gst/deinterlace2/: Add a deinterlacer plugin based on the tvtime/DScaler deinterlacer, which was relicensed to LGPL f...

Original commit message from CVS:
Based on a patch by: Martin Eikermann <meiker at upb dot de>
* gst/deinterlace2/Makefile.am:
* gst/deinterlace2/gstdeinterlace2.c:
(gst_deinterlace2_method_get_type),
(gst_deinterlace2_fields_get_type),
(gst_deinterlace2_field_layout_get_type),
(gst_deinterlace2_base_init), (gst_deinterlace2_class_init),
(gst_deinterlace2_init), (gst_deinterlace2_set_method),
(gst_deinterlace2_set_property), (gst_deinterlace2_get_property),
(gst_deinterlace2_finalize), (gst_deinterlace2_pop_history),
(gst_deinterlace2_head_history), (gst_deinterlace2_push_history),
(gst_deinterlace2_deinterlace_scanlines), (gst_deinterlace2_chain),
(gst_deinterlace2_setcaps), (gst_deinterlace2_sink_event),
(gst_deinterlace2_change_state), (gst_deinterlace2_src_event),
(gst_deinterlace2_src_query), (gst_deinterlace2_src_query_types),
(plugin_init):
* gst/deinterlace2/gstdeinterlace2.h:
* gst/deinterlace2/tvtime/greedy.c: (copy_scanline),
(deinterlace_greedy_packed422_scanline_mmxext),
(dscaler_greedyl_get_method):
* gst/deinterlace2/tvtime/greedyh.asm:
* gst/deinterlace2/tvtime/greedyh.c:
(deinterlace_frame_di_greedyh), (dscaler_greedyh_get_method),
(greedyh_init), (greedyh_filter_mmx), (greedyh_filter_3dnow),
(greedyh_filter_sse):
* gst/deinterlace2/tvtime/greedyh.h:
* gst/deinterlace2/tvtime/greedyhmacros.h:
* gst/deinterlace2/tvtime/mmx.h:
* gst/deinterlace2/tvtime/plugins.h:
* gst/deinterlace2/tvtime/speedtools.h:
* gst/deinterlace2/tvtime/speedy.c: (multiply_alpha), (clip255),
(comb_factor_packed422_scanline_mmx),
(diff_factor_packed422_scanline_c),
(diff_factor_packed422_scanline_mmx),
(diff_packed422_block8x8_mmx), (diff_packed422_block8x8_c),
(packed444_to_packed422_scanline_c),
(packed422_to_packed444_scanline_c),
(packed422_to_packed444_rec601_scanline_c),
(vfilter_chroma_121_packed422_scanline_mmx),
(vfilter_chroma_121_packed422_scanline_c),
(vfilter_chroma_332_packed422_scanline_mmx),
(vfilter_chroma_332_packed422_scanline_c),
(kill_chroma_packed422_inplace_scanline_mmx),
(kill_chroma_packed422_inplace_scanline_c),
(invert_colour_packed422_inplace_scanline_mmx),
(invert_colour_packed422_inplace_scanline_c),
(mirror_packed422_inplace_scanline_c),
(interpolate_packed422_scanline_c),
(convert_uyvy_to_yuyv_scanline_mmx),
(convert_uyvy_to_yuyv_scanline_c),
(interpolate_packed422_scanline_mmx),
(interpolate_packed422_scanline_mmxext),
(blit_colour_packed422_scanline_c),
(blit_colour_packed422_scanline_mmx),
(blit_colour_packed422_scanline_mmxext),
(blit_colour_packed4444_scanline_c),
(blit_colour_packed4444_scanline_mmx),
(blit_colour_packed4444_scanline_mmxext), (small_memcpy),
(speedy_memcpy_c), (speedy_memcpy_mmx), (speedy_memcpy_mmxext),
(blit_packed422_scanline_c), (blit_packed422_scanline_mmx),
(blit_packed422_scanline_mmxext),
(composite_colour4444_alpha_to_packed422_scanline_c),
(composite_colour4444_alpha_to_packed422_scanline_mmxext),
(composite_packed4444_alpha_to_packed422_scanline_c),
(composite_packed4444_alpha_to_packed422_scanline_mmxext),
(composite_packed4444_to_packed422_scanline_c),
(composite_packed4444_to_packed422_scanline_mmxext),
(composite_alphamask_to_packed4444_scanline_c),
(composite_alphamask_to_packed4444_scanline_mmxext),
(composite_alphamask_alpha_to_packed4444_scanline_c),
(premultiply_packed4444_scanline_c),
(premultiply_packed4444_scanline_mmxext),
(blend_packed422_scanline_c), (blend_packed422_scanline_mmxext),
(quarter_blit_vertical_packed422_scanline_mmxext),
(quarter_blit_vertical_packed422_scanline_c),
(subpix_blit_vertical_packed422_scanline_c),
(a8_subpix_blit_scanline_c), (myround), (init_RGB_to_YCbCr_tables),
(init_YCbCr_to_RGB_tables), (rgb24_to_packed444_rec601_scanline_c),
(rgba32_to_packed4444_rec601_scanline_c),
(packed444_to_rgb24_rec601_scanline_c),
(packed444_to_nonpremultiplied_packed4444_scanline_c),
(aspect_adjust_packed4444_scanline_c), (setup_speedy_calls),
(speedy_get_accel):
* gst/deinterlace2/tvtime/speedy.h:
* gst/deinterlace2/tvtime/sse.h:
* gst/deinterlace2/tvtime/tomsmocomp.c: (Fieldcopy),
(deinterlace_frame_di_tomsmocomp), (dscaler_tomsmocomp_get_method),
(tomsmocomp_init), (tomsmocomp_filter_mmx),
(tomsmocomp_filter_3dnow), (tomsmocomp_filter_sse):
* gst/deinterlace2/tvtime/tomsmocomp.h:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoop0A.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopBottom.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopEdgeA.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopEdgeA8.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopOddA.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopOddA2.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopOddA6.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopOddAH.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopOddAH2.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopTop.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopVA.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopVAH.inc:
* gst/deinterlace2/tvtime/tomsmocomp/StrangeBob.inc:
* gst/deinterlace2/tvtime/tomsmocomp/TomsMoCompAll.inc:
* gst/deinterlace2/tvtime/tomsmocomp/TomsMoCompAll2.inc:
* gst/deinterlace2/tvtime/tomsmocomp/WierdBob.inc:
* gst/deinterlace2/tvtime/vfir.c: (deinterlace_line),
(deinterlace_scanline_vfir), (copy_scanline),
(dscaler_vfir_get_method):
* gst/deinterlace2/tvtime/x86-64_macros.inc:
Add a deinterlacer plugin based on the tvtime/DScaler deinterlacer,
which was relicensed to LGPL for GStreamer and in theory provides
better and faster results than the simple deinterlace element.
Fixes bug #163578.
Ported to GStreamer 0.10 but still not enabled or included in the
build system by default because of bad artefacts caused by a bug
somewhere and as it can be only build on x86/amd64 ATM and requires
special CFLAGS. Will be fixed soon.

15 files changed, 6205 insertions, 0 deletions

diff --git a/gst/deinterlace2/tvtime/greedy.c b/gst/deinterlace2/tvtime/greedy.c
new file mode 100644
index 00000000..578eb711
--- /dev/null
+++ b/gst/deinterlace2/tvtime/greedy.c
@@ -0,0 +1,207 @@
+/*
+ *
+ * GStreamer
+ * Copyright (c) 2000 Tom Barry  All rights reserved.
+ * mmx.h port copyright (c) 2002 Billy Biggs <vektor@dumbterm.net>.
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Tom Barry
+ * and Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#include <stdio.h>
+#if defined (__SVR4) && defined (__sun)
+# include <sys/int_types.h>
+#else
+# include <stdint.h>
+#endif
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "mmx.h"
+#include "sse.h"
+#include "gstdeinterlace2.h"
+#include "speedtools.h"
+#include "speedy.h"
+
+// This is a simple lightweight DeInterlace method that uses little CPU time
+// but gives very good results for low or intermedite motion.
+// It defers frames by one field, but that does not seem to produce noticeable
+// lip sync problems.
+//
+// The method used is to take either the older or newer weave pixel depending
+// upon which give the smaller comb factor, and then clip to avoid large damage
+// when wrong.
+//
+// I'd intended this to be part of a larger more elaborate method added to 
+// Blended Clip but this give too good results for the CPU to ignore here.
+
+static void
+copy_scanline (GstDeinterlace2 * object,
+    deinterlace_scanline_data_t * data, uint8_t * output)
+{
+  blit_packed422_scanline (output, data->m1, object->frame_width);
+}
+
+static int GreedyMaxComb = 15;
+
+static void
+deinterlace_greedy_packed422_scanline_mmxext (GstDeinterlace2 * object,
+    deinterlace_scanline_data_t * data, uint8_t * output)
+{
+#ifdef HAVE_CPU_I386
+  mmx_t MaxComb;
+
+  uint8_t *m0 = data->m0;
+
+  uint8_t *t1 = data->t1;
+
+  uint8_t *b1 = data->b1;
+
+  uint8_t *m2 = data->m2;
+
+  int width = object->frame_width;
+
+  // How badly do we let it weave? 0-255
+  MaxComb.ub[0] = GreedyMaxComb;
+  MaxComb.ub[1] = GreedyMaxComb;
+  MaxComb.ub[2] = GreedyMaxComb;
+  MaxComb.ub[3] = GreedyMaxComb;
+  MaxComb.ub[4] = GreedyMaxComb;
+  MaxComb.ub[5] = GreedyMaxComb;
+  MaxComb.ub[6] = GreedyMaxComb;
+  MaxComb.ub[7] = GreedyMaxComb;
+
+  // L2 == m0
+  // L1 == t1
+  // L3 == b1
+  // LP2 == m2
+
+  width /= 4;
+  while (width--) {
+    movq_m2r (*t1, mm1);        // L1
+    movq_m2r (*m0, mm2);        // L2
+    movq_m2r (*b1, mm3);        // L3
+    movq_m2r (*m2, mm0);        // LP2
+
+    // average L1 and L3 leave result in mm4
+    movq_r2r (mm1, mm4);        // L1
+    pavgb_r2r (mm3, mm4);       // (L1 + L3)/2
+
+
+    // get abs value of possible L2 comb
+    movq_r2r (mm2, mm7);        // L2
+    psubusb_r2r (mm4, mm7);     // L2 - avg
+    movq_r2r (mm4, mm5);        // avg
+    psubusb_r2r (mm2, mm5);     // avg - L2
+    por_r2r (mm7, mm5);         // abs(avg-L2)
+    movq_r2r (mm4, mm6);        // copy of avg for later
+
+
+    // get abs value of possible LP2 comb
+    movq_r2r (mm0, mm7);        // LP2
+    psubusb_r2r (mm4, mm7);     // LP2 - avg
+    psubusb_r2r (mm0, mm4);     // avg - LP2
+    por_r2r (mm7, mm4);         // abs(avg-LP2)
+
+    // use L2 or LP2 depending upon which makes smaller comb
+    psubusb_r2r (mm5, mm4);     // see if it goes to zero
+    psubusb_r2r (mm5, mm5);     // 0
+    pcmpeqb_r2r (mm5, mm4);     // if (mm4=0) then FF else 0
+    pcmpeqb_r2r (mm4, mm5);     // opposite of mm4
+
+    // if Comb(LP2) <= Comb(L2) then mm4=ff, mm5=0 else mm4=0, mm5 = 55
+    pand_r2r (mm2, mm5);        // use L2 if mm5 == ff, else 0
+    pand_r2r (mm0, mm4);        // use LP2 if mm4 = ff, else 0
+    por_r2r (mm5, mm4);         // may the best win
+
+    // Now lets clip our chosen value to be not outside of the range
+    // of the high/low range L1-L3 by more than abs(L1-L3)
+    // This allows some comb but limits the damages and also allows more
+    // detail than a boring oversmoothed clip.
+
+    movq_r2r (mm1, mm2);        // copy L1
+    psubusb_r2r (mm3, mm2);     // - L3, with saturation
+    paddusb_r2r (mm3, mm2);     // now = Max(L1,L3)
+
+    pcmpeqb_r2r (mm7, mm7);     // all ffffffff
+    psubusb_r2r (mm1, mm7);     // - L1 
+    paddusb_r2r (mm7, mm3);     // add, may sat at fff..
+    psubusb_r2r (mm7, mm3);     // now = Min(L1,L3)
+
+    // allow the value to be above the high or below the low by amt of MaxComb
+    paddusb_m2r (MaxComb, mm2); // increase max by diff
+    psubusb_m2r (MaxComb, mm3); // lower min by diff
+
+    psubusb_r2r (mm3, mm4);     // best - Min
+    paddusb_r2r (mm3, mm4);     // now = Max(best,Min(L1,L3)
+
+    pcmpeqb_r2r (mm7, mm7);     // all ffffffff
+    psubusb_r2r (mm4, mm7);     // - Max(best,Min(best,L3) 
+    paddusb_r2r (mm7, mm2);     // add may sat at FFF..
+    psubusb_r2r (mm7, mm2);     // now = Min( Max(best, Min(L1,L3), L2 )=L2 clipped
+
+    movntq_r2m (mm2, *output);  // move in our clipped best
+
+    // Advance to the next set of pixels.
+    output += 8;
+    m0 += 8;
+    t1 += 8;
+    b1 += 8;
+    m2 += 8;
+  }
+  sfence ();
+  emms ();
+#endif
+}
+
+
+static deinterlace_method_t greedyl_method = {
+  0,                            //DEINTERLACE_PLUGIN_API_VERSION,
+  "Motion Adaptive: Simple Detection",
+  "AdaptiveSimple",
+  3,
+  OIL_IMPL_FLAG_MMXEXT,
+  0,
+  0,
+  0,
+  1,
+  copy_scanline,
+  deinterlace_greedy_packed422_scanline_mmxext,
+  0,
+  {"Uses heuristics to detect motion in the input",
+        "frames and reconstruct image detail where",
+        "possible.  Use this for high quality output",
+        "even on monitors set to an arbitrary refresh",
+        "rate.",
+        "",
+        "Simple detection uses linear interpolation",
+        "where motion is detected, using a two-field",
+        "buffer.  This is the Greedy: Low Motion",
+      "deinterlacer from DScaler."}
+};
+
+deinterlace_method_t *
+dscaler_greedyl_get_method (void)
+{
+  return &greedyl_method;
+}
diff --git a/gst/deinterlace2/tvtime/greedyh.asm b/gst/deinterlace2/tvtime/greedyh.asm
new file mode 100644
index 00000000..92ad1fe1
--- /dev/null
+++ b/gst/deinterlace2/tvtime/greedyh.asm
@@ -0,0 +1,307 @@
+/*
+ *
+ * GStreamer
+ * Copyright (c) 2001 Tom Barry.  All rights reserved.
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Tom Barry.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+
+#include "x86-64_macros.inc"
+
+void FUNCT_NAME( GstDeinterlace2 *object)
+{
+    int64_t i;
+    int InfoIsOdd = 0;
+
+    // in tight loop some vars are accessed faster in local storage
+    int64_t YMask        = 0x00ff00ff00ff00ffull; // to keep only luma
+    int64_t UVMask       = 0xff00ff00ff00ff00ull; // to keep only chroma
+    int64_t ShiftMask    = 0xfefffefffefffeffull; // to avoid shifting chroma to luma
+    int64_t QW256        = 0x0100010001000100ull; // 4 256's
+
+    // Set up our two parms that are actually evaluated for each pixel
+    i=GreedyMaxComb;
+    int64_t MaxComb = i << 56 | i << 48 | i << 40 | i << 32 | i << 24 | i << 16 | i << 8 | i;
+    
+    i = GreedyMotionThreshold;		// scale to range of 0-257
+    int64_t MotionThreshold = i << 48 | i << 32 | i << 16 | i | UVMask;
+    
+    i = GreedyMotionSense;		// scale to range of 0-257
+    int64_t MotionSense = i << 48 | i << 32 | i << 16 | i;
+
+    int Line;
+    long LoopCtr;
+    unsigned int Pitch = object->field_stride;
+
+    unsigned char* L1;					// ptr to Line1, of 3
+    unsigned char* L2;					// ptr to Line2, the weave line
+    unsigned char* L3;					// ptr to Line3
+
+    unsigned char* L2P;					// ptr to prev Line2
+    unsigned char* Dest = GST_BUFFER_DATA(object->out_buf);
+
+    int64_t QW256B;
+    int64_t LastAvg=0;			//interp value from left qword
+
+    i = 0xffffffff - 256;
+    QW256B =  i << 48 |  i << 32 | i << 16 | i;  // save a couple instr on PMINSW instruct.
+
+
+    // copy first even line no matter what, and the first odd line if we're
+    // processing an EVEN field. (note diff from other deint rtns.)
+
+    if (object->field_history[object->history_count-1].flags == PICTURE_INTERLACED_BOTTOM) {
+      InfoIsOdd = 1;
+
+      L1 = GST_BUFFER_DATA(object->field_history[object->history_count-2].buf);
+      L2 = GST_BUFFER_DATA(object->field_history[object->history_count-1].buf); 
+      L3 = L1 + Pitch;
+      L2P = GST_BUFFER_DATA(object->field_history[object->history_count-3].buf);
+
+      // copy first even line
+      object->pMemcpy(Dest, L1, object->line_length);
+      Dest += object->output_stride;
+    } 
+    else {
+      InfoIsOdd = 0;
+      L1 = GST_BUFFER_DATA(object->field_history[object->history_count-2].buf);
+      L2 = GST_BUFFER_DATA(object->field_history[object->history_count-1].buf) + Pitch; 
+      L3 = L1 + Pitch;
+      L2P = GST_BUFFER_DATA(object->field_history[object->history_count-3].buf) + Pitch;
+
+      // copy first even line
+      object->pMemcpy(Dest, GST_BUFFER_DATA(object->field_history[0].buf), object->line_length);
+      Dest += object->output_stride;
+      // then first odd line
+      object->pMemcpy(Dest, L1, object->line_length);
+      Dest += object->output_stride;
+    }
+
+
+    long oldbx;
+
+    for (Line = 0; Line < (object->field_height - 1); ++Line) {
+        LoopCtr = object->line_length / 8 - 1; // there are LineLength / 8 qwords per line but do 1 less, adj at end of loop
+
+        // For ease of reading, the comments below assume that we're operating on an odd
+        // field (i.e., that InfoIsOdd is true).  Assume the obvious for even lines..
+        __asm__ __volatile__
+            (
+             // save ebx (-fPIC)
+	     MOVX" %%"XBX", %[oldbx]\n\t"
+
+             MOVX"  %[L1],          %%"XAX"\n\t"
+             LEAX"  8(%%"XAX"),     %%"XBX"\n\t"    // next qword needed by DJR
+             MOVX"  %[L3],          %%"XCX"\n\t"
+             SUBX"  %%"XAX",        %%"XCX"\n\t"    // carry L3 addr as an offset
+             MOVX"  %[L2P],         %%"XDX"\n\t"
+             MOVX"  %[L2],          %%"XSI"\n\t"
+             MOVX"  %[Dest],        %%"XDI"\n\t"    // DL1 if Odd or DL2 if Even
+
+             ".align 8\n\t"
+             "1:\n\t"
+
+             "movq  (%%"XSI"),      %%mm0\n\t"      // L2 - the newest weave pixel value
+             "movq  (%%"XAX"),      %%mm1\n\t"      // L1 - the top pixel
+             "movq  (%%"XDX"),      %%mm2\n\t"      // L2P - the prev weave pixel
+             "movq  (%%"XAX", %%"XCX"), %%mm3\n\t"  // L3, next odd row
+             "movq  %%mm1,          %%mm6\n\t"      // L1 - get simple single pixel interp
+             //	pavgb   mm6, mm3                    // use macro below
+             V_PAVGB ("%%mm6", "%%mm3", "%%mm4", "%[ShiftMask]")
+
+             // DJR - Diagonal Jaggie Reduction
+             // In the event that we are going to use an average (Bob) pixel we do not want a jagged
+             // stair step effect.  To combat this we avg in the 2 horizontally adjacen pixels into the
+             // interpolated Bob mix. This will do horizontal smoothing for only the Bob'd pixels.
+
+             "movq  %[LastAvg],     %%mm4\n\t"      // the bob value from prev qword in row
+             "movq  %%mm6,          %[LastAvg]\n\t" // save for next pass
+             "psrlq $48,            %%mm4\n\t"      // right justify 1 pixel
+             "movq  %%mm6,          %%mm7\n\t"      // copy of simple bob pixel
+             "psllq $16,            %%mm7\n\t"      // left justify 3 pixels
+             "por   %%mm7,          %%mm4\n\t"      // and combine
+
+             "movq  (%%"XBX"),      %%mm5\n\t"      // next horiz qword from L1
+             //			pavgb   mm5, qword ptr[ebx+ecx] // next horiz qword from L3, use macro below
+             V_PAVGB ("%%mm5", "(%%"XBX",%%"XCX")", "%%mm7", "%[ShiftMask]")
+             "psllq $48,            %%mm5\n\t"      // left just 1 pixel
+             "movq  %%mm6,          %%mm7\n\t"      // another copy of simple bob pixel
+             "psrlq $16,            %%mm7\n\t"      // right just 3 pixels
+             "por   %%mm7,          %%mm5\n\t"      // combine
+             //			pavgb	mm4, mm5			// avg of forward and prev by 1 pixel, use macro
+             V_PAVGB ("%%mm4", "%%mm5", "%%mm5", "%[ShiftMask]")   // mm5 gets modified if MMX
+             //			pavgb	mm6, mm4			// avg of center and surround interp vals, use macro
+             V_PAVGB ("%%mm6", "%%mm4", "%%mm7", "%[ShiftMask]")
+
+             // Don't do any more averaging than needed for mmx. It hurts performance and causes rounding errors.
+#ifndef IS_MMX
+             //          pavgb	mm4, mm6			// 1/4 center, 3/4 adjacent
+             V_PAVGB ("%%mm4", "%%mm6", "%%mm7", "%[ShiftMask]")
+             //    		pavgb	mm6, mm4			// 3/8 center, 5/8 adjacent
+             V_PAVGB ("%%mm6", "%%mm4", "%%mm7", "%[ShiftMask]")
+#endif
+
+             // get abs value of possible L2 comb
+             "movq    %%mm6,        %%mm4\n\t"      // work copy of interp val
+             "movq    %%mm2,        %%mm7\n\t"      // L2
+             "psubusb %%mm4,        %%mm7\n\t"      // L2 - avg
+             "movq    %%mm4,        %%mm5\n\t"      // avg
+             "psubusb %%mm2,        %%mm5\n\t"      // avg - L2
+             "por     %%mm7,        %%mm5\n\t"      // abs(avg-L2)
+
+             // get abs value of possible L2P comb
+             "movq    %%mm0,        %%mm7\n\t"      // L2P
+             "psubusb %%mm4,        %%mm7\n\t"      // L2P - avg
+             "psubusb %%mm0,        %%mm4\n\t"      // avg - L2P
+             "por     %%mm7,        %%mm4\n\t"      // abs(avg-L2P)
+
+             // use L2 or L2P depending upon which makes smaller comb
+             "psubusb %%mm5,        %%mm4\n\t"      // see if it goes to zero
+             "psubusb %%mm5,        %%mm5\n\t"      // 0
+             "pcmpeqb %%mm5,        %%mm4\n\t"      // if (mm4=0) then FF else 0
+             "pcmpeqb %%mm4,        %%mm5\n\t"      // opposite of mm4
+
+             // if Comb(L2P) <= Comb(L2) then mm4=ff, mm5=0 else mm4=0, mm5 = 55
+             "pand    %%mm2,        %%mm5\n\t"      // use L2 if mm5 == ff, else 0
+             "pand    %%mm0,        %%mm4\n\t"      // use L2P if mm4 = ff, else 0
+             "por     %%mm5,        %%mm4\n\t"      // may the best win
+
+             // Inventory: at this point we have the following values:
+             // mm0 = L2P (or L2)
+             // mm1 = L1
+             // mm2 = L2 (or L2P)
+             // mm3 = L3
+             // mm4 = the best of L2,L2P weave pixel, base upon comb
+             // mm6 = the avg interpolated value, if we need to use it
+
+             // Let's measure movement, as how much the weave pixel has changed
+             "movq    %%mm2,        %%mm7\n\t"
+             "psubusb %%mm0,        %%mm2\n\t"
+             "psubusb %%mm7,        %%mm0\n\t"
+             "por     %%mm2,        %%mm0\n\t"      // abs value of change, used later
+
+             // Now lets clip our chosen value to be not outside of the range
+             // of the high/low range L1-L3 by more than MaxComb.
+             // This allows some comb but limits the damages and also allows more
+             // detail than a boring oversmoothed clip.
+             "movq    %%mm1,        %%mm2\n\t"      // copy L1
+             //	pmaxub mm2, mm3                     // use macro
+             V_PMAXUB ("%%mm2", "%%mm3")            // now = Max(L1,L3)
+             "movq    %%mm1,        %%mm5\n\t"      // copy L1
+             // pminub	mm5, mm3                    // now = Min(L1,L3), use macro
+             V_PMINUB ("%%mm5", "%%mm3", "%%mm7")
+             // allow the value to be above the high or below the low by amt of MaxComb
+             "psubusb %[MaxComb],   %%mm5\n\t"      // lower min by diff
+             "paddusb %[MaxComb],   %%mm2\n\t"      // increase max by diff
+             // pmaxub	mm4, mm5                    // now = Max(best,Min(L1,L3) use macro
+             V_PMAXUB ("%%mm4", "%%mm5")
+             // pminub	mm4, mm2                    // now = Min( Max(best, Min(L1,L3), L2 )=L2 clipped
+             V_PMINUB ("%%mm4", "%%mm2", "%%mm7")
+
+             // Blend weave pixel with bob pixel, depending on motion val in mm0
+             "psubusb %[MotionThreshold], %%mm0\n\t"// test Threshold, clear chroma change >>>??
+             "pmullw  %[MotionSense], %%mm0\n\t"    // mul by user factor, keep low 16 bits
+             "movq    %[QW256], %%mm7\n\t"
+#ifdef HAVE_SSE
+             "pminsw  %%mm7,        %%mm0\n\t"      // max = 256
+#else
+             "paddusw %[QW256B],    %%mm0\n\t"      // add, may sat at fff..
