-static double getSplineCoeff(double a, double b, double c, double d, double dist)
-{
-// printf("%f %f %f %f %f\n", a,b,c,d,dist);
- if (dist<=1.0) return ((d*dist + c)*dist + b)*dist +a;
- else return getSplineCoeff( 0.0,
- b+ 2.0*c + 3.0*d,
- c + 3.0*d,
- -b- 3.0*c - 6.0*d,
- dist-1.0);
-}
-
-static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outFilterSize, int xInc,
- int srcW, int dstW, int filterAlign, int one, int flags,
- SwsVector *srcFilter, SwsVector *dstFilter, double param[2])
-{
- int i;
- int filterSize;
- int filter2Size;
- int minFilterSize;
- int64_t *filter=NULL;
- int64_t *filter2=NULL;
- const int64_t fone= 1LL<<54;
- int ret= -1;
-#if ARCH_X86
- if (flags & SWS_CPU_CAPS_MMX)
- __asm__ volatile("emms\n\t"::: "memory"); //FIXME this should not be required but it IS (even for non-MMX versions)
-#endif
-
- // NOTE: the +1 is for the MMX scaler which reads over the end
- CHECKED_ALLOC(*filterPos, (dstW+1)*sizeof(int16_t));
-
- if (FFABS(xInc - 0x10000) <10) { // unscaled
- int i;
- filterSize= 1;
- CHECKED_ALLOCZ(filter, dstW*sizeof(*filter)*filterSize);
-
- for (i=0; i<dstW; i++) {
- filter[i*filterSize]= fone;
- (*filterPos)[i]=i;
- }
-
- } else if (flags&SWS_POINT) { // lame looking point sampling mode
- int i;
- int xDstInSrc;
- filterSize= 1;
- CHECKED_ALLOC(filter, dstW*sizeof(*filter)*filterSize);
-
- xDstInSrc= xInc/2 - 0x8000;
- for (i=0; i<dstW; i++) {
- int xx= (xDstInSrc - ((filterSize-1)<<15) + (1<<15))>>16;
-
- (*filterPos)[i]= xx;
- filter[i]= fone;
- xDstInSrc+= xInc;
- }
- } else if ((xInc <= (1<<16) && (flags&SWS_AREA)) || (flags&SWS_FAST_BILINEAR)) { // bilinear upscale
- int i;
- int xDstInSrc;
- filterSize= 2;
- CHECKED_ALLOC(filter, dstW*sizeof(*filter)*filterSize);
-
- xDstInSrc= xInc/2 - 0x8000;
- for (i=0; i<dstW; i++) {
- int xx= (xDstInSrc - ((filterSize-1)<<15) + (1<<15))>>16;
- int j;
-
- (*filterPos)[i]= xx;
- //bilinear upscale / linear interpolate / area averaging
- for (j=0; j<filterSize; j++) {
- int64_t coeff= fone - FFABS((xx<<16) - xDstInSrc)*(fone>>16);
- if (coeff<0) coeff=0;
- filter[i*filterSize + j]= coeff;
- xx++;
- }
- xDstInSrc+= xInc;
- }
- } else {
- int xDstInSrc;
- int sizeFactor;
-
- if (flags&SWS_BICUBIC) sizeFactor= 4;
- else if (flags&SWS_X) sizeFactor= 8;
- else if (flags&SWS_AREA) sizeFactor= 1; //downscale only, for upscale it is bilinear
- else if (flags&SWS_GAUSS) sizeFactor= 8; // infinite ;)
- else if (flags&SWS_LANCZOS) sizeFactor= param[0] != SWS_PARAM_DEFAULT ? ceil(2*param[0]) : 6;
- else if (flags&SWS_SINC) sizeFactor= 20; // infinite ;)
- else if (flags&SWS_SPLINE) sizeFactor= 20; // infinite ;)
- else if (flags&SWS_BILINEAR) sizeFactor= 2;
- else {
- sizeFactor= 0; //GCC warning killer
- assert(0);
- }
-
- if (xInc <= 1<<16) filterSize= 1 + sizeFactor; // upscale
- else filterSize= 1 + (sizeFactor*srcW + dstW - 1)/ dstW;
-
- if (filterSize > srcW-2) filterSize=srcW-2;
-
- CHECKED_ALLOC(filter, dstW*sizeof(*filter)*filterSize);
-
- xDstInSrc= xInc - 0x10000;
- for (i=0; i<dstW; i++) {
- int xx= (xDstInSrc - ((filterSize-2)<<16)) / (1<<17);
