FFmpeg: libavcodec/x86/idct

00001 /*
00002  * idct_mmx.c
00003  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
00004  *
00005  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
00006  * See http://libmpeg2.sourceforge.net/ for updates.
00007  *
00008  * mpeg2dec is free software; you can redistribute it and/or modify
00009  * it under the terms of the GNU General Public License as published by
00010  * the Free Software Foundation; either version 2 of the License, or
00011  * (at your option) any later version.
00012  *
00013  * mpeg2dec is distributed in the hope that it will be useful,
00014  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00015  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
00016  * GNU General Public License for more details.
00017  *
00018  * You should have received a copy of the GNU General Public License
00019  * along with mpeg2dec; if not, write to the Free Software
00020  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00021  */
00022 
00023 #include "libavutil/common.h"
00024 #include "libavcodec/dsputil.h"
00025 
00026 #include "mmx.h"
00027 
00028 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
00029 
00030 #define ROW_SHIFT 11
00031 #define COL_SHIFT 6
00032 
00033 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
00034 #define rounder(bias) {round (bias), round (bias)}
00035 
00036 
00037 #if 0
00038 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
00039 static inline void idct_row (int16_t * row, int offset,
00040                              int16_t * table, int32_t * rounder)
00041 {
00042     int C1, C2, C3, C4, C5, C6, C7;
00043     int a0, a1, a2, a3, b0, b1, b2, b3;
00044 
00045     row += offset;
00046 
00047     C1 = table[1];
00048     C2 = table[2];
00049     C3 = table[3];
00050     C4 = table[4];
00051     C5 = table[5];
00052     C6 = table[6];
00053     C7 = table[7];
00054 
00055     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
00056     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
00057     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
00058     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
00059 
00060     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
00061     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
00062     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
00063     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
00064 
00065     row[0] = (a0 + b0) >> ROW_SHIFT;
00066     row[1] = (a1 + b1) >> ROW_SHIFT;
00067     row[2] = (a2 + b2) >> ROW_SHIFT;
00068     row[3] = (a3 + b3) >> ROW_SHIFT;
00069     row[4] = (a3 - b3) >> ROW_SHIFT;
00070     row[5] = (a2 - b2) >> ROW_SHIFT;
00071     row[6] = (a1 - b1) >> ROW_SHIFT;
00072     row[7] = (a0 - b0) >> ROW_SHIFT;
00073 }
00074 #endif
00075 
00076 
00077 /* MMXEXT row IDCT */
00078 
00079 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
00080                                                    c4,  c6,  c4,  c6,   \
00081                                                    c1,  c3, -c1, -c5,   \
00082                                                    c5,  c7,  c3, -c7,   \
00083                                                    c4, -c6,  c4, -c6,   \
00084                                                   -c4,  c2,  c4, -c2,   \
00085                                                    c5, -c1,  c3, -c1,   \
00086                                                    c7,  c3,  c7, -c5 }
00087 
00088 static inline void mmxext_row_head (int16_t * const row, const int offset,
00089                                     const int16_t * const table)
00090 {
00091     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00092 
00093     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00094     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00095 
00096     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
00097     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00098 
00099     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
00100     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00101 
00102     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
00103 }
00104 
00105 static inline void mmxext_row (const int16_t * const table,
00106                                const int32_t * const rounder)
00107 {
00108     movq_m2r (*(table+8), mm1);         /* mm1 = -C5 -C1 C3 C1 */
00109     pmaddwd_r2r (mm2, mm4);             /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
00110 
00111     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
00112     pshufw_r2r (mm6, mm6, 0x4e);        /* mm6 = x3 x1 x7 x5 */
00113 
00114     movq_m2r (*(table+12), mm7);        /* mm7 = -C7 C3 C7 C5 */
00115     pmaddwd_r2r (mm5, mm1);             /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
00116 
00117     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
00118     pmaddwd_r2r (mm6, mm7);             /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
00119 
00120     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
00121     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
00122 
00123     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