+             "psubusw %[QW256B],    %%mm0\n\t"      // now = Min(L1,256)
+#endif
+             "psubusw %%mm0,        %%mm7\n\t"      // so the 2 sum to 256, weighted avg
+             "movq    %%mm4,        %%mm2\n\t"      // save weave chroma info before trashing
+             "pand    %[YMask],     %%mm4\n\t"      // keep only luma from calc'd value
+             "pmullw  %%mm7,        %%mm4\n\t"      // use more weave for less motion
+             "pand    %[YMask],     %%mm6\n\t"      // keep only luma from calc'd value
+             "pmullw  %%mm0,        %%mm6\n\t"      // use more bob for large motion
+             "paddusw %%mm6,        %%mm4\n\t"      // combine
+             "psrlw   $8,           %%mm4\n\t"      // div by 256 to get weighted avg
+
+             // chroma comes from weave pixel
+             "pand    %[UVMask],    %%mm2\n\t"      // keep chroma
+             "por     %%mm4,        %%mm2\n\t"      // and combine
+
+             V_MOVNTQ ("(%%"XDI")", "%%mm2")        // move in our clipped best, use macro
+
+             // bump ptrs and loop
+             LEAX"    8(%%"XAX"),   %%"XAX"\n\t"
+             LEAX"    8(%%"XBX"),   %%"XBX"\n\t"
+             LEAX"    8(%%"XDX"),   %%"XDX"\n\t"
+             LEAX"    8(%%"XDI"),   %%"XDI"\n\t"
+             LEAX"    8(%%"XSI"),   %%"XSI"\n\t"
+             DECX"    %[LoopCtr]\n\t"
+             "jg      1b\n\t"                       // loop if not to last line
+                                                    // note P-III default assumes backward branches taken
+             "jl      1f\n\t"                       // done
+             MOVX"    %%"XAX",      %%"XBX"\n\t"  // sharpness lookahead 1 byte only, be wrong on 1
+             "jmp     1b\n\t"
+
+             "1:\n\t"
+	     MOVX" %[oldbx], %%"XBX"\n\t"
+
+             : /* no outputs */
+
+             : [LastAvg]         "m"(LastAvg),
+               [L1]              "m"(L1),
+               [L3]              "m"(L3),
+               [L2P]             "m"(L2P),
+               [L2]              "m"(L2),
+               [Dest]            "m"(Dest),
+               [ShiftMask]       "m"(ShiftMask),
+               [MaxComb]         "m"(MaxComb),
+               [MotionThreshold] "m"(MotionThreshold),
+               [MotionSense]     "m"(MotionSense),
+               [QW256B]          "m"(QW256B),
+               [YMask]           "m"(YMask),
+               [UVMask]          "m"(UVMask),
+               [LoopCtr]         "m"(LoopCtr),
+               [QW256]           "m"(QW256),
+	       [oldbx]           "m"(oldbx)
+
+             : XAX, XCX, XDX, XSI, XDI,
+#ifdef HAVE_CPU_I386
+               "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",
+#endif
+               "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
+               "memory", "cc"
+            );
+
+        Dest += object->output_stride;
+        object->pMemcpy(Dest, L3, object->line_length);
+        Dest += object->output_stride;
+
+        L1  += Pitch;
+        L2  += Pitch;
+        L3  += Pitch;
+        L2P += Pitch;
+	}
+
+    if (InfoIsOdd) {
+        object->pMemcpy(Dest, L2, object->line_length);
+    }
+
+    // clear out the MMX registers ready for doing floating point again
+#ifdef HAVE_CPU_I386
+    __asm__ __volatile__ ("emms\n\t");
+#endif
+}
diff --git a/gst/deinterlace2/tvtime/greedyh.c b/gst/deinterlace2/tvtime/greedyh.c
new file mode 100644
index 00000000..623c2d8b
--- /dev/null
+++ b/gst/deinterlace2/tvtime/greedyh.c
@@ -0,0 +1,148 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "greedyh.h"
+#include "greedyhmacros.h"
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <string.h>
+
+#include "gst/gst.h"
+#include "plugins.h"
+#include "gstdeinterlace2.h"
+#include "speedy.h"
+
+
+#define MAXCOMB_DEFAULT          5
+#define MOTIONTHRESHOLD_DEFAULT 25
+#define MOTIONSENSE_DEFAULT     30
+
+unsigned int GreedyMaxComb;
+
+unsigned int GreedyMotionThreshold;
+
+unsigned int GreedyMotionSense;
+
+
+#define IS_SSE
+#define SSE_TYPE SSE
+#define FUNCT_NAME greedyDScaler_SSE
+#include "greedyh.asm"
+#undef SSE_TYPE
+#undef IS_SSE
+#undef FUNCT_NAME
+
+#define IS_3DNOW
+#define FUNCT_NAME greedyDScaler_3DNOW
+#define SSE_TYPE 3DNOW
+#include "greedyh.asm"
+#undef SSE_TYPE
+#undef IS_3DNOW
+#undef FUNCT_NAME
+
+#define IS_MMX
+#define SSE_TYPE MMX
+#define FUNCT_NAME greedyDScaler_MMX
+#include "greedyh.asm"
+#undef SSE_TYPE
+#undef IS_MMX
+#undef FUNCT_NAME
+
+void
+deinterlace_frame_di_greedyh (GstDeinterlace2 * object)
+{
+  if (object->cpu_feature_flags & OIL_IMPL_FLAG_SSE) {
+    greedyh_filter_sse (object);
+  } else if (object->cpu_feature_flags & OIL_IMPL_FLAG_3DNOW) {
+    greedyh_filter_3dnow (object);
+  } else {
+    greedyh_filter_mmx (object);
+  }
+}
+
+static deinterlace_method_t greedyh_method = {
+  0,                            //DEINTERLACE_PLUGIN_API_VERSION,
+  "Motion Adaptive: Advanced Detection",
+  "AdaptiveAdvanced",
+  4,
+  OIL_IMPL_FLAG_MMX,
+  0,
+  0,
+  0,
+  0,
+  0,
+  0,
+  deinterlace_frame_di_greedyh,
+  {"Uses heuristics to detect motion in the input",
+        "frames and reconstruct image detail where",
+        "possible.  Use this for high quality output",
+        "even on monitors set to an arbitrary refresh",
+        "rate.",
+        "",
+        "Advanced detection uses linear interpolation",
+        "where motion is detected, using a four-field",
+        "buffer.  This is the Greedy: High Motion",
+      "deinterlacer from DScaler."}
+};
+
+deinterlace_method_t *
+dscaler_greedyh_get_method (void)
+{
+  greedyh_init ();
+  return &greedyh_method;
+}
+
+void
+greedyh_init (void)
+{
+  GreedyMaxComb = MAXCOMB_DEFAULT;
+  GreedyMotionThreshold = MOTIONTHRESHOLD_DEFAULT;
+  GreedyMotionSense = MOTIONSENSE_DEFAULT;
+}
+
+void
+greedyh_filter_mmx (GstDeinterlace2 * object)
+{
+  greedyDScaler_MMX (object);
+}
+
+void
+greedyh_filter_3dnow (GstDeinterlace2 * object)
+{
+  greedyDScaler_3DNOW (object);
+}
+
+void
+greedyh_filter_sse (GstDeinterlace2 * object)
+{
+  greedyDScaler_SSE (object);
+}
diff --git a/gst/deinterlace2/tvtime/greedyh.h b/gst/deinterlace2/tvtime/greedyh.h
new file mode 100644
index 00000000..1156836a
--- /dev/null
+++ b/gst/deinterlace2/tvtime/greedyh.h
@@ -0,0 +1,45 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#ifndef GREEDYH_H_INCLUDED
+#define GREEDYH_H_INCLUDED
+
+#include "gstdeinterlace2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void greedyh_init( void );
+void greedyh_filter_mmx( GstDeinterlace2 *object );
+void greedyh_filter_3dnow( GstDeinterlace2 *object );
+void greedyh_filter_sse( GstDeinterlace2 *object );
+
+#ifdef __cplusplus
+};
+#endif
+
+#endif /* GREEDYH_H_INCLUDED */
diff --git a/gst/deinterlace2/tvtime/greedyhmacros.h b/gst/deinterlace2/tvtime/greedyhmacros.h
new file mode 100644
index 00000000..5f65959c
--- /dev/null
+++ b/gst/deinterlace2/tvtime/greedyhmacros.h
@@ -0,0 +1,74 @@
+/////////////////////////////////////////////////////////////////////////////
+// Copyright (c) 2001 Tom Barry.  All rights reserved.
+/////////////////////////////////////////////////////////////////////////////
+//
+//	This file is subject to the terms of the GNU General Public License as
+//	published by the Free Software Foundation.  A copy of this license is
+//	included with this software distribution in the file COPYING.  If you
+//	do not have a copy, you may obtain a copy by writing to the Free
+//	Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
+//
+//	This software 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
+//
+/////////////////////////////////////////////////////////////////////////////
+
+// Define a few macros for CPU dependent instructions. 
+// I suspect I don't really understand how the C macro preprocessor works but
+// this seems to get the job done.          // TRB 7/01
+
+// BEFORE USING THESE YOU MUST SET:
+
+// #define SSE_TYPE SSE            (or MMX or 3DNOW)
+
+// some macros for pavgb instruction
+//      V_PAVGB(mmr1, mmr2, mmr work register, smask) mmr2 may = mmrw if you can trash it
+
+#define V_PAVGB_MMX(mmr1, mmr2, mmrw, smask) \
+	"movq    "mmr2",  "mmrw"\n\t"            \
+	"pand    "smask", "mmrw"\n\t"            \
+	"psrlw   $1,      "mmrw"\n\t"            \
+	"pand    "smask", "mmr1"\n\t"            \
+	"psrlw   $1,      "mmr1"\n\t"            \
+	"paddusb "mmrw",  "mmr1"\n\t"
+#define V_PAVGB_SSE(mmr1, mmr2, mmrw, smask)      "pavgb   "mmr2", "mmr1"\n\t"
+#define V_PAVGB_3DNOW(mmr1, mmr2, mmrw, smask)    "pavgusb "mmr2", "mmr1"\n\t"
+#define V_PAVGB(mmr1, mmr2, mmrw, smask)          V_PAVGB2(mmr1, mmr2, mmrw, smask, SSE_TYPE) 
+#define V_PAVGB2(mmr1, mmr2, mmrw, smask, ssetyp) V_PAVGB3(mmr1, mmr2, mmrw, smask, ssetyp) 
+#define V_PAVGB3(mmr1, mmr2, mmrw, smask, ssetyp) V_PAVGB_##ssetyp(mmr1, mmr2, mmrw, smask) 
+
+// some macros for pmaxub instruction
+#define V_PMAXUB_MMX(mmr1, mmr2) \
+    "psubusb "mmr2", "mmr1"\n\t" \
+    "paddusb "mmr2", "mmr1"\n\t"
+#define V_PMAXUB_SSE(mmr1, mmr2)      "pmaxub "mmr2", "mmr1"\n\t"
+#define V_PMAXUB_3DNOW(mmr1, mmr2)    V_PMAXUB_MMX(mmr1, mmr2)  // use MMX version
+#define V_PMAXUB(mmr1, mmr2)          V_PMAXUB2(mmr1, mmr2, SSE_TYPE) 
+#define V_PMAXUB2(mmr1, mmr2, ssetyp) V_PMAXUB3(mmr1, mmr2, ssetyp) 
+#define V_PMAXUB3(mmr1, mmr2, ssetyp) V_PMAXUB_##ssetyp(mmr1, mmr2) 
+
+// some macros for pminub instruction
+//      V_PMINUB(mmr1, mmr2, mmr work register)     mmr2 may NOT = mmrw
+#define V_PMINUB_MMX(mmr1, mmr2, mmrw) \
+    "pcmpeqb "mmrw", "mmrw"\n\t"       \
+    "psubusb "mmr2", "mmrw"\n\t"       \
+    "paddusb "mmrw", "mmr1"\n\t"       \
+    "psubusb "mmrw", "mmr1"\n\t"
+#define V_PMINUB_SSE(mmr1, mmr2, mmrw)      "pminub "mmr2", "mmr1"\n\t"
+#define V_PMINUB_3DNOW(mmr1, mmr2, mmrw)    V_PMINUB_MMX(mmr1, mmr2, mmrw)  // use MMX version
+#define V_PMINUB(mmr1, mmr2, mmrw)          V_PMINUB2(mmr1, mmr2, mmrw, SSE_TYPE) 
+#define V_PMINUB2(mmr1, mmr2, mmrw, ssetyp) V_PMINUB3(mmr1, mmr2, mmrw, ssetyp) 
+#define V_PMINUB3(mmr1, mmr2, mmrw, ssetyp) V_PMINUB_##ssetyp(mmr1, mmr2, mmrw) 
+
+// some macros for movntq instruction
+//      V_MOVNTQ(mmr1, mmr2) 
+#define V_MOVNTQ_MMX(mmr1, mmr2)      "movq   "mmr2", "mmr1"\n\t"
+#define V_MOVNTQ_3DNOW(mmr1, mmr2)    "movq   "mmr2", "mmr1"\n\t"
+#define V_MOVNTQ_SSE(mmr1, mmr2)      "movntq "mmr2", "mmr1"\n\t"
+#define V_MOVNTQ(mmr1, mmr2)          V_MOVNTQ2(mmr1, mmr2, SSE_TYPE) 
+#define V_MOVNTQ2(mmr1, mmr2, ssetyp) V_MOVNTQ3(mmr1, mmr2, ssetyp) 
+#define V_MOVNTQ3(mmr1, mmr2, ssetyp) V_MOVNTQ_##ssetyp(mmr1, mmr2)
+
+// end of macros
diff --git a/gst/deinterlace2/tvtime/mmx.h b/gst/deinterlace2/tvtime/mmx.h
new file mode 100644
index 00000000..3627e61b
--- /dev/null
+++ b/gst/deinterlace2/tvtime/mmx.h
@@ -0,0 +1,723 @@
+/*	mmx.h
+
+	MultiMedia eXtensions GCC interface library for IA32.
+
+	To use this library, simply include this header file
+	and compile with GCC.  You MUST have inlining enabled
+	in order for mmx_ok() to work; this can be done by
+	simply using -O on the GCC command line.
+
+	Compiling with -DMMX_TRACE will cause detailed trace
+	output to be sent to stderr for each mmx operation.
+	This adds lots of code, and obviously slows execution to
+	a crawl, but can be very useful for debugging.
+
+	THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
+	EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT
+	LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY
+	AND FITNESS FOR ANY PARTICULAR PURPOSE.
+
+	1997-98 by H. Dietz and R. Fisher
+
+ History:
+	97-98*	R.Fisher	Early versions
+	980501	R.Fisher	Original Release
+	980611*	H.Dietz		Rewrite, correctly implementing inlines, and
+		R.Fisher	 including direct register accesses.
+	980616	R.Fisher	Release of 980611 as 980616.
+	980714	R.Fisher	Minor corrections to Makefile, etc.
+	980715	R.Fisher	mmx_ok() now prevents optimizer from using
+				 clobbered values.
+				mmx_ok() now checks if cpuid instruction is
+				 available before trying to use it.
+	980726*	R.Fisher	mm_support() searches for AMD 3DNow, Cyrix
+				 Extended MMX, and standard MMX.  It returns a
+				 value which is positive if any of these are
+				 supported, and can be masked with constants to
+				 see which.  mmx_ok() is now a call to this
+	980726*	R.Fisher	Added i2r support for shift functions
+	980919	R.Fisher	Fixed AMD extended feature recognition bug.
+	980921	R.Fisher	Added definition/check for _MMX_H.
+				Added "float s[2]" to mmx_t for use with
+				  3DNow and EMMX.  So same mmx_t can be used.
+	981013	R.Fisher	Fixed cpuid function 1 bug (looked at wrong reg)
+				Fixed psllq_i2r error in mmxtest.c
+
+	* Unreleased (internal or interim) versions
+
+ Notes:
+	It appears that the latest gas has the pand problem fixed, therefore
+	  I'll undefine BROKEN_PAND by default.
+	String compares may be quicker than the multiple test/jumps in vendor
+	  test sequence in mmx_ok(), but I'm not concerned with that right now.
+
+ Acknowledgments:
+	Jussi Laako for pointing out the errors ultimately found to be
+	  connected to the failure to notify the optimizer of clobbered values.
+	Roger Hardiman for reminding us that CPUID isn't everywhere, and that
+	  someone may actually try to use this on a machine without CPUID.
+	  Also for suggesting code for checking this.
+	Robert Dale for pointing out the AMD recognition bug.
+	Jimmy Mayfield and Carl Witty for pointing out the Intel recognition
+	  bug.
+	Carl Witty for pointing out the psllq_i2r test bug.
+*/
+
+#ifndef _MMX_H
+#define _MMX_H
+
+/*#define MMX_TRACE */
+
+/*	Warning:  at this writing, the version of GAS packaged
+	with most Linux distributions does not handle the
+	parallel AND operation mnemonic correctly.  If the
+	symbol BROKEN_PAND is defined, a slower alternative
+	coding will be used.  If execution of mmxtest results
+	in an illegal instruction fault, define this symbol.
+*/
+#undef	BROKEN_PAND
+
+
+/*	The type of an value that fits in an MMX register
+	(note that long long constant values MUST be suffixed
+	 by LL and unsigned long long values by ULL, lest
+	 they be truncated by the compiler)
+*/
+typedef	union {
+	long long		q;	/* Quadword (64-bit) value */
+	unsigned long long	uq;	/* Unsigned Quadword */
+	int			d[2];	/* 2 Doubleword (32-bit) values */
+	unsigned int		ud[2];	/* 2 Unsigned Doubleword */
+	short			w[4];	/* 4 Word (16-bit) values */
+	unsigned short		uw[4];	/* 4 Unsigned Word */
+	char			b[8];	/* 8 Byte (8-bit) values */
+	unsigned char		ub[8];	/* 8 Unsigned Byte */
+	float			s[2];	/* Single-precision (32-bit) value */
+} mmx_t;
+
+
+/*	Function to test if multimedia instructions are supported...
+*/
+inline extern int
+mm_support(void)
+{
+	/* Returns 1 if MMX instructions are supported,
+	   3 if Cyrix MMX and Extended MMX instructions are supported
+	   5 if AMD MMX and 3DNow! instructions are supported
+	   0 if hardware does not support any of these
+	*/
+	register int rval = 0;
+
+	__asm__ __volatile__ (
+		/* See if CPUID instruction is supported ... */
+		/* ... Get copies of EFLAGS into eax and ecx */
+		"pushf\n\t"
+		"popl %%eax\n\t"
+		"movl %%eax, %%ecx\n\t"
+
+		/* ... Toggle the ID bit in one copy and store */
+		/*     to the EFLAGS reg */
+		"xorl $0x200000, %%eax\n\t"
+		"push %%eax\n\t"
+		"popf\n\t"
+
+		/* ... Get the (hopefully modified) EFLAGS */
+		"pushf\n\t"
+		"popl %%eax\n\t"
+
+		/* ... Compare and test result */
+		"xorl %%eax, %%ecx\n\t"
+		"testl $0x200000, %%ecx\n\t"
+		"jz NotSupported1\n\t"		/* Nothing supported */
+
+
+		/* Get standard CPUID information, and
+		       go to a specific vendor section */
+		"movl $0, %%eax\n\t"
+		"cpuid\n\t"
+
+		/* Check for Intel */
+		"cmpl $0x756e6547, %%ebx\n\t"
+		"jne TryAMD\n\t"
+		"cmpl $0x49656e69, %%edx\n\t"
+		"jne TryAMD\n\t"
+		"cmpl $0x6c65746e, %%ecx\n"
+		"jne TryAMD\n\t"
+		"jmp Intel\n\t"
+
+		/* Check for AMD */
+		"\nTryAMD:\n\t"
+		"cmpl $0x68747541, %%ebx\n\t"
+		"jne TryCyrix\n\t"
+		"cmpl $0x69746e65, %%edx\n\t"
+		"jne TryCyrix\n\t"
+		"cmpl $0x444d4163, %%ecx\n"
+		"jne TryCyrix\n\t"
+		"jmp AMD\n\t"
+
+		/* Check for Cyrix */
+		"\nTryCyrix:\n\t"
+		"cmpl $0x69727943, %%ebx\n\t"
+		"jne NotSupported2\n\t"
+		"cmpl $0x736e4978, %%edx\n\t"
+		"jne NotSupported3\n\t"
+		"cmpl $0x64616574, %%ecx\n\t"
+		"jne NotSupported4\n\t"
+		/* Drop through to Cyrix... */
+
+
+		/* Cyrix Section */
+		/* See if extended CPUID is supported */
+		"movl $0x80000000, %%eax\n\t"
+		"cpuid\n\t"
+		"cmpl $0x80000000, %%eax\n\t"
+		"jl MMXtest\n\t"	/* Try standard CPUID instead */
+
+		/* Extended CPUID supported, so get extended features */
+		"movl $0x80000001, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%eax\n\t"	/* Test for MMX */
+		"jz NotSupported5\n\t"		/* MMX not supported */
+		"testl $0x01000000, %%eax\n\t"	/* Test for Ext'd MMX */
+		"jnz EMMXSupported\n\t"
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\n"
+		"EMMXSupported:\n\t"
+		"movl $3, %0:\n\n\t"		/* EMMX and MMX Supported */
+		"jmp Return\n\t"
+
+
+		/* AMD Section */
+		"AMD:\n\t"
+
+		/* See if extended CPUID is supported */
+		"movl $0x80000000, %%eax\n\t"
+		"cpuid\n\t"
+		"cmpl $0x80000000, %%eax\n\t"
+		"jl MMXtest\n\t"	/* Try standard CPUID instead */
+
+		/* Extended CPUID supported, so get extended features */
+		"movl $0x80000001, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%edx\n\t"	/* Test for MMX */
+		"jz NotSupported6\n\t"		/* MMX not supported */
+		"testl $0x80000000, %%edx\n\t"	/* Test for 3DNow! */
+		"jnz ThreeDNowSupported\n\t"
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\n"
+		"ThreeDNowSupported:\n\t"
+		"movl $5, %0:\n\n\t"		/* 3DNow! and MMX Supported */
+		"jmp Return\n\t"
+
+
+		/* Intel Section */
+		"Intel:\n\t"
+
+		/* Check for MMX */
+		"MMXtest:\n\t"
+		"movl $1, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%edx\n\t"	/* Test for MMX */
+		"jz NotSupported7\n\t"		/* MMX Not supported */
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\t"
+
+		/* Nothing supported */
+		"\nNotSupported1:\n\t"
+		"#movl $101, %0:\n\n\t"
+		"\nNotSupported2:\n\t"
+		"#movl $102, %0:\n\n\t"
+		"\nNotSupported3:\n\t"
+		"#movl $103, %0:\n\n\t"
+		"\nNotSupported4:\n\t"
+		"#movl $104, %0:\n\n\t"
+		"\nNotSupported5:\n\t"
+		"#movl $105, %0:\n\n\t"
+		"\nNotSupported6:\n\t"
+		"#movl $106, %0:\n\n\t"
+		"\nNotSupported7:\n\t"
+		"#movl $107, %0:\n\n\t"
+		"movl $0, %0:\n\n\t"
+
+		"Return:\n\t"
+		: "=a" (rval)
+		: /* no input */
+		: "eax", "ebx", "ecx", "edx"
+	);
+
+	/* Return */
+	return(rval);
+}
+
+/*	Function to test if mmx instructions are supported...
+*/
+inline extern int
+mmx_ok(void)
+{
+	/* Returns 1 if MMX instructions are supported, 0 otherwise */
+	return ( mm_support() & 0x1 );
+}
+
+
+/*	Helper functions for the instruction macros that follow...
+	(note that memory-to-register, m2r, instructions are nearly
+	 as efficient as register-to-register, r2r, instructions;
+	 however, memory-to-memory instructions are really simulated
+	 as a convenience, and are only 1/3 as efficient)
+*/
+#ifdef	MMX_TRACE
+
+/*	Include the stuff for printing a trace to stderr...
+*/
+
+#include <stdio.h>
+
+#define	mmx_i2r(op, imm, reg) \
+	{ \
+		mmx_t mmx_trace; \
+		mmx_trace = (imm); \
+		fprintf(stderr, #op "_i2r(" #imm "=0x%016llx, ", mmx_trace.q); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%016llx) => ", mmx_trace.q); \
+		__asm__ __volatile__ (#op " %0, %%" #reg \
+				      : /* nothing */ \
+				      : "X" (imm)); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%016llx\n", mmx_trace.q); \
+	}
+
+#define	mmx_m2r(op, mem, reg) \
+	{ \
+		mmx_t mmx_trace; \
+		mmx_trace = (mem); \
+		fprintf(stderr, #op "_m2r(" #mem "=0x%016llx, ", mmx_trace.q); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%016llx) => ", mmx_trace.q); \
+		__asm__ __volatile__ (#op " %0, %%" #reg \
+				      : /* nothing */ \
+				      : "X" (mem)); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%016llx\n", mmx_trace.q); \
+	}
+
+#define	mmx_r2m(op, reg, mem) \
+	{ \
+		mmx_t mmx_trace; \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #op "_r2m(" #reg "=0x%016llx, ", mmx_trace.q); \
+		mmx_trace = (mem); \
+		fprintf(stderr, #mem "=0x%016llx) => ", mmx_trace.q); \
+		__asm__ __volatile__ (#op " %%" #reg ", %0" \
+				      : "=X" (mem) \
+				      : /* nothing */ ); \
+		mmx_trace = (mem); \
+		fprintf(stderr, #mem "=0x%016llx\n", mmx_trace.q); \
+	}
+
+#define	mmx_r2r(op, regs, regd) \
+	{ \
+		mmx_t mmx_trace; \
+		__asm__ __volatile__ ("movq %%" #regs ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #op "_r2r(" #regs "=0x%016llx, ", mmx_trace.q); \
+		__asm__ __volatile__ ("movq %%" #regd ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #regd "=0x%016llx) => ", mmx_trace.q); \
+		__asm__ __volatile__ (#op " %" #regs ", %" #regd); \
+		__asm__ __volatile__ ("movq %%" #regd ", %0" \
+				      : "=X" (mmx_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #regd "=0x%016llx\n", mmx_trace.q); \
+	}
+
+#define	mmx_m2m(op, mems, memd) \
+	{ \
+		mmx_t mmx_trace; \
+		mmx_trace = (mems); \
+		fprintf(stderr, #op "_m2m(" #mems "=0x%016llx, ", mmx_trace.q); \
+		mmx_trace = (memd); \
+		fprintf(stderr, #memd "=0x%016llx) => ", mmx_trace.q); \
+		__asm__ __volatile__ ("movq %0, %%mm0\n\t" \
+				      #op " %1, %%mm0\n\t" \
+				      "movq %%mm0, %0" \
+				      : "=X" (memd) \
+				      : "X" (mems)); \
+		mmx_trace = (memd); \
+		fprintf(stderr, #memd "=0x%016llx\n", mmx_trace.q); \
+	}
+
+#else
+
+/*	These macros are a lot simpler without the tracing...
+*/
+
+#define	mmx_i2r(op, imm, reg) \
+	__asm__ __volatile__ (#op " $" #imm ", %%" #reg \
+			      : /* nothing */ \
+			      : /* nothing */);
+
+#define	mmx_m2r(op, mem, reg) \
+	__asm__ __volatile__ (#op " %0, %%" #reg \
+			      : /* nothing */ \
+			      : "m" (mem))
+
+#define	mmx_r2m(op, reg, mem) \
+	__asm__ __volatile__ (#op " %%" #reg ", %0" \
+			      : "=m" (mem) \
+			      : /* nothing */ )
+
+#define	mmx_r2r(op, regs, regd) \
+	__asm__ __volatile__ (#op " %" #regs ", %" #regd)
+
+#define	mmx_m2m(op, mems, memd) \
+	__asm__ __volatile__ ("movq %0, %%mm0\n\t" \
+			      #op " %1, %%mm0\n\t" \
+			      "movq %%mm0, %0" \
+			      : "=m" (memd) \
+			      : "m" (mems))
+
+#endif
+
+
+/*	1x64 MOVe Quadword
+	(this is both a load and a store...
+	 in fact, it is the only way to store)
+*/
+#define	movq_m2r(var, reg)	mmx_m2r(movq, var, reg)
+#define	movq_r2m(reg, var)	mmx_r2m(movq, reg, var)
+#define	movq_r2r(regs, regd)	mmx_r2r(movq, regs, regd)
+#define	movq(vars, vard) \
+	__asm__ __volatile__ ("movq %1, %%mm0\n\t" \
+			      "movq %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	1x32 MOVe Doubleword
+	(like movq, this is both load and store...