- int j;
- (*filterPos)[i]= xx;
- for (j=0; j<filterSize; j++) {
- int64_t d= ((int64_t)FFABS((xx<<17) - xDstInSrc))<<13;
- double floatd;
- int64_t coeff;
-
- if (xInc > 1<<16)
- d= d*dstW/srcW;
- floatd= d * (1.0/(1<<30));
-
- if (flags & SWS_BICUBIC) {
- int64_t B= (param[0] != SWS_PARAM_DEFAULT ? param[0] : 0) * (1<<24);
- int64_t C= (param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6) * (1<<24);
- int64_t dd = ( d*d)>>30;
- int64_t ddd= (dd*d)>>30;
-
- if (d < 1LL<<30)
- coeff = (12*(1<<24)-9*B-6*C)*ddd + (-18*(1<<24)+12*B+6*C)*dd + (6*(1<<24)-2*B)*(1<<30);
- else if (d < 1LL<<31)
- coeff = (-B-6*C)*ddd + (6*B+30*C)*dd + (-12*B-48*C)*d + (8*B+24*C)*(1<<30);
- else
- coeff=0.0;
- coeff *= fone>>(30+24);
- }
-/* else if (flags & SWS_X) {
- double p= param ? param*0.01 : 0.3;
- coeff = d ? sin(d*PI)/(d*PI) : 1.0;
- coeff*= pow(2.0, - p*d*d);
- }*/
- else if (flags & SWS_X) {
- double A= param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
- double c;
-
- if (floatd<1.0)
- c = cos(floatd*PI);
- else
- c=-1.0;
- if (c<0.0) c= -pow(-c, A);
- else c= pow( c, A);
- coeff= (c*0.5 + 0.5)*fone;
- } else if (flags & SWS_AREA) {
- int64_t d2= d - (1<<29);
- if (d2*xInc < -(1LL<<(29+16))) coeff= 1.0 * (1LL<<(30+16));
- else if (d2*xInc < (1LL<<(29+16))) coeff= -d2*xInc + (1LL<<(29+16));
- else coeff=0.0;
- coeff *= fone>>(30+16);
- } else if (flags & SWS_GAUSS) {
- double p= param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
- coeff = (pow(2.0, - p*floatd*floatd))*fone;
- } else if (flags & SWS_SINC) {
- coeff = (d ? sin(floatd*PI)/(floatd*PI) : 1.0)*fone;
- } else if (flags & SWS_LANCZOS) {
- double p= param[0] != SWS_PARAM_DEFAULT ? param[0] : 3.0;
- coeff = (d ? sin(floatd*PI)*sin(floatd*PI/p)/(floatd*floatd*PI*PI/p) : 1.0)*fone;
- if (floatd>p) coeff=0;
- } else if (flags & SWS_BILINEAR) {
- coeff= (1<<30) - d;
- if (coeff<0) coeff=0;
- coeff *= fone >> 30;
- } else if (flags & SWS_SPLINE) {
- double p=-2.196152422706632;
- coeff = getSplineCoeff(1.0, 0.0, p, -p-1.0, floatd) * fone;
- } else {
- coeff= 0.0; //GCC warning killer
- assert(0);
- }
-
- filter[i*filterSize + j]= coeff;
- xx++;
- }
- xDstInSrc+= 2*xInc;
- }
- }
-
- /* apply src & dst Filter to filter -> filter2
- av_free(filter);
- */
- assert(filterSize>0);
- filter2Size= filterSize;
- if (srcFilter) filter2Size+= srcFilter->length - 1;
- if (dstFilter) filter2Size+= dstFilter->length - 1;
- assert(filter2Size>0);
- CHECKED_ALLOCZ(filter2, filter2Size*dstW*sizeof(*filter2));
-
- for (i=0; i<dstW; i++) {
- int j, k;
-
- if(srcFilter) {
- for (k=0; k<srcFilter->length; k++) {
- for (j=0; j<filterSize; j++)
- filter2[i*filter2Size + k + j] += srcFilter->coeff[k]*filter[i*filterSize + j];
- }
- } else {
- for (j=0; j<filterSize; j++)
- filter2[i*filter2Size + j]= filter[i*filterSize + j];
- }
- //FIXME dstFilter
-
- (*filterPos)[i]+= (filterSize-1)/2 - (filter2Size-1)/2;
- }
- av_freep(&filter);
-
- /* try to reduce the filter-size (step1 find size and shift left) */
- // Assume it is near normalized (*0.5 or *2.0 is OK but * 0.001 is not).
