vf_fspp.c

来自「君正早期ucos系统(只有早期的才不没有打包成库),MPLAYER,文件系统,图」· C语言 代码 · 共 2,126 行 · 第 1/5 页

            p->src[index         - x - 1]= p->src[index +         x    ];
            p->src[index + width + x    ]= p->src[index + width - x - 1];
        }
    }
    for(y=0; y<8; y++){
        fast_memcpy(p->src + (      7-y)*stride, p->src + (      y+8)*stride, stride);
        fast_memcpy(p->src + (height+8+y)*stride, p->src + (height-y+7)*stride, stride);
    }
    //FIXME (try edge emu)

    for(y=8; y<24; y++)
	memset(p->temp+ 8 +y*stride, 0,width*sizeof(int16_t));

    for(y=step; y<height+8; y+=step){    //step= 1,2
	qy=y-4;
	if (qy>height-1) qy=height-1;
	if (qy<0) qy=0;
	qy=(qy>>qps)*qp_stride;
	row_fdct_s(block, p->src + y*stride +2-(y&1), stride, 2);
	for(x0=0; x0<width+8-8*(BLOCKSZ-1); x0+=8*(BLOCKSZ-1)){
	    row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, 2*(BLOCKSZ-1));
	    if(p->qp)        
		column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+0*8, block3+0*8, 8*(BLOCKSZ-1)); //yes, this is a HOTSPOT
	    else
		for (x=0; x<8*(BLOCKSZ-1); x+=8) {
		    t=x+x0-2; //correct t=x+x0-2-(y&1), but its the same 
		    if (t<0) t=0;//t always < width-2
		    t=qp_store[qy+(t>>qps)];
		    if(p->mpeg2) t>>=1; //copy p->mpeg2,prev_q to locals?
		    if (t!=p->prev_q) p->prev_q=t, mul_thrmat_s(p, t);
		    column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block+x*8, block3+x*8, 8); //yes, this is a HOTSPOT
		}
	    row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, 2*(BLOCKSZ-1));
	    memcpy(block, block+(BLOCKSZ-1)*64, 8*8*sizeof(DCTELEM)); //cycling
	    memcpy(block3, block3+(BLOCKSZ-1)*64, 6*8*sizeof(DCTELEM));  
	}
	//
	es=width+8-x0; //  8, ...      
	if (es>8)
	    row_fdct_s(block+8*8, p->src + y*stride+8+x0 +2-(y&1), stride, (es-4)>>2);
	column_fidct_s((int16_t*)(&p->threshold_mtx[0]), block, block3, es&(~1));
	row_idct_s(block3+0*8, p->temp + (y&15)*stride+x0+2-(y&1), stride, es>>2);
	{const int y1=y-8+step;//l5-7  l4-6
	    if (!(y1&7) && y1) {
		if (y1&8) store_slice_s(dst + (y1-8)*dst_stride, p->temp+ 8 +8*stride, 
					dst_stride, stride, width, 8, 5-p->log2_count);
		else store_slice2_s(dst + (y1-8)*dst_stride, p->temp+ 8 +0*stride, 
				    dst_stride, stride, width, 8, 5-p->log2_count);    
	    } }
    }

    if (y&7) {  // == height & 7
	if (y&8) store_slice_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +8*stride, 
			       dst_stride, stride, width, y&7, 5-p->log2_count);
	else store_slice2_s(dst + ((y-8)&~7)*dst_stride, p->temp+ 8 +0*stride, 
			    dst_stride, stride, width, y&7, 5-p->log2_count);
    }
}

static int config(struct vf_instance_s* vf,
		  int width, int height, int d_width, int d_height,
		  unsigned int flags, unsigned int outfmt)
{
    int h= (height+16+15)&(~15);

    vf->priv->temp_stride= (width+16+15)&(~15);
    vf->priv->temp= (int16_t*)av_mallocz(vf->priv->temp_stride*3*8*sizeof(int16_t));
    //this can also be avoided, see above
    vf->priv->src = (uint8_t*)av_malloc(vf->priv->temp_stride*h*sizeof(uint8_t));

    return vf_next_config(vf,width,height,d_width,d_height,flags,outfmt);
}

static void get_image(struct vf_instance_s* vf, mp_image_t *mpi)
{
    if(mpi->flags&MP_IMGFLAG_PRESERVE) return; // don't change
    // ok, we can do pp in-place (or pp disabled):
    vf->dmpi=vf_get_image(vf->next,mpi->imgfmt,
			  mpi->type, mpi->flags, mpi->width, mpi->height);
    mpi->planes[0]=vf->dmpi->planes[0];
    mpi->stride[0]=vf->dmpi->stride[0];
    mpi->width=vf->dmpi->width;
    if(mpi->flags&MP_IMGFLAG_PLANAR){
        mpi->planes[1]=vf->dmpi->planes[1];
        mpi->planes[2]=vf->dmpi->planes[2];
        mpi->stride[1]=vf->dmpi->stride[1];
        mpi->stride[2]=vf->dmpi->stride[2];
    }
    mpi->flags|=MP_IMGFLAG_DIRECT;
}

