gstreamer/gst/deinterlace2/tvtime/tomsmocomp/SearchLoopTop.inc
Sebastian Dröge 81f8895cb4 gst/deinterlace2/tvtime/tomsmocomp/: Unroll the loop to handle two bytes at once. This should give a small speedup an...
Original commit message from CVS:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopBottom.inc:
* gst/deinterlace2/tvtime/tomsmocomp/SearchLoopTop.inc:
* gst/deinterlace2/tvtime/tomsmocomp/StrangeBob.inc:
* gst/deinterlace2/tvtime/tomsmocomp/WierdBob.inc:
Unroll the loop to handle two bytes at once. This should give
a small speedup and makes it possible to handle chroma and luma
different which is needed later.
2008-08-26 12:33:16 +00:00

254 lines
8.8 KiB
C++

// -*- c++ -*-
unsigned char* pDest;
const unsigned char* pSrcP;
const unsigned char* pSrc;
const unsigned char* pBob;
const unsigned char* pBobP;
// long is int32 on ARCH_368, int64 on ARCH_AMD64. Declaring it this way
// saves a lot of xor's to delete 64bit garbage.
#if defined(DBL_RESIZE) || defined(USE_FOR_DSCALER)
long src_pitch2 = src_pitch; // even & odd lines are not interleaved in DScaler
#else
long src_pitch2 = 2 * src_pitch; // even & odd lines are interleaved in Avisynth
#endif
long dst_pitch2 = 2 * dst_pitch;
long y;
long Last8;
pSrc = pWeaveSrc; // points 1 weave line above
pSrcP = pWeaveSrcP; // "
#ifdef DBL_RESIZE
#ifdef USE_VERTICAL_FILTER
pDest = pWeaveDest + dst_pitch2;
#else
pDest = pWeaveDest + 3*dst_pitch;
#endif
#else
#ifdef USE_VERTICAL_FILTER
pDest = pWeaveDest + dst_pitch;
#else
pDest = pWeaveDest + dst_pitch2;
#endif
#endif
if (TopFirst)
{
pBob = pCopySrc + src_pitch2; // remember one weave line just copied previously
pBobP = pCopySrcP + src_pitch2;
}
else
{
pBob = pCopySrc;
pBobP = pCopySrcP;
}
#ifndef IS_C
#ifndef _pBob
#define _pBob "%0"
#define _src_pitch2 "%1"
#define _ShiftMask "%2"
#define _pDest "%3"
#define _dst_pitchw "%4"
#define _Last8 "%5"
#define _pSrc "%6"
#define _pSrcP "%7"
#define _pBobP "%8"
#define _DiffThres "%9"
#define _Min_Vals "%10"
#define _Max_Vals "%11"
#define _FOURS "%12"
#define _TENS "%13"
#define _ONES "%14"
#define _UVMask "%15"
#define _Max_Mov "%16"
#define _YMask "%17"
#define _oldbx "%18"
#endif
Last8 = (rowsize-8);
for (y=1; y < FldHeight-1; y++)
{
long dst_pitchw = dst_pitch; // local stor so asm can ref
int64_t Max_Mov = 0x0404040404040404ull;
int64_t DiffThres = 0x0f0f0f0f0f0f0f0full;
int64_t YMask = 0x00ff00ff00ff00ffull; // keeps only luma
int64_t UVMask = 0xff00ff00ff00ff00ull; // keeps only chroma
int64_t TENS = 0x0a0a0a0a0a0a0a0aull;
int64_t FOURS = 0x0404040404040404ull;
int64_t ONES = 0x0101010101010101ull;
int64_t Min_Vals = 0x0000000000000000ull;
int64_t Max_Vals = 0x0000000000000000ull;
int64_t ShiftMask = 0xfefffefffefffeffull;
long oldbx;
// pretend it's indented -->>
