From e016034afc0ee81765527ec850759b5902677d7d Mon Sep 17 00:00:00 2001 From: michael Date: Sun, 20 Jan 2002 05:30:23 +0000 Subject: removed global vars so that multiple swscalers can be used experimental upscaling mode (-sws 3) general convolution filters support (unfinished) bugfix for bicubic upscaling assertion checking if defined MP_DEBUG checking of the input/output size instead of segfault if its very large git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@4277 b3059339-0415-0410-9bf9-f77b7e298cf2 --- postproc/swscale.c | 764 +++++++++++++++++++++++++++++++++++++++++++++++------ 1 file changed, 679 insertions(+), 85 deletions(-) (limited to 'postproc/swscale.c') diff --git a/postproc/swscale.c b/postproc/swscale.c index 0f51745032..2ba194afc0 100644 --- a/postproc/swscale.c +++ b/postproc/swscale.c @@ -5,6 +5,11 @@ // current version mostly by Michael Niedermayer (michaelni@gmx.at) // the parts written by michael are under GNU GPL +/* + supported Input formats: YV12 (grayscale soon too) + supported output formats: YV12, BGR15, BGR16, BGR24, BGR32 (grayscale soon too) +*/ + #include #include #include @@ -16,6 +21,7 @@ #endif #include "swscale.h" #include "../cpudetect.h" +#include "../libvo/img_format.h" #undef MOVNTQ #undef PAVGB @@ -23,14 +29,20 @@ //#undef HAVE_MMX //#undef ARCH_X86 #define DITHER1XBPP -int fullUVIpol=0; -//disables the unscaled height version -int allwaysIpol=0; #define RET 0xC3 //near return opcode -//#define ASSERT(x) if(!(x)) { printf("ASSERT " #x " failed\n"); *((int*)0)=0; } +#ifdef MP_DEBUG +#define ASSERT(x) if(!(x)) { printf("ASSERT " #x " failed\n"); *((int*)0)=0; } +#else #define ASSERT(x) ; +#endif + +#ifdef M_PI +#define PI M_PI +#else +#define PI 3.14159265358979323846 +#endif extern int verbose; // defined in mplayer.c /* @@ -50,7 +62,6 @@ change the distance of the u & v buffer Move static / global vars into a struct so multiple scalers can be used write special vertical cubic upscale version Optimize C code (yv12 / minmax) -dstStride[3] */ #define ABS(a) ((a) > 0 ? (a) : (-(a))) @@ -101,39 +112,9 @@ static uint64_t __attribute__((aligned(8))) M24A= 0x00FF0000FF0000FFLL; static uint64_t __attribute__((aligned(8))) M24B= 0xFF0000FF0000FF00LL; static uint64_t __attribute__((aligned(8))) M24C= 0x0000FF0000FF0000LL; -static uint64_t __attribute__((aligned(8))) temp0; +// FIXME remove static uint64_t __attribute__((aligned(8))) asm_yalpha1; static uint64_t __attribute__((aligned(8))) asm_uvalpha1; - -static int16_t __attribute__((aligned(8))) *lumPixBuf[2000]; -static int16_t __attribute__((aligned(8))) *chrPixBuf[2000]; -static int16_t __attribute__((aligned(8))) hLumFilter[8000]; -static int16_t __attribute__((aligned(8))) hLumFilterPos[2000]; -static int16_t __attribute__((aligned(8))) hChrFilter[8000]; -static int16_t __attribute__((aligned(8))) hChrFilterPos[2000]; -static int16_t __attribute__((aligned(8))) vLumFilter[8000]; -static int16_t __attribute__((aligned(8))) vLumFilterPos[2000]; -static int16_t __attribute__((aligned(8))) vChrFilter[8000]; -static int16_t __attribute__((aligned(8))) vChrFilterPos[2000]; - -// Contain simply the values from v(Lum|Chr)Filter just nicely packed for mmx -//FIXME these are very likely too small / 8000 caused problems with 480x480 -static int16_t __attribute__((aligned(8))) lumMmxFilter[16000]; -static int16_t __attribute__((aligned(8))) chrMmxFilter[16000]; -#else -static int16_t *lumPixBuf[2000]; -static int16_t *chrPixBuf[2000]; -static int16_t hLumFilter[8000]; -static int16_t hLumFilterPos[2000]; -static int16_t hChrFilter[8000]; -static int16_t hChrFilterPos[2000]; -static int16_t vLumFilter[8000]; -static int16_t vLumFilterPos[2000]; -static int16_t vChrFilter[8000]; -static int16_t vChrFilterPos[2000]; -//FIXME just dummy vars -static int16_t lumMmxFilter[1]; -static int16_t chrMmxFilter[1]; #endif // clipping helper table for C implementations: @@ -159,28 +140,22 @@ static int clip_yuvtab_0c92[768]; static int clip_yuvtab_1a1e[768]; static int clip_yuvtab_40cf[768]; -static int hLumFilterSize=0; -static int hChrFilterSize=0; -static int vLumFilterSize=0; -static int vChrFilterSize=0; -static int vLumBufSize=0; -static int vChrBufSize=0; - +//global sws_flags from the command line int sws_flags=0; -#ifdef CAN_COMPILE_X86_ASM -static uint8_t funnyYCode[10000]; -static uint8_t funnyUVCode[10000]; -#endif +/* cpuCaps combined from cpudetect and whats actually compiled in + (if there is no support for something compiled in it wont appear here) */ +static CpuCaps cpuCaps; -static int canMMX2BeUsed=0; +void (*swScale)(SwsContext *context, uint8_t* src[], int srcStride[], int srcSliceY, + int srcSliceH, uint8_t* dst[], int dstStride[])=NULL; #ifdef CAN_COMPILE_X86_ASM void in_asm_used_var_warning_killer() { volatile int i= yCoeff+vrCoeff+ubCoeff+vgCoeff+ugCoeff+bF8+bFC+w400+w80+w10+ - bm00001111+bm00000111+bm11111000+b16Mask+g16Mask+r16Mask+b15Mask+g15Mask+r15Mask+temp0+asm_yalpha1+ asm_uvalpha1+ - M24A+M24B+M24C+w02 + funnyYCode[0]+ funnyUVCode[0]+b5Dither+g5Dither+r5Dither+g6Dither+dither4[0]+dither8[0]; + bm00001111+bm00000111+bm11111000+b16Mask+g16Mask+r16Mask+b15Mask+g15Mask+r15Mask+asm_yalpha1+ asm_uvalpha1+ + M24A+M24B+M24C+w02 + b5Dither+g5Dither+r5Dither+g6Dither+dither4[0]+dither8[0]; if(i) i=0; } #endif @@ -220,9 +195,9 @@ static inline void yuv2yuvXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilt static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize, int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize, - uint8_t *dest, int dstW, int dstbpp) + uint8_t *dest, int dstW, int dstFormat) { - if(dstbpp==32) + if(dstFormat==IMGFMT_BGR32) { int i; for(i=0; i<(dstW>>1); i++){ @@ -260,7 +235,7 @@ static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilt dest[8*i+6]=clip_table[((Y2 + Cr) >>13)]; } } - else if(dstbpp==24) + else if(dstFormat==IMGFMT_BGR24) { int i; for(i=0; i<(dstW>>1); i++){ @@ -299,7 +274,7 @@ static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilt dest+=6; } } - else if(dstbpp==16) + else if(dstFormat==IMGFMT_BGR16) { int i; for(i=0; i<(dstW>>1); i++){ @@ -339,7 +314,7 @@ static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilt clip_table16r[(Y2 + Cr) >>13]; } } - else if(dstbpp==15) + else if(dstFormat==IMGFMT_BGR15) { int i; for(i=0; i<(dstW>>1); i++){ @@ -467,42 +442,320 @@ static inline void yuv2rgbXinC(int16_t *lumFilter, int16_t **lumSrc, int lumFilt // minor note: the HAVE_xyz is messed up after that line so dont use it -// *** bilinear scaling and yuv->rgb or yuv->yuv conversion of yv12 slices: -// *** Note: it's called multiple times while decoding a frame, first time y==0 -// switching the cpu type during a sliced drawing can have bad effects, like sig11 -void SwScale_YV12slice(unsigned char* srcptr[],int stride[], int srcSliceY , - int srcSliceH, uint8_t* dstptr[], int dststride, int dstbpp, +// old global scaler, dont use for new code +// will use sws_flags from the command line +void SwScale_YV12slice(unsigned char* src[], int srcStride[], int srcSliceY , + int srcSliceH, uint8_t* dst[], int dstStride, int dstbpp, int srcW, int srcH, int dstW, int dstH){ -#ifdef RUNTIME_CPUDETECT -#ifdef CAN_COMPILE_X86_ASM - // ordered per speed fasterst first - if(gCpuCaps.hasMMX2) - SwScale_YV12slice_MMX2(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); - else if(gCpuCaps.has3DNow) - SwScale_YV12slice_3DNow(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); - else if(gCpuCaps.hasMMX) - SwScale_YV12slice_MMX(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); - else - SwScale_YV12slice_C(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); -#else - SwScale_YV12slice_C(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); -#endif -#else //RUNTIME_CPUDETECT -#ifdef HAVE_MMX2 - SwScale_YV12slice_MMX2(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); -#elif defined (HAVE_3DNOW) - SwScale_YV12slice_3DNow(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); -#elif defined (HAVE_MMX) - SwScale_YV12slice_MMX(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); -#else - SwScale_YV12slice_C(srcptr, stride, srcSliceY, srcSliceH, dstptr, dststride, dstbpp, srcW, srcH, dstW, dstH); + static SwsContext *context=NULL; + int dstFormat; + int flags=0; + static int firstTime=1; + int dstStride3[3]= {dstStride, dstStride>>1, dstStride>>1}; + + if(firstTime) + { + flags= SWS_PRINT_INFO; + firstTime=0; + } + + switch(dstbpp) + { + case 8 : dstFormat= IMGFMT_Y8; break; + case 12: dstFormat= IMGFMT_YV12; break; + case 15: dstFormat= IMGFMT_BGR15; break; + case 16: dstFormat= IMGFMT_BGR16; break; + case 24: dstFormat= IMGFMT_BGR24; break; + case 32: dstFormat= IMGFMT_BGR32; break; + default: return; + } + + switch(sws_flags) + { + case 0: flags|= SWS_FAST_BILINEAR; break; + case 1: flags|= SWS_BILINEAR; break; + case 2: flags|= SWS_BICUBIC; break; + case 3: flags|= SWS_X; break; + default:flags|= SWS_BILINEAR; break; + } + + if(!context) context=getSwsContext(srcW, srcH, IMGFMT_YV12, dstW, dstH, dstFormat, flags, NULL, NULL); + + + swScale(context, src, srcStride, srcSliceY, srcSliceH, dst, dstStride3); +} + +static inline void initFilter(int16_t *dstFilter, int16_t *filterPos, int *filterSize, int xInc, + int srcW, int dstW, int filterAlign, int one, int flags) +{ + int i; + double filter[10000]; +#ifdef ARCH_X86 + if(gCpuCaps.hasMMX) + asm volatile("emms\n\t"::: "memory"); //FIXME this shouldnt be required but it IS (even for non mmx versions) #endif -#endif //!RUNTIME_CPUDETECT + if(ABS(xInc - 0x10000) <10) // unscaled + { + int i; + *filterSize= (1 +(filterAlign-1)) & (~(filterAlign-1)); // 1 or 4 normaly + for(i=0; i>16) - (*filterSize>>1) + 1; + int j; + + filterPos[i]= xx; + if((flags & SWS_BICUBIC) || (flags & SWS_X)) + { + double d= ABS(((xx+1)<<16) - xDstInSrc)/(double)(1<<16); + double y1,y2,y3,y4; + double A= -0.6; + if(flags & SWS_BICUBIC){ + // Equation is from VirtualDub + y1 = ( + A*d - 2.0*A*d*d + A*d*d*d); + y2 = (+ 1.0 - (A+3.0)*d*d + (A+2.0)*d*d*d); + y3 = ( - A*d + (2.0*A+3.0)*d*d - (A+2.0)*d*d*d); + y4 = ( + A*d*d - A*d*d*d); + }else{ + // cubic interpolation (derived it myself) + y1 = ( -2.0*d + 3.0*d*d - 1.0*d*d*d)/6.0; + y2 = (6.0 -3.0*d - 6.0*d*d + 3.0*d*d*d)/6.0; + y3 = ( +6.0*d + 3.0*d*d - 3.0*d*d*d)/6.0; + y4 = ( -1.0*d + 1.0*d*d*d)/6.0; + } + +// printf("%d %d %d \n", coeff, (int)d, xDstInSrc); + filter[i*(*filterSize) + 0]= y1; + filter[i*(*filterSize) + 1]= y2; + filter[i*(*filterSize) + 2]= y3; + filter[i*(*filterSize) + 3]= y4; +// printf("%1.3f %1.3f %1.3f %1.3f %1.3f\n",d , y1, y2, y3, y4); + } + else + { + for(j=0; j<*filterSize; j++) + { + double d= ABS((xx<<16) - xDstInSrc)/(double)(1<<16); + double coeff= 1.0 - d; + if(coeff<0) coeff=0; + // printf("%d %d %d \n", coeff, (int)d, xDstInSrc); + filter[i*(*filterSize) + j]= coeff; + xx++; + } + } + xDstInSrc+= xInc; + } + } + else // downscale + { + int xDstInSrc; + if(flags&SWS_BICUBIC) *filterSize= (int)ceil(1 + 4.0*srcW / (double)dstW); + else if(flags&SWS_X) *filterSize= (int)ceil(1 + 4.0*srcW / (double)dstW); + else *filterSize= (int)ceil(1 + 2.0*srcW / (double)dstW); +// printf("%d %d %d\n", *filterSize, srcW, dstW); + *filterSize= (*filterSize +(filterAlign-1)) & (~(filterAlign-1)); + + xDstInSrc= xInc/2 - 0x8000; + for(i=0; i srcW) + { + int shift= filterPos[i] + (*filterSize) - srcW; + // Move filter coeffs right to compensate for filterPos + for(j=(*filterSize)-2; j>=0; j--) + { + int right= MIN(j + shift, (*filterSize)-1); + filter[i*(*filterSize) +right] += filter[i*(*filterSize) +j]; + filter[i*(*filterSize) +j]=0; + } + filterPos[i]= srcW - (*filterSize); + } + } + + //FIXME try to align filterpos if possible / try to shift filterpos to put zeros at the end + // and skip these than later + + //Normalize + for(i=0; i>16 + carry + "pshufw $0xFF, %%mm0, %%mm0 \n\t" + "2: \n\t" + "psrlw $9, %%mm3 \n\t" + "psubw %%mm1, %%mm0 \n\t" + "pmullw %%mm3, %%mm0 \n\t" + "paddw %%mm6, %%mm2 \n\t" // 2*alpha += xpos&0xFFFF + "psllw $7, %%mm1 \n\t" + "paddw %%mm1, %%mm0 \n\t" + + "movq %%mm0, (%%edi, %%eax) \n\t" + + "addl $8, %%eax \n\t" + // End + "9: \n\t" +// "int $3\n\t" + "leal 0b, %0 \n\t" + "leal 1b, %1 \n\t" + "leal 2b, %2 \n\t" + "decl %1 \n\t" + "decl %2 \n\t" + "subl %0, %1 \n\t" + "subl %0, %2 \n\t" + "leal 9b, %3 \n\t" + "subl %0, %3 \n\t" + :"=r" (fragment), "=r" (imm8OfPShufW1), "=r" (imm8OfPShufW2), + "=r" (fragmentLength) + ); + + xpos= 0; //lumXInc/2 - 0x8000; // difference between pixel centers + + for(i=0; i>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; + + memcpy(funnyCode + fragmentLength*i/4, fragment, fragmentLength); + + funnyCode[fragmentLength*i/4 + imm8OfPShufW1]= + funnyCode[fragmentLength*i/4 + imm8OfPShufW2]= + a | (b<<2) | (c<<4) | (d<<6); + + // if we dont need to read 8 bytes than dont :), reduces the chance of + // crossing a cache line + if(d<3) funnyCode[fragmentLength*i/4 + 1]= 0x6E; + + funnyCode[fragmentLength*(i+4)/4]= RET; + } + xpos+=xInc; + } +} +#endif // ARCH_X86 + +//FIXME remove void SwScale_Init(){ +} + +static void globalInit(){ // generating tables: int i; for(i=0; i<768; i++){ @@ -517,7 +770,7 @@ void SwScale_Init(){ for(i=0; i<768; i++) { - int v= clip_table[i]; + int v= clip_table[i]; clip_table16b[i]= v>>3; clip_table16g[i]= (v<<3)&0x07E0; clip_table16r[i]= (v<<8)&0xF800; @@ -526,5 +779,346 @@ void SwScale_Init(){ clip_table15r[i]= (v<<7)&0x7C00; } +cpuCaps= gCpuCaps; + +#ifdef RUNTIME_CPUDETECT +#ifdef CAN_COMPILE_X86_ASM + // ordered per speed fasterst first + if(gCpuCaps.hasMMX2) + swScale= swScale_MMX2; + else if(gCpuCaps.has3DNow) + swScale= swScale_3DNOW; + else if(gCpuCaps.hasMMX) + swScale= swScale_MMX; + else + swScale= swScale_C; + +#else + swScale= swScale_C; + cpuCaps.hasMMX2 = cpuCaps.hasMMX = cpuCaps.has3DNow = 0; +#endif +#else //RUNTIME_CPUDETECT +#ifdef HAVE_MMX2 + swScale= swScale_MMX2; + cpuCaps.has3DNow = 0; +#elif defined (HAVE_3DNOW) + swScale= swScale_3DNOW; + cpuCaps.hasMMX2 = 0; +#elif defined (HAVE_MMX) + swScale= swScale_MMX; + cpuCaps.hasMMX2 = cpuCaps.has3DNow = 0; +#else + swScale= swScale_C; + cpuCaps.hasMMX2 = cpuCaps.hasMMX = cpuCaps.has3DNow = 0; +#endif +#endif //!RUNTIME_CPUDETECT } + +SwsContext *getSwsContext(int srcW, int srcH, int srcFormat, int dstW, int dstH, int dstFormat, int flags, + SwsFilter *srcFilter, SwsFilter *dstFilter){ + + const int widthAlign= dstFormat==IMGFMT_YV12 ? 