static int put_image(struct vf_instance_s* vf, mp_image_t *mpi, double pts)
{
    mp_image_t *dmpi;
    if(!(mpi->flags&MP_IMGFLAG_DIRECT)){
	// no DR, so get a new image! hope we'll get DR buffer:
	dmpi=vf_get_image(vf->next,mpi->imgfmt,
			  MP_IMGTYPE_TEMP,
			  MP_IMGFLAG_ACCEPT_STRIDE|MP_IMGFLAG_PREFER_ALIGNED_STRIDE,
			  mpi->width,mpi->height);
	vf_clone_mpi_attributes(dmpi, mpi);
    }else{
	dmpi=vf->dmpi;
    }

    vf->priv->mpeg2= mpi->qscale_type;
    if(mpi->pict_type != 3 && mpi->qscale && !vf->priv->qp){
	if(!vf->priv->non_b_qp)
	    vf->priv->non_b_qp= malloc(mpi->qstride * ((mpi->h + 15) >> 4));
	fast_memcpy(vf->priv->non_b_qp, mpi->qscale, mpi->qstride * ((mpi->h + 15) >> 4));
    }
    if(vf->priv->log2_count || !(mpi->flags&MP_IMGFLAG_DIRECT)){
	char *qp_tab= vf->priv->non_b_qp;
	if(vf->priv->bframes || !qp_tab)
	    qp_tab= mpi->qscale;

	if(qp_tab || vf->priv->qp){
	    filter(vf->priv, dmpi->planes[0], mpi->planes[0], dmpi->stride[0], mpi->stride[0],
		   mpi->w, mpi->h, qp_tab, mpi->qstride, 1);
	    filter(vf->priv, dmpi->planes[1], mpi->planes[1], dmpi->stride[1], mpi->stride[1],
		   mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
	    filter(vf->priv, dmpi->planes[2], mpi->planes[2], dmpi->stride[2], mpi->stride[2],
		   mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, qp_tab, mpi->qstride, 0);
	}else{
	    memcpy_pic(dmpi->planes[0], mpi->planes[0], mpi->w, mpi->h, dmpi->stride[0], mpi->stride[0]);
	    memcpy_pic(dmpi->planes[1], mpi->planes[1], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[1], mpi->stride[1]);
	    memcpy_pic(dmpi->planes[2], mpi->planes[2], mpi->w>>mpi->chroma_x_shift, mpi->h>>mpi->chroma_y_shift, dmpi->stride[2], mpi->stride[2]);
	}
    }

#ifdef HAVE_MMX
    if(gCpuCaps.hasMMX) asm volatile ("emms\n\t");
#endif
#ifdef HAVE_MMX2
    if(gCpuCaps.hasMMX2) asm volatile ("sfence\n\t");
#endif
    return vf_next_put_image(vf,dmpi, pts);
}

static void uninit(struct vf_instance_s* vf)
{
    if(!vf->priv) return;

    if(vf->priv->temp) av_free(vf->priv->temp);
    vf->priv->temp= NULL;
    if(vf->priv->src)  av_free(vf->priv->src);
    vf->priv->src= NULL;
    //if(vf->priv->avctx) free(vf->priv->avctx);
    //vf->priv->avctx= NULL;
    if(vf->priv->non_b_qp) free(vf->priv->non_b_qp);
    vf->priv->non_b_qp= NULL;
        
    av_free(vf->priv);
    vf->priv=NULL;
}

//===========================================================================//

static int query_format(struct vf_instance_s* vf, unsigned int fmt)
{
    switch(fmt){
    case IMGFMT_YVU9:
    case IMGFMT_IF09:
    case IMGFMT_YV12:
    case IMGFMT_I420:
    case IMGFMT_IYUV:
    case IMGFMT_CLPL:
    case IMGFMT_Y800:
    case IMGFMT_Y8:
    case IMGFMT_444P:
    case IMGFMT_422P:
    case IMGFMT_411P:
	return vf_next_query_format(vf,fmt);
    }
    return 0;
}