__asm__ __volatile__
(
// Loop general reg usage
//
// XAX - pBobP, then pDest
// XBX - pBob
// XCX - src_pitch2
// XDX - current offset
// XDI - prev weave pixels, 1 line up
// XSI - next weave pixels, 1 line up
// Save "XBX" (-fPIC)
MOVX" %%"XBX", "_oldbx"\n\t"
// simple bob first 8 bytes
MOVX" "_pBob", %%"XBX"\n\t"
MOVX" "_src_pitch2", %%"XCX"\n\t"
#ifdef USE_VERTICAL_FILTER
"movq (%%"XBX"), %%mm0\n\t"
"movq (%%"XBX", %%"XCX"), %%mm1\n\t" //, qword ptr["XBX"+"XCX"]
"movq %%mm0, %%mm2\n\t"
V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between
V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way
V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way
MOVX" "_pDest", %%"XDI"\n\t"
MOVX" "_dst_pitchw", %%"XAX"\n\t"
V_MOVNTQ ("(%%"XDI")", "%%mm0")
V_MOVNTQ ("(%%"XDI", %%"XAX")", "%%mm1") // qword ptr["XDI"+"XAX"], mm1
// simple bob last 8 bytes
MOVX" "_Last8", %%"XDX"\n\t"
LEAX" (%%"XBX", %%"XDX"), %%"XSI"\n\t" // ["XBX"+"XDX"]
"movq (%%"XSI"), %%mm0\n\t"
"movq (%%"XSI", %%"XCX"), %%mm1\n\t" // qword ptr["XSI"+"XCX"]
"movq %%mm0, %%mm2\n\t"
V_PAVGB ("%%mm2", "%%mm1", "%%mm3", _ShiftMask) // halfway between
V_PAVGB ("%%mm0", "%%mm2", "%%mm3", _ShiftMask) // 1/4 way
V_PAVGB ("%%mm1", "%%mm2", "%%mm3", _ShiftMask) // 3/4 way
ADDX" %%"XDX", %%"XDI"\n\t" // last 8 bytes of dest
V_MOVNTQ ("%%"XDI"", "%%mm0")
V_MOVNTQ ("(%%"XDI", %%"XAX")", "%%mm1") // qword ptr["XDI"+"XAX"], mm1)
#else
"movq (%%"XBX"), %%mm0\n\t"
// pavgb mm0, qword ptr["XBX"+"XCX"]
V_PAVGB ("%%mm0", "(%%"XBX", %%"XCX")", "%%mm2", _ShiftMask) // qword ptr["XBX"+"XCX"], mm2, ShiftMask)
MOVX" "_pDest", %%"XDI"\n\t"
V_MOVNTQ ("(%%"XDI")", "%%mm0")
// simple bob last 8 bytes
MOVX" "_Last8", %%"XDX"\n\t"
LEAX" (%%"XBX", %%"XDX"), %%"XSI"\n\t" //"XSI", ["XBX"+"XDX"]
"movq (%%"XSI"), %%mm0\n\t"
// pavgb mm0, qword ptr["XSI"+"XCX"]
V_PAVGB ("%%mm0", "(%%"XSI", %%"XCX")", "%%mm2", _ShiftMask) // qword ptr["XSI"+"XCX"], mm2, ShiftMask)
V_MOVNTQ ("(%%"XDI", %%"XDX")", "%%mm0") // qword ptr["XDI"+"XDX"], mm0)
#endif
// now loop and get the middle qwords
MOVX" "_pSrc", %%"XSI"\n\t"
MOVX" "_pSrcP", %%"XDI"\n\t"
MOVX" $8, %%"XDX"\n\t" // curr offset longo all lines
"1:\n\t"
MOVX" "_pBobP", %%"XAX"\n\t"
ADDX" $8, %%"XDI"\n\t"
ADDX" $8, %%"XSI"\n\t"
ADDX" $8, %%"XBX"\n\t"
ADDX" %%"XDX", %%"XAX"\n\t"
#ifdef USE_STRANGE_BOB
#include "StrangeBob.inc"
#else
#include "WierdBob.inc"
#endif
// For non-SSE2:
// through out most of the rest of this loop we will maintain
// mm4 our min bob value
// mm5 best weave pixels so far
// mm6 our max Bob value
// mm7 best weighted pixel ratings so far
// We will keep a slight bias to using the weave pixels
// from the current location, by rating them by the min distance
// from the Bob value instead of the avg distance from that value.
