MythTV: zoom_filter_xmmx.c Source File

00001 #include <mythtv/mythconfig.h>
00002 
00003 /* Prototypes to keep gcc from spewing warnings */
00004 void    zoom_filter_xmmx (int prevX, int prevY, unsigned int *expix1, unsigned int *expix2, int *brutS, int *brutD, int buffratio, int precalCoef[16][16]);
00005 int     zoom_filter_xmmx_supported (void);
00006 
00007 
00008 #if defined(MMX) && !defined(ARCH_X86_64)
00009 /* a definir pour avoir exactement le meme resultat que la fonction C
00010  * (un chouillat plus lent)
00011  */
00012 #define STRICT_COMPAT
00013 //#define HAVE_ATHLON
00014 
00015 #define BUFFPOINTNB 16
00016 #define BUFFPOINTMASK 0xffff
00017 #define BUFFINCR 0xff
00018 
00019 #define sqrtperte 16
00020 // faire : a % sqrtperte <=> a & pertemask
00021 #define PERTEMASK 0xf
00022 // faire : a / sqrtperte <=> a >> PERTEDEC
00023 #define PERTEDEC 4
00024 
00025 
00026 //#define MMX_TRACE
00027 #include "mmx.h"
00028 
00029 int zoom_filter_xmmx_supported () {
00030     return (av_get_cpu_flags() & AV_CPU_FLAG_SSE) >> 3;
00031 }
00032 
00033 void zoom_filter_xmmx (int prevX, int prevY,
00034                                                                                          unsigned int *expix1, unsigned int *expix2,
00035                                                                                          int *lbruS, int *lbruD, int buffratio,
00036                                                                                          int precalCoef[16][16])
00037 {
00038   int bufsize = prevX * prevY; /* taille du buffer */
00039   volatile int loop;                    /* variable de boucle */
00040 
00041         mmx_t *brutS = (mmx_t*)lbruS; /* buffer de transformation source */
00042         mmx_t *brutD = (mmx_t*)lbruD; /* buffer de transformation dest */
00043 
00044   volatile mmx_t prevXY;
00045         volatile mmx_t ratiox;
00046         //volatile mmx_t interpix;
00047 
00048         expix1[0]=expix1[prevX-1]=expix1[prevX*prevY-1]=expix1[prevX*prevY-prevX]=0;
00049 
00050         prevXY.ud[0] = (prevX-1)<<PERTEDEC;
00051         prevXY.ud[1] = (prevY-1)<<PERTEDEC;
00052         
00053         ratiox.d[0] = buffratio;
00054         ratiox.d[1] = buffratio;
00055         movq_m2r (ratiox, mm6);
00056         pslld_i2r (16,mm6);
00057 
00058   pxor_r2r (mm7,mm7); /* mise a zero de mm7 */
00059                 
00060   loop=0;
00061 
00062         /*
00063          * NOTE : mm6 et mm7 ne sont pas modifies dans la boucle.
00064          */
00065         while (loop < bufsize)
00066                 {
00067                         /*
00068                          * pre : mm6 = [buffratio<<16|buffratio<<16]
00069                          * post : mm0 = S + ((D-S)*buffratio)>>16 format [X|Y]
00070                          * modified = mm0,mm1,mm2
00071                          */
00072                         
00073                         __asm__ __volatile__ (
00074                                 "movq %0,%%mm0\n"
00075                                 "movq %1,%%mm1\n"
00076                                 : :"m"(brutS[loop]),"m"(brutD[loop])
00077                                 );               /* mm0 = S */
00078                         
00079                         psubd_r2r (mm0,mm1);           /* mm1 = D - S */
00080                         movq_r2r (mm1, mm2);           /* mm2 = D - S */
00081                         
00082                         pslld_i2r (16,mm1);
00083                         mmx_r2r (pmulhuw, mm6, mm1);   /* mm1 = ?? */
00084                         pmullw_r2r (mm6, mm2);
00085                 
00086                         paddd_r2r (mm2, mm1);     /* mm1 = (D - S) * buffratio >> 16 */
00087                         pslld_i2r (16,mm0);
00088                         
00089                         paddd_r2r (mm1, mm0);     /* mm0 = S + mm1 */
00090                         psrld_i2r (16, mm0);
00091                         
00092                         /*
00093                          * pre : mm0 : position vector on screen
00094                          *       prevXY : coordinate of the lower-right point on screen
00095                          * post : clipped mm0
00096                          * modified : mm0,mm1,mm2
00097                          */
00098                         movq_m2r (prevXY,mm1);
00099                         pcmpgtd_r2r (mm0, mm1); /* mm0 en X contient :
