gmc.c

从FFMPEG转换而来的H264解码程序,VC下编译..

		vDst[i] = (uint8_t)f0;
	}
	uDst += dststride;
	vDst += dststride;
	}
}

static
void get_average_mv_1pt_C(const NEW_GMC_DATA * const Dsp, VECTOR * const mv,
							int x, int y, int qpel)
{
	mv->x = RSHIFT(Dsp->Uo<<qpel, 3);
	mv->y = RSHIFT(Dsp->Vo<<qpel, 3);
}

#if defined(ARCH_IS_IA32)
/* *************************************************************
 * MMX core function
 */

static
void (*GMC_Core_Lin_8)(uint8_t *Dst, const uint16_t * Offsets, 
                       const uint8_t * const Src0, const int BpS, const int Rounder) = 0;

extern void xvid_GMC_Core_Lin_8_mmx(uint8_t *Dst, const uint16_t * Offsets, 
                                    const uint8_t * const Src0, const int BpS, const int Rounder);

extern void xvid_GMC_Core_Lin_8_sse2(uint8_t *Dst, const uint16_t * Offsets, 
                                     const uint8_t * const Src0, const int BpS, const int Rounder);

/* *************************************************************/

static void GMC_Core_Non_Lin_8(uint8_t *Dst, 
                               const uint16_t * Offsets,
                               const uint8_t * const Src0, const int srcstride,
                               const int Rounder)
{
  int i;
  for(i=0; i<8; ++i)
  {
    uint32_t u = Offsets[i   ];
    uint32_t v = Offsets[i+16];
    const uint32_t ri = MTab[u&0x0f];
    const uint32_t rj = MTab[v&0x0f];
    uint32_t f0, f1;
    const uint8_t * const Src = Src0 + (u>>4) + (v>>4)*srcstride;
    f0  = Src[0];
    f0 |= Src[1] << 16;
    f1  = Src[srcstride +0];
    f1 |= Src[srcstride +1] << 16;
    f0 = (ri*f0)>>16;
    f1 = (ri*f1) & 0x0fff0000;
    f0 |= f1;
    f0 = ( rj*f0 + Rounder ) >> 24;
    Dst[i] = (uint8_t)f0;
  }
}

//////////////////////////////////////////////////////////

static
void Predict_16x16_mmx(const NEW_GMC_DATA * const This,
                       uint8_t *dst, const uint8_t *src,
                       int dststride, int srcstride, int x, int y, int rounding)
{
  const int W = This->sW;
  const int H = This->sH;
  const int rho = 3 - This->accuracy;
  const int Rounder = ( 128 - (rounding<<(2*rho)) ) << 16;
  const uint32_t W2 = W<<(16-rho);
  const uint32_t H2 = H<<(16-rho);
  
  const int dUx = This->dU[0];
  const int dVx = This->dV[0];
  const int dUy = This->dU[1];
  const int dVy = This->dV[1];

  int Uo = This->Uo + 16*(dUy*y + dUx*x);
  int Vo = This->Vo + 16*(dVy*y + dVx*x);

  int i, j;

  DECLARE_ALIGNED_MATRIX(Offsets, 2,16, uint16_t, CACHE_LINE);
  for(j=16; j>0; --j)
  {
    int32_t U = Uo, V = Vo;
    Uo += dUy; Vo += dVy;
    if ( W2>(uint32_t)U && W2>(uint32_t)(U+15*dUx) &&
         H2>(uint32_t)V && H2>(uint32_t)(V+15*dVx) )
    {
      uint32_t UV1, UV2;
      for(i=0; i<16; ++i)
      {
        uint32_t u = ( U >> 16 ) << rho;
        uint32_t v = ( V >> 16 ) << rho;
        U += dUx;  V += dVx;
        Offsets[   i] = u;
        Offsets[16+i] = v;
      }
          // batch 8 input pixels when linearity says it's ok