- minFilterSize= 0;
- for (i=dstW-1; i>=0; i--) {
- int min= filter2Size;
- int j;
- int64_t cutOff=0.0;
-
- /* get rid off near zero elements on the left by shifting left */
- for (j=0; j<filter2Size; j++) {
- int k;
- cutOff += FFABS(filter2[i*filter2Size]);
-
- if (cutOff > SWS_MAX_REDUCE_CUTOFF*fone) break;
-
- /* preserve monotonicity because the core can't handle the filter otherwise */
- if (i<dstW-1 && (*filterPos)[i] >= (*filterPos)[i+1]) break;
-
- // move filter coefficients left
- for (k=1; k<filter2Size; k++)
- filter2[i*filter2Size + k - 1]= filter2[i*filter2Size + k];
- filter2[i*filter2Size + k - 1]= 0;
- (*filterPos)[i]++;
- }
-
- cutOff=0;
- /* count near zeros on the right */
- for (j=filter2Size-1; j>0; j--) {
- cutOff += FFABS(filter2[i*filter2Size + j]);
-
- if (cutOff > SWS_MAX_REDUCE_CUTOFF*fone) break;
- min--;
- }
-
- if (min>minFilterSize) minFilterSize= min;
- }
-
- if (flags & SWS_CPU_CAPS_ALTIVEC) {
- // we can handle the special case 4,
- // so we don't want to go to the full 8
- if (minFilterSize < 5)
- filterAlign = 4;
-
- // We really don't want to waste our time
- // doing useless computation, so fall back on
- // the scalar C code for very small filters.
- // Vectorizing is worth it only if you have a
- // decent-sized vector.
- if (minFilterSize < 3)
- filterAlign = 1;
- }
-
- if (flags & SWS_CPU_CAPS_MMX) {
- // special case for unscaled vertical filtering
- if (minFilterSize == 1 && filterAlign == 2)
- filterAlign= 1;
- }
-
- assert(minFilterSize > 0);
- filterSize= (minFilterSize +(filterAlign-1)) & (~(filterAlign-1));
- assert(filterSize > 0);
- filter= av_malloc(filterSize*dstW*sizeof(*filter));
- if (filterSize >= MAX_FILTER_SIZE*16/((flags&SWS_ACCURATE_RND) ? APCK_SIZE : 16) || !filter)
- goto fail;
- *outFilterSize= filterSize;
-
- if (flags&SWS_PRINT_INFO)
- av_log(NULL, AV_LOG_VERBOSE, "SwScaler: reducing / aligning filtersize %d -> %d\n", filter2Size, filterSize);
- /* try to reduce the filter-size (step2 reduce it) */
- for (i=0; i<dstW; i++) {
- int j;
-
- for (j=0; j<filterSize; j++) {
- if (j>=filter2Size) filter[i*filterSize + j]= 0;
- else filter[i*filterSize + j]= filter2[i*filter2Size + j];
- if((flags & SWS_BITEXACT) && j>=minFilterSize)
- filter[i*filterSize + j]= 0;
- }
- }
-
-
- //FIXME try to align filterPos if possible
-
- //fix borders
- for (i=0; i<dstW; i++) {
- int j;
- if ((*filterPos)[i] < 0) {
- // move filter coefficients left to compensate for filterPos
- for (j=1; j<filterSize; j++) {
- int left= FFMAX(j + (*filterPos)[i], 0);
- filter[i*filterSize + left] += filter[i*filterSize + j];
- filter[i*filterSize + j]=0;
- }
- (*filterPos)[i]= 0;
- }
-
- if ((*filterPos)[i] + filterSize > srcW) {
- int shift= (*filterPos)[i] + filterSize - srcW;