00124     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
00125 
00126     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
00127     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
00128 
00129     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
00130     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
00131 
00132     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
00133     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
00134 
00135     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
00136     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
00137 
00138     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
00139     movq_r2r (mm0, mm4);                /* mm4 = a3 a2 + rounder */
00140 
00141     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
00142     psubd_r2r (mm5, mm4);               /* mm4 = a3-b3 a2-b2 + rounder */
00143 }
00144 
00145 static inline void mmxext_row_tail (int16_t * const row, const int store)
00146 {
00147     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00148 
00149     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
00150 
00151     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00152 
00153     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
00154 
00155     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00156     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
00157 
00158     /* slot */
00159 
00160     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
00161 }
00162 
00163 static inline void mmxext_row_mid (int16_t * const row, const int store,
00164                                    const int offset,
00165                                    const int16_t * const table)
00166 {
00167     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00168     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00169 
00170     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00171     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
00172 
00173     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00174     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00175 
00176     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
00177     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00178 
00179     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00180     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
00181 
00182     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
00183     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
00184 
00185     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00186 
00187     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
00188     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
00189 }
00190 
00191 
00192 /* MMX row IDCT */
00193 
00194 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00195                                            c4,  c6, -c4, -c2,   \
00196                                            c1,  c3,  c3, -c7,   \
00197                                            c5,  c7, -c1, -c5,   \
00198                                            c4, -c6,  c4, -c2,   \
00199                                           -c4,  c2,  c4, -c6,   \
00200                                            c5, -c1,  c7, -c5,   \
00201                                            c7,  c3,  c3, -c1 }
00202 
00203 static inline void mmx_row_head (int16_t * const row, const int offset,
00204                                  const int16_t * const table)
00205 {
00206     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00207 
00208     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00209     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00210 
00211     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
00212     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00213 
00214     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
00215 
00216     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
00217     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00218 
00219     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
00220     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
00221 }
00222 
00223 static inline void mmx_row (const int16_t * const table,
00224                             const int32_t * const rounder)
00225 {
00226     pmaddwd_r2r (mm2, mm4);             /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
00227     punpckldq_r2r (mm5, mm5);           /* mm5 = x3 x1 x3 x1 */
00228 
00229     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
00230     punpckhdq_r2r (mm6, mm6);           /* mm6 = x7 x5 x7 x5 */
00231 
00232     movq_m2r (*(table+12), mm7);        /* mm7 = -C5 -C1 C7 C5 */
00233     pmaddwd_r2r (mm5, mm1);             /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
00234 
00235     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
00236     pmaddwd_r2r (mm6, mm7);             /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
00237 
00238     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
00239     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
00240 
00241     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