      UV1 = (Offsets[0] | (Offsets[16]<<16)) & 0xfff0fff0U;
      UV2 = (Offsets[7] | (Offsets[23]<<16)) & 0xfff0fff0U;
      if (UV1+7*16==UV2)
        GMC_Core_Lin_8(dst,    Offsets,    src + (Offsets[0]>>4) + (Offsets[16]>>4)*srcstride, srcstride, Rounder);
      else
        GMC_Core_Non_Lin_8(dst,   Offsets,   src, srcstride, Rounder);
      UV1 = (Offsets[ 8] | (Offsets[24]<<16)) & 0xfff0fff0U;
      UV2 = (Offsets[15] | (Offsets[31]<<16)) & 0xfff0fff0U;
      if (UV1+7*16==UV2)
        GMC_Core_Lin_8(dst+8,  Offsets+8,  src + (Offsets[8]>>4) + (Offsets[24]>>4)*srcstride, srcstride, Rounder);
      else
        GMC_Core_Non_Lin_8(dst+8, Offsets+8, src, srcstride, Rounder);
    }
    else
    {
      for(i=0; i<16; ++i)
      {
        int u = ( U >> 16 ) << rho;
        int v = ( V >> 16 ) << rho;
        U += dUx; V += dVx;

        Offsets[   i] = (u<0) ? 0 : (u>=W) ? W : u;
        Offsets[16+i] = (v<0) ? 0 : (v>=H) ? H : v;
      }
        // due to boundary clipping, we cannot infer the 8-pixels batchability
        // simply by using the linearity. Oh well, not a big deal...
      GMC_Core_Non_Lin_8(dst,   Offsets,   src, srcstride, Rounder);
      GMC_Core_Non_Lin_8(dst+8, Offsets+8, src, srcstride, Rounder);
    }
    dst += dststride;
  }
}

static
void Predict_8x8_mmx(const NEW_GMC_DATA * const This,
                     uint8_t *uDst, const uint8_t *uSrc,
                     uint8_t *vDst, const uint8_t *vSrc,
                     int dststride, int srcstride, int x, int y, int rounding)
{
  const int W   = This->sW >> 1;
  const int H   = This->sH >> 1;
  const int rho = 3-This->accuracy;
  const int32_t Rounder = ( 128 - (rounding<<(2*rho)) ) << 16;
  const uint32_t W2 = W<<(16-rho);
  const uint32_t H2 = H<<(16-rho);

  const int dUx = This->dU[0];
  const int dVx = This->dV[0];
  const int dUy = This->dU[1];
  const int dVy = This->dV[1];

  int Uo = This->Uco + 8*(dUy*y + dUx*x);
  int Vo = This->Vco + 8*(dVy*y + dVx*x);

  DECLARE_ALIGNED_MATRIX(Offsets, 2,16, uint16_t, CACHE_LINE);
  int i, j;
  for(j=8; j>0; --j)
  {
    int32_t U = Uo, V = Vo;
    Uo += dUy; Vo += dVy;
    if ( W2>(uint32_t)U && W2>(uint32_t)(U+15*dUx) &&
         H2>(uint32_t)V && H2>(uint32_t)(V+15*dVx) )
    {
      uint32_t UV1, UV2;
      for(i=0; i<8; ++i)
      {
        int32_t u = ( U >> 16 ) << rho;
        int32_t v = ( V >> 16 ) << rho;
        U += dUx; V += dVx;
        Offsets[   i] = u;
        Offsets[16+i] = v;
      }

          // batch 8 input pixels when linearity says it's ok
			UV1 = (Offsets[ 0] | (Offsets[16]<<16)) & 0xfff0fff0U;
			UV2 = (Offsets[ 7] | (Offsets[23]<<16)) & 0xfff0fff0U;
			if (UV1+7*16==UV2)
      {
				const uint32_t Off = (Offsets[0]>>4) + (Offsets[16]>>4)*srcstride;
				GMC_Core_Lin_8(uDst, Offsets, uSrc+Off, srcstride, Rounder);
				GMC_Core_Lin_8(vDst, Offsets, vSrc+Off, srcstride, Rounder);
      }
      else {
        GMC_Core_Non_Lin_8(uDst, Offsets, uSrc, srcstride, Rounder);
        GMC_Core_Non_Lin_8(vDst, Offsets, vSrc, srcstride, Rounder);
      }
    }
    else
    {
      for(i=0; i<8; ++i)
      {
        int u = ( U >> 16 ) << rho;
        int v = ( V >> 16 ) << rho;
        U += dUx; V += dVx;
        Offsets[   i] = (u<0) ? 0 : (u>=W) ? W : u;
        Offsets[16+i] = (v<0) ? 0 : (v>=H) ? H : v;
      }
      GMC_Core_Non_Lin_8(uDst, Offsets, uSrc, srcstride, Rounder);
      GMC_Core_Non_Lin_8(vDst, Offsets, vSrc, srcstride, Rounder);
    }
    uDst += dststride;
    vDst += dststride;
  }
}