- // move filter coefficients right to compensate for filterPos
- for (j=filterSize-2; j>=0; j--) {
- int right= FFMIN(j + shift, filterSize-1);
- filter[i*filterSize +right] += filter[i*filterSize +j];
- filter[i*filterSize +j]=0;
- }
- (*filterPos)[i]= srcW - filterSize;
- }
- }
-
- // Note the +1 is for the MMX scaler which reads over the end
- /* align at 16 for AltiVec (needed by hScale_altivec_real) */
- CHECKED_ALLOCZ(*outFilter, *outFilterSize*(dstW+1)*sizeof(int16_t));
-
- /* normalize & store in outFilter */
- for (i=0; i<dstW; i++) {
- int j;
- int64_t error=0;
- int64_t sum=0;
-
- for (j=0; j<filterSize; j++) {
- sum+= filter[i*filterSize + j];
- }
- sum= (sum + one/2)/ one;
- for (j=0; j<*outFilterSize; j++) {
- int64_t v= filter[i*filterSize + j] + error;
- int intV= ROUNDED_DIV(v, sum);
- (*outFilter)[i*(*outFilterSize) + j]= intV;
- error= v - intV*sum;
- }
- }
-
- (*filterPos)[dstW]= (*filterPos)[dstW-1]; // the MMX scaler will read over the end
- for (i=0; i<*outFilterSize; i++) {
- int j= dstW*(*outFilterSize);
- (*outFilter)[j + i]= (*outFilter)[j + i - (*outFilterSize)];
- }
-
- ret=0;
-fail:
- av_free(filter);
- av_free(filter2);
- return ret;
-}
-
-#ifdef COMPILE_MMX2
-static int initMMX2HScaler(int dstW, int xInc, uint8_t *filterCode, int16_t *filter, int32_t *filterPos, int numSplits)
-{
- uint8_t *fragmentA;
- x86_reg imm8OfPShufW1A;
- x86_reg imm8OfPShufW2A;
- x86_reg fragmentLengthA;
- uint8_t *fragmentB;
- x86_reg imm8OfPShufW1B;
- x86_reg imm8OfPShufW2B;
- x86_reg fragmentLengthB;
- int fragmentPos;
-
- int xpos, i;
-
- // create an optimized horizontal scaling routine
-
- //code fragment
-
- __asm__ volatile(
- "jmp 9f \n\t"
- // Begin
- "0: \n\t"
- "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
- "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
- "movd 1(%%"REG_c", %%"REG_S"), %%mm1 \n\t"
- "punpcklbw %%mm7, %%mm1 \n\t"
- "punpcklbw %%mm7, %%mm0 \n\t"
- "pshufw $0xFF, %%mm1, %%mm1 \n\t"
- "1: \n\t"
- "pshufw $0xFF, %%mm0, %%mm0 \n\t"
- "2: \n\t"
- "psubw %%mm1, %%mm0 \n\t"
- "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
- "pmullw %%mm3, %%mm0 \n\t"
- "psllw $7, %%mm1 \n\t"
- "paddw %%mm1, %%mm0 \n\t"
-
- "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
-
- "add $8, %%"REG_a" \n\t"
- // End
- "9: \n\t"
-// "int $3 \n\t"
- "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
- "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
- "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
- "dec %1 \n\t"
- "dec %2 \n\t"
- "sub %0, %1 \n\t"
- "sub %0, %2 \n\t"
- "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
- "sub %0, %3 \n\t"
-
-
- :"=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
- "=r" (fragmentLengthA)
- );