/*
  static unsigned int fmt_list[]={
  IMGFMT_YVU9,
  IMGFMT_IF09,
  IMGFMT_YV12,
  IMGFMT_I420,
  IMGFMT_IYUV,
  IMGFMT_CLPL,
  IMGFMT_Y800,
  IMGFMT_Y8,
  IMGFMT_444P,
  IMGFMT_422P,
  IMGFMT_411P,
  0
  };
*/

static int control(struct vf_instance_s* vf, int request, void* data)
{
    switch(request){
    case VFCTRL_QUERY_MAX_PP_LEVEL:
        return 5;
    case VFCTRL_SET_PP_LEVEL:
        vf->priv->log2_count= *((unsigned int*)data);
        if (vf->priv->log2_count < 4) vf->priv->log2_count=4;
        return CONTROL_TRUE;
    }
    return vf_next_control(vf,request,data);
}

static int open(vf_instance_t *vf, char* args)
{
    int i=0, bias;
    int custom_threshold_m[64];
    int log2c=-1;
    
    vf->config=config;
    vf->put_image=put_image;
    vf->get_image=get_image;
    vf->query_format=query_format;
    vf->uninit=uninit;
    vf->control= control;
    vf->priv=av_mallocz(sizeof(struct vf_priv_s));//assumes align 16 ! 
    
    avcodec_init();

    //vf->priv->avctx= avcodec_alloc_context();
    //dsputil_init(&vf->priv->dsp, vf->priv->avctx);
    
    vf->priv->log2_count= 4;
    vf->priv->bframes = 0;
    
    if (args) sscanf(args, "%d:%d:%d:%d", &log2c, &vf->priv->qp, &i, &vf->priv->bframes);

    if( log2c >=4 && log2c <=5 )
        vf->priv->log2_count = log2c;
    else if( log2c >= 6 )
	vf->priv->log2_count = 5;

    if(vf->priv->qp < 0)
        vf->priv->qp = 0;

    if (i < -15) i = -15;
    if (i > 32) i = 32;
    
    bias= (1<<4)+i; //regulable
    vf->priv->prev_q=0;
    //
    for(i=0;i<64;i++) //FIXME: tune custom_threshold[] and remove this !
	custom_threshold_m[i]=(int)(custom_threshold[i]*(bias/71.)+ 0.5);
    for(i=0;i<8;i++){
	vf->priv->threshold_mtx_noq[2*i]=(uint64_t)custom_threshold_m[i*8+2]
	    |(((uint64_t)custom_threshold_m[i*8+6])<<16)
	    |(((uint64_t)custom_threshold_m[i*8+0])<<32)
	    |(((uint64_t)custom_threshold_m[i*8+4])<<48);
	vf->priv->threshold_mtx_noq[2*i+1]=(uint64_t)custom_threshold_m[i*8+5]
	    |(((uint64_t)custom_threshold_m[i*8+3])<<16)
	    |(((uint64_t)custom_threshold_m[i*8+1])<<32)
	    |(((uint64_t)custom_threshold_m[i*8+7])<<48);
    }

    if (vf->priv->qp) vf->priv->prev_q=vf->priv->qp, mul_thrmat_s(vf->priv, vf->priv->qp);

    return 1;
}

vf_info_t vf_info_fspp = {
    "fast simple postprocess",
    "fspp",
    "Michael Niedermayer, Nikolaj Poroshin",
    "",
    open,
    NULL
};

//====================================================================
//Specific spp's dct, idct and threshold functions
//I'd prefer to have them in the separate file.

#include "mangle.h"
//#define MANGLE(a) #a

//typedef int16_t DCTELEM; //! only int16_t

#define DCTSIZE 8
#define DCTSIZE_S "8"

#define FIX(x,s)  ((int) ((x) * (1<<s) + 0.5)&0xffff)
#define C64(x)    ((uint64_t)((x)|(x)<<16))<<32 | (uint64_t)(x) | (uint64_t)(x)<<16
#define FIX64(x,s)  C64(FIX(x,s))

#define MULTIPLY16H(x,k)   (((x)*(k))>>16)
#define THRESHOLD(r,x,t) if(((unsigned)((x)+t))>t*2) r=(x);else r=0;
#define DESCALE(x,n)  (((x) + (1 << ((n)-1))) >> n)

#ifdef HAVE_MMX

static uint64_t attribute_used __attribute__((aligned(8))) temps[4];//!!

static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_382683433=FIX64(0.382683433, 14); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_541196100=FIX64(0.541196100, 14); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_707106781=FIX64(0.707106781, 14); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_306562965=FIX64(1.306562965, 14); 

static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562_A=FIX64(1.414213562, 14); 

static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_847759065=FIX64(1.847759065, 13); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_2_613125930=FIX64(-2.613125930, 13); //-
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_414213562=FIX64(1.414213562, 13); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_1_082392200=FIX64(1.082392200, 13);
//for t3,t5,t7 == 0 shortcut
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_847759065=FIX64(0.847759065, 14); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_566454497=FIX64(0.566454497, 14); 
static uint64_t attribute_used __attribute__((aligned(8))) MM_FIX_0_198912367=FIX64(0.198912367, 14); 

static uint64_t attribute_used __attribute__((aligned(8))) MM_DESCALE_RND=C64(4);
static uint64_t attribute_used __attribute__((aligned(8))) MM_2=C64(2);