#endif /* ARCH_IS_IA32 */

/* *************************************************************
 * will initialize internal pointers
 */

void init_GMC(const unsigned int cpu_flags)
{
      Predict_16x16_func = Predict_16x16_C;
      Predict_8x8_func   = Predict_8x8_C;

#if defined(ARCH_IS_IA32)
      if ((cpu_flags & XVID_CPU_MMX)   || (cpu_flags & XVID_CPU_MMXEXT)   ||
          (cpu_flags & XVID_CPU_3DNOW) || (cpu_flags & XVID_CPU_3DNOWEXT) ||
          (cpu_flags & XVID_CPU_SSE)   || (cpu_flags & XVID_CPU_SSE2))
	{
	   Predict_16x16_func = Predict_16x16_mmx;
	   Predict_8x8_func   = Predict_8x8_mmx;
	   if (cpu_flags & XVID_CPU_SSE2)
	     GMC_Core_Lin_8 = xvid_GMC_Core_Lin_8_sse2;
	   else
             GMC_Core_Lin_8 = xvid_GMC_Core_Lin_8_mmx;
	}
#endif
}

/* *************************************************************
 * Warning! It's Accuracy being passed, not 'resolution'!
 */

void generate_GMCparameters( int nb_pts, const int accuracy,
								 const WARPPOINTS *const pts,
								 const int width, const int height,
								 NEW_GMC_DATA *const gmc)
{
	gmc->sW = width	<< 4;
	gmc->sH = height << 4;
	gmc->accuracy = accuracy;
	gmc->num_wp = nb_pts;

	/* reduce the number of points, if possible */
	if (nb_pts<2 || (pts->duv[2].x==0 && pts->duv[2].y==0 && pts->duv[1].x==0 && pts->duv[1].y==0 )) {
  	if (nb_pts<2 || (pts->duv[1].x==0 && pts->duv[1].y==0)) {
	  	if (nb_pts<1 || (pts->duv[0].x==0 && pts->duv[0].y==0)) {
		    nb_pts = 0;
  		}
	  	else nb_pts = 1;
  	}
	  else nb_pts = 2;
  }

	/* now, nb_pts stores the actual number of points required for interpolation */

	if (nb_pts<=1)
	{
	if (nb_pts==1) {
		/* store as 4b fixed point */
		gmc->Uo = pts->duv[0].x << accuracy;
		gmc->Vo = pts->duv[0].y << accuracy;
		gmc->Uco = ((pts->duv[0].x>>1) | (pts->duv[0].x&1)) << accuracy;	 /* DIV2RND() */
		gmc->Vco = ((pts->duv[0].y>>1) | (pts->duv[0].y&1)) << accuracy;	 /* DIV2RND() */
	}
	else {	/* zero points?! */
		gmc->Uo	= gmc->Vo	= 0;
		gmc->Uco = gmc->Vco = 0;
	}

	gmc->predict_16x16	= Predict_1pt_16x16_C;
	gmc->predict_8x8	= Predict_1pt_8x8_C;
	gmc->get_average_mv = get_average_mv_1pt_C;
	}
	else {		/* 2 or 3 points */
	const int rho	 = 3 - accuracy;	/* = {3,2,1,0} for Acc={0,1,2,3} */
	int Alpha = log2bin(width-1);
	int Ws = 1 << Alpha;