16 : 8; + SwsContext *c; + int i; +//const int bytespp= (dstbpp+1)/8; //(12->1, 15&16->2, 24->3, 32->4) +//const int over= dstFormat==IMGFMT_YV12 ? (((dstW+15)&(~15))) - dststride +// : (((dstW+7)&(~7)))*bytespp - dststride; + if(swScale==NULL) globalInit(); + + /* sanity check */ + if(srcW<1 || srcH<1 || dstW<1 || dstH<1) return NULL; + if(srcW>=SWS_MAX_SIZE || dstW>=SWS_MAX_SIZE || srcH>=SWS_MAX_SIZE || dstH>=SWS_MAX_SIZE) + { + fprintf(stderr, "size is too large, increase SWS_MAX_SIZE\n"); + return NULL; + } + +/* FIXME + if(dstStride[0]%widthAlign !=0 ) + { + if(flags & SWS_PRINT_INFO) + fprintf(stderr, "SwScaler: Warning: dstStride is not a multiple of %d!\n" + "SwScaler: ->cannot do aligned memory acesses anymore\n", + widthAlign); + } +*/ + c= memalign(64, sizeof(SwsContext)); + + c->srcW= srcW; + c->srcH= srcH; + c->dstW= dstW; + c->dstH= dstH; + c->lumXInc= ((srcW<<16) + (1<<15))/dstW; + c->lumYInc= ((srcH<<16) + (1<<15))/dstH; + c->flags= flags; + c->dstFormat= dstFormat; + c->srcFormat= srcFormat; + + if(cpuCaps.hasMMX2) + { + c->canMMX2BeUsed= (dstW >=srcW && (dstW&31)==0 && (srcW&15)==0) ? 1 : 0; + if(!c->canMMX2BeUsed && dstW >=srcW && (srcW&15)==0 && (flags&SWS_FAST_BILINEAR)) + { + if(flags&SWS_PRINT_INFO) + fprintf(stderr, "SwScaler: output Width is not a multiple of 32 -> no MMX2 scaler\n"); + } + } + else + c->canMMX2BeUsed=0; + + // match pixel 0 of the src to pixel 0 of dst and match pixel n-2 of src to pixel n-2 of dst + // but only for the FAST_BILINEAR mode otherwise do correct scaling + // n-2 is the last chrominance sample available + // this is not perfect, but noone shuld notice the difference, the more correct variant + // would be like the vertical one, but that would require some special code for the + // first and last pixel + if(flags&SWS_FAST_BILINEAR) + { + if(c->canMMX2BeUsed) c->lumXInc+= 20; + //we dont use the x86asm scaler if mmx is available + else if(cpuCaps.hasMMX) c->lumXInc = ((srcW-2)<<16)/(dstW-2) - 20; + } + + /* set chrXInc & chrDstW */ + if((flags&SWS_FULL_UV_IPOL) && dstFormat!=IMGFMT_YV12) + c->chrXInc= c->lumXInc>>1, c->chrDstW= dstW; + else + c->chrXInc= c->lumXInc, c->chrDstW= (dstW+1)>>1; + + /* set chrYInc & chrDstH */ + if(dstFormat==IMGFMT_YV12) c->chrYInc= c->lumYInc, c->chrDstH= (dstH+1)>>1; + else c->chrYInc= c->lumYInc>>1, c->chrDstH= dstH; + + /* precalculate horizontal scaler filter coefficients */ + { + const int filterAlign= cpuCaps.hasMMX ? 4 : 1; + + initFilter(c->hLumFilter, c->hLumFilterPos, &c->hLumFilterSize, c->lumXInc, + srcW , dstW, filterAlign, 1<<14, flags); + initFilter(c->hChrFilter, c->hChrFilterPos, &c->hChrFilterSize, c->chrXInc, + (srcW+1)>>1, c->chrDstW, filterAlign, 1<<14, flags); + +#ifdef ARCH_X86 +// cant downscale !!! + if(c->canMMX2BeUsed && (flags & SWS_FAST_BILINEAR)) + { + initMMX2HScaler( dstW, c->lumXInc, c->funnyYCode); + initMMX2HScaler(c->chrDstW, c->chrXInc, c->funnyUVCode); + } +#endif + } // Init