#else /* !HAVE_MMX */

typedef int32_t int_simd16_t;
static int16_t FIX_0_382683433=FIX(0.382683433, 14); 
static int16_t FIX_0_541196100=FIX(0.541196100, 14); 
static int16_t FIX_0_707106781=FIX(0.707106781, 14); 
static int16_t FIX_1_306562965=FIX(1.306562965, 14); 
static int16_t FIX_1_414213562_A=FIX(1.414213562, 14); 
static int16_t FIX_1_847759065=FIX(1.847759065, 13); 
static int16_t FIX_2_613125930=FIX(-2.613125930, 13); //-
static int16_t FIX_1_414213562=FIX(1.414213562, 13); 
static int16_t FIX_1_082392200=FIX(1.082392200, 13);

#endif

#ifndef HAVE_MMX

static void column_fidct_c(int16_t* thr_adr, DCTELEM *data, DCTELEM *output, int cnt)
{
    int_simd16_t tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    int_simd16_t tmp10, tmp11, tmp12, tmp13;
    int_simd16_t z1,z2,z3,z4,z5, z10, z11, z12, z13;
    int_simd16_t d0, d1, d2, d3, d4, d5, d6, d7;

    DCTELEM* dataptr;
    DCTELEM* wsptr;
    int16_t *threshold;
    int ctr;
  
    dataptr = data;
    wsptr = output;

    for (; cnt > 0; cnt-=2) { //start positions
	threshold=(int16_t*)thr_adr;//threshold_mtx
	for (ctr = DCTSIZE; ctr > 0; ctr--) { 
	    // Process columns from input, add to output. 
	    tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
	    tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
    
	    tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
	    tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
    
	    tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
	    tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
    
	    tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
	    tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];

	    // Even part of FDCT
    
	    tmp10 = tmp0 + tmp3;
	    tmp13 = tmp0 - tmp3;
	    tmp11 = tmp1 + tmp2;
	    tmp12 = tmp1 - tmp2;

	    d0 = tmp10 + tmp11; 
	    d4 = tmp10 - tmp11;
    
	    z1 = MULTIPLY16H((tmp12 + tmp13) <<2, FIX_0_707106781); 
	    d2 = tmp13 + z1; 
	    d6 = tmp13 - z1;    

	    // Even part of IDCT

	    THRESHOLD(tmp0, d0, threshold[0*8]);
	    THRESHOLD(tmp1, d2, threshold[2*8]);
	    THRESHOLD(tmp2, d4, threshold[4*8]);
	    THRESHOLD(tmp3, d6, threshold[6*8]);     
	    tmp0+=2;
	    tmp10 = (tmp0 + tmp2)>>2;
	    tmp11 = (tmp0 - tmp2)>>2;

	    tmp13 = (tmp1 + tmp3)>>2; //+2 !  (psnr decides)
	    tmp12 = MULTIPLY16H((tmp1 - tmp3), FIX_1_414213562_A) - tmp13; //<<2

	    tmp0 = tmp10 + tmp13; //->temps
	    tmp3 = tmp10 - tmp13; //->temps
	    tmp1 = tmp11 + tmp12; //->temps
	    tmp2 = tmp11 - tmp12; //->temps

	    // Odd part of FDCT

	    tmp10 = tmp4 + tmp5;  
	    tmp11 = tmp5 + tmp6;
	    tmp12 = tmp6 + tmp7;
        
	    z5 = MULTIPLY16H((tmp10 - tmp12)<<2, FIX_0_382683433); 
	    z2 = MULTIPLY16H(tmp10 <<2, FIX_0_541196100) + z5; 
	    z4 = MULTIPLY16H(tmp12 <<2, FIX_1_306562965) + z5; 
	    z3 = MULTIPLY16H(tmp11 <<2, FIX_0_707106781); 

	    z11 = tmp7 + z3;        
	    z13 = tmp7 - z3;

	    d5 = z13 + z2; 
	    d3 = z13 - z2;
	    d1 = z11 + z4;
	    d7 = z11 - z4;    

	    // Odd part of IDCT

	    THRESHOLD(tmp4, d1, threshold[1*8]);
	    THRESHOLD(tmp5, d3, threshold[3*8]);
	    THRESHOLD(tmp6, d5, threshold[5*8]);
	    THRESHOLD(tmp7, d7, threshold[7*8]);

	    //Simd version uses here a shortcut for the tmp5,tmp6,tmp7 == 0
	    z13 = tmp6 + tmp5;
	    z10 = (tmp6 - tmp5)<<1;
	    z11 = tmp4 + tmp7;
	    z12 = (tmp4 - tmp7)<<1;