+	 but is most useful for moving things between
+	 mmx registers and ordinary registers)
+*/
+#define	movd_m2r(var, reg)	mmx_m2r(movd, var, reg)
+#define	movd_r2m(reg, var)	mmx_r2m(movd, reg, var)
+#define	movd_r2r(regs, regd)	mmx_r2r(movd, regs, regd)
+#define	movd(vars, vard) \
+	__asm__ __volatile__ ("movd %1, %%mm0\n\t" \
+			      "movd %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	2x32, 4x16, and 8x8 Parallel ADDs
+*/
+#define	paddd_m2r(var, reg)	mmx_m2r(paddd, var, reg)
+#define	paddd_r2r(regs, regd)	mmx_r2r(paddd, regs, regd)
+#define	paddd(vars, vard)	mmx_m2m(paddd, vars, vard)
+
+#define	paddw_m2r(var, reg)	mmx_m2r(paddw, var, reg)
+#define	paddw_r2r(regs, regd)	mmx_r2r(paddw, regs, regd)
+#define	paddw(vars, vard)	mmx_m2m(paddw, vars, vard)
+
+#define	paddb_m2r(var, reg)	mmx_m2r(paddb, var, reg)
+#define	paddb_r2r(regs, regd)	mmx_r2r(paddb, regs, regd)
+#define	paddb(vars, vard)	mmx_m2m(paddb, vars, vard)
+
+
+/*	4x16 and 8x8 Parallel ADDs using Saturation arithmetic
+*/
+#define	paddsw_m2r(var, reg)	mmx_m2r(paddsw, var, reg)
+#define	paddsw_r2r(regs, regd)	mmx_r2r(paddsw, regs, regd)
+#define	paddsw(vars, vard)	mmx_m2m(paddsw, vars, vard)
+
+#define	paddsb_m2r(var, reg)	mmx_m2r(paddsb, var, reg)
+#define	paddsb_r2r(regs, regd)	mmx_r2r(paddsb, regs, regd)
+#define	paddsb(vars, vard)	mmx_m2m(paddsb, vars, vard)
+
+
+/*	4x16 and 8x8 Parallel ADDs using Unsigned Saturation arithmetic
+*/
+#define	paddusw_m2r(var, reg)	mmx_m2r(paddusw, var, reg)
+#define	paddusw_r2r(regs, regd)	mmx_r2r(paddusw, regs, regd)
+#define	paddusw(vars, vard)	mmx_m2m(paddusw, vars, vard)
+
+#define	paddusb_m2r(var, reg)	mmx_m2r(paddusb, var, reg)
+#define	paddusb_r2r(regs, regd)	mmx_r2r(paddusb, regs, regd)
+#define	paddusb(vars, vard)	mmx_m2m(paddusb, vars, vard)
+
+
+/*	2x32, 4x16, and 8x8 Parallel SUBs
+*/
+#define	psubd_m2r(var, reg)	mmx_m2r(psubd, var, reg)
+#define	psubd_r2r(regs, regd)	mmx_r2r(psubd, regs, regd)
+#define	psubd(vars, vard)	mmx_m2m(psubd, vars, vard)
+
+#define	psubw_m2r(var, reg)	mmx_m2r(psubw, var, reg)
+#define	psubw_r2r(regs, regd)	mmx_r2r(psubw, regs, regd)
+#define	psubw(vars, vard)	mmx_m2m(psubw, vars, vard)
+
+#define	psubb_m2r(var, reg)	mmx_m2r(psubb, var, reg)
+#define	psubb_r2r(regs, regd)	mmx_r2r(psubb, regs, regd)
+#define	psubb(vars, vard)	mmx_m2m(psubb, vars, vard)
+
+
+/*	4x16 and 8x8 Parallel SUBs using Saturation arithmetic
+*/
+#define	psubsw_m2r(var, reg)	mmx_m2r(psubsw, var, reg)
+#define	psubsw_r2r(regs, regd)	mmx_r2r(psubsw, regs, regd)
+#define	psubsw(vars, vard)	mmx_m2m(psubsw, vars, vard)
+
+#define	psubsb_m2r(var, reg)	mmx_m2r(psubsb, var, reg)
+#define	psubsb_r2r(regs, regd)	mmx_r2r(psubsb, regs, regd)
+#define	psubsb(vars, vard)	mmx_m2m(psubsb, vars, vard)
+
+
+/*	4x16 and 8x8 Parallel SUBs using Unsigned Saturation arithmetic
+*/
+#define	psubusw_m2r(var, reg)	mmx_m2r(psubusw, var, reg)
+#define	psubusw_r2r(regs, regd)	mmx_r2r(psubusw, regs, regd)
+#define	psubusw(vars, vard)	mmx_m2m(psubusw, vars, vard)
+
+#define	psubusb_m2r(var, reg)	mmx_m2r(psubusb, var, reg)
+#define	psubusb_r2r(regs, regd)	mmx_r2r(psubusb, regs, regd)
+#define	psubusb(vars, vard)	mmx_m2m(psubusb, vars, vard)
+
+
+/*	4x16 Parallel MULs giving Low 4x16 portions of results
+*/
+#define	pmullw_m2r(var, reg)	mmx_m2r(pmullw, var, reg)
+#define	pmullw_r2r(regs, regd)	mmx_r2r(pmullw, regs, regd)
+#define	pmullw(vars, vard)	mmx_m2m(pmullw, vars, vard)
+
+
+/*	4x16 Parallel MULs giving High 4x16 portions of results
+*/
+#define	pmulhw_m2r(var, reg)	mmx_m2r(pmulhw, var, reg)
+#define	pmulhw_r2r(regs, regd)	mmx_r2r(pmulhw, regs, regd)
+#define	pmulhw(vars, vard)	mmx_m2m(pmulhw, vars, vard)
+
+
+/*	4x16->2x32 Parallel Mul-ADD
+	(muls like pmullw, then adds adjacent 16-bit fields
+	 in the multiply result to make the final 2x32 result)
+*/
+#define	pmaddwd_m2r(var, reg)	mmx_m2r(pmaddwd, var, reg)
+#define	pmaddwd_r2r(regs, regd)	mmx_r2r(pmaddwd, regs, regd)
+#define	pmaddwd(vars, vard)	mmx_m2m(pmaddwd, vars, vard)
+
+
+/*	1x64 bitwise AND
+*/
+#ifdef	BROKEN_PAND
+#define	pand_m2r(var, reg) \
+	{ \
+		mmx_m2r(pandn, (mmx_t) -1LL, reg); \
+		mmx_m2r(pandn, var, reg); \
+	}
+#define	pand_r2r(regs, regd) \
+	{ \
+		mmx_m2r(pandn, (mmx_t) -1LL, regd); \
+		mmx_r2r(pandn, regs, regd); \
+	}
+#define	pand(vars, vard) \
+	{ \
+		movq_m2r(vard, mm0); \
+		mmx_m2r(pandn, (mmx_t) -1LL, mm0); \
+		mmx_m2r(pandn, vars, mm0); \
+		movq_r2m(mm0, vard); \
+	}
+#else
+#define	pand_m2r(var, reg)	mmx_m2r(pand, var, reg)
+#define	pand_r2r(regs, regd)	mmx_r2r(pand, regs, regd)
+#define	pand(vars, vard)	mmx_m2m(pand, vars, vard)
+#endif
+
+
+/*	1x64 bitwise AND with Not the destination
+*/
+#define	pandn_m2r(var, reg)	mmx_m2r(pandn, var, reg)
+#define	pandn_r2r(regs, regd)	mmx_r2r(pandn, regs, regd)
+#define	pandn(vars, vard)	mmx_m2m(pandn, vars, vard)
+
+
+/*	1x64 bitwise OR
+*/
+#define	por_m2r(var, reg)	mmx_m2r(por, var, reg)
+#define	por_r2r(regs, regd)	mmx_r2r(por, regs, regd)
+#define	por(vars, vard)	mmx_m2m(por, vars, vard)
+
+
+/*	1x64 bitwise eXclusive OR
+*/
+#define	pxor_m2r(var, reg)	mmx_m2r(pxor, var, reg)
+#define	pxor_r2r(regs, regd)	mmx_r2r(pxor, regs, regd)
+#define	pxor(vars, vard)	mmx_m2m(pxor, vars, vard)
+
+
+/*	2x32, 4x16, and 8x8 Parallel CoMPare for EQuality
+	(resulting fields are either 0 or -1)
+*/
+#define	pcmpeqd_m2r(var, reg)	mmx_m2r(pcmpeqd, var, reg)
+#define	pcmpeqd_r2r(regs, regd)	mmx_r2r(pcmpeqd, regs, regd)
+#define	pcmpeqd(vars, vard)	mmx_m2m(pcmpeqd, vars, vard)
+
+#define	pcmpeqw_m2r(var, reg)	mmx_m2r(pcmpeqw, var, reg)
+#define	pcmpeqw_r2r(regs, regd)	mmx_r2r(pcmpeqw, regs, regd)
+#define	pcmpeqw(vars, vard)	mmx_m2m(pcmpeqw, vars, vard)
+
+#define	pcmpeqb_m2r(var, reg)	mmx_m2r(pcmpeqb, var, reg)
+#define	pcmpeqb_r2r(regs, regd)	mmx_r2r(pcmpeqb, regs, regd)
+#define	pcmpeqb(vars, vard)	mmx_m2m(pcmpeqb, vars, vard)
+
+
+/*	2x32, 4x16, and 8x8 Parallel CoMPare for Greater Than
+	(resulting fields are either 0 or -1)
+*/
+#define	pcmpgtd_m2r(var, reg)	mmx_m2r(pcmpgtd, var, reg)
+#define	pcmpgtd_r2r(regs, regd)	mmx_r2r(pcmpgtd, regs, regd)
+#define	pcmpgtd(vars, vard)	mmx_m2m(pcmpgtd, vars, vard)
+
+#define	pcmpgtw_m2r(var, reg)	mmx_m2r(pcmpgtw, var, reg)
+#define	pcmpgtw_r2r(regs, regd)	mmx_r2r(pcmpgtw, regs, regd)
+#define	pcmpgtw(vars, vard)	mmx_m2m(pcmpgtw, vars, vard)
+
+#define	pcmpgtb_m2r(var, reg)	mmx_m2r(pcmpgtb, var, reg)
+#define	pcmpgtb_r2r(regs, regd)	mmx_r2r(pcmpgtb, regs, regd)
+#define	pcmpgtb(vars, vard)	mmx_m2m(pcmpgtb, vars, vard)
+
+
+/*	1x64, 2x32, and 4x16 Parallel Shift Left Logical
+*/
+#define	psllq_i2r(imm, reg)	mmx_i2r(psllq, imm, reg)
+#define	psllq_m2r(var, reg)	mmx_m2r(psllq, var, reg)
+#define	psllq_r2r(regs, regd)	mmx_r2r(psllq, regs, regd)
+#define	psllq(vars, vard)	mmx_m2m(psllq, vars, vard)
+
+#define	pslld_i2r(imm, reg)	mmx_i2r(pslld, imm, reg)
+#define	pslld_m2r(var, reg)	mmx_m2r(pslld, var, reg)
+#define	pslld_r2r(regs, regd)	mmx_r2r(pslld, regs, regd)
+#define	pslld(vars, vard)	mmx_m2m(pslld, vars, vard)
+
+#define	psllw_i2r(imm, reg)	mmx_i2r(psllw, imm, reg)
+#define	psllw_m2r(var, reg)	mmx_m2r(psllw, var, reg)
+#define	psllw_r2r(regs, regd)	mmx_r2r(psllw, regs, regd)
+#define	psllw(vars, vard)	mmx_m2m(psllw, vars, vard)
+
+
+/*	1x64, 2x32, and 4x16 Parallel Shift Right Logical
+*/
+#define	psrlq_i2r(imm, reg)	mmx_i2r(psrlq, imm, reg)
+#define	psrlq_m2r(var, reg)	mmx_m2r(psrlq, var, reg)
+#define	psrlq_r2r(regs, regd)	mmx_r2r(psrlq, regs, regd)
+#define	psrlq(vars, vard)	mmx_m2m(psrlq, vars, vard)
+
+#define	psrld_i2r(imm, reg)	mmx_i2r(psrld, imm, reg)
+#define	psrld_m2r(var, reg)	mmx_m2r(psrld, var, reg)
+#define	psrld_r2r(regs, regd)	mmx_r2r(psrld, regs, regd)
+#define	psrld(vars, vard)	mmx_m2m(psrld, vars, vard)
+
+#define	psrlw_i2r(imm, reg)	mmx_i2r(psrlw, imm, reg)
+#define	psrlw_m2r(var, reg)	mmx_m2r(psrlw, var, reg)
+#define	psrlw_r2r(regs, regd)	mmx_r2r(psrlw, regs, regd)
+#define	psrlw(vars, vard)	mmx_m2m(psrlw, vars, vard)
+
+
+/*	2x32 and 4x16 Parallel Shift Right Arithmetic
+*/
+#define	psrad_i2r(imm, reg)	mmx_i2r(psrad, imm, reg)
+#define	psrad_m2r(var, reg)	mmx_m2r(psrad, var, reg)
+#define	psrad_r2r(regs, regd)	mmx_r2r(psrad, regs, regd)
+#define	psrad(vars, vard)	mmx_m2m(psrad, vars, vard)
+
+#define	psraw_i2r(imm, reg)	mmx_i2r(psraw, imm, reg)
+#define	psraw_m2r(var, reg)	mmx_m2r(psraw, var, reg)
+#define	psraw_r2r(regs, regd)	mmx_r2r(psraw, regs, regd)
+#define	psraw(vars, vard)	mmx_m2m(psraw, vars, vard)
+
+
+/*	2x32->4x16 and 4x16->8x8 PACK and Signed Saturate
+	(packs source and dest fields into dest in that order)
+*/
+#define	packssdw_m2r(var, reg)	mmx_m2r(packssdw, var, reg)
+#define	packssdw_r2r(regs, regd) mmx_r2r(packssdw, regs, regd)
+#define	packssdw(vars, vard)	mmx_m2m(packssdw, vars, vard)
+
+#define	packsswb_m2r(var, reg)	mmx_m2r(packsswb, var, reg)
+#define	packsswb_r2r(regs, regd) mmx_r2r(packsswb, regs, regd)
+#define	packsswb(vars, vard)	mmx_m2m(packsswb, vars, vard)
+
+
+/*	4x16->8x8 PACK and Unsigned Saturate
+	(packs source and dest fields into dest in that order)
+*/
+#define	packuswb_m2r(var, reg)	mmx_m2r(packuswb, var, reg)
+#define	packuswb_r2r(regs, regd) mmx_r2r(packuswb, regs, regd)
+#define	packuswb(vars, vard)	mmx_m2m(packuswb, vars, vard)
+
+
+/*	2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK Low
+	(interleaves low half of dest with low half of source
+	 as padding in each result field)
+*/
+#define	punpckldq_m2r(var, reg)	mmx_m2r(punpckldq, var, reg)
+#define	punpckldq_r2r(regs, regd) mmx_r2r(punpckldq, regs, regd)
+#define	punpckldq(vars, vard)	mmx_m2m(punpckldq, vars, vard)
+
+#define	punpcklwd_m2r(var, reg)	mmx_m2r(punpcklwd, var, reg)
+#define	punpcklwd_r2r(regs, regd) mmx_r2r(punpcklwd, regs, regd)
+#define	punpcklwd(vars, vard)	mmx_m2m(punpcklwd, vars, vard)
+
+#define	punpcklbw_m2r(var, reg)	mmx_m2r(punpcklbw, var, reg)
+#define	punpcklbw_r2r(regs, regd) mmx_r2r(punpcklbw, regs, regd)
+#define	punpcklbw(vars, vard)	mmx_m2m(punpcklbw, vars, vard)
+
+
+/*	2x32->1x64, 4x16->2x32, and 8x8->4x16 UNPaCK High
+	(interleaves high half of dest with high half of source
+	 as padding in each result field)
+*/
+#define	punpckhdq_m2r(var, reg)	mmx_m2r(punpckhdq, var, reg)
+#define	punpckhdq_r2r(regs, regd) mmx_r2r(punpckhdq, regs, regd)
+#define	punpckhdq(vars, vard)	mmx_m2m(punpckhdq, vars, vard)
+
+#define	punpckhwd_m2r(var, reg)	mmx_m2r(punpckhwd, var, reg)
+#define	punpckhwd_r2r(regs, regd) mmx_r2r(punpckhwd, regs, regd)
+#define	punpckhwd(vars, vard)	mmx_m2m(punpckhwd, vars, vard)
+
+#define	punpckhbw_m2r(var, reg)	mmx_m2r(punpckhbw, var, reg)
+#define	punpckhbw_r2r(regs, regd) mmx_r2r(punpckhbw, regs, regd)
+#define	punpckhbw(vars, vard)	mmx_m2m(punpckhbw, vars, vard)
+
+
+/*	Empty MMx State
+	(used to clean-up when going from mmx to float use
+	 of the registers that are shared by both; note that
+	 there is no float-to-mmx operation needed, because
+	 only the float tag word info is corruptible)
+*/
+#ifdef	MMX_TRACE
+
+#define	emms() \
+	{ \
+		fprintf(stderr, "emms()\n"); \
+		__asm__ __volatile__ ("emms"); \
+	}
+
+#else
+
+#define	emms()			__asm__ __volatile__ ("emms")
+
+#endif
+
+#endif
diff --git a/gst/deinterlace2/tvtime/plugins.h b/gst/deinterlace2/tvtime/plugins.h
new file mode 100644
index 00000000..0eb90c0b
--- /dev/null
+++ b/gst/deinterlace2/tvtime/plugins.h
@@ -0,0 +1,42 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#ifndef TVTIME_PLUGINS_H_INCLUDED
+#define TVTIME_PLUGINS_H_INCLUDED
+
+deinterlace_method_t* dscaler_tomsmocomp_get_method( void );
+deinterlace_method_t* dscaler_greedyh_get_method( void );
+deinterlace_method_t* dscaler_greedyl_get_method( void );
+deinterlace_method_t* dscaler_vfir_get_method( void );
+
+//void linear_plugin_init( void );
+//void scalerbob_plugin_init( void );
+//void linearblend_plugin_init( void );
+//void weave_plugin_init( void );
+//void weavetff_plugin_init( void );
+//void weavebff_plugin_init( void );
+
+#endif /* TVTIME_PLUGINS_H_INCLUDED */
diff --git a/gst/deinterlace2/tvtime/speedtools.h b/gst/deinterlace2/tvtime/speedtools.h
new file mode 100644
index 00000000..677bb5e3
--- /dev/null
+++ b/gst/deinterlace2/tvtime/speedtools.h
@@ -0,0 +1,54 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#ifndef SPEEDTOOLS_H_INCLUDED
+#define SPEEDTOOLS_H_INCLUDED
+
+#define PREFETCH_2048(x) \
+    { int *pfetcha = (int *) x; \
+        prefetchnta( pfetcha ); \
+        prefetchnta( pfetcha + 64 ); \
+        prefetchnta( pfetcha + 128 ); \
+        prefetchnta( pfetcha + 192 ); \
+        pfetcha += 256; \
+        prefetchnta( pfetcha ); \
+        prefetchnta( pfetcha + 64 ); \
+        prefetchnta( pfetcha + 128 ); \
+        prefetchnta( pfetcha + 192 ); }
+
+#define READ_PREFETCH_2048(x) \
+    { int *pfetcha = (int *) x; int pfetchtmp; \
+        pfetchtmp = pfetcha[ 0 ] + pfetcha[ 16 ] + pfetcha[ 32 ] + pfetcha[ 48 ] + \
+            pfetcha[ 64 ] + pfetcha[ 80 ] + pfetcha[ 96 ] + pfetcha[ 112 ] + \
+            pfetcha[ 128 ] + pfetcha[ 144 ] + pfetcha[ 160 ] + pfetcha[ 176 ] + \
+            pfetcha[ 192 ] + pfetcha[ 208 ] + pfetcha[ 224 ] + pfetcha[ 240 ]; \
+        pfetcha += 256; \
+        pfetchtmp = pfetcha[ 0 ] + pfetcha[ 16 ] + pfetcha[ 32 ] + pfetcha[ 48 ] + \
+            pfetcha[ 64 ] + pfetcha[ 80 ] + pfetcha[ 96 ] + pfetcha[ 112 ] + \
+            pfetcha[ 128 ] + pfetcha[ 144 ] + pfetcha[ 160 ] + pfetcha[ 176 ] + \
+            pfetcha[ 192 ] + pfetcha[ 208 ] + pfetcha[ 224 ] + pfetcha[ 240 ]; }
+
+#endif /* SPEEDTOOLS_H_INCLUDED */
diff --git a/gst/deinterlace2/tvtime/speedy.c b/gst/deinterlace2/tvtime/speedy.c
new file mode 100644
index 00000000..821cc254
--- /dev/null
+++ b/gst/deinterlace2/tvtime/speedy.c
@@ -0,0 +1,2791 @@
+/**
+ * Copyright (c) 2002, 2003 Billy Biggs <vektor@dumbterm.net>.
+ * Copyright (C) 2001 Matthew J. Marjanovic <maddog@mir.com>
+ *
+ * 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, 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., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ */
+
+/**
+ * Includes 420to422, 422to444 scaling filters from the MPEG2 reference
+ * implementation.  The v12 source code indicates that they were written
+ * by Cheung Auyeung <auyeung@mot.com>.  The file they were in was:
+ *
+ * store.c, picture output routines
+ * Copyright (C) 1996, MPEG Software Simulation Group. All Rights Reserved.
+ *
+ * Disclaimer of Warranty
+ *
+ * These software programs are available to the user without any license fee or
+ * royalty on an "as is" basis.  The MPEG Software Simulation Group disclaims
+ * any and all warranties, whether express, implied, or statuary, including any
+ * implied warranties or merchantability or of fitness for a particular
+ * purpose.  In no event shall the copyright-holder be liable for any
+ * incidental, punitive, or consequential damages of any kind whatsoever
+ * arising from the use of these programs.
+ *
+ * This disclaimer of warranty extends to the user of these programs and user's
+ * customers, employees, agents, transferees, successors, and assigns.
+ *
+ * The MPEG Software Simulation Group does not represent or warrant that the
+ * programs furnished hereunder are free of infringement of any third-party
+ * patents.
+ *
+ * Commercial implementations of MPEG-1 and MPEG-2 video, including shareware,
+ * are subject to royalty fees to patent holders.  Many of these patents are
+ * general enough such that they are unavoidable regardless of implementation
+ * design.
+ *
+ */
+
+/**
+ * Code for the UYVY to YUYV routine comes from rivatv:
+ *
+ *   rivatv-convert.c video image conversion routines
+ *
+ *   Copyright (C) 2002 Stefan Jahn <stefan@lkcc.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., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+#include <stdio.h>
+#include <string.h>
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "gst/gst.h"
+#include "gstdeinterlace2.h"
+#include "speedy.h"
+#include "speedtools.h"
+#include "mmx.h"
+#include "sse.h"
+
+// TODO: remove includes
+//#include "attributes.h"
+//#include "mm_accel.h"
+
+/* Function pointer definitions. */
+void (*interpolate_packed422_scanline) (uint8_t * output, uint8_t * top,
+    uint8_t * bot, int width);
+void (*blit_colour_packed422_scanline) (uint8_t * output,
+    int width, int y, int cb, int cr);
+void (*blit_colour_packed4444_scanline) (uint8_t * output,
+    int width, int alpha, int luma, int cb, int cr);
+void (*blit_packed422_scanline) (uint8_t * dest, const uint8_t * src,
+    int width);
+void (*composite_packed4444_to_packed422_scanline) (uint8_t * output,
+    uint8_t * input, uint8_t * foreground, int width);
+void (*composite_packed4444_alpha_to_packed422_scanline) (uint8_t * output,
+    uint8_t * input, uint8_t * foreground, int width, int alpha);
+void (*composite_alphamask_to_packed4444_scanline) (uint8_t * output,
+    uint8_t * input, uint8_t * mask, int width, int textluma, int textcb,
+    int textcr);
+void (*composite_alphamask_alpha_to_packed4444_scanline) (uint8_t * output,
+    uint8_t * input, uint8_t * mask, int width, int textluma, int textcb,
+    int textcr, int alpha);
+void (*premultiply_packed4444_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*blend_packed422_scanline) (uint8_t * output, uint8_t * src1,
+    uint8_t * src2, int width, int pos);
+unsigned int (*diff_factor_packed422_scanline) (uint8_t * cur, uint8_t * old,
+    int width);
+unsigned int (*comb_factor_packed422_scanline) (uint8_t * top, uint8_t * mid,
+    uint8_t * bot, int width);
+void (*kill_chroma_packed422_inplace_scanline) (uint8_t * data, int width);
+
+void (*mirror_packed422_inplace_scanline) (uint8_t * data, int width);
+
+void (*speedy_memcpy) (void *output, const void *input, size_t size);
+
+void (*diff_packed422_block8x8) (pulldown_metrics_t * m, uint8_t * old,
+    uint8_t * new, int os, int ns);
+void (*a8_subpix_blit_scanline) (uint8_t * output, uint8_t * input,
+    int lasta, int startpos, int width);
+void (*quarter_blit_vertical_packed422_scanline) (uint8_t * output,
+    uint8_t * one, uint8_t * three, int width);
+void (*subpix_blit_vertical_packed422_scanline) (uint8_t * output,
+    uint8_t * top, uint8_t * bot, int subpixpos, int width);
+void (*packed444_to_nonpremultiplied_packed4444_scanline) (uint8_t * output,
+    uint8_t * input, int width, int alpha);
+void (*aspect_adjust_packed4444_scanline) (uint8_t * output, uint8_t * input,
+    int width, double pixel_aspect);
+void (*packed444_to_packed422_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*packed422_to_packed444_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*packed422_to_packed444_rec601_scanline) (uint8_t * dest, uint8_t * src,
+    int width);
+void (*packed444_to_rgb24_rec601_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*rgb24_to_packed444_rec601_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*rgba32_to_packed4444_rec601_scanline) (uint8_t * output, uint8_t * input,
+    int width);
+void (*invert_colour_packed422_inplace_scanline) (uint8_t * data, int width);
+
+void (*vfilter_chroma_121_packed422_scanline) (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b);
+void (*vfilter_chroma_332_packed422_scanline) (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b);
+void (*convert_uyvy_to_yuyv_scanline) (uint8_t * uyvy_buf, uint8_t * yuyv_buf,
+    int width);
+void (*composite_colour4444_alpha_to_packed422_scanline) (uint8_t * output,
+    uint8_t * input, int af, int y, int cb, int cr, int width, int alpha);
+
+/**
+ * result = (1 - alpha)B + alpha*F
+ *        =  B - alpha*B + alpha*F
+ *        =  B + alpha*(F - B)
+ */
+
+static inline __attribute__ ((always_inline, const))
+     int multiply_alpha (int a, int r)
+{
+  int temp;
+
+  temp = (r * a) + 0x80;
+  return ((temp + (temp >> 8)) >> 8);
+}
+
+static inline __attribute__ ((always_inline, const))
+     uint8_t clip255 (int x)
+{
+  if (x > 255) {
+    return 255;
+  } else if (x < 0) {
+    return 0;
+  } else {
+    return x;
+  }
+}
+
+unsigned long CombJaggieThreshold = 73;
+
+#ifdef HAVE_CPU_I386
+static unsigned int
+comb_factor_packed422_scanline_mmx (uint8_t * top, uint8_t * mid,
+    uint8_t * bot, int width)
+{
+  const mmx_t qwYMask = { 0x00ff00ff00ff00ffULL };
+  const mmx_t qwOnes = { 0x0001000100010001ULL };
+  mmx_t qwThreshold;
+
+  unsigned int temp1, temp2;
+
+  width /= 4;
+
+  qwThreshold.uw[0] = CombJaggieThreshold;
+  qwThreshold.uw[1] = CombJaggieThreshold;
+  qwThreshold.uw[2] = CombJaggieThreshold;
+  qwThreshold.uw[3] = CombJaggieThreshold;
+
+  movq_m2r (qwThreshold, mm0);
+  movq_m2r (qwYMask, mm1);
+  movq_m2r (qwOnes, mm2);
+  pxor_r2r (mm7, mm7);          /* mm7 = 0. */
+
+  while (width--) {
+    /* Load and keep just the luma. */
+    movq_m2r (*top, mm3);
+    movq_m2r (*mid, mm4);
+    movq_m2r (*bot, mm5);
+
+    pand_r2r (mm1, mm3);
+    pand_r2r (mm1, mm4);
+    pand_r2r (mm1, mm5);
+
+    /* Work out mm6 = (top - mid) * (bot - mid) - ( (top - mid)^2 >> 7 ) */
+    psrlw_i2r (1, mm3);
+    psrlw_i2r (1, mm4);
+    psrlw_i2r (1, mm5);
+
+    /* mm6 = (top - mid) */
+    movq_r2r (mm3, mm6);
+    psubw_r2r (mm4, mm6);
+
+    /* mm3 = (top - bot) */
+    psubw_r2r (mm5, mm3);
+
+    /* mm5 = (bot - mid) */
+    psubw_r2r (mm4, mm5);
+
+    /* mm6 = (top - mid) * (bot - mid) */
+    pmullw_r2r (mm5, mm6);
+
+    /* mm3 = (top - bot)^2 >> 7 */
+    pmullw_r2r (mm3, mm3);      /* mm3 = (top - bot)^2 */
+    psrlw_i2r (7, mm3);         /* mm3 = ((top - bot)^2 >> 7) */
+
+    /* mm6 is what we want. */
+    psubw_r2r (mm3, mm6);
+
+    /* FF's if greater than qwTheshold */
+    pcmpgtw_r2r (mm0, mm6);
+
+    /* Add to count if we are greater than threshold */
+    pand_r2r (mm2, mm6);
+    paddw_r2r (mm6, mm7);
+
+    top += 8;
+    mid += 8;
+    bot += 8;
+  }
+
+  movd_r2m (mm7, temp1);
+  psrlq_i2r (32, mm7);
+  movd_r2m (mm7, temp2);
+  temp1 += temp2;
+  temp2 = temp1;
+  temp1 >>= 16;
+  temp1 += temp2 & 0xffff;
+
+  emms ();
+
+  return temp1;
+}
+#endif
+
+static unsigned long BitShift = 6;
+
+static unsigned int
+diff_factor_packed422_scanline_c (uint8_t * cur, uint8_t * old, int width)
+{
+  unsigned int ret = 0;
+
+  width /= 4;
+
+  while (width--) {
+    unsigned int tmp1 = (cur[0] + cur[2] + cur[4] + cur[6] + 2) >> 2;
+
+    unsigned int tmp2 = (old[0] + old[2] + old[4] + old[6] + 2) >> 2;
+
+    tmp1 = (tmp1 - tmp2);
+    tmp1 *= tmp1;
+    tmp1 >>= BitShift;
+    ret += tmp1;
+    cur += 8;
+    old += 8;
+  }
+
+  return ret;
+}
+
+/*
+static unsigned int diff_factor_packed422_scanline_test_c( uint8_t *cur, uint8_t *old, int width )
+{
+    unsigned int ret = 0;
+
+    width /= 16;
+
+    while( width-- ) {
+        unsigned int tmp1 = (cur[ 0 ] + cur[ 2 ] + cur[ 4 ] + cur[ 6 ])>>2;
+        unsigned int tmp2 = (old[ 0 ] + old[ 2 ] + old[ 4 ] + old[ 6 ])>>2;
+        tmp1  = (tmp1 - tmp2);
+        tmp1 *= tmp1;
+        tmp1 >>= BitShift;
+        ret += tmp1;
+        cur += (8*4);
+        old += (8*4);
+    }
+
+    return ret;
+}
+*/
+
+#ifdef HAVE_CPU_I386
+static unsigned int
+diff_factor_packed422_scanline_mmx (uint8_t * cur, uint8_t * old, int width)
+{
+  const mmx_t qwYMask = { 0x00ff00ff00ff00ffULL };
+  unsigned int temp1, temp2;
+
+  width /= 4;
+
+  movq_m2r (qwYMask, mm1);
+  movd_m2r (BitShift, mm7);
+  pxor_r2r (mm0, mm0);
+
+  while (width--) {
+    movq_m2r (*cur, mm4);
+    movq_m2r (*old, mm5);
+
+    pand_r2r (mm1, mm4);
+    pand_r2r (mm1, mm5);
+
+    psubw_r2r (mm5, mm4);       /* mm4 = Y1 - Y2            */
+    pmaddwd_r2r (mm4, mm4);     /* mm4 = (Y1 - Y2)^2        */
+    psrld_r2r (mm7, mm4);       /* divide mm4 by 2^BitShift */
+    paddd_r2r (mm4, mm0);       /* keep total in mm0        */
+
+    cur += 8;
+    old += 8;
+  }
+
+  movd_r2m (mm0, temp1);
+  psrlq_i2r (32, mm0);
+  movd_r2m (mm0, temp2);
+  temp1 += temp2;
+
+  emms ();
+
+  return temp1;
+}
+#endif
+
+// defined in glib/gmacros.h #define ABS(a) (((a) < 0)?-(a):(a))
+
+#ifdef HAVE_CPU_I386
+static void
+diff_packed422_block8x8_mmx (pulldown_metrics_t * m, uint8_t * old,
+    uint8_t * new, int os, int ns)
+{
+  const mmx_t ymask = { 0x00ff00ff00ff00ffULL };
+  short out[24];                /* Output buffer for the partial metrics from the mmx code. */
+
+  uint8_t *outdata = (uint8_t *) out;
+
+  uint8_t *oldp, *newp;
+
+  int i;
+
+  pxor_r2r (mm4, mm4);          // 4 even difference sums.
+  pxor_r2r (mm5, mm5);          // 4 odd difference sums.