00100                                                                                                                                  1111 si prevXY > px
00101                                                                                                                                  0000 si prevXY <= px
00102                                                                                                                                  (idem pour y) */
00103 #ifdef STRICT_COMPAT
00104                         movq_r2r (mm1,mm2);
00105                         punpckhdq_r2r (mm2,mm2);
00106                         punpckldq_r2r (mm1,mm1);
00107                         pand_r2r (mm2, mm0);
00108 #endif
00109                         pand_r2r (mm1, mm0);    /* on met a zero la partie qui deborde */
00110                         
00111                         /*
00112                          * pre : mm0 : clipped position on screen
00113                          *
00114                          * post : mm3 & mm4 : coefs for this position
00115                          *              mm1 : X vector [0|X]
00116                          *
00117                          * modif : eax,ecx
00118                          */
00119                         __asm__ __volatile__ (
00120                                 "movd %%mm0,%%ecx\n"
00121                                 "movq %%mm0,%%mm1\n"
00122 
00123                                 "andl $15,%%ecx\n"
00124                                 "psrlq $32,%%mm1\n"
00125 
00126                                 "shll $6,%%ecx\n"
00127                                 "movd %%mm1,%%eax\n"
00128 
00129                                 "addl %0,%%ecx\n"
00130                                 "andl $15,%%eax\n"
00131 
00132                                 "movd (%%ecx,%%eax,4),%%mm3\n"
00133                                 /* ::"X"(precalCoef):"eax","ecx"); */
00134                                 ::"m"(precalCoef):"eax","ecx");
00135                                 
00136 
00137                         /*
00138                          * extraction des coefficients...
00139                          *
00140                          * pre : coef dans mm3
00141                          *
00142                          * post : coef extraits dans mm3 (c1 & c2)
00143                          *                        et mm4 (c3 & c4)
00144                          *
00145                          * modif : mm5
00146                          */
00147 // entrelace avec portion d'apres (cf les '^')
00148 //                      movq_r2r (mm3, mm5);            /* ??-??-??-??-c4-c3-c2-c1 */
00149 //                      punpcklbw_r2r (mm5, mm3);       /* c4-c4-c3-c3-c2-c2-c1-c1 */
00150 //                      movq_r2r (mm3, mm4);            /* c4-c4-c3-c3-c2-c2-c1-c1 */
00151 //                      movq_r2r (mm3, mm5);            /* c4-c4-c3-c3-c2-c2-c1-c1 */
00152                         
00153 //                      punpcklbw_r2r (mm5, mm3);       /* c2-c2-c2-c2-c1-c1-c1-c1 */
00154 //                      punpckhbw_r2r (mm5, mm4);       /* c4-c4-c4-c4-c3-c3-c3-c3 */
00155                         
00156                         /*
00157                          * pre : mm0 : Y pos [*|Y]
00158                          *       mm1 : X pos [*|X]
00159                          *
00160                          * post : mm0 : expix1[position]
00161                          *        mm2 : expix1[position+largeur]
00162                          *
00163                          * modif : eax,ecx
00164                          */
00165                         psrld_i2r (PERTEDEC,mm0);
00166                         psrld_i2r (PERTEDEC,mm1);
00167                         __asm__ __volatile__ (
00168                                 "movd %%mm1,%%eax\n"
00169                                 /*^*/ "movq %%mm3,%%mm5\n"       /*^*/
00170 
00171                                 "mull %1\n"
00172                                 "movd %%mm0,%%ecx\n"
00173                                 /*^*/ "punpcklbw %%mm5, %%mm3\n" /*^*/
00174 
00175                                 "addl %%ecx,%%eax\n"
00176                                 /*^*/ "movq %%mm3,%%mm4\n"       /*^*/
00177                                 /*^*/ "movq %%mm3,%%mm5\n"       /*^*/
00178 
00179                                 "movl %0,%%ecx\n"
00180                                 /*^*/ "punpcklbw %%mm5,%%mm3\n"  /*^*/