	gmc->dU[0] = 16*Ws + RDIV( 8*Ws*pts->duv[1].x, width );	 /* dU/dx */
	gmc->dV[0] =		 RDIV( 8*Ws*pts->duv[1].y, width );	 /* dV/dx */

	if (nb_pts==2) {
		gmc->dU[1] = -gmc->dV[0];	/* -Sin */
		gmc->dV[1] =	gmc->dU[0] ;	/* Cos */
	}
	else
	{
		const int Beta = log2bin(height-1);
		const int Hs = 1<<Beta;
		gmc->dU[1] =		 RDIV( 8*Hs*pts->duv[2].x, height );	 /* dU/dy */
		gmc->dV[1] = 16*Hs + RDIV( 8*Hs*pts->duv[2].y, height );	 /* dV/dy */
		if (Beta>Alpha) {
		gmc->dU[0] <<= (Beta-Alpha);
		gmc->dV[0] <<= (Beta-Alpha);
		Alpha = Beta;
		Ws = Hs;
		}
		else {
		gmc->dU[1] <<= Alpha - Beta;
		gmc->dV[1] <<= Alpha - Beta;
		}
	}
	/* upscale to 16b fixed-point */
	gmc->dU[0] <<= (16-Alpha - rho);
	gmc->dU[1] <<= (16-Alpha - rho);
	gmc->dV[0] <<= (16-Alpha - rho);
	gmc->dV[1] <<= (16-Alpha - rho);

	gmc->Uo	= ( pts->duv[0].x	 <<(16+ accuracy)) + (1<<15);
	gmc->Vo	= ( pts->duv[0].y	 <<(16+ accuracy)) + (1<<15);
	gmc->Uco = ((pts->duv[0].x-1)<<(17+ accuracy)) + (1<<17);
	gmc->Vco = ((pts->duv[0].y-1)<<(17+ accuracy)) + (1<<17);
	gmc->Uco = (gmc->Uco + gmc->dU[0] + gmc->dU[1])>>2;
	gmc->Vco = (gmc->Vco + gmc->dV[0] + gmc->dV[1])>>2;

	gmc->predict_16x16	= Predict_16x16_func;
	gmc->predict_8x8	= Predict_8x8_func;
	gmc->get_average_mv = get_average_mv_C;
	}
}

/* *******************************************************************
 * quick and dirty routine to generate the full warped image
 * (pGMC != NULL) or just all average Motion Vectors (pGMC == NULL) */

void
generate_GMCimage(	const NEW_GMC_DATA *const gmc_data, /* [input] precalculated data */
					const IMAGE *const pRef,		/* [input] */
					const int mb_width,
					const int mb_height,
					const int stride,
					const int stride2,
					const int fcode, 				/* [input] some parameters... */
						const int32_t quarterpel,		/* [input] for rounding avgMV */
					const int reduced_resolution,	/* [input] ignored */
					const int32_t rounding,			/* [input] for rounding image data */
					MACROBLOCK *const pMBs, 		/* [output] average motion vectors */
					IMAGE *const pGMC)				/* [output] full warped image */
{

	unsigned int mj,mi;
	VECTOR avgMV;

	for (mj = 0; mj < (unsigned int)mb_height; mj++)
		for (mi = 0; mi < (unsigned int)mb_width; mi++) {
			const int mbnum = mj*mb_width+mi;
			if (pGMC)
			{
				gmc_data->predict_16x16(gmc_data,
							pGMC->y + mj*16*stride + mi*16, pRef->y,
							stride, stride, mi, mj, rounding);

				gmc_data->predict_8x8(gmc_data,
					pGMC->u + mj*8*stride2 + mi*8, pRef->u,
					pGMC->v + mj*8*stride2 + mi*8, pRef->v,
					stride2, stride2, mi, mj, rounding);
			}
			gmc_data->get_average_mv(gmc_data, &avgMV, mi, mj, quarterpel);

			pMBs[mbnum].amv.x = gmc_sanitize(avgMV.x, quarterpel, fcode);
			pMBs[mbnum].amv.y = gmc_sanitize(avgMV.y, quarterpel, fcode);

			pMBs[mbnum].mcsel = 0; /* until mode decision */
	}
  emms();
}