+  pxor_r2r (mm7, mm7);          // zeros
+
+  oldp = old;
+  newp = new;
+  for (i = 4; i; --i) {
+    // Even difference.
+    movq_m2r (oldp[0], mm0);
+    movq_m2r (oldp[8], mm2);
+    pand_m2r (ymask, mm0);
+    pand_m2r (ymask, mm2);
+    oldp += os;
+
+    movq_m2r (newp[0], mm1);
+    movq_m2r (newp[8], mm3);
+    pand_m2r (ymask, mm1);
+    pand_m2r (ymask, mm3);
+    newp += ns;
+
+    movq_r2r (mm0, mm6);
+    psubusb_r2r (mm1, mm0);
+    psubusb_r2r (mm6, mm1);
+    movq_r2r (mm2, mm6);
+    psubusb_r2r (mm3, mm2);
+    psubusb_r2r (mm6, mm3);
+
+    paddw_r2r (mm0, mm4);
+    paddw_r2r (mm1, mm4);
+    paddw_r2r (mm2, mm4);
+    paddw_r2r (mm3, mm4);
+
+    // Odd difference.
+    movq_m2r (oldp[0], mm0);
+    movq_m2r (oldp[8], mm2);
+    pand_m2r (ymask, mm0);
+    pand_m2r (ymask, mm2);
+    oldp += os;
+
+    movq_m2r (newp[0], mm1);
+    movq_m2r (newp[8], mm3);
+    pand_m2r (ymask, mm1);
+    pand_m2r (ymask, mm3);
+    newp += ns;
+
+    movq_r2r (mm0, mm6);
+    psubusb_r2r (mm1, mm0);
+    psubusb_r2r (mm6, mm1);
+    movq_r2r (mm2, mm6);
+    psubusb_r2r (mm3, mm2);
+    psubusb_r2r (mm6, mm3);
+
+    paddw_r2r (mm0, mm5);
+    paddw_r2r (mm1, mm5);
+    paddw_r2r (mm2, mm5);
+    paddw_r2r (mm3, mm5);
+  }
+  movq_r2m (mm4, outdata[0]);
+  movq_r2m (mm5, outdata[8]);
+
+  m->e = out[0] + out[1] + out[2] + out[3];
+  m->o = out[4] + out[5] + out[6] + out[7];
+  m->d = m->e + m->o;
+
+  pxor_r2r (mm4, mm4);          // Past spacial noise.
+  pxor_r2r (mm5, mm5);          // Temporal noise.
+  pxor_r2r (mm6, mm6);          // Current spacial noise.
+
+  // First loop to measure first four columns
+  oldp = old;
+  newp = new;
+  for (i = 4; i; --i) {
+    movq_m2r (oldp[0], mm0);
+    movq_m2r (oldp[os], mm1);
+    pand_m2r (ymask, mm0);
+    pand_m2r (ymask, mm1);
+    oldp += (os * 2);
+
+    movq_m2r (newp[0], mm2);
+    movq_m2r (newp[ns], mm3);
+    pand_m2r (ymask, mm2);
+    pand_m2r (ymask, mm3);
+    newp += (ns * 2);
+
+    paddw_r2r (mm1, mm4);
+    paddw_r2r (mm1, mm5);
+    paddw_r2r (mm3, mm6);
+    psubw_r2r (mm0, mm4);
+    psubw_r2r (mm2, mm5);
+    psubw_r2r (mm2, mm6);
+  }
+  movq_r2m (mm4, outdata[0]);
+  movq_r2m (mm5, outdata[16]);
+  movq_r2m (mm6, outdata[32]);
+
+  pxor_r2r (mm4, mm4);
+  pxor_r2r (mm5, mm5);
+  pxor_r2r (mm6, mm6);
+
+  // Second loop for the last four columns
+  oldp = old;
+  newp = new;
+  for (i = 4; i; --i) {
+    movq_m2r (oldp[8], mm0);
+    movq_m2r (oldp[os + 8], mm1);
+    pand_m2r (ymask, mm0);
+    pand_m2r (ymask, mm1);
+    oldp += (os * 2);
+
+    movq_m2r (newp[8], mm2);
+    movq_m2r (newp[ns + 8], mm3);
+    pand_m2r (ymask, mm2);
+    pand_m2r (ymask, mm3);
+    newp += (ns * 2);
+
+    paddw_r2r (mm1, mm4);
+    paddw_r2r (mm1, mm5);
+    paddw_r2r (mm3, mm6);
+    psubw_r2r (mm0, mm4);
+    psubw_r2r (mm2, mm5);
+    psubw_r2r (mm2, mm6);
+  }
+  movq_r2m (mm4, outdata[8]);
+  movq_r2m (mm5, outdata[24]);
+  movq_r2m (mm6, outdata[40]);
+
+  m->p = m->t = m->s = 0;
+  for (i = 0; i < 8; i++) {
+    // FIXME: move abs() into the mmx code!
+    m->p += ABS (out[i]);
+    m->t += ABS (out[8 + i]);
+    m->s += ABS (out[16 + i]);
+  }
+
+  emms ();
+}
+#endif
+
+static void
+diff_packed422_block8x8_c (pulldown_metrics_t * m, uint8_t * old,
+    uint8_t * new, int os, int ns)
+{
+  int x, y, e = 0, o = 0, s = 0, p = 0, t = 0;
+
+  uint8_t *oldp, *newp;
+
+  m->s = m->p = m->t = 0;
+  for (x = 8; x; x--) {
+    oldp = old;
+    old += 2;
+    newp = new;
+    new += 2;
+    s = p = t = 0;
+    for (y = 4; y; y--) {
+      e += ABS (newp[0] - oldp[0]);
+      o += ABS (newp[ns] - oldp[os]);
+      s += newp[ns] - newp[0];
+      p += oldp[os] - oldp[0];
+      t += oldp[os] - newp[0];
+      oldp += os << 1;
+      newp += ns << 1;
+    }
+    m->s += ABS (s);
+    m->p += ABS (p);
+    m->t += ABS (t);
+  }
+  m->e = e;
+  m->o = o;
+  m->d = e + o;
+}
+
+static void
+packed444_to_packed422_scanline_c (uint8_t * output, uint8_t * input, int width)
+{
+  width /= 2;
+  while (width--) {
+    output[0] = input[0];
+    output[1] = input[1];
+    output[2] = input[3];
+    output[3] = input[2];
+    output += 4;
+    input += 6;
+  }
+}
+
+static void
+packed422_to_packed444_scanline_c (uint8_t * output, uint8_t * input, int width)
+{
+  width /= 2;
+  while (width--) {
+    output[0] = input[0];
+    output[1] = input[1];
+    output[2] = input[3];
+    output[3] = input[2];
+    output[4] = input[1];
+    output[5] = input[3];
+    output += 6;
+    input += 4;
+  }
+}
+
+/**
+ * For the middle pixels, the filter kernel is:
+ *
+ * [-1 3 -6 12 -24 80 80 -24 12 -6 3 -1]
+ */
+static void
+packed422_to_packed444_rec601_scanline_c (uint8_t * dest, uint8_t * src,
+    int width)
+{
+  int i;
+
+  /* Process two input pixels at a time.  Input is [Y'][Cb][Y'][Cr]. */
+  for (i = 0; i < width / 2; i++) {
+    dest[(i * 6) + 0] = src[(i * 4) + 0];
+    dest[(i * 6) + 1] = src[(i * 4) + 1];
+    dest[(i * 6) + 2] = src[(i * 4) + 3];
+
+    dest[(i * 6) + 3] = src[(i * 4) + 2];
+    if (i > (5 * 2) && i < ((width / 2) - (6 * 2))) {
+      dest[(i * 6) + 4] =
+          clip255 ((((80 * (src[(i * 4) + 1] + src[(i * 4) + 5]))
+                  - (24 * (src[(i * 4) - 3] + src[(i * 4) + 9]))
+                  + (12 * (src[(i * 4) - 7] + src[(i * 4) + 13]))
+                  - (6 * (src[(i * 4) - 11] + src[(i * 4) + 17]))
+                  + (3 * (src[(i * 4) - 15] + src[(i * 4) + 21]))
+                  - ((src[(i * 4) - 19] + src[(i * 4) + 25]))) + 64) >> 7);
+      dest[(i * 6) + 5] =
+          clip255 ((((80 * (src[(i * 4) + 3] + src[(i * 4) + 7]))
+                  - (24 * (src[(i * 4) - 1] + src[(i * 4) + 11]))
+                  + (12 * (src[(i * 4) - 5] + src[(i * 4) + 15]))
+                  - (6 * (src[(i * 4) - 9] + src[(i * 4) + 19]))
+                  + (3 * (src[(i * 4) - 13] + src[(i * 4) + 23]))
+                  - ((src[(i * 4) - 17] + src[(i * 4) + 27]))) + 64) >> 7);
+    } else if (i < ((width / 2) - 1)) {
+      dest[(i * 6) + 4] = (src[(i * 4) + 1] + src[(i * 4) + 5] + 1) >> 1;
+      dest[(i * 6) + 5] = (src[(i * 4) + 3] + src[(i * 4) + 7] + 1) >> 1;
+    } else {
+      dest[(i * 6) + 4] = src[(i * 4) + 1];
+      dest[(i * 6) + 5] = src[(i * 4) + 3];
+    }
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+vfilter_chroma_121_packed422_scanline_mmx (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b)
+{
+  int i;
+  const mmx_t ymask = { 0x00ff00ff00ff00ffULL };
+  const mmx_t cmask = { 0xff00ff00ff00ff00ULL };
+
+  // Get width in bytes.
+  width *= 2;
+  i = width / 8;
+  width -= i * 8;
+
+  movq_m2r (ymask, mm7);
+  movq_m2r (cmask, mm6);
+
+  while (i--) {
+    movq_m2r (*t, mm0);
+    movq_m2r (*b, mm1);
+    movq_m2r (*m, mm2);
+
+    movq_r2r (mm2, mm3);
+    pand_r2r (mm7, mm3);
+
+    pand_r2r (mm6, mm0);
+    pand_r2r (mm6, mm1);
+    pand_r2r (mm6, mm2);
+
+    psrlq_i2r (8, mm0);
+    psrlq_i2r (8, mm1);
+    psrlq_i2r (7, mm2);
+
+    paddw_r2r (mm0, mm2);
+    paddw_r2r (mm1, mm2);
+
+    psllw_i2r (6, mm2);
+    pand_r2r (mm6, mm2);
+
+    por_r2r (mm3, mm2);
+
+    movq_r2m (mm2, *output);
+    output += 8;
+    t += 8;
+    b += 8;
+    m += 8;
+  }
+  output++;
+  t++;
+  b++;
+  m++;
+  while (width--) {
+    *output = (*t + *b + (*m << 1)) >> 2;
+    output += 2;
+    t += 2;
+    b += 2;
+    m += 2;
+  }
+
+  emms ();
+}
+#endif
+
+static void
+vfilter_chroma_121_packed422_scanline_c (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b)
+{
+  output++;
+  t++;
+  b++;
+  m++;
+  while (width--) {
+    *output = (*t + *b + (*m << 1)) >> 2;
+    output += 2;
+    t += 2;
+    b += 2;
+    m += 2;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+vfilter_chroma_332_packed422_scanline_mmx (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b)
+{
+  int i;
+  const mmx_t ymask = { 0x00ff00ff00ff00ffULL };
+  const mmx_t cmask = { 0xff00ff00ff00ff00ULL };
+
+  // Get width in bytes.
+  width *= 2;
+  i = width / 8;
+  width -= i * 8;
+
+  movq_m2r (ymask, mm7);
+  movq_m2r (cmask, mm6);
+
+  while (i--) {
+    movq_m2r (*t, mm0);
+    movq_m2r (*b, mm1);
+    movq_m2r (*m, mm2);
+
+    movq_r2r (mm2, mm3);
+    pand_r2r (mm7, mm3);
+
+    pand_r2r (mm6, mm0);
+    pand_r2r (mm6, mm1);
+    pand_r2r (mm6, mm2);
+
+    psrlq_i2r (8, mm0);
+    psrlq_i2r (7, mm1);
+    psrlq_i2r (8, mm2);
+
+    movq_r2r (mm0, mm4);
+    psllw_i2r (1, mm4);
+    paddw_r2r (mm4, mm0);
+
+    movq_r2r (mm2, mm4);
+    psllw_i2r (1, mm4);
+    paddw_r2r (mm4, mm2);
+
+    paddw_r2r (mm0, mm2);
+    paddw_r2r (mm1, mm2);
+
+    psllw_i2r (5, mm2);
+    pand_r2r (mm6, mm2);
+
+    por_r2r (mm3, mm2);
+
+    movq_r2m (mm2, *output);
+    output += 8;
+    t += 8;
+    b += 8;
+    m += 8;
+  }
+  output++;
+  t++;
+  b++;
+  m++;
+  while (width--) {
+    *output = (3 * *t + 3 * *m + 2 * *b) >> 3;
+    output += 2;
+    t += 2;
+    b += 2;
+    m += 2;
+  }
+
+  emms ();
+}
+#endif
+
+static void
+vfilter_chroma_332_packed422_scanline_c (uint8_t * output, int width,
+    uint8_t * m, uint8_t * t, uint8_t * b)
+{
+  output++;
+  t++;
+  b++;
+  m++;
+  while (width--) {
+    *output = (3 * *t + 3 * *m + 2 * *b) >> 3;
+    output += 2;
+    t += 2;
+    b += 2;
+    m += 2;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+kill_chroma_packed422_inplace_scanline_mmx (uint8_t * data, int width)
+{
+  const mmx_t ymask = { 0x00ff00ff00ff00ffULL };
+  const mmx_t nullchroma = { 0x8000800080008000ULL };
+
+  movq_m2r (ymask, mm7);
+  movq_m2r (nullchroma, mm6);
+  for (; width > 4; width -= 4) {
+    movq_m2r (*data, mm0);
+    pand_r2r (mm7, mm0);
+    paddb_r2r (mm6, mm0);
+    movq_r2m (mm0, *data);
+    data += 8;
+  }
+  emms ();
+
+  while (width--) {
+    data[1] = 128;
+    data += 2;
+  }
+}
+#endif
+
+static void
+kill_chroma_packed422_inplace_scanline_c (uint8_t * data, int width)
+{
+  while (width--) {
+    data[1] = 128;
+    data += 2;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+invert_colour_packed422_inplace_scanline_mmx (uint8_t * data, int width)
+{
+  const mmx_t allones = { 0xffffffffffffffffULL };
+
+  movq_m2r (allones, mm1);
+  for (; width > 4; width -= 4) {
+    movq_r2r (mm1, mm2);
+    movq_m2r (*data, mm0);
+    psubb_r2r (mm0, mm2);
+    movq_r2m (mm2, *data);
+    data += 8;
+  }
+  emms ();
+
+  width *= 2;
+  while (width--) {
+    *data = 255 - *data;
+    data++;
+  }
+}
+#endif
+
+static void
+invert_colour_packed422_inplace_scanline_c (uint8_t * data, int width)
+{
+  width *= 2;
+  while (width--) {
+    *data = 255 - *data;
+    data++;
+  }
+}
+
+static void
+mirror_packed422_inplace_scanline_c (uint8_t * data, int width)
+{
+  int x, tmp1, tmp2;
+
+  int width2 = width * 2;
+
+  for (x = 0; x < width; x += 2) {
+    tmp1 = data[x];
+    tmp2 = data[x + 1];
+    data[x] = data[width2 - x];
+    data[x + 1] = data[width2 - x + 1];
+    data[width2 - x] = tmp1;
+    data[width2 - x + 1] = tmp2;
+  }
+}
+
+static void
+interpolate_packed422_scanline_c (uint8_t * output, uint8_t * top,
+    uint8_t * bot, int width)
+{
+  int i;
+
+  for (i = width * 2; i; --i) {
+    *output++ = ((*top++) + (*bot++)) >> 1;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+convert_uyvy_to_yuyv_scanline_mmx (uint8_t * uyvy_buf, uint8_t * yuyv_buf,
+    int width)
+{
+#if defined(HAVE_CPU_I386) && !defined(HAVE_CPU_X86_64)
+  __asm__ __volatile__ ("   movl      %0, %%esi         \n"
+      "   movl      %1, %%edi         \n"
+      "   movl      %2, %%edx         \n" "   shrl      $3, %%edx         \n"
+      /* Process 8 pixels at once */
+      "1: movq      (%%esi), %%mm0    \n"       /* mm0 = Y3V2Y2U2Y1V0Y0U0 */
+      "   movq      8(%%esi), %%mm2   \n"       /* mm2 = Y7V6Y6U6Y5V4Y4U4 */
+      "   movq      %%mm0, %%mm1      \n"       /* mm1 = Y3V2Y2U2Y1V0Y0U0 */
+      "   movq      %%mm2, %%mm3      \n"       /* mm3 = Y7V6Y6U6Y5V4Y4U4 */
+      "   psllw     $8, %%mm0         \n"       /* mm0 = V2__U2__V0__U0__ */
+      "   psrlw     $8, %%mm1         \n"       /* mm1 = __Y3__Y2__Y1__Y0 */
+      "   psllw     $8, %%mm2         \n"       /* mm2 = V6__U6__V4__U4__ */
+      "   psrlw     $8, %%mm3         \n"       /* mm3 = __Y7__Y6__Y5__Y4 */
+      "   por       %%mm1, %%mm0      \n"       /* mm0 = V2Y3U2Y2V0Y1U0Y0 */
+      "   por       %%mm3, %%mm2      \n"       /* mm2 = V6Y7U6Y6V4Y5U4Y4 */
+      "   movq      %%mm0, (%%edi)    \n"
+      "   movq      %%mm2, 8(%%edi)   \n"
+      "   addl      $16, %%esi        \n"
+      "   addl      $16, %%edi        \n"
+      "   decl      %%edx             \n"
+      "   jnz       1b                \n" "   emms                        \n"
+      /* output */ :
+      /* input */ :"g" (uyvy_buf), "g" (yuyv_buf), "g" (width)
+      /* clobber registers */
+      :"cc", "edx", "esi", "edi");
+#endif
+#ifdef HAVE_CPU_X86_64
+  __asm__ __volatile__ ("   movq      %0, %%rsi         \n"
+      "   movq      %1, %%rdi         \n"
+      "   xorq      %%rdx, %%rdx      \n"
+      "   movl      %2, %%edx         \n" "   shrq      $3, %%rdx         \n"
+      /* Process 8 pixels at once */
+      "1: movq      (%%rsi), %%mm0    \n"       /* mm0 = Y3V2Y2U2Y1V0Y0U0 */
+      "   movq      8(%%rsi), %%mm2   \n"       /* mm2 = Y7V6Y6U6Y5V4Y4U4 */
+      "   movq      %%mm0, %%mm1      \n"       /* mm1 = Y3V2Y2U2Y1V0Y0U0 */
+      "   movq      %%mm2, %%mm3      \n"       /* mm3 = Y7V6Y6U6Y5V4Y4U4 */
+      "   psllw     $8, %%mm0         \n"       /* mm0 = V2__U2__V0__U0__ */
+      "   psrlw     $8, %%mm1         \n"       /* mm1 = __Y3__Y2__Y1__Y0 */
+      "   psllw     $8, %%mm2         \n"       /* mm2 = V6__U6__V4__U4__ */
+      "   psrlw     $8, %%mm3         \n"       /* mm3 = __Y7__Y6__Y5__Y4 */
+      "   por       %%mm1, %%mm0      \n"       /* mm0 = V2Y3U2Y2V0Y1U0Y0 */
+      "   por       %%mm3, %%mm2      \n"       /* mm2 = V6Y7U6Y6V4Y5U4Y4 */
+      "   movq      %%mm0, (%%rdi)    \n"
+      "   movq      %%mm2, 8(%%rdi)   \n"
+      "   addq      $16, %%rsi        \n"
+      "   addq      $16, %%rdi        \n"
+      "   decq      %%rdx             \n"
+      "   jnz       1b                \n" "   emms                        \n"
+      /* output */ :
+      /* input */ :"g" (uyvy_buf), "g" (yuyv_buf), "g" (width)
+      /* clobber registers */
+      :"cc", "rdx", "rsi", "rdi");
+#endif
+  if (width & 7) {
+    uint32_t *uyvy = (uint32_t *) uyvy_buf;
+
+    uint32_t *yuyv = (uint32_t *) yuyv_buf;
+
+    uint32_t val;
+
+    width &= 7;
+    width >>= 1;
+    while (width--) {
+      val = *uyvy++;
+      val = ((val << 8) & ~0x00FF0000) | ((val >> 8) & ~0x0000FF00);
+      *yuyv++ = val;
+    }
+  }
+}
+#endif
+
+static void
+convert_uyvy_to_yuyv_scanline_c (uint8_t * uyvy_buf, uint8_t * yuyv_buf,
+    int width)
+{
+  uint32_t *uyvy = (uint32_t *) uyvy_buf;
+
+  uint32_t *yuyv = (uint32_t *) yuyv_buf;
+
+  uint32_t val;
+
+  width >>= 1;
+  while (width--) {
+    val = *uyvy++;
+    val = ((val << 8) & ~0x00FF0000) | ((val >> 8) & ~0x0000FF00);
+    *yuyv++ = val;
+  }
+}
+
+
+#ifdef HAVE_CPU_I386
+static void
+interpolate_packed422_scanline_mmx (uint8_t * output, uint8_t * top,
+    uint8_t * bot, int width)
+{
+  const mmx_t shiftmask = { 0xfefffefffefffeffULL };    /* To avoid shifting chroma to luma. */
+  int i;
+
+  for (i = width / 16; i; --i) {
+    movq_m2r (*bot, mm0);
+    movq_m2r (*top, mm1);
+    movq_m2r (*(bot + 8), mm2);
+    movq_m2r (*(top + 8), mm3);
+    movq_m2r (*(bot + 16), mm4);
+    movq_m2r (*(top + 16), mm5);
+    movq_m2r (*(bot + 24), mm6);
+    movq_m2r (*(top + 24), mm7);
+    pand_m2r (shiftmask, mm0);
+    pand_m2r (shiftmask, mm1);
+    pand_m2r (shiftmask, mm2);
+    pand_m2r (shiftmask, mm3);
+    pand_m2r (shiftmask, mm4);
+    pand_m2r (shiftmask, mm5);
+    pand_m2r (shiftmask, mm6);
+    pand_m2r (shiftmask, mm7);
+    psrlw_i2r (1, mm0);
+    psrlw_i2r (1, mm1);
+    psrlw_i2r (1, mm2);
+    psrlw_i2r (1, mm3);
+    psrlw_i2r (1, mm4);
+    psrlw_i2r (1, mm5);
+    psrlw_i2r (1, mm6);
+    psrlw_i2r (1, mm7);
+    paddb_r2r (mm1, mm0);
+    paddb_r2r (mm3, mm2);
+    paddb_r2r (mm5, mm4);
+    paddb_r2r (mm7, mm6);
+    movq_r2m (mm0, *output);
+    movq_r2m (mm2, *(output + 8));
+    movq_r2m (mm4, *(output + 16));
+    movq_r2m (mm6, *(output + 24));
+    output += 32;
+    top += 32;
+    bot += 32;
+  }
+  width = (width & 0xf);
+
+  for (i = width / 4; i; --i) {
+    movq_m2r (*bot, mm0);
+    movq_m2r (*top, mm1);
+    pand_m2r (shiftmask, mm0);
+    pand_m2r (shiftmask, mm1);
+    psrlw_i2r (1, mm0);
+    psrlw_i2r (1, mm1);
+    paddb_r2r (mm1, mm0);
+    movq_r2m (mm0, *output);
+    output += 8;
+    top += 8;
+    bot += 8;
+  }
+  width = width & 0x7;
+
+  /* Handle last few pixels. */
+  for (i = width * 2; i; --i) {
+    *output++ = ((*top++) + (*bot++)) >> 1;
+  }
+
+  emms ();
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+interpolate_packed422_scanline_mmxext (uint8_t * output, uint8_t * top,
+    uint8_t * bot, int width)
+{
+  int i;
+
+  for (i = width / 16; i; --i) {
+    movq_m2r (*bot, mm0);
+    movq_m2r (*top, mm1);
+    movq_m2r (*(bot + 8), mm2);
+    movq_m2r (*(top + 8), mm3);
+    movq_m2r (*(bot + 16), mm4);
+    movq_m2r (*(top + 16), mm5);
+    movq_m2r (*(bot + 24), mm6);
+    movq_m2r (*(top + 24), mm7);
+    pavgb_r2r (mm1, mm0);
+    pavgb_r2r (mm3, mm2);
+    pavgb_r2r (mm5, mm4);
+    pavgb_r2r (mm7, mm6);
+    movntq_r2m (mm0, *output);
+    movntq_r2m (mm2, *(output + 8));
+    movntq_r2m (mm4, *(output + 16));
+    movntq_r2m (mm6, *(output + 24));
+    output += 32;
+    top += 32;
+    bot += 32;
+  }
+  width = (width & 0xf);
+
+  for (i = width / 4; i; --i) {
+    movq_m2r (*bot, mm0);
+    movq_m2r (*top, mm1);
+    pavgb_r2r (mm1, mm0);
+    movntq_r2m (mm0, *output);
+    output += 8;
+    top += 8;
+    bot += 8;
+  }
+  width = width & 0x7;
+
+  /* Handle last few pixels. */
+  for (i = width * 2; i; --i) {
+    *output++ = ((*top++) + (*bot++)) >> 1;
+  }
+
+  sfence ();
+  emms ();
+}
+#endif
+
+static void
+blit_colour_packed422_scanline_c (uint8_t * output, int width, int y, int cb,
+    int cr)
+{
+  uint32_t colour = cr << 24 | y << 16 | cb << 8 | y;
+
+  uint32_t *o = (uint32_t *) output;
+
+  for (width /= 2; width; --width) {
+    *o++ = colour;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+blit_colour_packed422_scanline_mmx (uint8_t * output, int width, int y, int cb,
+    int cr)
+{
+  uint32_t colour = cr << 24 | y << 16 | cb << 8 | y;
+
+  int i;
+
+  movd_m2r (colour, mm1);
+  movd_m2r (colour, mm2);
+  psllq_i2r (32, mm1);
+  por_r2r (mm1, mm2);
+
+  for (i = width / 16; i; --i) {
+    movq_r2m (mm2, *output);
+    movq_r2m (mm2, *(output + 8));
+    movq_r2m (mm2, *(output + 16));
+    movq_r2m (mm2, *(output + 24));
+    output += 32;
+  }
+  width = (width & 0xf);
+
+  for (i = width / 4; i; --i) {
+    movq_r2m (mm2, *output);
+    output += 8;
+  }
+  width = (width & 0x7);
+
+  for (i = width / 2; i; --i) {
+    *((uint32_t *) output) = colour;
+    output += 4;
+  }
+
+  if (width & 1) {
+    *output = y;
+    *(output + 1) = cb;
+  }
+
+  emms ();
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+blit_colour_packed422_scanline_mmxext (uint8_t * output, int width, int y,
+    int cb, int cr)
+{
+  uint32_t colour = cr << 24 | y << 16 | cb << 8 | y;
+
+  int i;
+
+  movd_m2r (colour, mm1);
+  movd_m2r (colour, mm2);
+  psllq_i2r (32, mm1);
+  por_r2r (mm1, mm2);
+
+  for (i = width / 16; i; --i) {
+    movntq_r2m (mm2, *output);
+    movntq_r2m (mm2, *(output + 8));
+    movntq_r2m (mm2, *(output + 16));
+    movntq_r2m (mm2, *(output + 24));
+    output += 32;
+  }
+  width = (width & 0xf);
+
+  for (i = width / 4; i; --i) {
+    movntq_r2m (mm2, *output);
+    output += 8;
+  }
+  width = (width & 0x7);
+
+  for (i = width / 2; i; --i) {
+    *((uint32_t *) output) = colour;
+    output += 4;
+  }
+
+  if (width & 1) {
+    *output = y;
+    *(output + 1) = cb;
+  }
+
+  sfence ();
+  emms ();
+}
+#endif
+
+static void
+blit_colour_packed4444_scanline_c (uint8_t * output, int width,
+    int alpha, int luma, int cb, int cr)
+{
+  int j;
+
+  for (j = 0; j < width; j++) {
+    *output++ = alpha;
+    *output++ = luma;
+    *output++ = cb;
+    *output++ = cr;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+blit_colour_packed4444_scanline_mmx (uint8_t * output, int width,
+    int alpha, int luma, int cb, int cr)
+{
+  uint32_t colour = (cr << 24) | (cb << 16) | (luma << 8) | alpha;
+
+  int i;
+
+  movd_m2r (colour, mm1);
+  movd_m2r (colour, mm2);
+  psllq_i2r (32, mm1);
+  por_r2r (mm1, mm2);
+
+  for (i = width / 8; i; --i) {
+    movq_r2m (mm2, *output);
+    movq_r2m (mm2, *(output + 8));
+    movq_r2m (mm2, *(output + 16));
+    movq_r2m (mm2, *(output + 24));
+    output += 32;
+  }
+  width = (width & 0x7);
+
+  for (i = width / 2; i; --i) {
+    movq_r2m (mm2, *output);
+    output += 8;
+  }
+  width = (width & 0x1);
+
+  if (width) {
+    *((uint32_t *) output) = colour;
+    output += 4;
+  }
+
+  emms ();
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+blit_colour_packed4444_scanline_mmxext (uint8_t * output, int width,
+    int alpha, int luma, int cb, int cr)
+{
+  uint32_t colour = (cr << 24) | (cb << 16) | (luma << 8) | alpha;
+
+  int i;
+
+  movd_m2r (colour, mm1);
+  movd_m2r (colour, mm2);
+  psllq_i2r (32, mm1);
+  por_r2r (mm1, mm2);
+
+  for (i = width / 8; i; --i) {
+    movntq_r2m (mm2, *output);
+    movntq_r2m (mm2, *(output + 8));
+    movntq_r2m (mm2, *(output + 16));
+    movntq_r2m (mm2, *(output + 24));
+    output += 32;
+  }
+  width = (width & 0x7);
+
+  for (i = width / 2; i; --i) {
+    movntq_r2m (mm2, *output);
+    output += 8;
+  }
+  width = (width & 0x1);
+
+  if (width) {
+    *((uint32_t *) output) = colour;
+    output += 4;
+  }
+
+  sfence ();
+  emms ();
+}
+#endif
+
+
+/**
+ * Some memcpy code inspired by the xine code which originally came
+ * from mplayer.