00242     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
00243 
00244     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
00245     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
00246 
00247     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
00248     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
00249 
00250     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
00251     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
00252 
00253     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
00254     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
00255 
00256     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
00257     movq_r2r (mm0, mm7);                /* mm7 = a3 a2 + rounder */
00258 
00259     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
00260     psubd_r2r (mm5, mm7);               /* mm7 = a3-b3 a2-b2 + rounder */
00261 }
00262 
00263 static inline void mmx_row_tail (int16_t * const row, const int store)
00264 {
00265     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00266 
00267     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
00268 
00269     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00270 
00271     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
00272 
00273     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00274     movq_r2r (mm7, mm4);                /* mm4 = y6 y7 y4 y5 */
00275 
00276     pslld_i2r (16, mm7);                /* mm7 = y7 0 y5 0 */
00277 
00278     psrld_i2r (16, mm4);                /* mm4 = 0 y6 0 y4 */
00279 
00280     por_r2r (mm4, mm7);                 /* mm7 = y7 y6 y5 y4 */
00281 
00282     /* slot */
00283 
00284     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
00285 }
00286 
00287 static inline void mmx_row_mid (int16_t * const row, const int store,
00288                                 const int offset, const int16_t * const table)
00289 {
00290     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00291     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00292 
00293     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00294     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
00295 
00296     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00297     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00298 
00299     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
00300     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00301 
00302     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00303     movq_r2r (mm7, mm1);                /* mm1 = y6 y7 y4 y5 */
00304 
00305     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
00306     psrld_i2r (16, mm7);                /* mm7 = 0 y6 0 y4 */
00307 
00308     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
00309     pslld_i2r (16, mm1);                /* mm1 = y7 0 y5 0 */
00310 
00311     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
00312     por_r2r (mm1, mm7);                 /* mm7 = y7 y6 y5 y4 */
00313 
00314     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
00315     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
00316 
00317     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
00318     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00319 }
00320 
00321 
00322 #if 0
00323 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
00324 static inline void idct_col (int16_t * col, int offset)
00325 {
00326 /* multiplication - as implemented on mmx */
00327 #define F(c,x) (((c) * (x)) >> 16)
00328 
00329 /* saturation - it helps us handle torture test cases */
00330 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00331 
00332     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00333     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00334     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00335     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00336 
00337     col += offset;
00338 
00339     x0 = col[0*8];
00340     x1 = col[1*8];
00341     x2 = col[2*8];
00342     x3 = col[3*8];
00343     x4 = col[4*8];
00344     x5 = col[5*8];
00345     x6 = col[6*8];
00346     x7 = col[7*8];
00347 
00348     u04 = S (x0 + x4);
00349     v04 = S (x0 - x4);
00350     u26 = S (F (T2, x6) + x2);
00351     v26 = S (F (T2, x2) - x6);
00352 
00353     a0 = S (u04 + u26);
00354     a1 = S (v04 + v26);
00355     a2 = S (v04 - v26);
00356     a3 = S (u04 - u26);
00357 
00358     u17 = S (F (T1, x7) + x1);
00359     v17 = S (F (T1, x1) - x7);
00360     u35 = S (F (T3, x5) + x3);
00361     v35 = S (F (T3, x3) - x5);
00362 
00363     b0 = S (u17 + u35);
00364     b3 = S (v17 - v35);
00365     u12 = S (u17 - u35);
00366     v12 = S (v17 + v35);
00367     u12 = S (2 * F (C4, u12));
00368     v12 = S (2 * F (C4, v12));
00369     b1 = S (u12 + v12);
00370     b2 = S (u12 - v12);
00371 
00372     y0 = S (a0 + b0) >> COL_SHIFT;
00373     y1 = S (a1 + b1) >> COL_SHIFT;
00374     y2 = S (a2 + b2) >> COL_SHIFT;
00375     y3 = S (a3 + b3) >> COL_SHIFT;
00376 
00377     y4 = S (a3 - b3) >> COL_SHIFT;
00378     y5 = S (a2 - b2) >> COL_SHIFT;
00379     y6 = S (a1 - b1) >> COL_SHIFT;
