Skip to content

L
libswscale
Commit f438d5b2 authored by michael's avatar michael
Browse files
Options

Avoid using floating point for calculating filter coefficients. 


git-svn-id: file:///var/local/repositories/mplayer/trunk/libswscale@27612 b3059339-0415-0410-9bf9-f77b7e298cf2
parent 671a7a5a
Hide whitespace changes
Inline Side-by-side
Showing with 83 additions and 81 deletions
swscale.c
View file @ f438d5b2
......@@ -1048,8 +1048,9 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
int filterSize;
int filter2Size;
int minFilterSize;
double *filter=NULL;
double *filter2=NULL;
int64_t *filter=NULL;
int64_t *filter2=NULL;
const int64_t fone= 1LL<<54;
int ret= -1;
#if defined(ARCH_X86)
if (flags & SWS_CPU_CAPS_MMX)
......@@ -1067,7 +1068,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
for (i=0; i<dstW; i++)
{
filter[i*filterSize]=1;
filter[i*filterSize]= fone;
(*filterPos)[i]=i;
}
......@@ -1085,7 +1086,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
int xx= (xDstInSrc - ((filterSize-1)<<15) + (1<<15))>>16;
(*filterPos)[i]= xx;
filter[i]= 1.0;
filter[i]= fone;
xDstInSrc+= xInc;
}
}
......@@ -1108,8 +1109,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
//Bilinear upscale / linear interpolate / Area averaging
for (j=0; j<filterSize; j++)
{
double d= FFABS((xx<<16) - xDstInSrc)/(double)(1<<16);
double coeff= 1.0 - d;
int64_t coeff= fone - FFABS((xx<<16) - xDstInSrc)*(fone>>16);
if (coeff<0) coeff=0;
filter[i*filterSize + j]= coeff;
xx++;
......@@ -1119,52 +1119,59 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
}
else
{
double xDstInSrc;
double sizeFactor, filterSizeInSrc;
const double xInc1= (double)xInc / (double)(1<<16);
if (flags&SWS_BICUBIC) sizeFactor= 4.0;
else if (flags&SWS_X) sizeFactor= 8.0;
else if (flags&SWS_AREA) sizeFactor= 1.0; //downscale only, for upscale it is bilinear
else if (flags&SWS_GAUSS) sizeFactor= 8.0; // infinite ;)
else if (flags&SWS_LANCZOS) sizeFactor= param[0] != SWS_PARAM_DEFAULT ? 2.0*param[0] : 6.0;
else if (flags&SWS_SINC) sizeFactor= 20.0; // infinite ;)
else if (flags&SWS_SPLINE) sizeFactor= 20.0; // infinite ;)
else if (flags&SWS_BILINEAR) sizeFactor= 2.0;
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= 0.0; //GCC warning killer
sizeFactor= 0; //GCC warning killer
assert(0);
}
if (xInc1 <= 1.0) filterSizeInSrc= sizeFactor; // upscale
else filterSizeInSrc= sizeFactor*srcW / (double)dstW;
if (xInc <= 1<<16) filterSize= 1 + sizeFactor; // upscale
else filterSize= 1 + (sizeFactor*srcW + dstW - 1)/ dstW;
filterSize= (int)ceil(1 + filterSizeInSrc); // will be reduced later if possible
if (filterSize > srcW-2) filterSize=srcW-2;
filter= av_malloc(dstW*sizeof(*filter)*filterSize);
xDstInSrc= xInc1 / 2.0 - 0.5;
xDstInSrc= xInc - 0x10000;
for (i=0; i<dstW; i++)
{
int xx= (int)(xDstInSrc - (filterSize-1)*0.5 + 0.5);
int xx= (xDstInSrc - ((filterSize-2)<<16)) / (1<<17);
int j;
(*filterPos)[i]= xx;
for (j=0; j<filterSize; j++)
{
double d= FFABS(xx - xDstInSrc)/filterSizeInSrc*sizeFactor;
double coeff;
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)
{
double B= param[0] != SWS_PARAM_DEFAULT ? param[0] : 0.0;
double C= param[1] != SWS_PARAM_DEFAULT ? param[1] : 0.6;
if (d<1.0)
coeff = (12-9*B-6*C)*d*d*d + (-18+12*B+6*C)*d*d + 6-2*B;
else if (d<2.0)
coeff = (-B-6*C)*d*d*d + (6*B+30*C)*d*d + (-12*B-48*C)*d +8*B+24*C;
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)
{
......@@ -1175,46 +1182,49 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
else if (flags & SWS_X)
{
double A= param[0] != SWS_PARAM_DEFAULT ? param[0] : 1.0;
double c;
if (d<1.0)
coeff = cos(d*PI);
if (floatd<1.0)