+ */
+
+/* linux kernel __memcpy (from: /include/asm/string.h) */
+#ifdef HAVE_CPU_I386
+static inline __attribute__ ((always_inline, const))
+     void small_memcpy (void *to, const void *from, size_t n)
+{
+  int d0, d1, d2;
+
+  __asm__ __volatile__ ("rep ; movsl\n\t"
+      "testb $2,%b4\n\t"
+      "je 1f\n\t"
+      "movsw\n"
+      "1:\ttestb $1,%b4\n\t"
+      "je 2f\n\t" "movsb\n" "2:":"=&c" (d0), "=&D" (d1), "=&S" (d2)
+      :"0" (n / 4), "q" (n), "1" ((long) to), "2" ((long) from)
+      :"memory");
+}
+#endif
+
+static void
+speedy_memcpy_c (void *dest, const void *src, size_t n)
+{
+  if (dest != src) {
+    memcpy (dest, src, n);
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+speedy_memcpy_mmx (void *d, const void *s, size_t n)
+{
+  const uint8_t *src = s;
+
+  uint8_t *dest = d;
+
+  if (dest != src) {
+    while (n > 64) {
+      movq_m2r (src[0], mm0);
+      movq_m2r (src[8], mm1);
+      movq_m2r (src[16], mm2);
+      movq_m2r (src[24], mm3);
+      movq_m2r (src[32], mm4);
+      movq_m2r (src[40], mm5);
+      movq_m2r (src[48], mm6);
+      movq_m2r (src[56], mm7);
+      movq_r2m (mm0, dest[0]);
+      movq_r2m (mm1, dest[8]);
+      movq_r2m (mm2, dest[16]);
+      movq_r2m (mm3, dest[24]);
+      movq_r2m (mm4, dest[32]);
+      movq_r2m (mm5, dest[40]);
+      movq_r2m (mm6, dest[48]);
+      movq_r2m (mm7, dest[56]);
+      dest += 64;
+      src += 64;
+      n -= 64;
+    }
+
+    while (n > 8) {
+      movq_m2r (src[0], mm0);
+      movq_r2m (mm0, dest[0]);
+      dest += 8;
+      src += 8;
+      n -= 8;
+    }
+
+    if (n)
+      small_memcpy (dest, src, n);
+
+    emms ();
+  }
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+speedy_memcpy_mmxext (void *d, const void *s, size_t n)
+{
+  const uint8_t *src = s;
+
+  uint8_t *dest = d;
+
+  if (dest != src) {
+    while (n > 64) {
+      movq_m2r (src[0], mm0);
+      movq_m2r (src[8], mm1);
+      movq_m2r (src[16], mm2);
+      movq_m2r (src[24], mm3);
+      movq_m2r (src[32], mm4);
+      movq_m2r (src[40], mm5);
+      movq_m2r (src[48], mm6);
+      movq_m2r (src[56], mm7);
+      movntq_r2m (mm0, dest[0]);
+      movntq_r2m (mm1, dest[8]);
+      movntq_r2m (mm2, dest[16]);
+      movntq_r2m (mm3, dest[24]);
+      movntq_r2m (mm4, dest[32]);
+      movntq_r2m (mm5, dest[40]);
+      movntq_r2m (mm6, dest[48]);
+      movntq_r2m (mm7, dest[56]);
+      dest += 64;
+      src += 64;
+      n -= 64;
+    }
+
+    while (n > 8) {
+      movq_m2r (src[0], mm0);
+      movntq_r2m (mm0, dest[0]);
+      dest += 8;
+      src += 8;
+      n -= 8;
+    }
+
+    if (n)
+      small_memcpy (dest, src, n);
+
+    sfence ();
+    emms ();
+  }
+}
+#endif
+
+static void
+blit_packed422_scanline_c (uint8_t * dest, const uint8_t * src, int width)
+{
+  speedy_memcpy_c (dest, src, width * 2);
+}
+
+#ifdef HAVE_CPU_I386
+static void
+blit_packed422_scanline_mmx (uint8_t * dest, const uint8_t * src, int width)
+{
+  speedy_memcpy_mmx (dest, src, width * 2);
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+blit_packed422_scanline_mmxext (uint8_t * dest, const uint8_t * src, int width)
+{
+  speedy_memcpy_mmxext (dest, src, width * 2);
+}
+#endif
+
+static void
+composite_colour4444_alpha_to_packed422_scanline_c (uint8_t * output,
+    uint8_t * input, int af, int y, int cb, int cr, int width, int alpha)
+{
+  int a = ((af * alpha) + 0x80) >> 8;
+
+  if (a == 0xff) {
+    blit_colour_packed422_scanline (output, width, y, cb, cr);
+  } else if (a) {
+    int i;
+
+    for (i = 0; i < width; i++) {
+            /**
+             * (1 - alpha)*B + alpha*F
+             * (1 - af*a)*B + af*a*F
+             *  B - af*a*B + af*a*F
+             *  B + a*(af*F - af*B)
+             */
+
+      output[0] =
+          input[0] + ((alpha * (y - multiply_alpha (af,
+                      input[0])) + 0x80) >> 8);
+
+      if ((i & 1) == 0) {
+
+                    /**
+                     * At first I thought I was doing this incorrectly, but
+                     * the following math has convinced me otherwise.
+                     *
+                     * C_r = (1 - alpha)*B + alpha*F
+                     * C_r = B - af*a*B + af*a*F
+                     *
+                     * C_r = 128 + ((1 - af*a)*(B - 128) + a*af*(F - 128))
+                     * C_r = 128 + (B - af*a*B - 128 + af*a*128 + a*af*F - a*af*128)
+                     * C_r = B - af*a*B + a*af*F
+                     */
+
+        output[1] =
+            input[1] + ((alpha * (cb - multiply_alpha (af,
+                        input[1])) + 0x80) >> 8);
+        output[3] =
+            input[3] + ((alpha * (cr - multiply_alpha (af,
+                        input[3])) + 0x80) >> 8);
+      }
+      output += 2;
+      input += 2;
+    }
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+composite_colour4444_alpha_to_packed422_scanline_mmxext (uint8_t * output,
+    uint8_t * input, int af, int y, int cb, int cr, int width, int alpha)
+{
+  const mmx_t alpha2 = { 0x0000FFFF00000000ULL };
+  const mmx_t alpha1 = { 0xFFFF0000FFFFFFFFULL };
+  const mmx_t round = { 0x0080008000800080ULL };
+  mmx_t foreground;
+
+  int i;
+
+  if (!alpha) {
+    blit_packed422_scanline (output, input, width);
+    return;
+  }
+
+  foreground.ub[0] = foreground.ub[4] = af;
+  foreground.ub[1] = foreground.ub[5] = y;
+  foreground.ub[2] = foreground.ub[6] = cb;
+  foreground.ub[3] = foreground.ub[7] = cr;
+
+  movq_m2r (alpha, mm2);
+  pshufw_r2r (mm2, mm2, 0);
+  pxor_r2r (mm7, mm7);
+
+  for (i = width / 2; i; i--) {
+    /* mm1 = [ cr ][ y ][ cb ][ y ] */
+    movd_m2r (*input, mm1);
+    punpcklbw_r2r (mm7, mm1);
+
+    movq_m2r (foreground, mm3);
+    movq_r2r (mm3, mm4);
+    punpcklbw_r2r (mm7, mm3);
+    punpckhbw_r2r (mm7, mm4);
+    /* mm3 and mm4 will be the appropriate colours, mm5 and mm6 for alpha. */
+
+    /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0 a ][ 0 a ][ 0 a ][ 0 a ] */
+    pshufw_r2r (mm3, mm5, 0);
+    pshufw_r2r (mm4, mm6, 0);
+    /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 3 cr ][ 0 a ][ 2 cb ][ 1 y ]  == 11001000 == 201 */
+    pshufw_r2r (mm3, mm3, 201);
+    /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0 a ][ 1 y ][ 0 a ][ 0 a ]  == 00010000 == 16 */
+    pshufw_r2r (mm4, mm4, 16);
+
+    pand_m2r (alpha1, mm3);
+    pand_m2r (alpha2, mm4);
+    pand_m2r (alpha1, mm5);
+    pand_m2r (alpha2, mm6);
+    por_r2r (mm4, mm3);
+    por_r2r (mm6, mm5);
+
+    /* now, mm5 is af and mm1 is B.  Need to multiply them. */
+    pmullw_r2r (mm1, mm5);
+
+    /* Multiply by appalpha. */
+    pmullw_r2r (mm2, mm3);
+    paddw_m2r (round, mm3);
+    psrlw_i2r (8, mm3);
+    /* Result is now B + F. */
+    paddw_r2r (mm3, mm1);
+
+    /* Round up appropriately. */
+    paddw_m2r (round, mm5);
+
+    /* mm6 contains our i>>8; */
+    movq_r2r (mm5, mm6);
+    psrlw_i2r (8, mm6);
+
+    /* Add mm6 back into mm5.  Now our result is in the high bytes. */
+    paddw_r2r (mm6, mm5);
+
+    /* Shift down. */
+    psrlw_i2r (8, mm5);
+
+    /* Multiply by appalpha. */
+    pmullw_r2r (mm2, mm5);
+    paddw_m2r (round, mm5);
+    psrlw_i2r (8, mm5);
+
+    psubusw_r2r (mm5, mm1);
+
+    /* mm1 = [ B + F - af*B ] */
+    packuswb_r2r (mm1, mm1);
+    movd_r2m (mm1, *output);
+
+    output += 4;
+    input += 4;
+  }
+  sfence ();
+  emms ();
+}
+#endif
+
+
+static void
+composite_packed4444_alpha_to_packed422_scanline_c (uint8_t * output,
+    uint8_t * input, uint8_t * foreground, int width, int alpha)
+{
+  int i;
+
+  for (i = 0; i < width; i++) {
+    int af = foreground[0];
+
+    if (af) {
+      int a = ((af * alpha) + 0x80) >> 8;
+
+
+      if (a == 0xff) {
+        output[0] = foreground[1];
+
+        if ((i & 1) == 0) {
+          output[1] = foreground[2];
+          output[3] = foreground[3];
+        }
+      } else if (a) {
+                /**
+                 * (1 - alpha)*B + alpha*F
+                 * (1 - af*a)*B + af*a*F
+                 *  B - af*a*B + af*a*F
+                 *  B + a*(af*F - af*B)
+                 */
+
+        output[0] = input[0]
+            + ((alpha * (foreground[1]
+                    - multiply_alpha (foreground[0], input[0])) + 0x80) >> 8);
+
+        if ((i & 1) == 0) {
+
+                    /**
+                     * At first I thought I was doing this incorrectly, but
+                     * the following math has convinced me otherwise.
+                     *
+                     * C_r = (1 - alpha)*B + alpha*F
+                     * C_r = B - af*a*B + af*a*F
+                     *
+                     * C_r = 128 + ((1 - af*a)*(B - 128) + a*af*(F - 128))
+                     * C_r = 128 + (B - af*a*B - 128 + af*a*128 + a*af*F - a*af*128)
+                     * C_r = B - af*a*B + a*af*F
+                     */
+
+          output[1] = input[1] + ((alpha * (foreground[2]
+                      - multiply_alpha (foreground[0], input[1])) + 0x80) >> 8);
+          output[3] = input[3] + ((alpha * (foreground[3]
+                      - multiply_alpha (foreground[0], input[3])) + 0x80) >> 8);
+        }
+      }
+    }
+    foreground += 4;
+    output += 2;
+    input += 2;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+composite_packed4444_alpha_to_packed422_scanline_mmxext (uint8_t * output,
+    uint8_t * input, uint8_t * foreground, int width, int alpha)
+{
+  const mmx_t alpha2 = { 0x0000FFFF00000000ULL };
+  const mmx_t alpha1 = { 0xFFFF0000FFFFFFFFULL };
+  const mmx_t round = { 0x0080008000800080ULL };
+  int i;
+
+  if (!alpha) {
+    blit_packed422_scanline (output, input, width);
+    return;
+  }
+
+  if (alpha == 256) {
+    composite_packed4444_to_packed422_scanline (output, input, foreground,
+        width);
+    return;
+  }
+
+  READ_PREFETCH_2048 (input);
+  READ_PREFETCH_2048 (foreground);
+
+  movq_m2r (alpha, mm2);
+  pshufw_r2r (mm2, mm2, 0);
+  pxor_r2r (mm7, mm7);
+
+  for (i = width / 2; i; i--) {
+    int fg1 = *((uint32_t *) foreground);
+
+    int fg2 = *(((uint32_t *) foreground) + 1);
+
+    if (fg1 || fg2) {
+      /* mm1 = [ cr ][ y ][ cb ][ y ] */
+      movd_m2r (*input, mm1);
+      punpcklbw_r2r (mm7, mm1);
+
+      movq_m2r (*foreground, mm3);
+      movq_r2r (mm3, mm4);
+      punpcklbw_r2r (mm7, mm3);
+      punpckhbw_r2r (mm7, mm4);
+      /* mm3 and mm4 will be the appropriate colours, mm5 and mm6 for alpha. */
+
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0 a ][ 0 a ][ 0 a ][ 0 a ] */
+      pshufw_r2r (mm3, mm5, 0);
+      pshufw_r2r (mm4, mm6, 0);
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 3 cr ][ 0 a ][ 2 cb ][ 1 y ]  == 11001000 == 201 */
+      pshufw_r2r (mm3, mm3, 201);
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0 a ][ 1 y ][ 0 a ][ 0 a ]  == 00010000 == 16 */
+      pshufw_r2r (mm4, mm4, 16);
+
+      pand_m2r (alpha1, mm3);
+      pand_m2r (alpha2, mm4);
+      pand_m2r (alpha1, mm5);
+      pand_m2r (alpha2, mm6);
+      por_r2r (mm4, mm3);
+      por_r2r (mm6, mm5);
+
+      /* now, mm5 is af and mm1 is B.  Need to multiply them. */
+      pmullw_r2r (mm1, mm5);
+
+      /* Multiply by appalpha. */
+      pmullw_r2r (mm2, mm3);
+      paddw_m2r (round, mm3);
+      psrlw_i2r (8, mm3);
+      /* Result is now B + F. */
+      paddw_r2r (mm3, mm1);
+
+      /* Round up appropriately. */
+      paddw_m2r (round, mm5);
+
+      /* mm6 contains our i>>8; */
+      movq_r2r (mm5, mm6);
+      psrlw_i2r (8, mm6);
+
+      /* Add mm6 back into mm5.  Now our result is in the high bytes. */
+      paddw_r2r (mm6, mm5);
+
+      /* Shift down. */
+      psrlw_i2r (8, mm5);
+
+      /* Multiply by appalpha. */
+      pmullw_r2r (mm2, mm5);
+      paddw_m2r (round, mm5);
+      psrlw_i2r (8, mm5);
+
+      psubusw_r2r (mm5, mm1);
+
+      /* mm1 = [ B + F - af*B ] */
+      packuswb_r2r (mm1, mm1);
+      movd_r2m (mm1, *output);
+    }
+
+    foreground += 8;
+    output += 4;
+    input += 4;
+  }
+  sfence ();
+  emms ();
+}
+#endif
+
+static void
+composite_packed4444_to_packed422_scanline_c (uint8_t * output, uint8_t * input,
+    uint8_t * foreground, int width)
+{
+  int i;
+
+  for (i = 0; i < width; i++) {
+    int a = foreground[0];
+
+    if (a == 0xff) {
+      output[0] = foreground[1];
+
+      if ((i & 1) == 0) {
+        output[1] = foreground[2];
+        output[3] = foreground[3];
+      }
+    } else if (a) {
+            /**
+             * (1 - alpha)*B + alpha*F
+             *  B + af*F - af*B
+             */
+
+      output[0] =
+          input[0] + foreground[1] - multiply_alpha (foreground[0], input[0]);
+
+      if ((i & 1) == 0) {
+
+                /**
+                 * C_r = (1 - af)*B + af*F
+                 * C_r = B - af*B + af*F
+                 */
+
+        output[1] =
+            input[1] + foreground[2] - multiply_alpha (foreground[0], input[1]);
+        output[3] =
+            input[3] + foreground[3] - multiply_alpha (foreground[0], input[3]);
+      }
+    }
+    foreground += 4;
+    output += 2;
+    input += 2;
+  }
+}
+
+
+#ifdef HAVE_CPU_I386
+static void
+composite_packed4444_to_packed422_scanline_mmxext (uint8_t * output,
+    uint8_t * input, uint8_t * foreground, int width)
+{
+  const mmx_t alpha2 = { 0x0000FFFF00000000ULL };
+  const mmx_t alpha1 = { 0xFFFF0000FFFFFFFFULL };
+  const mmx_t round = { 0x0080008000800080ULL };
+  int i;
+
+  READ_PREFETCH_2048 (input);
+  READ_PREFETCH_2048 (foreground);
+
+  pxor_r2r (mm7, mm7);
+  for (i = width / 2; i; i--) {
+    int fg1 = *((uint32_t *) foreground);
+
+    int fg2 = *(((uint32_t *) foreground) + 1);
+
+    if ((fg1 & 0xff) == 0xff && (fg2 & 0xff) == 0xff) {
+      movq_m2r (*foreground, mm3);
+      movq_r2r (mm3, mm4);
+      punpcklbw_r2r (mm7, mm3);
+      punpckhbw_r2r (mm7, mm4);
+      /* mm3 and mm4 will be the appropriate colours, mm5 and mm6 for alpha. */
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 3 cr ][ 0 a ][ 2 cb ][ 1 y ]  == 11001000 == 201 */
+      pshufw_r2r (mm3, mm3, 201);
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0  a ][ 1 y ][ 0  a ][ 0 a ]  == 00010000 == 16 */
+      pshufw_r2r (mm4, mm4, 16);
+      pand_m2r (alpha1, mm3);
+      pand_m2r (alpha2, mm4);
+      por_r2r (mm4, mm3);
+      /* mm1 = [ B + F - af*B ] */
+      packuswb_r2r (mm3, mm3);
+      movd_r2m (mm3, *output);
+    } else if (fg1 || fg2) {
+
+      /* mm1 = [ cr ][ y ][ cb ][ y ] */
+      movd_m2r (*input, mm1);
+      punpcklbw_r2r (mm7, mm1);
+
+      movq_m2r (*foreground, mm3);
+      movq_r2r (mm3, mm4);
+      punpcklbw_r2r (mm7, mm3);
+      punpckhbw_r2r (mm7, mm4);
+      /* mm3 and mm4 will be the appropriate colours, mm5 and mm6 for alpha. */
+
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0 a ][ 0 a ][ 0 a ][ 0 a ] */
+      pshufw_r2r (mm3, mm5, 0);
+      pshufw_r2r (mm4, mm6, 0);
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 3 cr ][ 0 a ][ 2 cb ][ 1 y ]  == 11001000 == 201 */
+      pshufw_r2r (mm3, mm3, 201);
+      /* [ 3 cr ][ 2 cb ][ 1 y ][ 0 a ]  -> [ 0  a ][ 1 y ][ 0  a ][ 0 a ]  == 00010000 == 16 */
+      pshufw_r2r (mm4, mm4, 16);
+
+      pand_m2r (alpha1, mm3);
+      pand_m2r (alpha2, mm4);
+      pand_m2r (alpha1, mm5);
+      pand_m2r (alpha2, mm6);
+      por_r2r (mm4, mm3);
+      por_r2r (mm6, mm5);
+
+      /* now, mm5 is af and mm1 is B.  Need to multiply them. */
+      pmullw_r2r (mm1, mm5);
+
+      /* Result is now B + F. */
+      paddw_r2r (mm3, mm1);
+
+      /* Round up appropriately. */
+      paddw_m2r (round, mm5);
+
+      /* mm6 contains our i>>8; */
+      movq_r2r (mm5, mm6);
+      psrlw_i2r (8, mm6);
+
+      /* Add mm6 back into mm5.  Now our result is in the high bytes. */
+      paddw_r2r (mm6, mm5);
+
+      /* Shift down. */
+      psrlw_i2r (8, mm5);
+
+      psubusw_r2r (mm5, mm1);
+
+      /* mm1 = [ B + F - af*B ] */
+      packuswb_r2r (mm1, mm1);
+      movd_r2m (mm1, *output);
+    }
+
+    foreground += 8;
+    output += 4;
+    input += 4;
+  }
+  sfence ();
+  emms ();
+}
+#endif
+
+/**
+ * um... just need some scrap paper...
+ *   D = (1 - alpha)*B + alpha*F
+ *   D = (1 - a)*B + a*textluma
+ *     = B - a*B + a*textluma
+ *     = B + a*(textluma - B)
+ *   Da = (1 - a)*b + a
+ */
+static void
+composite_alphamask_to_packed4444_scanline_c (uint8_t * output,
+    uint8_t * input,
+    uint8_t * mask, int width, int textluma, int textcb, int textcr)
+{
+  uint32_t opaque = (textcr << 24) | (textcb << 16) | (textluma << 8) | 0xff;
+
+  int i;
+
+  for (i = 0; i < width; i++) {
+    int a = *mask;
+
+    if (a == 0xff) {
+      *((uint32_t *) output) = opaque;
+    } else if ((input[0] == 0x00)) {
+      *((uint32_t *) output) = (multiply_alpha (a, textcr) << 24)
+          | (multiply_alpha (a, textcb) << 16)
+          | (multiply_alpha (a, textluma) << 8) | a;
+    } else if (a) {
+      *((uint32_t *) output) =
+          ((input[3] + multiply_alpha (a, textcr - input[3])) << 24)
+          | ((input[2] + multiply_alpha (a, textcb - input[2])) << 16)
+          | ((input[1] + multiply_alpha (a, textluma - input[1])) << 8)
+          | (input[0] + multiply_alpha (a, 0xff - input[0]));
+    }
+    mask++;
+    output += 4;
+    input += 4;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+composite_alphamask_to_packed4444_scanline_mmxext (uint8_t * output,
+    uint8_t * input,
+    uint8_t * mask, int width, int textluma, int textcb, int textcr)
+{
+  uint32_t opaque = (textcr << 24) | (textcb << 16) | (textluma << 8) | 0xff;
+  const mmx_t round = { 0x0080008000800080ULL };
+  const mmx_t fullalpha = { 0x00000000000000ffULL };
+  mmx_t colour;
+
+  colour.w[0] = 0x00;
+  colour.w[1] = textluma;
+  colour.w[2] = textcb;
+  colour.w[3] = textcr;
+
+  movq_m2r (colour, mm1);
+  movq_r2r (mm1, mm0);
+
+  /* mm0 = [ cr ][ cb ][ y ][ 0xff ] */
+  paddw_m2r (fullalpha, mm0);
+
+  /* mm7 = 0 */
+  pxor_r2r (mm7, mm7);
+
+  /* mm6 = round */
+  movq_m2r (round, mm6);
+
+  while (width--) {
+    int a = *mask;
+
+    if (a == 0xff) {
+      *((uint32_t *) output) = opaque;
+    } else if ((input[0] == 0x00)) {
+      /* We just need to multiply our colour by the alpha value. */
+
+      /* mm2 = [ a ][ a ][ a ][ a ] */
+      movd_m2r (a, mm2);
+      movq_r2r (mm2, mm3);
+      pshufw_r2r (mm2, mm2, 0);
+
+      /* mm5 = [ cr ][ cb ][ y ][ 0 ] */
+      movq_r2r (mm1, mm5);
+
+      /* Multiply by alpha. */
+      pmullw_r2r (mm2, mm5);
+      paddw_m2r (round, mm5);
+      movq_r2r (mm5, mm6);
+      psrlw_i2r (8, mm6);
+      paddw_r2r (mm6, mm5);
+      psrlw_i2r (8, mm5);
+
+      /* Set alpha to a. */
+      por_r2r (mm3, mm5);
+
+      /* Pack and write our result. */
+      packuswb_r2r (mm5, mm5);
+      movd_r2m (mm5, *output);
+    } else if (a) {
+      /* mm2 = [ a ][ a ][ a ][ a ] */
+      movd_m2r (a, mm2);
+      pshufw_r2r (mm2, mm2, 0);
+
+      /* mm3 = [ cr ][ cb ][ y ][ 0xff ] */
+      movq_r2r (mm0, mm3);
+
+      /* mm4 = [ i_cr ][ i_cb ][ i_y ][ i_a ] */
+      movd_m2r (*input, mm4);
+      punpcklbw_r2r (mm7, mm4);
+
+      /* Subtract input and colour. */
+      psubw_r2r (mm4, mm3);     /* mm3 = mm3 - mm4 */
+
+      /* Multiply alpha. */
+      pmullw_r2r (mm2, mm3);
+      paddw_r2r (mm6, mm3);
+      movq_r2r (mm3, mm2);
+      psrlw_i2r (8, mm3);
+      paddw_r2r (mm2, mm3);
+      psrlw_i2r (8, mm3);
+
+      /* Add back in the input. */
+      paddb_r2r (mm3, mm4);
+
+      /* Write result. */
+      packuswb_r2r (mm4, mm4);
+      movd_r2m (mm4, *output);
+    }
+    mask++;
+    output += 4;
+    input += 4;
+  }
+  sfence ();
+  emms ();
+}
+#endif
+
+static void
+composite_alphamask_alpha_to_packed4444_scanline_c (uint8_t * output,
+    uint8_t * input,
+    uint8_t * mask, int width, int textluma, int textcb, int textcr, int alpha)
+{
+  uint32_t opaque = (textcr << 24) | (textcb << 16) | (textluma << 8) | 0xff;
+
+  int i;
+
+  for (i = 0; i < width; i++) {
+    int af = *mask;
+
+    if (af) {
+      int a = ((af * alpha) + 0x80) >> 8;
+
+      if (a == 0xff) {
+        *((uint32_t *) output) = opaque;
+      } else if (input[0] == 0x00) {
+        *((uint32_t *) output) = (multiply_alpha (a, textcr) << 24)
+            | (multiply_alpha (a, textcb) << 16)
+            | (multiply_alpha (a, textluma) << 8) | a;
+      } else if (a) {
+        *((uint32_t *) output) =
+            ((input[3] + multiply_alpha (a, textcr - input[3])) << 24)
+            | ((input[2] + multiply_alpha (a, textcb - input[2])) << 16)
+            | ((input[1] + multiply_alpha (a, textluma - input[1])) << 8)
+            | (a + multiply_alpha (0xff - a, input[0]));
+      }
+    }
+    mask++;
+    output += 4;
+    input += 4;
+  }
+}
+
+static void
+premultiply_packed4444_scanline_c (uint8_t * output, uint8_t * input, int width)
+{
+  while (width--) {
+    unsigned int cur_a = input[0];
+
+    *((uint32_t *) output) = (multiply_alpha (cur_a, input[3]) << 24)
+        | (multiply_alpha (cur_a, input[2]) << 16)
+        | (multiply_alpha (cur_a, input[1]) << 8)
+        | cur_a;
+
+    output += 4;
+    input += 4;
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+premultiply_packed4444_scanline_mmxext (uint8_t * output, uint8_t * input,
+    int width)
+{
+  const mmx_t round = { 0x0080008000800080ULL };
+  const mmx_t alpha = { 0x00000000000000ffULL };
+  const mmx_t noalp = { 0xffffffffffff0000ULL };
+
+  pxor_r2r (mm7, mm7);
+  while (width--) {
+    movd_m2r (*input, mm0);
+    punpcklbw_r2r (mm7, mm0);
+
+    movq_r2r (mm0, mm2);
+    pshufw_r2r (mm2, mm2, 0);
+    movq_r2r (mm2, mm4);
+    pand_m2r (alpha, mm4);
+
+    pmullw_r2r (mm2, mm0);
+    paddw_m2r (round, mm0);
+
+    movq_r2r (mm0, mm3);
+    psrlw_i2r (8, mm3);
+    paddw_r2r (mm3, mm0);
+    psrlw_i2r (8, mm0);
+
+    pand_m2r (noalp, mm0);
+    paddw_r2r (mm4, mm0);
+
+    packuswb_r2r (mm0, mm0);
+    movd_r2m (mm0, *output);
+
+    output += 4;
+    input += 4;
+  }
+  sfence ();
+  emms ();
+}
+#endif
+
+static void
+blend_packed422_scanline_c (uint8_t * output, uint8_t * src1,
+    uint8_t * src2, int width, int pos)
+{
+  if (pos == 0) {
+    blit_packed422_scanline (output, src1, width);
+  } else if (pos == 256) {
+    blit_packed422_scanline (output, src2, width);
+  } else if (pos == 128) {
+    interpolate_packed422_scanline (output, src1, src2, width);
+  } else {
+    width *= 2;
+    while (width--) {
+      *output++ = ((*src1++ * (256 - pos)) + (*src2++ * pos) + 0x80) >> 8;
+    }
+  }
+}
+
+#ifdef HAVE_CPU_I386
+static void
+blend_packed422_scanline_mmxext (uint8_t * output, uint8_t * src1,
+    uint8_t * src2, int width, int pos)
+{
+  if (pos <= 0) {
+    blit_packed422_scanline (output, src1, width);
+  } else if (pos >= 256) {
+    blit_packed422_scanline (output, src2, width);
+  } else if (pos == 128) {
+    interpolate_packed422_scanline (output, src1, src2, width);
+  } else {
+    const mmx_t all256 = { 0x0100010001000100ULL };
+    const mmx_t round = { 0x0080008000800080ULL };
+
+    movd_m2r (pos, mm0);
+    pshufw_r2r (mm0, mm0, 0);
+    movq_m2r (all256, mm1);
+    psubw_r2r (mm0, mm1);
+    pxor_r2r (mm7, mm7);
+
+    for (width /= 2; width; width--) {
+      movd_m2r (*src1, mm3);
+      movd_m2r (*src2, mm4);
+      punpcklbw_r2r (mm7, mm3);
+      punpcklbw_r2r (mm7, mm4);
+
+      pmullw_r2r (mm1, mm3);
+      pmullw_r2r (mm0, mm4);
+      paddw_r2r (mm4, mm3);
+      paddw_m2r (round, mm3);
+      psrlw_i2r (8, mm3);
+
+      packuswb_r2r (mm3, mm3);
+      movd_r2m (mm3, *output);
+
+      output += 4;
+      src1 += 4;
+      src2 += 4;
+    }
+    sfence ();
+    emms ();
+  }
+}
+#endif
+
+#ifdef HAVE_CPU_I386
+static void
+quarter_blit_vertical_packed422_scanline_mmxext (uint8_t * output,
+    uint8_t * one, uint8_t * three, int width)
+{
+  int i;
+
+  for (i = width / 16; i; --i) {
+    movq_m2r (*one, mm0);
+    movq_m2r (*three, mm1);
+    movq_m2r (*(one + 8), mm2);
+    movq_m2r (*(three + 8), mm3);
+    movq_m2r (*(one + 16), mm4);
+    movq_m2r (*(three + 16), mm5);
+    movq_m2r (*(one + 24), mm6);
+    movq_m2r (*(three + 24), mm7);
+    pavgb_r2r (mm1, mm0);
+    pavgb_r2r (mm1, mm0);
+    pavgb_r2r (mm3, mm2);
+    pavgb_r2r (mm3, mm2);
+    pavgb_r2r (mm5, mm4);
+    pavgb_r2r (mm5, mm4);
+    pavgb_r2r (mm7, mm6);
+    pavgb_r2r (mm7, mm6);
+    movntq_r2m (mm0, *output);
+    movntq_r2m (mm2, *(output + 8));
+    movntq_r2m (mm4, *(output + 16));
+    movntq_r2m (mm6, *(output + 24));
+    output += 32;
+    one += 32;
+    three += 32;
+  }
+  width = (width & 0xf);
+
+  for (i = width / 4; i; --i) {
+    movq_m2r (*one, mm0);
+    movq_m2r (*three, mm1);
+    pavgb_r2r (mm1, mm0);
+    pavgb_r2r (mm1, mm0);
+    movntq_r2m (mm0, *output);
+    output += 8;
+    one += 8;
+    three += 8;
+  }
+  width = width & 0x7;
+
+  /* Handle last few pixels. */
+  for (i = width * 2; i; --i) {
+    *output++ = (*one + *three + *three + *three + 2) / 4;
+    one++;
+    three++;
+  }
+
+  sfence ();
+  emms ();
+}
+#endif
+
+
+static void
+quarter_blit_vertical_packed422_scanline_c (uint8_t * output, uint8_t * one,
+    uint8_t * three, int width)
+{
+  width *= 2;
+  while (width--) {
+    *output++ = (*one + *three + *three + *three + 2) / 4;
+    one++;
+    three++;
+  }
+}
+
+static void
+subpix_blit_vertical_packed422_scanline_c (uint8_t * output, uint8_t * top,
+    uint8_t * bot, int subpixpos, int width)
+{
+  if (subpixpos == 32768) {
+    interpolate_packed422_scanline (output, top, bot, width);
+  } else if (subpixpos == 16384) {
+    quarter_blit_vertical_packed422_scanline (output, top, bot, width);
+  } else if (subpixpos == 49152) {
+    quarter_blit_vertical_packed422_scanline (output, bot, top, width);
+  } else {
+    int x;
+
+    width *= 2;
+    for (x = 0; x < width; x++) {
+      output[x] =
+          ((top[x] * subpixpos) + (bot[x] * (0xffff - subpixpos))) >> 16;
+    }
+  }
+}
+
+static void
+a8_subpix_blit_scanline_c (uint8_t * output, uint8_t * input,
+    int lasta, int startpos, int width)
+{
+  int pos = 0xffff - (startpos & 0xffff);
+
+  int prev = lasta;
+
+  int x;
+
+  for (x = 0; x < width; x++) {
+    output[x] = ((prev * pos) + (input[x] * (0xffff - pos))) >> 16;
+    prev = input[x];
+  }
+}
+
+/**
+ * These are from lavtools in mjpegtools:
+ *
+ * colorspace.c:  Routines to perform colorspace conversions.