00380     y7 = S (a0 - b0) >> COL_SHIFT;
00381 
00382     col[0*8] = y0;
00383     col[1*8] = y1;
00384     col[2*8] = y2;
00385     col[3*8] = y3;
00386     col[4*8] = y4;
00387     col[5*8] = y5;
00388     col[6*8] = y6;
00389     col[7*8] = y7;
00390 }
00391 #endif
00392 
00393 
00394 /* MMX column IDCT */
00395 static inline void idct_col (int16_t * const col, const int offset)
00396 {
00397 #define T1 13036
00398 #define T2 27146
00399 #define T3 43790
00400 #define C4 23170
00401 
00402     static const short t1_vector[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
00403     static const short t2_vector[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
00404     static const short t3_vector[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
00405     static const short c4_vector[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
00406 
00407     /* column code adapted from Peter Gubanov */
00408     /* http://www.elecard.com/peter/idct.shtml */
00409 
00410     movq_m2r (*t1_vector, mm0);         /* mm0 = T1 */
00411 
00412     movq_m2r (*(col+offset+1*8), mm1);  /* mm1 = x1 */
00413     movq_r2r (mm0, mm2);                /* mm2 = T1 */
00414 
00415     movq_m2r (*(col+offset+7*8), mm4);  /* mm4 = x7 */
00416     pmulhw_r2r (mm1, mm0);              /* mm0 = T1*x1 */
00417 
00418     movq_m2r (*t3_vector, mm5);         /* mm5 = T3 */
00419     pmulhw_r2r (mm4, mm2);              /* mm2 = T1*x7 */
00420 
00421     movq_m2r (*(col+offset+5*8), mm6);  /* mm6 = x5 */
00422     movq_r2r (mm5, mm7);                /* mm7 = T3-1 */
00423 
00424     movq_m2r (*(col+offset+3*8), mm3);  /* mm3 = x3 */
00425     psubsw_r2r (mm4, mm0);              /* mm0 = v17 */
00426 
00427     movq_m2r (*t2_vector, mm4);         /* mm4 = T2 */
00428     pmulhw_r2r (mm3, mm5);              /* mm5 = (T3-1)*x3 */
00429 
00430     paddsw_r2r (mm2, mm1);              /* mm1 = u17 */
00431     pmulhw_r2r (mm6, mm7);              /* mm7 = (T3-1)*x5 */
00432 
00433     /* slot */
00434 
00435     movq_r2r (mm4, mm2);                /* mm2 = T2 */
00436     paddsw_r2r (mm3, mm5);              /* mm5 = T3*x3 */
00437 
00438     pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
00439     paddsw_r2r (mm6, mm7);              /* mm7 = T3*x5 */
00440 
00441     psubsw_r2r (mm6, mm5);              /* mm5 = v35 */
00442     paddsw_r2r (mm3, mm7);              /* mm7 = u35 */
00443 
00444     movq_m2r (*(col+offset+6*8), mm3);  /* mm3 = x6 */
00445     movq_r2r (mm0, mm6);                /* mm6 = v17 */
00446 
00447     pmulhw_r2r (mm3, mm2);              /* mm2 = T2*x6 */
00448     psubsw_r2r (mm5, mm0);              /* mm0 = b3 */
00449 
00450     psubsw_r2r (mm3, mm4);              /* mm4 = v26 */
00451     paddsw_r2r (mm6, mm5);              /* mm5 = v12 */
00452 
00453     movq_r2m (mm0, *(col+offset+3*8));  /* save b3 in scratch0 */
00454     movq_r2r (mm1, mm6);                /* mm6 = u17 */
00455 
00456     paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
00457     paddsw_r2r (mm7, mm6);              /* mm6 = b0 */
00458 
00459     psubsw_r2r (mm7, mm1);              /* mm1 = u12 */
00460     movq_r2r (mm1, mm7);                /* mm7 = u12 */
00461 
00462     movq_m2r (*(col+offset+0*8), mm3);  /* mm3 = x0 */
00463     paddsw_r2r (mm5, mm1);              /* mm1 = u12+v12 */
00464 
00465     movq_m2r (*c4_vector, mm0);         /* mm0 = C4/2 */
00466     psubsw_r2r (mm5, mm7);              /* mm7 = u12-v12 */
00467 
00468     movq_r2m (mm6, *(col+offset+5*8));  /* save b0 in scratch1 */
00469     pmulhw_r2r (mm0, mm1);              /* mm1 = b1/2 */
00470 
00471     movq_r2r (mm4, mm6);                /* mm6 = v26 */
00472     pmulhw_r2r (mm0, mm7);              /* mm7 = b2/2 */
00473 
00474     movq_m2r (*(col+offset+4*8), mm5);  /* mm5 = x4 */
00475     movq_r2r (mm3, mm0);                /* mm0 = x0 */
00476 
00477     psubsw_r2r (mm5, mm3);              /* mm3 = v04 */
00478     paddsw_r2r (mm5, mm0);              /* mm0 = u04 */
00479 
00480     paddsw_r2r (mm3, mm4);              /* mm4 = a1 */
00481     movq_r2r (mm0, mm5);                /* mm5 = u04 */
00482 
00483     psubsw_r2r (mm6, mm3);              /* mm3 = a2 */
00484     paddsw_r2r (mm2, mm5);              /* mm5 = a0 */
00485 
00486     paddsw_r2r (mm1, mm1);              /* mm1 = b1 */
00487     psubsw_r2r (mm2, mm0);              /* mm0 = a3 */
00488 
00489     paddsw_r2r (mm7, mm7);              /* mm7 = b2 */
00490     movq_r2r (mm3, mm2);                /* mm2 = a2 */
00491 
00492     movq_r2r (mm4, mm6);                /* mm6 = a1 */
00493     paddsw_r2r (mm7, mm3);              /* mm3 = a2+b2 */
00494 
00495     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y2 */
00496     paddsw_r2r (mm1, mm4);              /* mm4 = a1+b1 */
00497 
00498     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y1 */
00499     psubsw_r2r (mm1, mm6);              /* mm6 = a1-b1 */
00500 
00501     movq_m2r (*(col+offset+5*8), mm1);  /* mm1 = b0 */
00502     psubsw_r2r (mm7, mm2);              /* mm2 = a2-b2 */
00503 
00504     psraw_i2r (COL_SHIFT, mm6);         /* mm6 = y6 */
00505     movq_r2r (mm5, mm7);                /* mm7 = a0 */
