gstreamer/subprojects/gst-plugins-good/gst/deinterlace/tvtime/greedyh.asm

/*
 *
 * GStreamer
 * Copyright (c) 2001 Tom Barry.  All rights reserved.
 * Copyright (C) 2008,2010 Sebastian Dröge <slomo@collabora.co.uk>
 *
 * 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 aglong with this library; if not, write to the
 * Free Software Foundation, Inc., 51 Franklin St, Fifth Floor,
 * Boston, MA 02110-1301, 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"

static void
FUNCT_NAME_YUY2 (GstDeinterlaceMethodGreedyH *self, const guint8 * L1, const guint8 * L2, const guint8 * L3, const guint8 * L2P, guint8 * Dest, gint width)
{

  // in tight loop some vars are accessed faster in local storage
  gint64 YMask = 0x00ff00ff00ff00ffull;        // to keep only luma
  gint64 UVMask = 0xff00ff00ff00ff00ull;       // to keep only chroma
  gint64 ShiftMask = 0xfefefefefefefefeull;    // to avoid shifting chroma to luma
  gint64 QW256 = 0x0100010001000100ull;        // 4 256's
  gint64 MaxComb;
  gint64 MotionThreshold;
  gint64 MotionSense;
  gint64 i;
  glong LoopCtr;
  glong oldbx = 0;

  gint64 QW256B;
  gint64 LastAvg = 0;          //interp value from left qword

  // FIXME: Use C implementation if the width is not a multiple of 4
  // Do something more optimal later
  if (width % 4 != 0)
    C_FUNCT_YUY2 (self, L1, L2, L3, L2P, Dest, width);

  // Set up our two parms that are actually evaluated for each pixel
  i = self->max_comb;
  MaxComb =
      i << 56 | i << 48 | i << 40 | i << 32 | i << 24 | i << 16 | i << 8 | i;

  i = self->motion_threshold;    // scale to range of 0-257
  MotionThreshold = i << 48 | i << 32 | i << 16 | i | UVMask;

  i = self->motion_sense;        // scale to range of 0-257
  MotionSense = i << 48 | i << 32 | i << 16 | i;

  i = 0xffffffff - 256;
  QW256B = i << 48 | i << 32 | i << 16 | i;     // save a couple instr on PMINSW instruct.

  LoopCtr = width / 8 - 1;       // there are LineLength / 4 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 IS_MMXEXT
      "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"
      "emms\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,
      "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",
#ifdef __MMX__
      "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
#endif
      "memory", "cc");
}

static void
FUNCT_NAME_UYVY (GstDeinterlaceMethodGreedyH *self, const guint8 * L1, const guint8 * L2, const guint8 * L3, const guint8 * L2P, guint8 * Dest, gint width)
{

  // in tight loop some vars are accessed faster in local storage
  gint64 YMask = 0xff00ff00ff00ff00ull;       // to keep only luma
  gint64 UVMask = 0x00ff00ff00ff00ffull;        // to keep only chroma
  gint64 ShiftMask = 0xfefefefefefefefeull;    // to avoid shifting chroma to luma
  gint64 QW256 = 0x0100010001000100ull;        // 4 256's
  gint64 MaxComb;
  gint64 MotionThreshold;
  gint64 MotionSense;
  gint64 i;
  glong LoopCtr;
  glong oldbx = 0;

  gint64 QW256B;
  gint64 LastAvg = 0;          //interp value from left qword

  // FIXME: Use C implementation if the width is not a multiple of 4
  // Do something more optimal later
  if (width % 4 != 0)
    C_FUNCT_UYVY (self, L1, L2, L3, L2P, Dest, width);

  // Set up our two parms that are actually evaluated for each pixel
  i = self->max_comb;
  MaxComb =
      i << 56 | i << 48 | i << 40 | i << 32 | i << 24 | i << 16 | i << 8 | i;

  i = self->motion_threshold;    // scale to range of 0-257
  MotionThreshold = i << 48 | i << 32 | i << 16 | i | UVMask;

  i = self->motion_sense;        // scale to range of 0-257
  MotionSense = i << 48 | i << 32 | i << 16 | i;

  i = 0xffffffff - 256;
  QW256B = i << 48 | i << 32 | i << 16 | i;     // save a couple instr on PMINSW instruct.

  LoopCtr = width / 8 - 1;       // there are LineLength / 4 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
      "psrlw   $8,           %%mm0\n\t" // div by 256 to get weighted avg
      "pmullw  %[MotionSense], %%mm0\n\t"       // mul by user factor, keep low 16 bits
      "movq    %[QW256], %%mm7\n\t"
#ifdef IS_MMXEXT
      "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
      "psrlw   $8,           %%mm4\n\t" // div by 256 to get weighted avg
      "pmullw  %%mm7,        %%mm4\n\t" // use more weave for less motion
      "pand    %[YMask],     %%mm6\n\t" // keep only luma from calc'd value
      "psrlw   $8,           %%mm6\n\t" // div by 256 to get weighted avg
      "pmullw  %%mm0,        %%mm6\n\t" // use more bob for large motion
      "paddusw %%mm6,        %%mm4\n\t" // combine
      "pand    %[YMask],     %%mm4\n\t" // keep only luma from calc'd value
      // 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"
      "emms\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,
      "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)",
#ifdef __MMX__
      "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7",
#endif
      "memory", "cc");
}