+ *
+ *  Copyright (C) 2001 Matthew J. Marjanovic <maddog@mir.com>
+ *
+ *  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., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
+ */
+
+#define FP_BITS 18
+
+/* precomputed tables */
+
+static int Y_R[256];
+
+static int Y_G[256];
+
+static int Y_B[256];
+
+static int Cb_R[256];
+
+static int Cb_G[256];
+
+static int Cb_B[256];
+
+static int Cr_R[256];
+
+static int Cr_G[256];
+
+static int Cr_B[256];
+
+static int conv_RY_inited = 0;
+
+static int RGB_Y[256];
+
+static int R_Cr[256];
+
+static int G_Cb[256];
+
+static int G_Cr[256];
+
+static int B_Cb[256];
+
+static int conv_YR_inited = 0;
+
+static int
+myround (double n)
+{
+  if (n >= 0)
+    return (int) (n + 0.5);
+  else
+    return (int) (n - 0.5);
+}
+
+static void
+init_RGB_to_YCbCr_tables (void)
+{
+  int i;
+
+  /*
+   * Q_Z[i] =   (coefficient * i
+   *             * (Q-excursion) / (Z-excursion) * fixed-point-factor)
+   *
+   * to one of each, add the following:
+   *             + (fixed-point-factor / 2)         --- for rounding later
+   *             + (Q-offset * fixed-point-factor)  --- to add the offset
+   *             
+   */
+  for (i = 0; i < 256; i++) {
+    Y_R[i] =
+        myround (0.299 * (double) i * 219.0 / 255.0 * (double) (1 << FP_BITS));
+    Y_G[i] =
+        myround (0.587 * (double) i * 219.0 / 255.0 * (double) (1 << FP_BITS));
+    Y_B[i] =
+        myround ((0.114 * (double) i * 219.0 / 255.0 * (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)) + (16.0 * (double) (1 << FP_BITS)));
+
+    Cb_R[i] =
+        myround (-0.168736 * (double) i * 224.0 / 255.0 *
+        (double) (1 << FP_BITS));
+    Cb_G[i] =
+        myround (-0.331264 * (double) i * 224.0 / 255.0 *
+        (double) (1 << FP_BITS));
+    Cb_B[i] =
+        myround ((0.500 * (double) i * 224.0 / 255.0 * (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)) + (128.0 * (double) (1 << FP_BITS)));
+
+    Cr_R[i] =
+        myround (0.500 * (double) i * 224.0 / 255.0 * (double) (1 << FP_BITS));
+    Cr_G[i] =
+        myround (-0.418688 * (double) i * 224.0 / 255.0 *
+        (double) (1 << FP_BITS));
+    Cr_B[i] =
+        myround ((-0.081312 * (double) i * 224.0 / 255.0 *
+            (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)) + (128.0 * (double) (1 << FP_BITS)));
+  }
+  conv_RY_inited = 1;
+}
+
+static void
+init_YCbCr_to_RGB_tables (void)
+{
+  int i;
+
+  /*
+   * Q_Z[i] =   (coefficient * i
+   *             * (Q-excursion) / (Z-excursion) * fixed-point-factor)
+   *
+   * to one of each, add the following:
+   *             + (fixed-point-factor / 2)         --- for rounding later
+   *             + (Q-offset * fixed-point-factor)  --- to add the offset
+   *             
+   */
+
+  /* clip Y values under 16 */
+  for (i = 0; i < 16; i++) {
+    RGB_Y[i] =
+        myround ((1.0 * (double) (16) * 255.0 / 219.0 * (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)));
+  }
+  for (i = 16; i < 236; i++) {
+    RGB_Y[i] =
+        myround ((1.0 * (double) (i -
+                16) * 255.0 / 219.0 * (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)));
+  }
+  /* clip Y values above 235 */
+  for (i = 236; i < 256; i++) {
+    RGB_Y[i] =
+        myround ((1.0 * (double) (235) * 255.0 / 219.0 *
+            (double) (1 << FP_BITS))
+        + (double) (1 << (FP_BITS - 1)));
+  }
+
+  /* clip Cb/Cr values below 16 */
+  for (i = 0; i < 16; i++) {
+    R_Cr[i] =
+        myround (1.402 * (double) (-112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+    G_Cr[i] =
+        myround (-0.714136 * (double) (-112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+    G_Cb[i] =
+        myround (-0.344136 * (double) (-112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+    B_Cb[i] =
+        myround (1.772 * (double) (-112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+  }
+  for (i = 16; i < 241; i++) {
+    R_Cr[i] =
+        myround (1.402 * (double) (i -
+            128) * 255.0 / 224.0 * (double) (1 << FP_BITS));
+    G_Cr[i] =
+        myround (-0.714136 * (double) (i -
+            128) * 255.0 / 224.0 * (double) (1 << FP_BITS));
+    G_Cb[i] =
+        myround (-0.344136 * (double) (i -
+            128) * 255.0 / 224.0 * (double) (1 << FP_BITS));
+    B_Cb[i] =
+        myround (1.772 * (double) (i -
+            128) * 255.0 / 224.0 * (double) (1 << FP_BITS));
+  }
+  /* clip Cb/Cr values above 240 */
+  for (i = 241; i < 256; i++) {
+    R_Cr[i] =
+        myround (1.402 * (double) (112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+    G_Cr[i] =
+        myround (-0.714136 * (double) (112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+    G_Cb[i] =
+        myround (-0.344136 * (double) (i -
+            128) * 255.0 / 224.0 * (double) (1 << FP_BITS));
+    B_Cb[i] =
+        myround (1.772 * (double) (112) * 255.0 / 224.0 *
+        (double) (1 << FP_BITS));
+  }
+  conv_YR_inited = 1;
+}
+
+static void
+rgb24_to_packed444_rec601_scanline_c (uint8_t * output, uint8_t * input,
+    int width)
+{
+  if (!conv_RY_inited)
+    init_RGB_to_YCbCr_tables ();
+
+  while (width--) {
+    int r = input[0];
+
+    int g = input[1];
+
+    int b = input[2];
+
+    output[0] = (Y_R[r] + Y_G[g] + Y_B[b]) >> FP_BITS;
+    output[1] = (Cb_R[r] + Cb_G[g] + Cb_B[b]) >> FP_BITS;
+    output[2] = (Cr_R[r] + Cr_G[g] + Cr_B[b]) >> FP_BITS;
+    output += 3;
+    input += 3;
+  }
+}
+
+static void
+rgba32_to_packed4444_rec601_scanline_c (uint8_t * output, uint8_t * input,
+    int width)
+{
+  if (!conv_RY_inited)
+    init_RGB_to_YCbCr_tables ();
+
+  while (width--) {
+    int r = input[0];
+
+    int g = input[1];
+
+    int b = input[2];
+
+    int a = input[3];
+
+    output[0] = a;
+    output[1] = (Y_R[r] + Y_G[g] + Y_B[b]) >> FP_BITS;
+    output[2] = (Cb_R[r] + Cb_G[g] + Cb_B[b]) >> FP_BITS;
+    output[3] = (Cr_R[r] + Cr_G[g] + Cr_B[b]) >> FP_BITS;
+    output += 4;
+    input += 4;
+  }
+}
+
+static void
+packed444_to_rgb24_rec601_scanline_c (uint8_t * output, uint8_t * input,
+    int width)
+{
+  if (!conv_YR_inited)
+    init_YCbCr_to_RGB_tables ();
+
+  while (width--) {
+    int luma = input[0];
+
+    int cb = input[1];
+
+    int cr = input[2];
+
+    output[0] = clip255 ((RGB_Y[luma] + R_Cr[cr]) >> FP_BITS);
+    output[1] = clip255 ((RGB_Y[luma] + G_Cb[cb] + G_Cr[cr]) >> FP_BITS);
+    output[2] = clip255 ((RGB_Y[luma] + B_Cb[cb]) >> FP_BITS);
+
+    output += 3;
+    input += 3;
+  }
+}
+
+/**
+ * 601 numbers:
+ *
+ * Y' =  0.299*R' + 0.587*G' + 0.114*B' (in  0.0 to  1.0)
+ * Cb = -0.169*R' - 0.331*G' + 0.500*B' (in -0.5 to +0.5)
+ * Cr =  0.500*R' - 0.419*G' - 0.081*B' (in -0.5 to +0.5)
+ *
+ * Inverse:
+ *      Y         Cb        Cr
+ * R  1.0000   -0.0009    1.4017
+ * G  1.0000   -0.3437   -0.7142
+ * B  1.0000    1.7722    0.0010
+ *
+ * S170M numbers:
+ * Y'   =  0.299*R' + 0.587*G' + 0.114*B' (in  0.0 to 1.0)
+ * B-Y' = -0.299*R' - 0.587*G' + 0.886*B'
+ * R-Y' =  0.701*R' - 0.587*G' - 0.114*B'
+ */
+/*
+static void packed444_to_rgb24_rec601_reference_scanline( uint8_t *output, uint8_t *input, int width )
+{
+    while( width-- ) {
+        double yp = (((double) input[ 0 ]) - 16.0) / 255.0;
+        double cb = (((double) input[ 1 ]) - 128.0) / 255.0;
+        double cr = (((double) input[ 2 ]) - 128.0) / 255.0;
+        double r, g, b;
+
+        r = yp - (0.0009*cb) + (1.4017*cr);
+        g = yp - (0.3437*cb) - (0.7142*cr);
+        b = yp + (1.7722*cb) + (0.0010*cr);
+
+        if( r > 1.0 ) r = 1.0; else if( r < 0.0 ) r = 0.0;
+        if( g > 1.0 ) g = 1.0; else if( g < 0.0 ) g = 0.0;
+        if( b > 1.0 ) b = 1.0; else if( b < 0.0 ) b = 0.0;
+
+        output[ 0 ] = (int) ((r * 255.0) + 0.5);
+        output[ 1 ] = (int) ((g * 255.0) + 0.5);
+        output[ 2 ] = (int) ((b * 255.0) + 0.5);
+
+        output += 3;
+        input += 3;
+    }
+}
+*/
+
+static void
+packed444_to_nonpremultiplied_packed4444_scanline_c (uint8_t * output,
+    uint8_t * input, int width, int alpha)
+{
+  int i;
+
+  for (i = 0; i < width; i++) {
+    output[0] = alpha & 0xff;
+    output[1] = input[0] & 0xff;
+    output[2] = input[1] & 0xff;
+    output[3] = input[2] & 0xff;
+
+    output += 4;
+    input += 3;
+  }
+}
+
+static void
+aspect_adjust_packed4444_scanline_c (uint8_t * output,
+    uint8_t * input, int width, double pixel_aspect)
+{
+  double i;
+
+  int prev_i = 0;
+
+  int w = 0;
+
+  pixel_aspect = 1.0 / pixel_aspect;
+
+  for (i = 0.0; i < width; i += pixel_aspect) {
+    uint8_t *curin = input + ((int) i) * 4;
+
+    if (!prev_i) {
+      output[0] = curin[0];
+      output[1] = curin[1];
+      output[2] = curin[2];
+      output[3] = curin[3];
+    } else {
+      int avg_a = 0;
+
+      int avg_y = 0;
+
+      int avg_cb = 0;
+
+      int avg_cr = 0;
+
+      int pos = prev_i * 4;
+
+      int c = 0;
+
+      int j;
+
+      for (j = prev_i; j <= (int) i; j++) {
+        avg_a += input[pos++];
+        avg_y += input[pos++];
+        avg_cb += input[pos++];
+        avg_cr += input[pos++];
+        c++;
+      }
+      output[0] = avg_a / c;
+      output[1] = avg_y / c;
+      output[2] = avg_cb / c;
+      output[3] = avg_cr / c;
+    }
+    output += 4;
+    prev_i = (int) i;
+    w++;
+  }
+}
+
+static uint32_t speedy_accel;
+
+void
+setup_speedy_calls (uint32_t accel, int verbose)
+{
+  speedy_accel = accel;
+
+  interpolate_packed422_scanline = interpolate_packed422_scanline_c;
+  blit_colour_packed422_scanline = blit_colour_packed422_scanline_c;
+  blit_colour_packed4444_scanline = blit_colour_packed4444_scanline_c;
+  blit_packed422_scanline = blit_packed422_scanline_c;
+  composite_packed4444_to_packed422_scanline =
+      composite_packed4444_to_packed422_scanline_c;
+  composite_packed4444_alpha_to_packed422_scanline =
+      composite_packed4444_alpha_to_packed422_scanline_c;
+  composite_alphamask_to_packed4444_scanline =
+      composite_alphamask_to_packed4444_scanline_c;
+  composite_alphamask_alpha_to_packed4444_scanline =
+      composite_alphamask_alpha_to_packed4444_scanline_c;
+  premultiply_packed4444_scanline = premultiply_packed4444_scanline_c;
+  blend_packed422_scanline = blend_packed422_scanline_c;
+  comb_factor_packed422_scanline = 0;
+  diff_factor_packed422_scanline = diff_factor_packed422_scanline_c;
+  kill_chroma_packed422_inplace_scanline =
+      kill_chroma_packed422_inplace_scanline_c;
+  mirror_packed422_inplace_scanline = mirror_packed422_inplace_scanline_c;
+  speedy_memcpy = speedy_memcpy_c;
+  diff_packed422_block8x8 = diff_packed422_block8x8_c;
+  a8_subpix_blit_scanline = a8_subpix_blit_scanline_c;
+  quarter_blit_vertical_packed422_scanline =
+      quarter_blit_vertical_packed422_scanline_c;
+  subpix_blit_vertical_packed422_scanline =
+      subpix_blit_vertical_packed422_scanline_c;
+  packed444_to_nonpremultiplied_packed4444_scanline =
+      packed444_to_nonpremultiplied_packed4444_scanline_c;
+  aspect_adjust_packed4444_scanline = aspect_adjust_packed4444_scanline_c;
+  packed444_to_packed422_scanline = packed444_to_packed422_scanline_c;
+  packed422_to_packed444_scanline = packed422_to_packed444_scanline_c;
+  packed422_to_packed444_rec601_scanline =
+      packed422_to_packed444_rec601_scanline_c;
+  packed444_to_rgb24_rec601_scanline = packed444_to_rgb24_rec601_scanline_c;
+  rgb24_to_packed444_rec601_scanline = rgb24_to_packed444_rec601_scanline_c;
+  rgba32_to_packed4444_rec601_scanline = rgba32_to_packed4444_rec601_scanline_c;
+  invert_colour_packed422_inplace_scanline =
+      invert_colour_packed422_inplace_scanline_c;
+  vfilter_chroma_121_packed422_scanline =
+      vfilter_chroma_121_packed422_scanline_c;
+  vfilter_chroma_332_packed422_scanline =
+      vfilter_chroma_332_packed422_scanline_c;
+  convert_uyvy_to_yuyv_scanline = convert_uyvy_to_yuyv_scanline_c;
+  composite_colour4444_alpha_to_packed422_scanline =
+      composite_colour4444_alpha_to_packed422_scanline_c;
+
+#ifdef HAVE_CPU_I386
+  if (speedy_accel & OIL_IMPL_FLAG_MMXEXT) {
+    if (verbose) {
+      fprintf (stderr, "speedycode: Using MMXEXT optimized functions.\n");
+    }
+    interpolate_packed422_scanline = interpolate_packed422_scanline_mmxext;
+    blit_colour_packed422_scanline = blit_colour_packed422_scanline_mmxext;
+    blit_colour_packed4444_scanline = blit_colour_packed4444_scanline_mmxext;
+    blit_packed422_scanline = blit_packed422_scanline_mmxext;
+    composite_packed4444_to_packed422_scanline =
+        composite_packed4444_to_packed422_scanline_mmxext;
+    composite_packed4444_alpha_to_packed422_scanline =
+        composite_packed4444_alpha_to_packed422_scanline_mmxext;
+    composite_alphamask_to_packed4444_scanline =
+        composite_alphamask_to_packed4444_scanline_mmxext;
+    premultiply_packed4444_scanline = premultiply_packed4444_scanline_mmxext;
+    kill_chroma_packed422_inplace_scanline =
+        kill_chroma_packed422_inplace_scanline_mmx;
+    blend_packed422_scanline = blend_packed422_scanline_mmxext;
+    diff_factor_packed422_scanline = diff_factor_packed422_scanline_mmx;
+    comb_factor_packed422_scanline = comb_factor_packed422_scanline_mmx;
+    diff_packed422_block8x8 = diff_packed422_block8x8_mmx;
+    quarter_blit_vertical_packed422_scanline =
+        quarter_blit_vertical_packed422_scanline_mmxext;
+    invert_colour_packed422_inplace_scanline =
+        invert_colour_packed422_inplace_scanline_mmx;
+    vfilter_chroma_121_packed422_scanline =
+        vfilter_chroma_121_packed422_scanline_mmx;
+    vfilter_chroma_332_packed422_scanline =
+        vfilter_chroma_332_packed422_scanline_mmx;
+    convert_uyvy_to_yuyv_scanline = convert_uyvy_to_yuyv_scanline_mmx;
+    composite_colour4444_alpha_to_packed422_scanline =
+        composite_colour4444_alpha_to_packed422_scanline_mmxext;
+    speedy_memcpy = speedy_memcpy_mmxext;
+  } else if (speedy_accel & OIL_IMPL_FLAG_MMX) {
+    if (verbose) {
+      fprintf (stderr, "speedycode: Using MMX optimized functions.\n");
+    }
+    interpolate_packed422_scanline = interpolate_packed422_scanline_mmx;
+    blit_colour_packed422_scanline = blit_colour_packed422_scanline_mmx;
+    blit_colour_packed4444_scanline = blit_colour_packed4444_scanline_mmx;
+    blit_packed422_scanline = blit_packed422_scanline_mmx;
+    diff_factor_packed422_scanline = diff_factor_packed422_scanline_mmx;
+    comb_factor_packed422_scanline = comb_factor_packed422_scanline_mmx;
+    kill_chroma_packed422_inplace_scanline =
+        kill_chroma_packed422_inplace_scanline_mmx;
+    diff_packed422_block8x8 = diff_packed422_block8x8_mmx;
+    invert_colour_packed422_inplace_scanline =
+        invert_colour_packed422_inplace_scanline_mmx;
+    vfilter_chroma_121_packed422_scanline =
+        vfilter_chroma_121_packed422_scanline_mmx;
+    vfilter_chroma_332_packed422_scanline =
+        vfilter_chroma_332_packed422_scanline_mmx;
+    convert_uyvy_to_yuyv_scanline = convert_uyvy_to_yuyv_scanline_mmx;
+    speedy_memcpy = speedy_memcpy_mmx;
+  } else {
+    if (verbose) {
+      fprintf (stderr,
+          "speedycode: No MMX or MMXEXT support detected, using C fallbacks.\n");
+    }
+  }
+#endif
+}
+
+uint32_t
+speedy_get_accel (void)
+{
+  return speedy_accel;
+}
diff --git a/gst/deinterlace2/tvtime/speedy.h b/gst/deinterlace2/tvtime/speedy.h
new file mode 100644
index 00000000..fb833ff1
--- /dev/null
+++ b/gst/deinterlace2/tvtime/speedy.h
@@ -0,0 +1,308 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#ifndef SPEEDY_H_INCLUDED
+#define SPEEDY_H_INCLUDED
+
+#if defined (__SVR4) && defined (__sun)
+# include <sys/int_types.h>
+#else
+# include <stdint.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * Speedy is a collection of optimized functions plus their C fallbacks.
+ * This includes a simple system to select which functions to use
+ * at runtime.
+ *
+ * The optimizations are done with the help of the mmx.h system, from
+ * libmpeg2 by Michel Lespinasse and Aaron Holtzman.
+ *
+ * The library is a collection of function pointers which must be first
+ * initialized by setup_speedy_calls() to point at the fastest available
+ * implementation of each function.
+ */
+
+/**
+ * Struct for pulldown detection metrics.
+ */
+typedef struct pulldown_metrics_s {
+    /* difference: total, even lines, odd lines */
+    int d, e, o;
+    /* noise: temporal, spacial (current), spacial (past) */
+    int t, s, p;
+} pulldown_metrics_t;
+
+/**
+ * Interpolates a packed 4:2:2 scanline using linear interpolation.
+ */
+extern void (*interpolate_packed422_scanline)( uint8_t *output, uint8_t *top,
+                                               uint8_t *bot, int width );
+
+/**
+ * Blits a colour to a packed 4:2:2 scanline.
+ */
+extern void (*blit_colour_packed422_scanline)( uint8_t *output,
+                                               int width, int y, int cb, int cr );
+
+/**
+ * Blits a colour to a packed 4:4:4:4 scanline.  I use luma/cb/cr instead of
+ * RGB but this will of course work for either.
+ */
+extern void (*blit_colour_packed4444_scanline)( uint8_t *output,
+                                                int width, int alpha, int luma,
+                                                int cb, int cr );
+
+/**
+ * Blit from and to packed 4:2:2 scanline.
+ */
+extern void (*blit_packed422_scanline)( uint8_t *dest, const uint8_t *src, int width );
+
+/**
+ * Composites a premultiplied 4:4:4:4 pixel onto a packed 4:2:2 scanline.
+ */
+extern void (*composite_colour4444_alpha_to_packed422_scanline)( uint8_t *output, uint8_t *input,
+                                                                 int af, int y, int cb, int cr,
+                                                                 int width, int alpha );
+
+/**
+ * Composites a packed 4:4:4:4 scanline onto a packed 4:2:2 scanline.
+ * Chroma is downsampled by dropping samples (nearest neighbour).
+ */
+extern void (*composite_packed4444_to_packed422_scanline)( uint8_t *output,
+                                                           uint8_t *input,
+                                                           uint8_t *foreground,
+                                                           int width );
+
+/**
+ * Composites a packed 4:4:4:4 scanline onto a packed 4:2:2 scanline.
+ * Chroma is downsampled by dropping samples (nearest neighbour).  The
+ * alpha value provided is in the range 0-256 and is first applied to
+ * the input (for fadeouts).
+ */
+extern void (*composite_packed4444_alpha_to_packed422_scanline)( uint8_t *output,
+                                                                 uint8_t *input,
+                                                                 uint8_t *foreground,
+                                                                 int width, int alpha );
+
+/**
+ * Takes an alphamask and the given colour (in Y'CbCr) and composites it
+ * onto a packed 4:4:4:4 scanline.
+ */
+extern void (*composite_alphamask_to_packed4444_scanline)( uint8_t *output,
+                                                           uint8_t *input,
+                                                           uint8_t *mask, int width,
+                                                           int textluma, int textcb,
+                                                           int textcr );
+
+/**
+ * Takes an alphamask and the given colour (in Y'CbCr) and composites it
+ * onto a packed 4:4:4:4 scanline.  The alpha value provided is in the
+ * range 0-256 and is first applied to the input (for fadeouts).
+ */
+extern void (*composite_alphamask_alpha_to_packed4444_scanline)( uint8_t *output,
+                                                                 uint8_t *input,
+                                                                 uint8_t *mask, int width,
+                                                                 int textluma, int textcb,
+                                                                 int textcr, int alpha );
+
+/**
+ * Premultiplies the colour by the alpha channel in a packed 4:4:4:4
+ * scanline.
+ */
+extern void (*premultiply_packed4444_scanline)( uint8_t *output, uint8_t *input, int width );
+
+/**
+ * Blend between two packed 4:2:2 scanline.  Pos is the fade value in
+ * the range 0-256.  A value of 0 gives 100% src1, and a value of 256
+ * gives 100% src2.  Anything in between gives the appropriate faded
+ * version.
+ */
+extern void (*blend_packed422_scanline)( uint8_t *output, uint8_t *src1,
+                                         uint8_t *src2, int width, int pos );
+
+/**
+ * Calculates the 'difference factor' for two scanlines.  This is a
+ * metric where higher values indicate that the two scanlines are more
+ * different.
+ */
+extern unsigned int (*diff_factor_packed422_scanline)( uint8_t *cur, uint8_t *old, int width );
+
+/**
+ * Calculates the 'comb factor' for a set of three scanlines.  This is a
+ * metric where higher values indicate a more likely chance that the two
+ * fields are at separate points in time.
+ */
+extern unsigned int (*comb_factor_packed422_scanline)( uint8_t *top, uint8_t *mid,
+                                                       uint8_t *bot, int width );
+
+/**
+ * Vertical [1 2 1] chroma filter.
+ */
+extern void (*vfilter_chroma_121_packed422_scanline)( uint8_t *output, int width,
+                                                      uint8_t *m, uint8_t *t, uint8_t *b );
+
+/**
+ * Vertical [3 3 2] chroma filter.