00506 
00507     movq_r2m (mm4, *(col+offset+1*8));  /* save y1 */
00508     psraw_i2r (COL_SHIFT, mm2);         /* mm2 = y5 */
00509 
00510     movq_r2m (mm3, *(col+offset+2*8));  /* save y2 */
00511     paddsw_r2r (mm1, mm5);              /* mm5 = a0+b0 */
00512 
00513     movq_m2r (*(col+offset+3*8), mm4);  /* mm4 = b3 */
00514     psubsw_r2r (mm1, mm7);              /* mm7 = a0-b0 */
00515 
00516     psraw_i2r (COL_SHIFT, mm5);         /* mm5 = y0 */
00517     movq_r2r (mm0, mm3);                /* mm3 = a3 */
00518 
00519     movq_r2m (mm2, *(col+offset+5*8));  /* save y5 */
00520     psubsw_r2r (mm4, mm3);              /* mm3 = a3-b3 */
00521 
00522     psraw_i2r (COL_SHIFT, mm7);         /* mm7 = y7 */
00523     paddsw_r2r (mm0, mm4);              /* mm4 = a3+b3 */
00524 
00525     movq_r2m (mm5, *(col+offset+0*8));  /* save y0 */
00526     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y4 */
00527 
00528     movq_r2m (mm6, *(col+offset+6*8));  /* save y6 */
00529     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y3 */
00530 
00531     movq_r2m (mm7, *(col+offset+7*8));  /* save y7 */
00532 
00533     movq_r2m (mm3, *(col+offset+4*8));  /* save y4 */
00534 
00535     movq_r2m (mm4, *(col+offset+3*8));  /* save y3 */
00536 
00537 #undef T1
00538 #undef T2
00539 #undef T3
00540 #undef C4
00541 }
00542 
00543 
00544 static const int32_t rounder0[] ATTR_ALIGN(8) =
00545     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
00546 static const int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
00547 static const int32_t rounder1[] ATTR_ALIGN(8) =
00548     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
00549 static const int32_t rounder7[] ATTR_ALIGN(8) =
00550     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
00551 static const int32_t rounder2[] ATTR_ALIGN(8) =
00552     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
00553 static const int32_t rounder6[] ATTR_ALIGN(8) =
00554     rounder (-0.25);                /* C2 * (C6-C2)/2 */
00555 static const int32_t rounder3[] ATTR_ALIGN(8) =
00556     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
00557 static const int32_t rounder5[] ATTR_ALIGN(8) =
00558     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
00559 
00560 #undef COL_SHIFT
00561 #undef ROW_SHIFT
00562 
00563 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
00564 void idct (int16_t * const block)                                       \
00565 {                                                                       \
00566     static const int16_t table04[] ATTR_ALIGN(16) =                     \
00567         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00568     static const int16_t table17[] ATTR_ALIGN(16) =                     \
00569         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00570     static const int16_t table26[] ATTR_ALIGN(16) =                     \
00571         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00572     static const int16_t table35[] ATTR_ALIGN(16) =                     \
00573         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00574                                                                         \
00575     idct_row_head (block, 0*8, table04);                                \
00576     idct_row (table04, rounder0);                                       \
00577     idct_row_mid (block, 0*8, 4*8, table04);                            \
00578     idct_row (table04, rounder4);                                       \
00579     idct_row_mid (block, 4*8, 1*8, table17);                            \
00580     idct_row (table17, rounder1);                                       \
00581     idct_row_mid (block, 1*8, 7*8, table17);                            \
00582     idct_row (table17, rounder7);                                       \
00583     idct_row_mid (block, 7*8, 2*8, table26);                            \
00584     idct_row (table26, rounder2);                                       \
00585     idct_row_mid (block, 2*8, 6*8, table26);                            \
00586     idct_row (table26, rounder6);                                       \
00587     idct_row_mid (block, 6*8, 3*8, table35);                            \
00588     idct_row (table35, rounder3);                                       \
00589     idct_row_mid (block, 3*8, 5*8, table35);                            \
00590     idct_row (table35, rounder5);                                       \
00591     idct_row_tail (block, 5*8);                                         \
00592                                                                         \
00593     idct_col (block, 0);                                                \
00594     idct_col (block, 4);                                                \
00595 }
00596 
00597 void ff_mmx_idct(DCTELEM *block);
00598 void ff_mmxext_idct(DCTELEM *block);
00599 
00600 declare_idct (ff_mmxext_idct, mmxext_table,
00601               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00602 
00603 declare_idct (ff_mmx_idct, mmx_table,
00604               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00605
libavcodec/x86/idct_mmx.c