-
- __asm__ volatile(
- "jmp 9f \n\t"
- // Begin
- "0: \n\t"
- "movq (%%"REG_d", %%"REG_a"), %%mm3 \n\t"
- "movd (%%"REG_c", %%"REG_S"), %%mm0 \n\t"
- "punpcklbw %%mm7, %%mm0 \n\t"
- "pshufw $0xFF, %%mm0, %%mm1 \n\t"
- "1: \n\t"
- "pshufw $0xFF, %%mm0, %%mm0 \n\t"
- "2: \n\t"
- "psubw %%mm1, %%mm0 \n\t"
- "movl 8(%%"REG_b", %%"REG_a"), %%esi \n\t"
- "pmullw %%mm3, %%mm0 \n\t"
- "psllw $7, %%mm1 \n\t"
- "paddw %%mm1, %%mm0 \n\t"
-
- "movq %%mm0, (%%"REG_D", %%"REG_a") \n\t"
-
- "add $8, %%"REG_a" \n\t"
- // End
- "9: \n\t"
-// "int $3 \n\t"
- "lea " LOCAL_MANGLE(0b) ", %0 \n\t"
- "lea " LOCAL_MANGLE(1b) ", %1 \n\t"
- "lea " LOCAL_MANGLE(2b) ", %2 \n\t"
- "dec %1 \n\t"
- "dec %2 \n\t"
- "sub %0, %1 \n\t"
- "sub %0, %2 \n\t"
- "lea " LOCAL_MANGLE(9b) ", %3 \n\t"
- "sub %0, %3 \n\t"
-
-
- :"=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
- "=r" (fragmentLengthB)
- );
-
- xpos= 0; //lumXInc/2 - 0x8000; // difference between pixel centers
- fragmentPos=0;
-
- for (i=0; i<dstW/numSplits; i++) {
- int xx=xpos>>16;
-
- if ((i&3) == 0) {
- int a=0;
- int b=((xpos+xInc)>>16) - xx;
- int c=((xpos+xInc*2)>>16) - xx;
- int d=((xpos+xInc*3)>>16) - xx;
- int inc = (d+1<4);
- uint8_t *fragment = (d+1<4) ? fragmentB : fragmentA;
- x86_reg imm8OfPShufW1 = (d+1<4) ? imm8OfPShufW1B : imm8OfPShufW1A;
- x86_reg imm8OfPShufW2 = (d+1<4) ? imm8OfPShufW2B : imm8OfPShufW2A;
- x86_reg fragmentLength = (d+1<4) ? fragmentLengthB : fragmentLengthA;
- int maxShift= 3-(d+inc);
- int shift=0;
-
- if (filterCode) {
- filter[i ] = (( xpos & 0xFFFF) ^ 0xFFFF)>>9;
- filter[i+1] = (((xpos+xInc ) & 0xFFFF) ^ 0xFFFF)>>9;
- filter[i+2] = (((xpos+xInc*2) & 0xFFFF) ^ 0xFFFF)>>9;
- filter[i+3] = (((xpos+xInc*3) & 0xFFFF) ^ 0xFFFF)>>9;
- filterPos[i/2]= xx;
-
- memcpy(filterCode + fragmentPos, fragment, fragmentLength);
-
- filterCode[fragmentPos + imm8OfPShufW1]=
- (a+inc) | ((b+inc)<<2) | ((c+inc)<<4) | ((d+inc)<<6);
- filterCode[fragmentPos + imm8OfPShufW2]=
- a | (b<<2) | (c<<4) | (d<<6);
-
- if (i+4-inc>=dstW) shift=maxShift; //avoid overread
- else if ((filterPos[i/2]&3) <= maxShift) shift=filterPos[i/2]&3; //Align
-
- if (shift && i>=shift) {
- filterCode[fragmentPos + imm8OfPShufW1]+= 0x55*shift;
- filterCode[fragmentPos + imm8OfPShufW2]+= 0x55*shift;
- filterPos[i/2]-=shift;
- }
- }
-
- fragmentPos+= fragmentLength;
-
- if (filterCode)
- filterCode[fragmentPos]= RET;
- }
- xpos+=xInc;
- }
- if (filterCode)
- filterPos[((i/2)+1)&(~1)]= xpos>>16; // needed to jump to the next part
-
- return fragmentPos + 1;
-}
-#endif /* COMPILE_MMX2 */
-
-static void globalInit(void)
-{
- // generating tables:
- int i;
- for (i=0; i<768; i++) {
- int c= av_clip_uint8(i-256);
- clip_table[i]=c;
- }
-}
-
-static SwsFunc getSwsFunc(SwsContext *c)