Horizontal stuff + + + + /* precalculate vertical scaler filter coefficients */ + initFilter(c->vLumFilter, c->vLumFilterPos, &c->vLumFilterSize, c->lumYInc, + srcH , dstH, 1, (1<<12)-4, flags); + initFilter(c->vChrFilter, c->vChrFilterPos, &c->vChrFilterSize, c->chrYInc, + (srcH+1)>>1, c->chrDstH, 1, (1<<12)-4, flags); + + // Calculate Buffer Sizes so that they wont run out while handling these damn slices + c->vLumBufSize= c->vLumFilterSize; + c->vChrBufSize= c->vChrFilterSize; + for(i=0; ichrDstH / dstH; + int nextSlice= MAX(c->vLumFilterPos[i ] + c->vLumFilterSize - 1, + ((c->vChrFilterPos[chrI] + c->vChrFilterSize - 1)<<1)); + nextSlice&= ~1; // Slices start at even boundaries + if(c->vLumFilterPos[i ] + c->vLumBufSize < nextSlice) + c->vLumBufSize= nextSlice - c->vLumFilterPos[i ]; + if(c->vChrFilterPos[chrI] + c->vChrBufSize < (nextSlice>>1)) + c->vChrBufSize= (nextSlice>>1) - c->vChrFilterPos[chrI]; + } + + // allocate pixbufs (we use dynamic allocation because otherwise we would need to + // allocate several megabytes to handle all possible cases) + for(i=0; ivLumBufSize; i++) + c->lumPixBuf[i]= c->lumPixBuf[i+c->vLumBufSize]= (uint16_t*)memalign(8, 4000); + for(i=0; ivChrBufSize; i++) + c->chrPixBuf[i]= c->chrPixBuf[i+c->vChrBufSize]= (uint16_t*)memalign(8, 8000); + + //try to avoid drawing green stuff between the right end and the stride end + for(i=0; ivLumBufSize; i++) memset(c->lumPixBuf[i], 0, 4000); + for(i=0; ivChrBufSize; i++) memset(c->chrPixBuf[i], 64, 8000); + + ASSERT(c->chrDstH <= dstH) + ASSERT(c->vLumFilterSize* dstH*4 <= SWS_MAX_SIZE*20) + ASSERT(c->vChrFilterSize*c->chrDstH*4 <= SWS_MAX_SIZE*20) + + // pack filter data for mmx code + if(cpuCaps.hasMMX) + { + for(i=0; ivLumFilterSize*dstH; i++) + c->lumMmxFilter[4*i]=c->lumMmxFilter[4*i+1]=c->lumMmxFilter[4*i+2]=c->lumMmxFilter[4*i+3]= + c->vLumFilter[i]; + for(i=0; ivChrFilterSize*c->chrDstH; i++) + c->chrMmxFilter[4*i]=c->chrMmxFilter[4*i+1]=c->chrMmxFilter[4*i+2]=c->chrMmxFilter[4*i+3]= + c->vChrFilter[i]; + } + + if(flags&SWS_PRINT_INFO) + { +#ifdef DITHER1XBPP + char *dither= cpuCaps.hasMMX ? " dithered" : ""; +#endif + if(flags&SWS_FAST_BILINEAR) + fprintf(stderr, "\nSwScaler: FAST_BILINEAR scaler "); + else if(flags&SWS_BILINEAR) + fprintf(stderr, "\nSwScaler: BILINEAR scaler "); + else if(flags&SWS_BICUBIC) + fprintf(stderr, "\nSwScaler: BICUBIC scaler "); + else + fprintf(stderr, "\nSwScaler: ehh flags invalid?! "); + + if(dstFormat==IMGFMT_BGR15) + fprintf(stderr, "with%s BGR15 output ", dither); + else if(dstFormat==IMGFMT_BGR16) + fprintf(stderr, "with%s BGR16 output ", dither); + else if(dstFormat==IMGFMT_BGR24) + fprintf(stderr, "with BGR24 output "); + else if(dstFormat==IMGFMT_BGR32) + fprintf(stderr, "with BGR32 output "); + else if(dstFormat==IMGFMT_YV12) + fprintf(stderr, "with YV12 output "); + else + fprintf(stderr, "without output "); + + if(cpuCaps.hasMMX2) + fprintf(stderr, "using MMX2\n"); + else if(cpuCaps.has3DNow) + fprintf(stderr, "using 3DNOW\n"); + else if(cpuCaps.hasMMX) + fprintf(stderr, "using MMX\n"); + else + fprintf(stderr, "using C\n"); + } + + if((flags & SWS_PRINT_INFO) && verbose) + { + if(cpuCaps.hasMMX) + { + if(c->canMMX2BeUsed && (flags&SWS_FAST_BILINEAR)) + printf("SwScaler: using FAST_BILINEAR MMX2 scaler for horizontal scaling\n"); + else + { + if(c->hLumFilterSize==4) + printf("SwScaler: using 4-tap MMX scaler for horizontal luminance scaling\n"); + else if(c->hLumFilterSize==8) + printf("SwScaler: using 8-tap MMX scaler for horizontal luminance scaling\n"); + else + printf("SwScaler: using n-tap MMX scaler for horizontal luminance scaling\n"); + + if(c->hChrFilterSize==4) + printf("SwScaler: using 4-tap MMX scaler for horizontal chrominance scaling\n"); + else if(c->hChrFilterSize==8) + printf("SwScaler: using 8-tap MMX scaler for horizontal chrominance scaling\n"); + else + printf("SwScaler: using n-tap MMX scaler for horizontal chrominance scaling\n"); + } + } + else + { +#ifdef ARCH_X86 + printf("SwScaler: using X86-Asm scaler for horizontal scaling\n"); +#else + if(flags & SWS_FAST_BILINEAR) + printf("SwScaler: using FAST_BILINEAR C scaler for horizontal scaling\n"); + else + printf("SwScaler: using C scaler for horizontal scaling\n"); +#endif + } + if(dstFormat==IMGFMT_YV12) + { + if(c->vLumFilterSize==1) + printf("SwScaler: using 1-tap %s \"scaler\" for vertical scaling (YV12)\n", cpuCaps.hasMMX ? "MMX" : "C"); + else + printf("SwScaler: using n-tap %s scaler for vertical scaling (YV12)\n", cpuCaps.hasMMX ? "MMX" : "C"); + } + else + { + if(c->vLumFilterSize==1 && c->vChrFilterSize==2) + printf("SwScaler: using 1-tap %s \"scaler\" for vertical luminance scaling (BGR)\n" + "SwScaler: 2-tap scaler for vertical chrominance scaling (BGR)\n",cpuCaps.hasMMX ? "MMX" : "C"); + else if(c->vLumFilterSize==2 && c->vChrFilterSize==2) + printf("SwScaler: using 2-tap linear %s scaler for vertical scaling (BGR)\n", cpuCaps.hasMMX ? "MMX" : "C"); + else + printf("SwScaler: using n-tap %s scaler for vertical scaling (BGR)\n", cpuCaps.hasMMX ? "MMX" : "C"); + } + + if(dstFormat==IMGFMT_BGR24) + printf("SwScaler: using %s YV12->BGR24 Converter\n", + cpuCaps.hasMMX2 ? "MMX2" : (cpuCaps.hasMMX ? "MMX" : "C")); + else + printf("SwScaler: using %s YV12->BGR Converter\n", cpuCaps.hasMMX ? "MMX" : "C");//FIXME print format + + printf("SwScaler: %dx%d -> %dx%d\n", srcW, srcH, dstW, dstH); + } + + return c; +} + +/** + * returns a normalized gaussian curve used to filter stuff + * quality=3 is high quality, lowwer is lowwer quality + */ +double *getGaussian(double variance, double quality){ + const int length= (int)(variance*quality + 0.5) | 1; + int i; + double *coeff= memalign(sizeof(double), length*sizeof(double)); + double middle= (length-1)*0.5; + + for(i=0; i