00181 
00182                                 "movq (%%ecx,%%eax,4),%%mm0\n"
00183                                 /*^*/ "punpckhbw %%mm5,%%mm4\n"  /*^*/
00184 
00185                                 "addl %1,%%eax\n"
00186                                 "movq (%%ecx,%%eax,4),%%mm2\n"
00187                                 
00188                                 : : "X"(expix1), "X"(prevX):"eax","ecx"
00189                                 );
00190 
00191                         /*
00192                          * pre :       mm0 : expix1[position]
00193                          *             mm2 : expix1[position+largeur]
00194                          *       mm3 & mm4 : coefs
00195                          */
00196 
00197                         /* recopie des deux premiers pixels dans mm0 et mm1 */
00198                         movq_r2r (mm0, mm1);            /* b1-v1-r1-a1-b2-v2-r2-a2 */
00199 
00200                         /* depackage du premier pixel */
00201                         punpcklbw_r2r (mm7, mm0);       /* 00-b2-00-v2-00-r2-00-a2 */
00202 
00203                         /* extraction des coefficients... */
00204 
00205                         movq_r2r (mm3, mm5);      /* c2-c2-c2-c2-c1-c1-c1-c1 */
00206 
00207                         /*^en parrallele^*/ /* depackage du 2ieme pixel */
00208                         /*^*/ punpckhbw_r2r (mm7, mm1); /* 00-b1-00-v1-00-r1-00-a1 */
00209 
00210                         punpcklbw_r2r (mm7, mm5);       /* 00-c1-00-c1-00-c1-00-c1 */
00211                         punpckhbw_r2r (mm7, mm3);       /* 00-c2-00-c2-00-c2-00-c2 */
00212                         
00213                         /* multiplication des pixels par les coefficients */
00214                         pmullw_r2r (mm5, mm0);          /* c1*b2-c1*v2-c1*r2-c1*a2 */
00215                         pmullw_r2r (mm3, mm1);          /* c2*b1-c2*v1-c2*r1-c2*a1 */
00216                         paddw_r2r (mm1, mm0);
00217                                 
00218                         /* ...extraction des 2 derniers coefficients */
00219                         movq_r2r (mm4, mm5);                    /* c4-c4-c4-c4-c3-c3-c3-c3 */
00220                         punpcklbw_r2r (mm7, mm4);       /* 00-c3-00-c3-00-c3-00-c3 */
00221                         punpckhbw_r2r (mm7, mm5);       /* 00-c4-00-c4-00-c4-00-c4 */
00222                         
00223                         /* recuperation des 2 derniers pixels */
00224                         movq_r2r (mm2, mm1);
00225                         
00226                         /* depackage des pixels */
00227                         punpcklbw_r2r (mm7, mm1);
00228                         punpckhbw_r2r (mm7, mm2);
00229                         
00230                         /* multiplication pas les coeffs */
00231                         pmullw_r2r (mm4, mm1);
00232                         pmullw_r2r (mm5, mm2);
00233                            
00234                         /* ajout des valeurs obtenues � la valeur finale */
00235                         paddw_r2r (mm1, mm0);
00236                         paddw_r2r (mm2, mm0);
00237                            
00238                         /* division par 256 = 16+16+16+16, puis repackage du pixel final */
00239                         psrlw_i2r (8, mm0);
00240                         packuswb_r2r (mm7, mm0);
00241 
00242                         movd_r2m (mm0,expix2[loop]);
00243                         
00244                         ++loop;
00245                 }
00246 #ifdef HAVE_ATHLON
00247         __asm__ __volatile__ ("femms\n");
00248 #else
00249         emms();
00250 #endif
00251 }
00252 #else
00253 int zoom_filter_xmmx_supported () {
00254         return 0;
00255 }
00256 void zoom_filter_xmmx (int prevX, int prevY,
00257                                                                                          unsigned int *expix1, unsigned int *expix2,
00258                                                                                          int *lbruS, int *lbruD, int buffratio,
00259                                                                                          int precalCoef[16][16])
00260 {
00261     (void) prevX;     (void) prevY;
00262     (void) expix1;    (void) expix2;
00263     (void) lbruS;     (void) lbruD;
00264     (void) buffratio; (void) precalCoef;
00265         return;
00266 }
00267 #endif