// our best and only rating so far
"pcmpeqb %%mm7, %%mm7\n\t" // ffff, say we didn't find anything good yet
#else
Last8 = (rowsize - 4);
for (y=1; y < FldHeight-1; y++)
{
#ifdef USE_STRANGE_BOB
long DiffThres = 0x0f;
#endif
#ifndef SKIP_SEARCH
long weave[2], MaxVals[2], MinVals[2];
#endif
long diff[2], best[2], avg[2], diff2[2], out[2], x;
#ifdef USE_VERTICAL_FILTER
pDest[0] = (3 * pBob[0] + pBob[src_pitch2]) / 4;
pDest[1] = (3 * pBob[1] + pBob[src_pitch2 + 1]) / 4;
pDest[2] = (3 * pBob[2] + pBob[src_pitch2 + 2]) / 4;
pDest[3] = (3 * pBob[3] + pBob[src_pitch2 + 3]) / 4;
pDest[dst_pitchw] = (pBob[0] + 3 * pBob[src_pitch2]) / 4;
pDest[dst_pitchw + 1] = (pBob[1] + 3 * pBob[src_pitch2 + 1]) / 4;
pDest[dst_pitchw + 2] = (pBob[2] + 3 * pBob[src_pitch2 + 2]) / 4;
pDest[dst_pitchw + 3] = (pBob[3] + 3 * pBob[src_pitch2 + 3]) / 4;
// simple bob last byte
pDest[Last8] = (3 * pBob[Last8] + pBob[Last8 + src_pitch2]) / 4;
pDest[Last8 + 1] = (3 * pBob[Last8 + 1] + pBob[Last8 + src_pitch2 + 1]) / 4;
pDest[Last8 + 2] = (3 * pBob[Last8 + 2] + pBob[Last8 + src_pitch2 + 2]) / 4;
pDest[Last8 + 3] = (3 * pBob[Last8 + 3] + pBob[Last8 + src_pitch2 + 3]) / 4;
pDest[Last8 + src_pitch2] = (pBob[Last8] + 3 * pBob[Last8 + src_pitch2]) / 4;
pDest[Last8 + src_pitch2 + 1] = (pBob[Last8 + 1] + 3 * pBob[Last8 + src_pitch2 + 1]) / 4;
pDest[Last8 + src_pitch2 + 2] = (pBob[Last8 + 2] + 3 * pBob[Last8 + src_pitch2 + 2]) / 4;
pDest[Last8 + src_pitch2 + 3] = (pBob[Last8 + 3] + 3 * pBob[Last8 + src_pitch2 + 3]) / 4;
#else
pDest[0] = (pBob[0] + pBob[src_pitch2 + 1]) / 2;
pDest[1] = (pBob[1] + pBob[src_pitch2 + 1]) / 2;
pDest[2] = (pBob[2] + pBob[src_pitch2 + 2]) / 2;
pDest[3] = (pBob[3] + pBob[src_pitch2 + 3]) / 2;
// simple bob last byte
pDest[Last8] = (pBob[Last8] + pBob[Last8 + src_pitch2]) / 2;
pDest[Last8 + 1] = (pBob[Last8 + 1] + pBob[Last8 + src_pitch2 + 1]) / 2;
pDest[Last8 + 2] = (pBob[Last8 + 2] + pBob[Last8 + src_pitch2 + 2]) / 2;
pDest[Last8 + 3] = (pBob[Last8 + 3] + pBob[Last8 + src_pitch2 + 3]) / 2;
#endif
pBob += 4;
pBobP += 4;
pSrc += 4;
pSrcP += 4;
for (x=4; x < Last8; x += 2) {
#ifdef USE_STRANGE_BOB
#include "StrangeBob.inc"
#else
#include "WierdBob.inc"
#endif
// We will keep a slight bias to using the weave pixels
// from the current location, by rating them by the min distance
// from the Bob value instead of the avg distance from that value.
// our best and only rating so far
diff[0] = diff[1] = 255;
#endif