+ */
+extern void (*vfilter_chroma_332_packed422_scanline)( uint8_t *output, int width,
+                                                      uint8_t *m, uint8_t *t, uint8_t *b );
+
+/**
+ * Sets the chroma of the scanline to neutral (128) in-place.
+ */
+extern void (*kill_chroma_packed422_inplace_scanline)( uint8_t *data, int width );
+
+/**
+ * Mirrors the scanline in-place.
+ */
+extern void (*mirror_packed422_inplace_scanline)( uint8_t *data, int width );
+
+/**
+ * Inverts the colours on a scanline in-place.
+ */
+extern void (*invert_colour_packed422_inplace_scanline)( uint8_t *data, int width );
+
+/**
+ * Fast memcpy function, used by all of the blit functions.  Won't blit
+ * anything if dest == src.
+ */
+extern void (*speedy_memcpy)( void *output, const void *input, size_t size );
+
+/**
+ * Calculates the block difference metrics for dalias' pulldown
+ * detection algorithm.
+ */
+extern void (*diff_packed422_block8x8)( pulldown_metrics_t *m, uint8_t *old,
+                                        uint8_t *new, int os, int ns );
+
+/**
+ * Takes an alpha mask and subpixelly blits it using linear
+ * interpolation.
+ */
+extern void (*a8_subpix_blit_scanline)( uint8_t *output, uint8_t *input,
+                                        int lasta, int startpos, int width );
+
+/**
+ * 1/4 vertical subpixel blit for packed 4:2:2 scanlines using linear
+ * interpolation.
+ */
+extern void (*quarter_blit_vertical_packed422_scanline)( uint8_t *output, uint8_t *one,
+                                                         uint8_t *three, int width );
+
+/**
+ * Vertical subpixel blit for packed 4:2:2 scanlines using linear
+ * interpolation.
+ */
+extern void (*subpix_blit_vertical_packed422_scanline)( uint8_t *output, uint8_t *top,
+                                                        uint8_t *bot, int subpixpos, int width );
+
+/**
+ * Simple function to convert a 4:4:4 scanline to a 4:4:4:4 scanline by
+ * adding an alpha channel.  Result is non-premultiplied.
+ */
+extern void (*packed444_to_nonpremultiplied_packed4444_scanline)( uint8_t *output, 
+                                                                  uint8_t *input,
+                                                                  int width, int alpha );
+
+/**
+ * I think this function needs to be rethought and renamed, but here
+ * it is for now.  This function horizontally resamples a scanline
+ * using linear interpolation to compensate for a change in pixel
+ * aspect ratio.
+ */
+extern void (*aspect_adjust_packed4444_scanline)( uint8_t *output,
+                                                  uint8_t *input, 
+                                                  int width,
+                                                  double pixel_aspect );
+
+/**
+ * Convert a packed 4:4:4 surface to a packed 4:2:2 surface using
+ * nearest neighbour chroma downsampling.
+ */
+extern void (*packed444_to_packed422_scanline)( uint8_t *output,
+                                                uint8_t *input,
+                                                int width );
+
+/**
+ * Converts packed 4:2:2 to packed 4:4:4 scanlines using nearest
+ * neighbour chroma upsampling.
+ */
+extern void (*packed422_to_packed444_scanline)( uint8_t *output,
+                                                uint8_t *input,
+                                                int width );
+
+/**
+ * This filter actually does not meet the spec so calling it rec601
+ * is a bit of a lie.  I got the filter from Poynton's site.  This
+ * converts a scanline from packed 4:2:2 to packed 4:4:4.  But this
+ * function should point at some high quality to-the-spec resampler.
+ */
+extern void (*packed422_to_packed444_rec601_scanline)( uint8_t *dest,
+                                                       uint8_t *src,
+                                                       int width );
+
+/**
+ * Conversions between Y'CbCr and R'G'B'.  We use Rec.601 numbers
+ * since our source is broadcast video, but I think there is an
+ * argument to be made for switching to Rec.709.
+ */
+extern void (*packed444_to_rgb24_rec601_scanline)( uint8_t *output,
+                                                   uint8_t *input,
+                                                   int width );
+extern void (*rgb24_to_packed444_rec601_scanline)( uint8_t *output,
+                                                   uint8_t *input,
+                                                   int width );
+extern void (*rgba32_to_packed4444_rec601_scanline)( uint8_t *output,
+                                                     uint8_t *input,
+                                                     int width );
+
+/**
+ * Convert from 4:2:2 with UYVY ordering to 4:2:2 with YUYV ordering.
+ */
+extern void (*convert_uyvy_to_yuyv_scanline)( uint8_t *uyvy_buf,
+                                              uint8_t *yuyv_buf, int width );
+
+/**
+ * Sets up the function pointers to point at the fastest function
+ * available.  Requires accelleration settings (see mm_accel.h).
+ */
+void setup_speedy_calls( uint32_t accel, int verbose );
+
+/**
+ * Returns a bitfield of what accellerations were used when speedy was
+ * initialized.  See mm_accel.h.
+ */
+uint32_t speedy_get_accel( void );
+
+#ifdef __cplusplus
+};
+#endif
+#endif /* SPEEDY_H_INCLUDED */
diff --git a/gst/deinterlace2/tvtime/sse.h b/gst/deinterlace2/tvtime/sse.h
new file mode 100644
index 00000000..2e00ee0c
--- /dev/null
+++ b/gst/deinterlace2/tvtime/sse.h
@@ -0,0 +1,992 @@
+/*	sse.h
+
+	Streaming SIMD Extenstions (a.k.a. Katmai New Instructions)
+	GCC interface library for IA32.
+
+	To use this library, simply include this header file
+	and compile with GCC.  You MUST have inlining enabled
+	in order for sse_ok() to work; this can be done by
+	simply using -O on the GCC command line.
+
+	Compiling with -DSSE_TRACE will cause detailed trace
+	output to be sent to stderr for each sse operation.
+	This adds lots of code, and obviously slows execution to
+	a crawl, but can be very useful for debugging.
+
+	THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
+	EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT
+	LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY
+	AND FITNESS FOR ANY PARTICULAR PURPOSE.
+
+	1999 by R. Fisher
+	Based on libmmx by H. Dietz and R. Fisher
+
+ Notes:
+	This is still extremely alpha.
+	Because this library depends on an assembler which understands the
+	 SSE opcodes, you probably won't be able to use this yet.
+	For now, do not use TRACE versions.  These both make use
+	 of the MMX registers, not the SSE registers.  This will be resolved
+	 at a later date.
+ ToDo:
+	Rewrite TRACE macros
+	Major Debugging Work
+*/
+
+#ifndef _SSE_H
+#define _SSE_H
+
+
+
+/*	The type of an value that fits in an SSE register
+	(note that long long constant values MUST be suffixed
+	 by LL and unsigned long long values by ULL, lest
+	 they be truncated by the compiler)
+*/
+typedef	union {
+	float			sf[4];	/* Single-precision (32-bit) value */
+} __attribute__ ((aligned (16))) sse_t;	/* On a 16 byte (128-bit) boundary */
+
+
+#if 0
+/*	Function to test if multimedia instructions are supported...
+*/
+inline extern int
+mm_support(void)
+{
+	/* Returns 1 if MMX instructions are supported,
+	   3 if Cyrix MMX and Extended MMX instructions are supported
+	   5 if AMD MMX and 3DNow! instructions are supported
+	   9 if MMX and SSE instructions are supported
+	   0 if hardware does not support any of these
+	*/
+	register int rval = 0;
+
+	__asm__ __volatile__ (
+		/* See if CPUID instruction is supported ... */
+		/* ... Get copies of EFLAGS into eax and ecx */
+		"pushf\n\t"
+		"popl %%eax\n\t"
+		"movl %%eax, %%ecx\n\t"
+
+		/* ... Toggle the ID bit in one copy and store */
+		/*     to the EFLAGS reg */
+		"xorl $0x200000, %%eax\n\t"
+		"push %%eax\n\t"
+		"popf\n\t"
+
+		/* ... Get the (hopefully modified) EFLAGS */
+		"pushf\n\t"
+		"popl %%eax\n\t"
+
+		/* ... Compare and test result */
+		"xorl %%eax, %%ecx\n\t"
+		"testl $0x200000, %%ecx\n\t"
+		"jz NotSupported1\n\t"		/* CPUID not supported */
+
+
+		/* Get standard CPUID information, and
+		       go to a specific vendor section */
+		"movl $0, %%eax\n\t"
+		"cpuid\n\t"
+
+		/* Check for Intel */
+		"cmpl $0x756e6547, %%ebx\n\t"
+		"jne TryAMD\n\t"
+		"cmpl $0x49656e69, %%edx\n\t"
+		"jne TryAMD\n\t"
+		"cmpl $0x6c65746e, %%ecx\n"
+		"jne TryAMD\n\t"
+		"jmp Intel\n\t"
+
+		/* Check for AMD */
+		"\nTryAMD:\n\t"
+		"cmpl $0x68747541, %%ebx\n\t"
+		"jne TryCyrix\n\t"
+		"cmpl $0x69746e65, %%edx\n\t"
+		"jne TryCyrix\n\t"
+		"cmpl $0x444d4163, %%ecx\n"
+		"jne TryCyrix\n\t"
+		"jmp AMD\n\t"
+
+		/* Check for Cyrix */
+		"\nTryCyrix:\n\t"
+		"cmpl $0x69727943, %%ebx\n\t"
+		"jne NotSupported2\n\t"
+		"cmpl $0x736e4978, %%edx\n\t"
+		"jne NotSupported3\n\t"
+		"cmpl $0x64616574, %%ecx\n\t"
+		"jne NotSupported4\n\t"
+		/* Drop through to Cyrix... */
+
+
+		/* Cyrix Section */
+		/* See if extended CPUID level 80000001 is supported */
+		/* The value of CPUID/80000001 for the 6x86MX is undefined
+		   according to the Cyrix CPU Detection Guide (Preliminary
+		   Rev. 1.01 table 1), so we'll check the value of eax for
+		   CPUID/0 to see if standard CPUID level 2 is supported.
+		   According to the table, the only CPU which supports level
+		   2 is also the only one which supports extended CPUID levels.
+		*/
+		"cmpl $0x2, %%eax\n\t"
+		"jne MMXtest\n\t"	/* Use standard CPUID instead */
+
+		/* Extended CPUID supported (in theory), so get extended
+		   features */
+		"movl $0x80000001, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%eax\n\t"	/* Test for MMX */
+		"jz NotSupported5\n\t"		/* MMX not supported */
+		"testl $0x01000000, %%eax\n\t"	/* Test for Ext'd MMX */
+		"jnz EMMXSupported\n\t"
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\n"
+		"EMMXSupported:\n\t"
+		"movl $3, %0:\n\n\t"		/* EMMX and MMX Supported */
+		"jmp Return\n\t"
+
+
+		/* AMD Section */
+		"AMD:\n\t"
+
+		/* See if extended CPUID is supported */
+		"movl $0x80000000, %%eax\n\t"
+		"cpuid\n\t"
+		"cmpl $0x80000000, %%eax\n\t"
+		"jl MMXtest\n\t"	/* Use standard CPUID instead */
+
+		/* Extended CPUID supported, so get extended features */
+		"movl $0x80000001, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%edx\n\t"	/* Test for MMX */
+		"jz NotSupported6\n\t"		/* MMX not supported */
+		"testl $0x80000000, %%edx\n\t"	/* Test for 3DNow! */
+		"jnz ThreeDNowSupported\n\t"
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\n"
+		"ThreeDNowSupported:\n\t"
+		"movl $5, %0:\n\n\t"		/* 3DNow! and MMX Supported */
+		"jmp Return\n\t"
+
+
+		/* Intel Section */
+		"Intel:\n\t"
+
+		/* Check for SSE */
+		"SSEtest:\n\t"
+		"movl $1, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x02000000, %%edx\n\t"	/* Test for SSE */
+		"jz MMXtest\n\t"		/* SSE Not supported */
+		"movl $9, %0:\n\n\t"		/* SSE Supported */
+		"jmp Return\n\t"
+
+		/* Check for MMX */
+		"MMXtest:\n\t"
+		"movl $1, %%eax\n\t"
+		"cpuid\n\t"
+		"testl $0x00800000, %%edx\n\t"	/* Test for MMX */
+		"jz NotSupported7\n\t"		/* MMX Not supported */
+		"movl $1, %0:\n\n\t"		/* MMX Supported */
+		"jmp Return\n\t"
+
+		/* Nothing supported */
+		"\nNotSupported1:\n\t"
+		"#movl $101, %0:\n\n\t"
+		"\nNotSupported2:\n\t"
+		"#movl $102, %0:\n\n\t"
+		"\nNotSupported3:\n\t"
+		"#movl $103, %0:\n\n\t"
+		"\nNotSupported4:\n\t"
+		"#movl $104, %0:\n\n\t"
+		"\nNotSupported5:\n\t"
+		"#movl $105, %0:\n\n\t"
+		"\nNotSupported6:\n\t"
+		"#movl $106, %0:\n\n\t"
+		"\nNotSupported7:\n\t"
+		"#movl $107, %0:\n\n\t"
+		"movl $0, %0:\n\n\t"
+
+		"Return:\n\t"
+		: "=a" (rval)
+		: /* no input */
+		: "eax", "ebx", "ecx", "edx"
+	);
+
+	/* Return */
+	return(rval);
+}
+
+/*	Function to test if sse instructions are supported...
+*/
+inline extern int
+sse_ok(void)
+{
+	/* Returns 1 if SSE instructions are supported, 0 otherwise */
+	return ( (mm_support() & 0x8) >> 3  );
+}
+#endif
+
+
+
+/*	Helper functions for the instruction macros that follow...
+	(note that memory-to-register, m2r, instructions are nearly
+	 as efficient as register-to-register, r2r, instructions;
+	 however, memory-to-memory instructions are really simulated
+	 as a convenience, and are only 1/3 as efficient)
+*/
+#ifdef	SSE_TRACE
+
+/*	Include the stuff for printing a trace to stderr...
+*/
+
+#include <stdio.h>
+
+#define	sse_i2r(op, imm, reg) \
+	{ \
+		sse_t sse_trace; \
+		sse_trace.uq = (imm); \
+		fprintf(stderr, #op "_i2r(" #imm "=0x%08x%08x, ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%08x%08x) => ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ (#op " %0, %%" #reg \
+				      : /* nothing */ \
+				      : "X" (imm)); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%08x%08x\n", \
+			sse_trace.d[1], sse_trace.d[0]); \
+	}
+
+#define	sse_m2r(op, mem, reg) \
+	{ \
+		sse_t sse_trace; \
+		sse_trace = (mem); \
+		fprintf(stderr, #op "_m2r(" #mem "=0x%08x%08x, ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%08x%08x) => ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ (#op " %0, %%" #reg \
+				      : /* nothing */ \
+				      : "X" (mem)); \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #reg "=0x%08x%08x\n", \
+			sse_trace.d[1], sse_trace.d[0]); \
+	}
+
+#define	sse_r2m(op, reg, mem) \
+	{ \
+		sse_t sse_trace; \
+		__asm__ __volatile__ ("movq %%" #reg ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #op "_r2m(" #reg "=0x%08x%08x, ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		sse_trace = (mem); \
+		fprintf(stderr, #mem "=0x%08x%08x) => ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ (#op " %%" #reg ", %0" \
+				      : "=X" (mem) \
+				      : /* nothing */ ); \
+		sse_trace = (mem); \
+		fprintf(stderr, #mem "=0x%08x%08x\n", \
+			sse_trace.d[1], sse_trace.d[0]); \
+	}
+
+#define	sse_r2r(op, regs, regd) \
+	{ \
+		sse_t sse_trace; \
+		__asm__ __volatile__ ("movq %%" #regs ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #op "_r2r(" #regs "=0x%08x%08x, ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ ("movq %%" #regd ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #regd "=0x%08x%08x) => ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ (#op " %" #regs ", %" #regd); \
+		__asm__ __volatile__ ("movq %%" #regd ", %0" \
+				      : "=X" (sse_trace) \
+				      : /* nothing */ ); \
+		fprintf(stderr, #regd "=0x%08x%08x\n", \
+			sse_trace.d[1], sse_trace.d[0]); \
+	}
+
+#define	sse_m2m(op, mems, memd) \
+	{ \
+		sse_t sse_trace; \
+		sse_trace = (mems); \
+		fprintf(stderr, #op "_m2m(" #mems "=0x%08x%08x, ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		sse_trace = (memd); \
+		fprintf(stderr, #memd "=0x%08x%08x) => ", \
+			sse_trace.d[1], sse_trace.d[0]); \
+		__asm__ __volatile__ ("movq %0, %%mm0\n\t" \
+				      #op " %1, %%mm0\n\t" \
+				      "movq %%mm0, %0" \
+				      : "=X" (memd) \
+				      : "X" (mems)); \
+		sse_trace = (memd); \
+		fprintf(stderr, #memd "=0x%08x%08x\n", \
+			sse_trace.d[1], sse_trace.d[0]); \
+	}
+
+#else
+
+/*	These macros are a lot simpler without the tracing...
+*/
+
+#define	sse_i2r(op, imm, reg) \
+	__asm__ __volatile__ (#op " %0, %%" #reg \
+			      : /* nothing */ \
+			      : "X" (imm) )
+
+#define	sse_m2r(op, mem, reg) \
+	__asm__ __volatile__ (#op " %0, %%" #reg \
+			      : /* nothing */ \
+			      : "X" (mem))
+
+#define	sse_r2m(op, reg, mem) \
+	__asm__ __volatile__ (#op " %%" #reg ", %0" \
+			      : "=X" (mem) \
+			      : /* nothing */ )
+
+#define	sse_r2r(op, regs, regd) \
+	__asm__ __volatile__ (#op " %" #regs ", %" #regd)
+
+#define	sse_r2ri(op, regs, regd, imm) \
+	__asm__ __volatile__ (#op " %0, %%" #regs ", %%" #regd \
+			      : /* nothing */ \
+			      : "X" (imm) )
+
+/* Load data from mems to xmmreg, operate on xmmreg, and store data to memd */
+#define	sse_m2m(op, mems, memd, xmmreg) \
+	__asm__ __volatile__ ("movups %0, %%xmm0\n\t" \
+			      #op " %1, %%xmm0\n\t" \
+			      "movups %%mm0, %0" \
+			      : "=X" (memd) \
+			      : "X" (mems))
+
+#define	sse_m2ri(op, mem, reg, subop) \
+	__asm__ __volatile__ (#op " %0, %%" #reg ", " #subop \
+			      : /* nothing */ \
+			      : "X" (mem))
+
+#define	sse_m2mi(op, mems, memd, xmmreg, subop) \
+	__asm__ __volatile__ ("movups %0, %%xmm0\n\t" \
+			      #op " %1, %%xmm0, " #subop "\n\t" \
+			      "movups %%mm0, %0" \
+			      : "=X" (memd) \
+			      : "X" (mems))
+#endif
+
+
+
+
+/*	1x128 MOVe Aligned four Packed Single-fp
+*/
+#define	movaps_m2r(var, reg)	sse_m2r(movaps, var, reg)
+#define	movaps_r2m(reg, var)	sse_r2m(movaps, reg, var)
+#define	movaps_r2r(regs, regd)	sse_r2r(movaps, regs, regd)
+#define	movaps(vars, vard) \
+	__asm__ __volatile__ ("movaps %1, %%mm0\n\t" \
+			      "movaps %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	1x128 MOVe aligned Non-Temporal four Packed Single-fp
+*/
+#define	movntps_r2m(xmmreg, var)	sse_r2m(movntps, xmmreg, var)
+
+
+/*	1x64 MOVe Non-Temporal Quadword
+*/
+#define	movntq_r2m(mmreg, var)		sse_r2m(movntq, mmreg, var)
+
+
+/*	1x128 MOVe Unaligned four Packed Single-fp
+*/
+#define	movups_m2r(var, reg)	sse_m2r(movups, var, reg)
+#define	movups_r2m(reg, var)	sse_r2m(movups, reg, var)
+#define	movups_r2r(regs, regd)	sse_r2r(movups, regs, regd)
+#define	movups(vars, vard) \
+	__asm__ __volatile__ ("movups %1, %%mm0\n\t" \
+			      "movups %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	MOVe High to Low Packed Single-fp
+	high half of 4x32f (x) -> low half of 4x32f (y)
+*/
+#define	movhlps_r2r(regs, regd)	sse_r2r(movhlps, regs, regd)
+
+
+/*	MOVe Low to High Packed Single-fp
+	low half of 4x32f (x) -> high half of 4x32f (y)
+*/
+#define	movlhps_r2r(regs, regd)	sse_r2r(movlhps, regs, regd)
+
+
+/*	MOVe High Packed Single-fp
+	2x32f -> high half of 4x32f
+*/
+#define	movhps_m2r(var, reg)	sse_m2r(movhps, var, reg)
+#define	movhps_r2m(reg, var)	sse_r2m(movhps, reg, var)
+#define	movhps(vars, vard) \
+	__asm__ __volatile__ ("movhps %1, %%mm0\n\t" \
+			      "movhps %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	MOVe Low Packed Single-fp
+	2x32f -> low half of 4x32f
+*/
+#define	movlps_m2r(var, reg)	sse_m2r(movlps, var, reg)
+#define	movlps_r2m(reg, var)	sse_r2m(movlps, reg, var)
+#define	movlps(vars, vard) \
+	__asm__ __volatile__ ("movlps %1, %%mm0\n\t" \
+			      "movlps %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	MOVe Scalar Single-fp
+	lowest field of 4x32f (x) -> lowest field of 4x32f (y)
+*/
+#define	movss_m2r(var, reg)	sse_m2r(movss, var, reg)
+#define	movss_r2m(reg, var)	sse_r2m(movss, reg, var)
+#define	movss_r2r(regs, regd)	sse_r2r(movss, regs, regd)
+#define	movss(vars, vard) \
+	__asm__ __volatile__ ("movss %1, %%mm0\n\t" \
+			      "movss %%mm0, %0" \
+			      : "=X" (vard) \
+			      : "X" (vars))
+
+
+/*	4x16 Packed SHUFfle Word
+*/
+#define	pshufw_m2r(var, reg, index)	sse_m2ri(pshufw, var, reg, index)
+#define	pshufw_r2r(regs, regd, index)	sse_r2ri(pshufw, regs, regd, index)
+
+
+/*	1x128 SHUFfle Packed Single-fp
+*/
+#define	shufps_m2r(var, reg, index)	sse_m2ri(shufps, var, reg, index)
+#define	shufps_r2r(regs, regd, index)	sse_r2ri(shufps, regs, regd, index)
+
+
+/*	ConVerT Packed signed Int32 to(2) Packed Single-fp
+*/
+#define	cvtpi2ps_m2r(var, xmmreg)	sse_m2r(cvtpi2ps, var, xmmreg)
+#define	cvtpi2ps_r2r(mmreg, xmmreg)	sse_r2r(cvtpi2ps, mmreg, xmmreg)
+
+
+/*	ConVerT Packed Single-fp to(2) Packed signed Int32
+*/
+#define	cvtps2pi_m2r(var, mmreg)	sse_m2r(cvtps2pi, var, mmreg)
+#define	cvtps2pi_r2r(xmmreg, mmreg)	sse_r2r(cvtps2pi, mmreg, xmmreg)
+
+
+/*	ConVerT with Truncate Packed Single-fp to(2) Packed Int32
+*/
+#define	cvttps2pi_m2r(var, mmreg)	sse_m2r(cvttps2pi, var, mmreg)
+#define	cvttps2pi_r2r(xmmreg, mmreg)	sse_r2r(cvttps2pi, mmreg, xmmreg)
+
+
+/*	ConVerT Signed Int32 to(2) Single-fp (Scalar)
+*/
+#define	cvtsi2ss_m2r(var, xmmreg)	sse_m2r(cvtsi2ss, var, xmmreg)
+#define	cvtsi2ss_r2r(reg, xmmreg)	sse_r2r(cvtsi2ss, reg, xmmreg)
+
+
+/*	ConVerT Scalar Single-fp to(2) Signed Int32
+*/
+#define	cvtss2si_m2r(var, reg)		sse_m2r(cvtss2si, var, reg)
+#define	cvtss2si_r2r(xmmreg, reg)	sse_r2r(cvtss2si, xmmreg, reg)
+
+
+/*	ConVerT with Truncate Scalar Single-fp to(2) Signed Int32
+*/
+#define	cvttss2si_m2r(var, reg)		sse_m2r(cvtss2si, var, reg)
+#define	cvttss2si_r2r(xmmreg, reg)	sse_r2r(cvtss2si, xmmreg, reg)
+
+
+/*	Parallel EXTRact Word from 4x16
+*/
+#define	pextrw_r2r(mmreg, reg, field)	sse_r2ri(pextrw, mmreg, reg, field)
+
+
+/*	Parallel INSeRt Word from 4x16
+*/
+#define	pinsrw_r2r(reg, mmreg, field)	sse_r2ri(pinsrw, reg, mmreg, field)
+
+
+
+/*	MOVe MaSK from Packed Single-fp
+*/
+#ifdef	SSE_TRACE
+	#define	movmskps(xmmreg, reg) \
+	{ \
+		fprintf(stderr, "movmskps()\n"); \
+		__asm__ __volatile__ ("movmskps %" #xmmreg ", %" #reg) \
+	}
+#else
+	#define	movmskps(xmmreg, reg) \
+	__asm__ __volatile__ ("movmskps %" #xmmreg ", %" #reg)
+#endif
+
+
+/*	Parallel MOVe MaSK from mmx reg to 32-bit reg
+*/
+#ifdef	SSE_TRACE
+	#define	pmovmskb(mmreg, reg) \
+	{ \
+		fprintf(stderr, "movmskps()\n"); \
+		__asm__ __volatile__ ("movmskps %" #mmreg ", %" #reg) \
+	}
+#else
+	#define	pmovmskb(mmreg, reg) \
+	__asm__ __volatile__ ("movmskps %" #mmreg ", %" #reg)
+#endif
+
+
+/*	MASKed MOVe from 8x8 to memory pointed to by (e)di register
+*/
+#define	maskmovq(mmregs, fieldreg)	sse_r2ri(maskmovq, mmregs, fieldreg)
+
+
+
+
+/*	4x32f Parallel ADDs
+*/
+#define	addps_m2r(var, reg)		sse_m2r(addps, var, reg)
+#define	addps_r2r(regs, regd)		sse_r2r(addps, regs, regd)
+#define	addps(vars, vard, xmmreg)	sse_m2m(addps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel ADDs
+*/
+#define	addss_m2r(var, reg)		sse_m2r(addss, var, reg)
+#define	addss_r2r(regs, regd)		sse_r2r(addss, regs, regd)
+#define	addss(vars, vard, xmmreg)	sse_m2m(addss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel SUBs
+*/
+#define	subps_m2r(var, reg)		sse_m2r(subps, var, reg)
+#define	subps_r2r(regs, regd)		sse_r2r(subps, regs, regd)
+#define	subps(vars, vard, xmmreg)	sse_m2m(subps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel SUBs
+*/
+#define	subss_m2r(var, reg)		sse_m2r(subss, var, reg)
+#define	subss_r2r(regs, regd)		sse_r2r(subss, regs, regd)
+#define	subss(vars, vard, xmmreg)	sse_m2m(subss, vars, vard, xmmreg)
+
+
+/*	8x8u -> 4x16u Packed Sum of Absolute Differences
+*/
+#define	psadbw_m2r(var, reg)		sse_m2r(psadbw, var, reg)
+#define	psadbw_r2r(regs, regd)		sse_r2r(psadbw, regs, regd)
+#define	psadbw(vars, vard, mmreg)	sse_m2m(psadbw, vars, vard, mmreg)
+
+
+/*	4x16u Parallel MUL High Unsigned
+*/
+#define	pmulhuw_m2r(var, reg)		sse_m2r(pmulhuw, var, reg)
+#define	pmulhuw_r2r(regs, regd)		sse_r2r(pmulhuw, regs, regd)
+#define	pmulhuw(vars, vard, mmreg)	sse_m2m(pmulhuw, vars, vard, mmreg)
+
+
+/*	4x32f Parallel MULs
+*/
+#define	mulps_m2r(var, reg)		sse_m2r(mulps, var, reg)
+#define	mulps_r2r(regs, regd)		sse_r2r(mulps, regs, regd)
+#define	mulps(vars, vard, xmmreg)	sse_m2m(mulps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel MULs
+*/
+#define	mulss_m2r(var, reg)		sse_m2r(mulss, var, reg)
+#define	mulss_r2r(regs, regd)		sse_r2r(mulss, regs, regd)
+#define	mulss(vars, vard, xmmreg)	sse_m2m(mulss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel DIVs
+*/
+#define	divps_m2r(var, reg)		sse_m2r(divps, var, reg)
+#define	divps_r2r(regs, regd)		sse_r2r(divps, regs, regd)
+#define	divps(vars, vard, xmmreg)	sse_m2m(divps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel DIVs
+*/
+#define	divss_m2r(var, reg)		sse_m2r(divss, var, reg)
+#define	divss_r2r(regs, regd)		sse_r2r(divss, regs, regd)
+#define	divss(vars, vard, xmmreg)	sse_m2m(divss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel Reciprocals
+*/
+#define	rcpps_m2r(var, reg)		sse_m2r(rcpps, var, reg)
+#define	rcpps_r2r(regs, regd)		sse_r2r(rcpps, regs, regd)
+#define	rcpps(vars, vard, xmmreg)	sse_m2m(rcpps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel Reciprocals
+*/
+#define	rcpss_m2r(var, reg)		sse_m2r(rcpss, var, reg)
+#define	rcpss_r2r(regs, regd)		sse_r2r(rcpss, regs, regd)
+#define	rcpss(vars, vard, xmmreg)	sse_m2m(rcpss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel Square Root of Reciprocals
+*/
+#define	rsqrtps_m2r(var, reg)		sse_m2r(rsqrtps, var, reg)
+#define	rsqrtps_r2r(regs, regd)		sse_r2r(rsqrtps, regs, regd)
+#define	rsqrtps(vars, vard, xmmreg)	sse_m2m(rsqrtps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel Square Root of Reciprocals
+*/
+#define	rsqrtss_m2r(var, reg)		sse_m2r(rsqrtss, var, reg)
+#define	rsqrtss_r2r(regs, regd)		sse_r2r(rsqrtss, regs, regd)
+#define	rsqrtss(vars, vard, xmmreg)	sse_m2m(rsqrtss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel Square Roots
+*/
+#define	sqrtps_m2r(var, reg)		sse_m2r(sqrtps, var, reg)
+#define	sqrtps_r2r(regs, regd)		sse_r2r(sqrtps, regs, regd)
+#define	sqrtps(vars, vard, xmmreg)	sse_m2m(sqrtps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel Square Roots
+*/
+#define	sqrtss_m2r(var, reg)		sse_m2r(sqrtss, var, reg)
+#define	sqrtss_r2r(regs, regd)		sse_r2r(sqrtss, regs, regd)
+#define	sqrtss(vars, vard, xmmreg)	sse_m2m(sqrtss, vars, vard, xmmreg)
+
+
+/*	8x8u and 4x16u Parallel AVeraGe
+*/
+#define	pavgb_m2r(var, reg)		sse_m2r(pavgb, var, reg)
+#define	pavgb_r2r(regs, regd)		sse_r2r(pavgb, regs, regd)
+#define	pavgb(vars, vard, mmreg)	sse_m2m(pavgb, vars, vard, mmreg)
+
+#define	pavgw_m2r(var, reg)		sse_m2r(pavgw, var, reg)
+#define	pavgw_r2r(regs, regd)		sse_r2r(pavgw, regs, regd)
+#define	pavgw(vars, vard, mmreg)	sse_m2m(pavgw, vars, vard, mmreg)
+
+
+/*	1x128 bitwise AND
+*/
+#define	andps_m2r(var, reg)		sse_m2r(andps, var, reg)
+#define	andps_r2r(regs, regd)		sse_r2r(andps, regs, regd)
+#define	andps(vars, vard, xmmreg)	sse_m2m(andps, vars, vard, xmmreg)
+
+
+/*	1x128 bitwise AND with Not the destination
+*/
+#define	andnps_m2r(var, reg)		sse_m2r(andnps, var, reg)
+#define	andnps_r2r(regs, regd)		sse_r2r(andnps, regs, regd)
+#define	andnps(vars, vard, xmmreg)	sse_m2m(andnps, vars, vard, xmmreg)
+
+
+/*	1x128 bitwise OR
+*/
+#define	orps_m2r(var, reg)		sse_m2r(orps, var, reg)
+#define	orps_r2r(regs, regd)		sse_r2r(orps, regs, regd)
+#define	orps(vars, vard, xmmreg)	sse_m2m(orps, vars, vard, xmmreg)
+
+
+/*	1x128 bitwise eXclusive OR
+*/
+#define	xorps_m2r(var, reg)		sse_m2r(xorps, var, reg)
+#define	xorps_r2r(regs, regd)		sse_r2r(xorps, regs, regd)
+#define	xorps(vars, vard, xmmreg)	sse_m2m(xorps, vars, vard, xmmreg)
+
+
+/*	8x8u, 4x16, and 4x32f Parallel Maximum
+*/
+#define	pmaxub_m2r(var, reg)		sse_m2r(pmaxub, var, reg)
+#define	pmaxub_r2r(regs, regd)		sse_r2r(pmaxub, regs, regd)
+#define	pmaxub(vars, vard, mmreg)	sse_m2m(pmaxub, vars, vard, mmreg)
+
+#define	pmaxsw_m2r(var, reg)		sse_m2r(pmaxsw, var, reg)
+#define	pmaxsw_r2r(regs, regd)		sse_r2r(pmaxsw, regs, regd)
+#define	pmaxsw(vars, vard, mmreg)	sse_m2m(pmaxsw, vars, vard, mmreg)
+
+#define	maxps_m2r(var, reg)		sse_m2r(maxps, var, reg)
+#define	maxps_r2r(regs, regd)		sse_r2r(maxps, regs, regd)
+#define	maxps(vars, vard, xmmreg)	sse_m2m(maxps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel Maximum
+*/
+#define	maxss_m2r(var, reg)		sse_m2r(maxss, var, reg)
+#define	maxss_r2r(regs, regd)		sse_r2r(maxss, regs, regd)
+#define	maxss(vars, vard, xmmreg)	sse_m2m(maxss, vars, vard, xmmreg)
+
+
+/*	8x8u, 4x16, and 4x32f Parallel Minimum
+*/
+#define	pminub_m2r(var, reg)		sse_m2r(pminub, var, reg)
+#define	pminub_r2r(regs, regd)		sse_r2r(pminub, regs, regd)
+#define	pminub(vars, vard, mmreg)	sse_m2m(pminub, vars, vard, mmreg)
+
+#define	pminsw_m2r(var, reg)		sse_m2r(pminsw, var, reg)
+#define	pminsw_r2r(regs, regd)		sse_r2r(pminsw, regs, regd)
+#define	pminsw(vars, vard, mmreg)	sse_m2m(pminsw, vars, vard, mmreg)
+
+#define	minps_m2r(var, reg)		sse_m2r(minps, var, reg)
+#define	minps_r2r(regs, regd)		sse_r2r(minps, regs, regd)
+#define	minps(vars, vard, xmmreg)	sse_m2m(minps, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Parallel Minimum
+*/
+#define	minss_m2r(var, reg)		sse_m2r(minss, var, reg)
+#define	minss_r2r(regs, regd)		sse_r2r(minss, regs, regd)
+#define	minss(vars, vard, xmmreg)	sse_m2m(minss, vars, vard, xmmreg)
+
+
+/*	4x32f Parallel CoMPares
+	(resulting fields are either 0 or -1)
+*/
+#define	cmpps_m2r(var, reg, op)		sse_m2ri(cmpps, var, reg, op)
+#define	cmpps_r2r(regs, regd, op)	sse_r2ri(cmpps, regs, regd, op)
+#define	cmpps(vars, vard, op, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, op)
+
+#define	cmpeqps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 0)
+#define	cmpeqps_r2r(regs, regd)		sse_r2ri(cmpps, regs, regd, 0)
+#define	cmpeqps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 0)
+
+#define	cmpltps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 1)
+#define	cmpltps_r2r(regs, regd)		sse_r2ri(cmpps, regs, regd, 1)
+#define	cmpltps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 1)
+
+#define	cmpleps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 2)
+#define	cmpleps_r2r(regs, regd)		sse_r2ri(cmpps, regs, regd, 2)
+#define	cmpleps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 2)
+
+#define	cmpunordps_m2r(var, reg)	sse_m2ri(cmpps, var, reg, 3)
+#define	cmpunordps_r2r(regs, regd)	sse_r2ri(cmpps, regs, regd, 3)
+#define	cmpunordps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 3)
+
+#define	cmpneqps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 4)
+#define	cmpneqps_r2r(regs, regd)	sse_r2ri(cmpps, regs, regd, 4)
+#define	cmpneqps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 4)
+
+#define	cmpnltps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 5)
+#define	cmpnltps_r2r(regs, regd)	sse_r2ri(cmpps, regs, regd, 5)
+#define	cmpnltps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 5)
+
+#define	cmpnleps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 6)
+#define	cmpnleps_r2r(regs, regd)	sse_r2ri(cmpps, regs, regd, 6)
+#define	cmpnleps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 6)
+
+#define	cmpordps_m2r(var, reg)		sse_m2ri(cmpps, var, reg, 7)
+#define	cmpordps_r2r(regs, regd)	sse_r2ri(cmpps, regs, regd, 7)
+#define	cmpordps(vars, vard, xmmreg)	sse_m2mi(cmpps, vars, vard, xmmreg, 7)
+
+
+/*	Lowest Field of 4x32f Parallel CoMPares
+	(resulting fields are either 0 or -1)
+*/
+#define	cmpss_m2r(var, reg, op)		sse_m2ri(cmpss, var, reg, op)
+#define	cmpss_r2r(regs, regd, op)	sse_r2ri(cmpss, regs, regd, op)
+#define	cmpss(vars, vard, op, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, op)
+
+#define	cmpeqss_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 0)
+#define	cmpeqss_r2r(regs, regd)		sse_r2ri(cmpss, regs, regd, 0)
+#define	cmpeqss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 0)
+
+#define	cmpltss_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 1)
+#define	cmpltss_r2r(regs, regd)		sse_r2ri(cmpss, regs, regd, 1)
+#define	cmpltss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 1)
+
+#define	cmpless_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 2)
+#define	cmpless_r2r(regs, regd)		sse_r2ri(cmpss, regs, regd, 2)
+#define	cmpless(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 2)
+
+#define	cmpunordss_m2r(var, reg)	sse_m2ri(cmpss, var, reg, 3)
+#define	cmpunordss_r2r(regs, regd)	sse_r2ri(cmpss, regs, regd, 3)
+#define	cmpunordss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 3)
+
+#define	cmpneqss_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 4)
+#define	cmpneqss_r2r(regs, regd)	sse_r2ri(cmpss, regs, regd, 4)
+#define	cmpneqss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 4)
+
+#define	cmpnltss_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 5)
+#define	cmpnltss_r2r(regs, regd)	sse_r2ri(cmpss, regs, regd, 5)
+#define	cmpnltss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 5)
+
+#define	cmpnless_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 6)
+#define	cmpnless_r2r(regs, regd)	sse_r2ri(cmpss, regs, regd, 6)
+#define	cmpnless(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 6)
+
+#define	cmpordss_m2r(var, reg)		sse_m2ri(cmpss, var, reg, 7)
+#define	cmpordss_r2r(regs, regd)	sse_r2ri(cmpss, regs, regd, 7)
+#define	cmpordss(vars, vard, xmmreg)	sse_m2mi(cmpss, vars, vard, xmmreg, 7)
+
+
+/*	Lowest Field of 4x32f Parallel CoMPares to set EFLAGS
+	(resulting fields are either 0 or -1)
+*/
+#define	comiss_m2r(var, reg)		sse_m2r(comiss, var, reg)
+#define	comiss_r2r(regs, regd)		sse_r2r(comiss, regs, regd)
+#define	comiss(vars, vard, xmmreg)	sse_m2m(comiss, vars, vard, xmmreg)
+
+
+/*	Lowest Field of 4x32f Unordered Parallel CoMPares to set EFLAGS
+	(resulting fields are either 0 or -1)
+*/
+#define	ucomiss_m2r(var, reg)		sse_m2r(ucomiss, var, reg)
+#define	ucomiss_r2r(regs, regd)		sse_r2r(ucomiss, regs, regd)
+#define	ucomiss(vars, vard, xmmreg)	sse_m2m(ucomiss, vars, vard, xmmreg)
+
+
+/*	2-(4x32f) -> 4x32f UNPaCK Low Packed Single-fp
+	(interleaves low half of dest with low half of source
+	 as padding in each result field)
+*/
+#define	unpcklps_m2r(var, reg)		sse_m2r(unpcklps, var, reg)
+#define	unpcklps_r2r(regs, regd)	sse_r2r(unpcklps, regs, regd)
+
+
+/*	2-(4x32f) -> 4x32f UNPaCK High Packed Single-fp
+	(interleaves high half of dest with high half of source
+	 as padding in each result field)
+*/
+#define	unpckhps_m2r(var, reg)		sse_m2r(unpckhps, var, reg)
+#define	unpckhps_r2r(regs, regd)	sse_r2r(unpckhps, regs, regd)
+
+
+
+/*	Fp and mmX ReSTORe state
+*/
+#ifdef	SSE_TRACE
+	#define	fxrstor(mem) \
+	{ \
+		fprintf(stderr, "fxrstor()\n"); \
+		__asm__ __volatile__ ("fxrstor %0" \
+			      : /* nothing */ \
+			      : "X" (mem)) \
+	}
+#else
+	#define	fxrstor(mem) \
+	__asm__ __volatile__ ("fxrstor %0" \
+			      : /* nothing */ \
+			      : "X" (mem))
+#endif
+
+
+/*	Fp and mmX SAVE state
+*/
+#ifdef	SSE_TRACE
+	#define	fxsave(mem) \
+	{ \
+		fprintf(stderr, "fxsave()\n"); \
+		__asm__ __volatile__ ("fxsave %0" \
+			      : /* nothing */ \
+			      : "X" (mem)) \
+	}
+#else
+	#define	fxsave(mem) \
+	__asm__ __volatile__ ("fxsave %0" \
+			      : /* nothing */ \
+			      : "X" (mem))
+#endif
+
+
+/*	STore streaMing simd eXtensions Control/Status Register
+*/
+#ifdef	SSE_TRACE
+	#define	stmxcsr(mem) \
+	{ \
+		fprintf(stderr, "stmxcsr()\n"); \
+		__asm__ __volatile__ ("stmxcsr %0" \
+			      : /* nothing */ \
+			      : "X" (mem)) \
+	}
+#else
+	#define	stmxcsr(mem) \
+	__asm__ __volatile__ ("stmxcsr %0" \
+			      : /* nothing */ \
+			      : "X" (mem))
+#endif
+
+
+/*	LoaD streaMing simd eXtensions Control/Status Register
+*/
+#ifdef	SSE_TRACE
+	#define	ldmxcsr(mem) \
+	{ \
+		fprintf(stderr, "ldmxcsr()\n"); \
+		__asm__ __volatile__ ("ldmxcsr %0" \
+			      : /* nothing */ \
+			      : "X" (mem)) \
+	}
+#else
+	#define	ldmxcsr(mem) \
+	__asm__ __volatile__ ("ldmxcsr %0" \
+			      : /* nothing */ \
+			      : "X" (mem))
+#endif
+
+
+/*	Store FENCE - enforce ordering of stores before fence vs. stores
+	occuring after fence in source code.
+*/
+#ifdef	SSE_TRACE
+	#define	sfence() \
+	{ \
+		fprintf(stderr, "sfence()\n"); \
+		__asm__ __volatile__ ("sfence\n\t") \
+	}
+#else
+	#define	sfence() \
+	__asm__ __volatile__ ("sfence\n\t")
+#endif
+
+
+/*	PREFETCH data using T0, T1, T2, or NTA hint
+		T0  = Prefetch into all cache levels
+		T1  = Prefetch into all cache levels except 0th level
+		T2  = Prefetch into all cache levels except 0th and 1st levels
+		NTA = Prefetch data into non-temporal cache structure
+*/
+#ifdef	SSE_TRACE
+#else
+	#define	prefetch(mem, hint) \
+	__asm__ __volatile__ ("prefetch" #hint " %0" \
+			      : /* nothing */ \
+			      : "X" (mem))
+
+	#define	prefetcht0(mem)		prefetch(mem, t0)
+	#define	prefetcht1(mem)		prefetch(mem, t1)
+	#define	prefetcht2(mem)		prefetch(mem, t2)
+	#define	prefetchnta(mem)	prefetch(mem, nta)
+#endif
+
+
+
+#endif
diff --git a/gst/deinterlace2/tvtime/tomsmocomp.c b/gst/deinterlace2/tvtime/tomsmocomp.c
new file mode 100644
index 00000000..f0b73677
--- /dev/null
+++ b/gst/deinterlace2/tvtime/tomsmocomp.c
@@ -0,0 +1,187 @@
+/**
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * 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, 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., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+#include <string.h>
+
+#include "gst/gst.h"
+#include "gstdeinterlace2.h"
+#include "plugins.h"
+#include "speedy.h"
+
+#include "tomsmocomp.h"
+#include "tomsmocomp/tomsmocompmacros.h"
+#include "x86-64_macros.inc"
+
+
+#define SearchEffortDefault 5
+#define UseStrangeBobDefault 0
+
+long SearchEffort;
+
+int UseStrangeBob;
+
+MEMCPY_FUNC *pMyMemcpy;
+
+int IsOdd;
+
+const unsigned char *pWeaveSrc;
+
+const unsigned char *pWeaveSrcP;
+
+unsigned char *pWeaveDest;
+
+const unsigned char *pCopySrc;
+
+const unsigned char *pCopySrcP;
+
+unsigned char *pCopyDest;
+
+int src_pitch;
+
+int dst_pitch;
+
+int rowsize;
+
+int height;
+
+int FldHeight;
+
+int
+Fieldcopy (void *dest, const void *src, size_t count,
+    int rows, int dst_pitch, int src_pitch)
+{
+  unsigned char *pDest = (unsigned char *) dest;
+
+  unsigned char *pSrc = (unsigned char *) src;
+
+  int i;
+
+  for (i = 0; i < rows; i++) {
+    pMyMemcpy (pDest, pSrc, count);
+    pSrc += src_pitch;
+    pDest += dst_pitch;
+  }
+  return 0;
+}
+
+
+#define IS_MMX
+#define SSE_TYPE MMX
+#define FUNCT_NAME tomsmocompDScaler_MMX
+#include "tomsmocomp/TomsMoCompAll.inc"
+#undef  IS_MMX
+#undef  SSE_TYPE
+#undef  FUNCT_NAME
+
+#define IS_3DNOW
+#define SSE_TYPE 3DNOW
+#define FUNCT_NAME tomsmocompDScaler_3DNOW
+#include "tomsmocomp/TomsMoCompAll.inc"
+#undef  IS_3DNOW
+#undef  SSE_TYPE
+#undef  FUNCT_NAME
+
+#define IS_SSE
+#define SSE_TYPE SSE
+#define FUNCT_NAME tomsmocompDScaler_SSE
+#include "tomsmocomp/TomsMoCompAll.inc"
+#undef  IS_SSE
+#undef  SSE_TYPE
+#undef  FUNCT_NAME
+
+
+
+void
+deinterlace_frame_di_tomsmocomp (GstDeinterlace2 * object)
+{
+  if (object->cpu_feature_flags & OIL_IMPL_FLAG_SSE) {
+    tomsmocomp_filter_sse (object);
+  } else if (object->cpu_feature_flags & OIL_IMPL_FLAG_3DNOW) {
+    tomsmocomp_filter_3dnow (object);
+  } else {
+    tomsmocomp_filter_mmx (object);
+  }
+}
+
+static deinterlace_method_t tomsmocompmethod = {
+  0,                            //DEINTERLACE_PLUGIN_API_VERSION,
+  "Motion Adaptive: Motion Search",
+  "AdaptiveSearch",
+  4,
+  OIL_IMPL_FLAG_MMX,
+  0,
+  0,
+  0,
+  0,
+  0,
+  0,
+  deinterlace_frame_di_tomsmocomp,
+  {"Uses heuristics to detect motion in the input",
+        "frames and reconstruct image detail where",
+        "possible.  Use this for high quality output",
+        "even on monitors set to an arbitrary refresh",
+        "rate.",
+        "",
+        "Motion search mode finds and follows motion",
+        "vectors for accurate interpolation.  This is",
+        "the TomsMoComp deinterlacer from DScaler.",
+      ""}
+};
+
+
+
+deinterlace_method_t *
+dscaler_tomsmocomp_get_method (void)
+{
+  tomsmocomp_init ();
+  return &tomsmocompmethod;
+}
+
+
+
+void
+tomsmocomp_init (void)
+{
+  SearchEffort = SearchEffortDefault;
+  UseStrangeBob = UseStrangeBobDefault;
+}
+
+void
+tomsmocomp_filter_mmx (GstDeinterlace2 * object)
+{
+  tomsmocompDScaler_MMX (object);
+}
+
+void
+tomsmocomp_filter_3dnow (GstDeinterlace2 * object)
+{
+  tomsmocompDScaler_3DNOW (object);
+}
+
+void
+tomsmocomp_filter_sse (GstDeinterlace2 * object)
+{
+  tomsmocompDScaler_SSE (object);
+}
diff --git a/gst/deinterlace2/tvtime/tomsmocomp.h b/gst/deinterlace2/tvtime/tomsmocomp.h
new file mode 100644
index 00000000..12127800
--- /dev/null
+++ b/gst/deinterlace2/tvtime/tomsmocomp.h
@@ -0,0 +1,61 @@
+/**
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ *
+ * 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, 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., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
+ */
+
+#ifndef TOMSMOCOMP_H_INCLUDED
+#define TOMSMOCOMP_H_INCLUDED
+
+#include "gstdeinterlace2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int  Search_Effort_0();
+int  Search_Effort_1();
+int  Search_Effort_3();
+int  Search_Effort_5();
+int  Search_Effort_9();
+int  Search_Effort_11();
+int  Search_Effort_13();
+int  Search_Effort_15();
+int  Search_Effort_19();
+int  Search_Effort_21();
+int  Search_Effort_Max();
+
+int  Search_Effort_0_SB();
+int  Search_Effort_1_SB();
+int  Search_Effort_3_SB();
+int  Search_Effort_5_SB();
+int  Search_Effort_9_SB();
+int  Search_Effort_11_SB();
+int  Search_Effort_13_SB();
+int  Search_Effort_15_SB();
+int  Search_Effort_19_SB();
+int  Search_Effort_21_SB();
+int  Search_Effort_Max_SB();
+
+void tomsmocomp_init( void );
+void tomsmocomp_filter_mmx( GstDeinterlace2 *object );
+void tomsmocomp_filter_3dnow( GstDeinterlace2 *object );
+void tomsmocomp_filter_sse( GstDeinterlace2 *object );
+
+#ifdef __cplusplus
+};
+#endif
+
+#endif /* TOMSMOCOMP_H_INCLUDED */
diff --git a/gst/deinterlace2/tvtime/vfir.c b/gst/deinterlace2/tvtime/vfir.c
new file mode 100644
index 00000000..bb42f5d3
--- /dev/null
+++ b/gst/deinterlace2/tvtime/vfir.c
@@ -0,0 +1,184 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Billy Biggs <vektor@dumbterm.net>
+ * Copyright (c) 2001, 2002, 2003 Fabrice Bellard.
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * This file contains code from ffmpeg, see http://ffmpeg.org/ (LGPL)
+ * and modifications by Billy Biggs.
+ *
+ * Relicensed for GStreamer from GPL to LGPL with permit from Billy Biggs.
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+#include <stdio.h>
+#if defined (__SVR4) && defined (__sun)
+# include <sys/int_types.h>
+#else
+# include <stdint.h>
+#endif
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "mmx.h"
+#include "speedy.h"
+#include "gstdeinterlace2.h"
+
+/**
+ * The MPEG2 spec uses a slightly harsher filter, they specify
+ * [-1 8 2 8 -1].  ffmpeg uses a similar filter but with more of
+ * a tendancy to blur than to use the local information.  The
+ * filter taps here are: [-1 4 2 4 -1].
+ */
+
+static void
+deinterlace_line (uint8_t * dst, uint8_t * lum_m4,
+    uint8_t * lum_m3, uint8_t * lum_m2,
+    uint8_t * lum_m1, uint8_t * lum, int size)
+{
+#ifdef HAVE_CPU_I386
+  mmx_t rounder;
+
+  rounder.uw[0] = 4;
+  rounder.uw[1] = 4;
+  rounder.uw[2] = 4;
+  rounder.uw[3] = 4;
+  pxor_r2r (mm7, mm7);
+  movq_m2r (rounder, mm6);
+
+  for (; size > 3; size -= 4) {
+    movd_m2r (lum_m4[0], mm0);
+    movd_m2r (lum_m3[0], mm1);
+    movd_m2r (lum_m2[0], mm2);
+    movd_m2r (lum_m1[0], mm3);
+    movd_m2r (lum[0], mm4);
+    punpcklbw_r2r (mm7, mm0);
+    punpcklbw_r2r (mm7, mm1);
+    punpcklbw_r2r (mm7, mm2);
+    punpcklbw_r2r (mm7, mm3);
+    punpcklbw_r2r (mm7, mm4);
+    paddw_r2r (mm3, mm1);
+    psllw_i2r (1, mm2);
+    paddw_r2r (mm4, mm0);
+    psllw_i2r (2, mm1);         // 2
+    paddw_r2r (mm6, mm2);
+    paddw_r2r (mm2, mm1);
+    psubusw_r2r (mm0, mm1);
+    psrlw_i2r (3, mm1);         // 3
+    packuswb_r2r (mm7, mm1);
+    movd_r2m (mm1, dst[0]);
+    lum_m4 += 4;
+    lum_m3 += 4;
+    lum_m2 += 4;
+    lum_m1 += 4;
+    lum += 4;
+    dst += 4;
+  }
+  emms ();
+#else
+    /**
+     * C implementation.
+     */
+  int sum;
+
+  for (; size > 0; size--) {
+    sum = -lum_m4[0];
+    sum += lum_m3[0] << 2;
+    sum += lum_m2[0] << 1;
+    sum += lum_m1[0] << 2;
+    sum += -lum[0];
+    dst[0] = (sum + 4) >> 3;    // This needs to be clipped at 0 and 255: cm[(sum + 4) >> 3];
+    lum_m4++;
+    lum_m3++;
+    lum_m2++;
+    lum_m1++;
+    lum++;
+    dst++;
+  }
+#endif
+}
+
+
+/**
+ * The commented-out method below that uses the bottom_field member is more
+ * like the filter as specified in the MPEG2 spec, but it doesn't seem to
+ * have the desired effect.
+ */
+
+static void
+deinterlace_scanline_vfir (GstDeinterlace2 * object,
+    deinterlace_scanline_data_t * data, uint8_t * output)
+{
+  deinterlace_line (output, data->tt1, data->t0, data->m1, data->b0, data->bb1,
+      object->frame_width * 2);
+  // blit_packed422_scanline( output, data->m1, width );
+}
+
+static void
+copy_scanline (GstDeinterlace2 * object,
+    deinterlace_scanline_data_t * data, uint8_t * output)
+{
+  blit_packed422_scanline (output, data->m0, object->frame_width);
+  /*
+     if( data->bottom_field ) {
+     deinterlace_line( output, data->tt2, data->t1, data->m2, data->b1, data->bb2, width*2 );
+     } else {
+     deinterlace_line( output, data->tt0, data->t1, data->m0, data->b1, data->bb0, width*2 );
+     }
+   */
+}
+
+
+static deinterlace_method_t vfirmethod = {
+  0,                            //DEINTERLACE_PLUGIN_API_VERSION,
+  "Blur: Vertical",
+  "BlurVertical",
+  2,
+#ifdef HAVE_CPU_I386
+  OIL_IMPL_FLAG_MMXEXT,
+#else
+  0,
+#endif
+  0,
+  0,
+  0,
+  1,
+  deinterlace_scanline_vfir,
+  copy_scanline,
+  0,
+  {"Avoids flicker by blurring consecutive frames",
+        "of input.  Use this if you want to run your",
+        "monitor at an arbitrary refresh rate and not",
+        "use much CPU, and are willing to sacrifice",
+        "detail.",
+        "",
+        "Vertical mode blurs favouring the most recent",
+        "field for less visible trails.  From the",
+        "deinterlacer filter in ffmpeg.",
+      ""}
+};
+
+deinterlace_method_t *
+dscaler_vfir_get_method (void)
+{
+  return &vfirmethod;
+}
diff --git a/gst/deinterlace2/tvtime/x86-64_macros.inc b/gst/deinterlace2/tvtime/x86-64_macros.inc
new file mode 100644
index 00000000..2e9df758
--- /dev/null
+++ b/gst/deinterlace2/tvtime/x86-64_macros.inc
@@ -0,0 +1,82 @@
+/*
+ *
+ * GStreamer
+ * Copyright (C) 2004 Dirk Ziegelmeier <dziegel@gmx.de>
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Library General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library 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
+ * Library General Public License for more details.
+ *
+ * You should have received a copy of the GNU Library General Public
+ * License along with this library; if not, write to the
+ * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
+ * Boston, MA 02111-1307, USA.
+ */
+
+/*
+ * 
+ * See: http://bugzilla.gnome.org/show_bug.cgi?id=163578
+ */
+
+/*
+ * This file is copied from TVTIME's sources.
+ * Original author: Achim Schneider <batchall@mordor.ch>
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#ifndef XAX
+
+#if defined (HAVE_CPU_I386) && !defined(HAVE_CPU_X86_64)
+
+#define XAX   "eax"
+#define XBX   "ebx"
+#define XCX   "ecx"
+#define XDX   "edx"
+#define XSI   "esi"
+#define XDI   "edi"
+#define XSP   "esp"
+#define MOVX  "movl"
+#define LEAX  "leal"
+#define DECX  "decl"
+#define PUSHX "pushl"
+#define POPX  "popl"
+#define CMPX  "cmpl"
+#define ADDX  "addl"
+#define SHLX  "shll"
+#define SHRX  "shrl"
+#define SUBX  "subl"
+
+#elif defined (HAVE_CPU_X86_64)
+
+#define XAX   "rax"
+#define XBX   "rbx"
+#define XCX   "rcx"
+#define XDX   "rdx"
+#define XSI   "rsi"
+#define XDI   "rdi"
+#define XSP   "rsp"
+#define MOVX  "movq"
+#define LEAX  "leaq"
+#define DECX  "decq"
+#define PUSHX "pushq"
+#define POPX  "popq"
+#define CMPX  "cmpq"
+#define ADDX  "addq"
+#define SHLX  "shlq"
+#define SHRX  "shrq"
+#define SUBX  "subq"
+
+#else
+#error Undefined architecture. Define either ARCH_X86 or ARCH_X86_64.
+#endif
+
+#endif