c = cos(floatd*PI);
else
coeff=-1.0;
if (coeff<0.0) coeff= -pow(-coeff, A);
else coeff= pow( coeff, A);
coeff= coeff*0.5 + 0.5;
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)
{
double srcPixelSize= 1.0/xInc1;
if (d + srcPixelSize/2 < 0.5) coeff= 1.0;
else if (d - srcPixelSize/2 < 0.5) coeff= (0.5-d)/srcPixelSize + 0.5;
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*d*d);
coeff = (pow(2.0, - p*floatd*floatd))*fone;
}
else if (flags & SWS_SINC)
{
coeff = d ? sin(d*PI)/(d*PI) : 1.0;
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(d*PI)*sin(d*PI/p)/(d*d*PI*PI/p) : 1.0;
if (d>p) coeff=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= 1.0 - d;
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, d);
coeff = getSplineCoeff(1.0, 0.0, p, -p-1.0, d) * fone;
}
else {
coeff= 0.0; //GCC warning killer
......@@ -1224,7 +1234,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
filter[i*filterSize + j]= coeff;
xx++;
}
xDstInSrc+= xInc1;
xDstInSrc+= 2*xInc;
}
}
......@@ -1236,31 +1246,24 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
if (srcFilter) filter2Size+= srcFilter->length - 1;
if (dstFilter) filter2Size+= dstFilter->length - 1;
assert(filter2Size>0);
filter2= av_malloc(filter2Size*dstW*sizeof(*filter2));
filter2= av_mallocz(filter2Size*dstW*sizeof(*filter2));
for (i=0; i<dstW; i++)
{
int j;
SwsVector scaleFilter;
SwsVector *outVec;
scaleFilter.coeff= filter + i*filterSize;
scaleFilter.length= filterSize;
if (srcFilter) outVec= sws_getConvVec(srcFilter, &scaleFilter);
else outVec= &scaleFilter;
int j, k;
assert(outVec->length == filter2Size);
//FIXME dstFilter
for (j=0; j<outVec->length; j++)
{
filter2[i*filter2Size + j]= outVec->coeff[j];
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;
if (outVec != &scaleFilter) sws_freeVec(outVec);
}
av_freep(&filter);
......@@ -1271,7 +1274,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
{
int min= filter2Size;
int j;
double cutOff=0.0;
int64_t cutOff=0.0;
/* get rid off near zero elements on the left by shifting left */
for (j=0; j<filter2Size; j++)
......@@ -1279,7 +1282,7 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
int k;
cutOff += FFABS(filter2[i*filter2Size]);
if (cutOff > SWS_MAX_REDUCE_CUTOFF) break;
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;
......@@ -1287,17 +1290,17 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
// Move filter coeffs left
for (k=1; k<filter2Size; k++)
filter2[i*filter2Size + k - 1]= filter2[i*filter2Size + k];
filter2[i*filter2Size + k - 1]= 0.0;
filter2[i*filter2Size + k - 1]= 0;
(*filterPos)[i]++;
}
cutOff=0.0;
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) break;
if (cutOff > SWS_MAX_REDUCE_CUTOFF*fone) break;
min--;
}
......@@ -1342,10 +1345,10 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
for (j=0; j<filterSize; j++)
{
if (j>=filter2Size) filter[i*filterSize + j]= 0.0;
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]= 0.0;
filter[i*filterSize + j]= 0;
}
}
......@@ -1390,21 +1393,20 @@ static inline int initFilter(int16_t **outFilter, int16_t **filterPos, int *outF
for (i=0; i<dstW; i++)
{
int j;
double error=0;
double sum=0;
double scale= one;
int64_t error=0;
int64_t sum=0;
for (j=0; j<filterSize; j++)
{
sum+= filter[i*filterSize + j];
}
scale/= sum;
sum= (sum + one/2)/ one;
for (j=0; j<*outFilterSize; j++)
{
double v= filter[i*filterSize + j]*scale + error;
int intV= floor(v + 0.5);
int64_t v= filter[i*filterSize + j] + error;
int intV= ROUNDED_DIV(v, sum);
(*outFilter)[i*(*outFilterSize) + j]= intV;
error = v - intV;
error= v - intV*sum;
}
}
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment