00001 /*
00002  * idct_mmx.c
00003  * Copyright (C) 2000-2002 Michel Lespinasse <walken@zoy.org>
00004  * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
00005  *
00006  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
00007  * See http://libmpeg2.sourceforge.net/ for updates.
00008  *
00009  * mpeg2dec is free software; you can redistribute it and/or modify
00010  * it under the terms of the GNU General Public License as published by
00011  * the Free Software Foundation; either version 2 of the License, or
00012  * (at your option) any later version.
00013  *
00014  * mpeg2dec is distributed in the hope that it will be useful,
00015  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00016  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
00017  * GNU General Public License for more details.
00018  *
00019  * You should have received a copy of the GNU General Public License
00020  * along with this program; if not, write to the Free Software
00021  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
00022  */
00023 
00024 #include "config.h"
00025 
00026 #ifdef ARCH_X86
00027 
00028 #include <inttypes.h>
00029 
00030 #include "mpeg2.h"
00031 #include "mpeg2_internal.h"
00032 #include "attributes.h"
00033 #include "mmx.h"
00034 
00035 #define ROW_SHIFT 11
00036 #define COL_SHIFT 6
00037 
00038 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
00039 #define rounder(bias) {round (bias), round (bias)}
00040 
00041 
00042 #if 0
00043 /* C row IDCT - its just here to document the MMXEXT and MMX versions */
00044 static inline void idct_row (int16_t * row, int offset,
00045                              int16_t * table, int32_t * rounder)
00046 {
00047     int C1, C2, C3, C4, C5, C6, C7;
00048     int a0, a1, a2, a3, b0, b1, b2, b3;
00049 
00050     row += offset;
00051 
00052     C1 = table[1];
00053     C2 = table[2];
00054     C3 = table[3];
00055     C4 = table[4];
00056     C5 = table[5];
00057     C6 = table[6];
00058     C7 = table[7];
00059 
00060     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
00061     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
00062     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
00063     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
00064 
00065     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
00066     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
00067     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
00068     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
00069 
00070     row[0] = (a0 + b0) >> ROW_SHIFT;
00071     row[1] = (a1 + b1) >> ROW_SHIFT;
00072     row[2] = (a2 + b2) >> ROW_SHIFT;
00073     row[3] = (a3 + b3) >> ROW_SHIFT;
00074     row[4] = (a3 - b3) >> ROW_SHIFT;
00075     row[5] = (a2 - b2) >> ROW_SHIFT;
00076     row[6] = (a1 - b1) >> ROW_SHIFT;
00077     row[7] = (a0 - b0) >> ROW_SHIFT;
00078 }
00079 #endif
00080 
00081 
00082 /* MMXEXT row IDCT */
00083 
00084 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
00085                                                    c4,  c6,  c4,  c6,   \
00086                                                    c1,  c3, -c1, -c5,   \
00087                                                    c5,  c7,  c3, -c7,   \
00088                                                    c4, -c6,  c4, -c6,   \
00089                                                   -c4,  c2,  c4, -c2,   \
00090                                                    c5, -c1,  c3, -c1,   \
00091                                                    c7,  c3,  c7, -c5 }
00092 
00093 static inline void mmxext_row_head (int16_t * const row, const int offset,
00094                                     const int16_t * const table)
00095 {
00096     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00097 
00098     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00099     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00100 
00101     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
00102     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00103 
00104     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
00105     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00106 
00107     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
00108 }
00109 
00110 static inline void mmxext_row (const int16_t * const table,
00111                                const int32_t * const rounder)
00112 {
00113     movq_m2r (*(table+8), mm1);         /* mm1 = -C5 -C1 C3 C1 */
00114     pmaddwd_r2r (mm2, mm4);             /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
00115 
00116     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
00117     pshufw_r2r (mm6, mm6, 0x4e);        /* mm6 = x3 x1 x7 x5 */
00118 
00119     movq_m2r (*(table+12), mm7);        /* mm7 = -C7 C3 C7 C5 */
00120     pmaddwd_r2r (mm5, mm1);             /* mm1 = -C1*x5-C5*x7 C1*x1+C3*x3 */
00121 
00122     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
00123     pmaddwd_r2r (mm6, mm7);             /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
00124 
00125     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x0-C2*x2 -C4*x4+C2*x6 */
00126     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
00127 
00128     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
00129     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
00130 
00131     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
00132     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
00133 
00134     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
00135     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
00136 
00137     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
00138     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
00139 
00140     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
00141     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
00142 
00143     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
00144     movq_r2r (mm0, mm4);                /* mm4 = a3 a2 + rounder */
00145 
00146     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
00147     psubd_r2r (mm5, mm4);               /* mm4 = a3-b3 a2-b2 + rounder */
00148 }
00149 
00150 static inline void mmxext_row_tail (int16_t * const row, const int store)
00151 {
00152     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00153 
00154     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
00155 
00156     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00157 
00158     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
00159 
00160     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00161     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
00162 
00163     /* slot */
00164 
00165     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
00166 }
00167 
00168 static inline void mmxext_row_mid (int16_t * const row, const int store,
00169                                    const int offset,
00170                                    const int16_t * const table)
00171 {
00172     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00173     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00174 
00175     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00176     psrad_i2r (ROW_SHIFT, mm4);         /* mm4 = y4 y5 */
00177 
00178     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00179     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00180 
00181     packssdw_r2r (mm3, mm4);            /* mm4 = y6 y7 y4 y5 */
00182     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00183 
00184     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00185     pshufw_r2r (mm4, mm4, 0xb1);        /* mm4 = y7 y6 y5 y4 */
00186 
00187     movq_m2r (*table, mm3);             /* mm3 = -C2 -C4 C2 C4 */
00188     movq_r2m (mm4, *(row+store+4));     /* save y7 y6 y5 y4 */
00189 
00190     pmaddwd_r2r (mm0, mm3);             /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00191 
00192     movq_m2r (*(table+4), mm4);         /* mm4 = C6 C4 C6 C4 */
00193     pshufw_r2r (mm2, mm2, 0x4e);        /* mm2 = x2 x0 x6 x4 */
00194 }
00195 
00196 
00197 /* MMX row IDCT */
00198 
00199 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00200                                            c4,  c6, -c4, -c2,   \
00201                                            c1,  c3,  c3, -c7,   \
00202                                            c5,  c7, -c1, -c5,   \
00203                                            c4, -c6,  c4, -c2,   \
00204                                           -c4,  c2,  c4, -c6,   \
00205                                            c5, -c1,  c7, -c5,   \
00206                                            c7,  c3,  c3, -c1 }
00207 
00208 static inline void mmx_row_head (int16_t * const row, const int offset,
00209                                  const int16_t * const table)
00210 {
00211     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00212 
00213     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00214     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00215 
00216     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
00217     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00218 
00219     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
00220 
00221     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
00222     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00223 
00224     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
00225     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
00226 }
00227 
00228 static inline void mmx_row (const int16_t * const table,
00229                             const int32_t * const rounder)
00230 {
00231     pmaddwd_r2r (mm2, mm4);             /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
00232     punpckldq_r2r (mm5, mm5);           /* mm5 = x3 x1 x3 x1 */
00233 
00234     pmaddwd_m2r (*(table+16), mm0);     /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
00235     punpckhdq_r2r (mm6, mm6);           /* mm6 = x7 x5 x7 x5 */
00236 
00237     movq_m2r (*(table+12), mm7);        /* mm7 = -C5 -C1 C7 C5 */
00238     pmaddwd_r2r (mm5, mm1);             /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
00239 
00240     paddd_m2r (*rounder, mm3);          /* mm3 += rounder */
00241     pmaddwd_r2r (mm6, mm7);             /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
00242 
00243     pmaddwd_m2r (*(table+20), mm2);     /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
00244     paddd_r2r (mm4, mm3);               /* mm3 = a1 a0 + rounder */
00245 
00246     pmaddwd_m2r (*(table+24), mm5);     /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
00247     movq_r2r (mm3, mm4);                /* mm4 = a1 a0 + rounder */
00248 
00249     pmaddwd_m2r (*(table+28), mm6);     /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
00250     paddd_r2r (mm7, mm1);               /* mm1 = b1 b0 */
00251 
00252     paddd_m2r (*rounder, mm0);          /* mm0 += rounder */
00253     psubd_r2r (mm1, mm3);               /* mm3 = a1-b1 a0-b0 + rounder */
00254 
00255     psrad_i2r (ROW_SHIFT, mm3);         /* mm3 = y6 y7 */
00256     paddd_r2r (mm4, mm1);               /* mm1 = a1+b1 a0+b0 + rounder */
00257 
00258     paddd_r2r (mm2, mm0);               /* mm0 = a3 a2 + rounder */
00259     psrad_i2r (ROW_SHIFT, mm1);         /* mm1 = y1 y0 */
00260 
00261     paddd_r2r (mm6, mm5);               /* mm5 = b3 b2 */
00262     movq_r2r (mm0, mm7);                /* mm7 = a3 a2 + rounder */
00263 
00264     paddd_r2r (mm5, mm0);               /* mm0 = a3+b3 a2+b2 + rounder */
00265     psubd_r2r (mm5, mm7);               /* mm7 = a3-b3 a2-b2 + rounder */
00266 }
00267 
00268 static inline void mmx_row_tail (int16_t * const row, const int store)
00269 {
00270     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00271 
00272     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
00273 
00274     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00275 
00276     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
00277 
00278     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00279     movq_r2r (mm7, mm4);                /* mm4 = y6 y7 y4 y5 */
00280 
00281     pslld_i2r (16, mm7);                /* mm7 = y7 0 y5 0 */
00282 
00283     psrld_i2r (16, mm4);                /* mm4 = 0 y6 0 y4 */
00284 
00285     por_r2r (mm4, mm7);                 /* mm7 = y7 y6 y5 y4 */
00286 
00287     /* slot */
00288 
00289     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
00290 }
00291 
00292 static inline void mmx_row_mid (int16_t * const row, const int store,
00293                                 const int offset, const int16_t * const table)
00294 {
00295     movq_m2r (*(row+offset), mm2);      /* mm2 = x6 x4 x2 x0 */
00296     psrad_i2r (ROW_SHIFT, mm0);         /* mm0 = y3 y2 */
00297 
00298     movq_m2r (*(row+offset+4), mm5);    /* mm5 = x7 x5 x3 x1 */
00299     psrad_i2r (ROW_SHIFT, mm7);         /* mm7 = y4 y5 */
00300 
00301     packssdw_r2r (mm0, mm1);            /* mm1 = y3 y2 y1 y0 */
00302     movq_r2r (mm5, mm6);                /* mm6 = x7 x5 x3 x1 */
00303 
00304     packssdw_r2r (mm3, mm7);            /* mm7 = y6 y7 y4 y5 */
00305     movq_r2r (mm2, mm0);                /* mm0 = x6 x4 x2 x0 */
00306 
00307     movq_r2m (mm1, *(row+store));       /* save y3 y2 y1 y0 */
00308     movq_r2r (mm7, mm1);                /* mm1 = y6 y7 y4 y5 */
00309 
00310     punpckldq_r2r (mm0, mm0);           /* mm0 = x2 x0 x2 x0 */
00311     psrld_i2r (16, mm7);                /* mm7 = 0 y6 0 y4 */
00312 
00313     movq_m2r (*table, mm3);             /* mm3 = C6 C4 C2 C4 */
00314     pslld_i2r (16, mm1);                /* mm1 = y7 0 y5 0 */
00315 
00316     movq_m2r (*(table+4), mm4);         /* mm4 = -C2 -C4 C6 C4 */
00317     por_r2r (mm1, mm7);                 /* mm7 = y7 y6 y5 y4 */
00318 
00319     movq_m2r (*(table+8), mm1);         /* mm1 = -C7 C3 C3 C1 */
00320     punpckhdq_r2r (mm2, mm2);           /* mm2 = x6 x4 x6 x4 */
00321 
00322     movq_r2m (mm7, *(row+store+4));     /* save y7 y6 y5 y4 */
00323     pmaddwd_r2r (mm0, mm3);             /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00324 }
00325 
00326 
00327 #if 0
00328 /* C column IDCT - its just here to document the MMXEXT and MMX versions */
00329 static inline void idct_col (int16_t * col, int offset)
00330 {
00331 /* multiplication - as implemented on mmx */
00332 #define F(c,x) (((c) * (x)) >> 16)
00333 
00334 /* saturation - it helps us handle torture test cases */
00335 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00336 
00337     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00338     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00339     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00340     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00341 
00342     col += offset;
00343 
00344     x0 = col[0*8];
00345     x1 = col[1*8];
00346     x2 = col[2*8];
00347     x3 = col[3*8];
00348     x4 = col[4*8];
00349     x5 = col[5*8];
00350     x6 = col[6*8];
00351     x7 = col[7*8];
00352 
00353     u04 = S (x0 + x4);
00354     v04 = S (x0 - x4);
00355     u26 = S (F (T2, x6) + x2);
00356     v26 = S (F (T2, x2) - x6);
00357 
00358     a0 = S (u04 + u26);
00359     a1 = S (v04 + v26);
00360     a2 = S (v04 - v26);
00361     a3 = S (u04 - u26);
00362 
00363     u17 = S (F (T1, x7) + x1);
00364     v17 = S (F (T1, x1) - x7);
00365     u35 = S (F (T3, x5) + x3);
00366     v35 = S (F (T3, x3) - x5);
00367 
00368     b0 = S (u17 + u35);
00369     b3 = S (v17 - v35);
00370     u12 = S (u17 - u35);
00371     v12 = S (v17 + v35);
00372     u12 = S (2 * F (C4, u12));
00373     v12 = S (2 * F (C4, v12));
00374     b1 = S (u12 + v12);
00375     b2 = S (u12 - v12);
00376 
00377     y0 = S (a0 + b0) >> COL_SHIFT;
00378     y1 = S (a1 + b1) >> COL_SHIFT;
00379     y2 = S (a2 + b2) >> COL_SHIFT;
00380     y3 = S (a3 + b3) >> COL_SHIFT;
00381 
00382     y4 = S (a3 - b3) >> COL_SHIFT;
00383     y5 = S (a2 - b2) >> COL_SHIFT;
00384     y6 = S (a1 - b1) >> COL_SHIFT;
00385     y7 = S (a0 - b0) >> COL_SHIFT;
00386 
00387     col[0*8] = y0;
00388     col[1*8] = y1;
00389     col[2*8] = y2;
00390     col[3*8] = y3;
00391     col[4*8] = y4;
00392     col[5*8] = y5;
00393     col[6*8] = y6;
00394     col[7*8] = y7;
00395 }
00396 #endif
00397 
00398 
00399 /* MMX column IDCT */
00400 static inline void idct_col (int16_t * const col, const int offset)
00401 {
00402 #define T1 13036
00403 #define T2 27146
00404 #define T3 43790
00405 #define C4 23170
00406 
00407     static const short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
00408     static const short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
00409     static const short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
00410     static const short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
00411 
00412     /* column code adapted from peter gubanov */
00413     /* http://www.elecard.com/peter/idct.shtml */
00414 
00415     movq_m2r (*_T1, mm0);               /* mm0 = T1 */
00416 
00417     movq_m2r (*(col+offset+1*8), mm1);  /* mm1 = x1 */
00418     movq_r2r (mm0, mm2);                /* mm2 = T1 */
00419 
00420     movq_m2r (*(col+offset+7*8), mm4);  /* mm4 = x7 */
00421     pmulhw_r2r (mm1, mm0);              /* mm0 = T1*x1 */
00422 
00423     movq_m2r (*_T3, mm5);               /* mm5 = T3 */
00424     pmulhw_r2r (mm4, mm2);              /* mm2 = T1*x7 */
00425 
00426     movq_m2r (*(col+offset+5*8), mm6);  /* mm6 = x5 */
00427     movq_r2r (mm5, mm7);                /* mm7 = T3-1 */
00428 
00429     movq_m2r (*(col+offset+3*8), mm3);  /* mm3 = x3 */
00430     psubsw_r2r (mm4, mm0);              /* mm0 = v17 */
00431 
00432     movq_m2r (*_T2, mm4);               /* mm4 = T2 */
00433     pmulhw_r2r (mm3, mm5);              /* mm5 = (T3-1)*x3 */
00434 
00435     paddsw_r2r (mm2, mm1);              /* mm1 = u17 */
00436     pmulhw_r2r (mm6, mm7);              /* mm7 = (T3-1)*x5 */
00437 
00438     /* slot */
00439 
00440     movq_r2r (mm4, mm2);                /* mm2 = T2 */
00441     paddsw_r2r (mm3, mm5);              /* mm5 = T3*x3 */
00442 
00443     pmulhw_m2r (*(col+offset+2*8), mm4);/* mm4 = T2*x2 */
00444     paddsw_r2r (mm6, mm7);              /* mm7 = T3*x5 */
00445 
00446     psubsw_r2r (mm6, mm5);              /* mm5 = v35 */
00447     paddsw_r2r (mm3, mm7);              /* mm7 = u35 */
00448 
00449     movq_m2r (*(col+offset+6*8), mm3);  /* mm3 = x6 */
00450     movq_r2r (mm0, mm6);                /* mm6 = v17 */
00451 
00452     pmulhw_r2r (mm3, mm2);              /* mm2 = T2*x6 */
00453     psubsw_r2r (mm5, mm0);              /* mm0 = b3 */
00454 
00455     psubsw_r2r (mm3, mm4);              /* mm4 = v26 */
00456     paddsw_r2r (mm6, mm5);              /* mm5 = v12 */
00457 
00458     movq_r2m (mm0, *(col+offset+3*8));  /* save b3 in scratch0 */
00459     movq_r2r (mm1, mm6);                /* mm6 = u17 */
00460 
00461     paddsw_m2r (*(col+offset+2*8), mm2);/* mm2 = u26 */
00462     paddsw_r2r (mm7, mm6);              /* mm6 = b0 */
00463 
00464     psubsw_r2r (mm7, mm1);              /* mm1 = u12 */
00465     movq_r2r (mm1, mm7);                /* mm7 = u12 */
00466 
00467     movq_m2r (*(col+offset+0*8), mm3);  /* mm3 = x0 */
00468     paddsw_r2r (mm5, mm1);              /* mm1 = u12+v12 */
00469 
00470     movq_m2r (*_C4, mm0);               /* mm0 = C4/2 */
00471     psubsw_r2r (mm5, mm7);              /* mm7 = u12-v12 */
00472 
00473     movq_r2m (mm6, *(col+offset+5*8));  /* save b0 in scratch1 */
00474     pmulhw_r2r (mm0, mm1);              /* mm1 = b1/2 */
00475 
00476     movq_r2r (mm4, mm6);                /* mm6 = v26 */
00477     pmulhw_r2r (mm0, mm7);              /* mm7 = b2/2 */
00478 
00479     movq_m2r (*(col+offset+4*8), mm5);  /* mm5 = x4 */
00480     movq_r2r (mm3, mm0);                /* mm0 = x0 */
00481 
00482     psubsw_r2r (mm5, mm3);              /* mm3 = v04 */
00483     paddsw_r2r (mm5, mm0);              /* mm0 = u04 */
00484 
00485     paddsw_r2r (mm3, mm4);              /* mm4 = a1 */
00486     movq_r2r (mm0, mm5);                /* mm5 = u04 */
00487 
00488     psubsw_r2r (mm6, mm3);              /* mm3 = a2 */
00489     paddsw_r2r (mm2, mm5);              /* mm5 = a0 */
00490 
00491     paddsw_r2r (mm1, mm1);              /* mm1 = b1 */
00492     psubsw_r2r (mm2, mm0);              /* mm0 = a3 */
00493 
00494     paddsw_r2r (mm7, mm7);              /* mm7 = b2 */
00495     movq_r2r (mm3, mm2);                /* mm2 = a2 */
00496 
00497     movq_r2r (mm4, mm6);                /* mm6 = a1 */
00498     paddsw_r2r (mm7, mm3);              /* mm3 = a2+b2 */
00499 
00500     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y2 */
00501     paddsw_r2r (mm1, mm4);              /* mm4 = a1+b1 */
00502 
00503     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y1 */
00504     psubsw_r2r (mm1, mm6);              /* mm6 = a1-b1 */
00505 
00506     movq_m2r (*(col+offset+5*8), mm1);  /* mm1 = b0 */
00507     psubsw_r2r (mm7, mm2);              /* mm2 = a2-b2 */
00508 
00509     psraw_i2r (COL_SHIFT, mm6);         /* mm6 = y6 */
00510     movq_r2r (mm5, mm7);                /* mm7 = a0 */
00511 
00512     movq_r2m (mm4, *(col+offset+1*8));  /* save y1 */
00513     psraw_i2r (COL_SHIFT, mm2);         /* mm2 = y5 */
00514 
00515     movq_r2m (mm3, *(col+offset+2*8));  /* save y2 */
00516     paddsw_r2r (mm1, mm5);              /* mm5 = a0+b0 */
00517 
00518     movq_m2r (*(col+offset+3*8), mm4);  /* mm4 = b3 */
00519     psubsw_r2r (mm1, mm7);              /* mm7 = a0-b0 */
00520 
00521     psraw_i2r (COL_SHIFT, mm5);         /* mm5 = y0 */
00522     movq_r2r (mm0, mm3);                /* mm3 = a3 */
00523 
00524     movq_r2m (mm2, *(col+offset+5*8));  /* save y5 */
00525     psubsw_r2r (mm4, mm3);              /* mm3 = a3-b3 */
00526 
00527     psraw_i2r (COL_SHIFT, mm7);         /* mm7 = y7 */
00528     paddsw_r2r (mm0, mm4);              /* mm4 = a3+b3 */
00529 
00530     movq_r2m (mm5, *(col+offset+0*8));  /* save y0 */
00531     psraw_i2r (COL_SHIFT, mm3);         /* mm3 = y4 */
00532 
00533     movq_r2m (mm6, *(col+offset+6*8));  /* save y6 */
00534     psraw_i2r (COL_SHIFT, mm4);         /* mm4 = y3 */
00535 
00536     movq_r2m (mm7, *(col+offset+7*8));  /* save y7 */
00537 
00538     movq_r2m (mm3, *(col+offset+4*8));  /* save y4 */
00539 
00540     movq_r2m (mm4, *(col+offset+3*8));  /* save y3 */
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 
00561 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid)      \
00562 static inline void idct (int16_t * const block)                         \
00563 {                                                                       \
00564     static const int16_t table04[] ATTR_ALIGN(16) =                     \
00565         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00566     static const int16_t table17[] ATTR_ALIGN(16) =                     \
00567         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00568     static const int16_t table26[] ATTR_ALIGN(16) =                     \
00569         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00570     static const int16_t table35[] ATTR_ALIGN(16) =                     \
00571         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00572                                                                         \
00573     idct_row_head (block, 0*8, table04);                                \
00574     idct_row (table04, rounder0);                                       \
00575     idct_row_mid (block, 0*8, 4*8, table04);                            \
00576     idct_row (table04, rounder4);                                       \
00577     idct_row_mid (block, 4*8, 1*8, table17);                            \
00578     idct_row (table17, rounder1);                                       \
00579     idct_row_mid (block, 1*8, 7*8, table17);                            \
00580     idct_row (table17, rounder7);                                       \
00581     idct_row_mid (block, 7*8, 2*8, table26);                            \
00582     idct_row (table26, rounder2);                                       \
00583     idct_row_mid (block, 2*8, 6*8, table26);                            \
00584     idct_row (table26, rounder6);                                       \
00585     idct_row_mid (block, 6*8, 3*8, table35);                            \
00586     idct_row (table35, rounder3);                                       \
00587     idct_row_mid (block, 3*8, 5*8, table35);                            \
00588     idct_row (table35, rounder5);                                       \
00589     idct_row_tail (block, 5*8);                                         \
00590                                                                         \
00591     idct_col (block, 0);                                                \
00592     idct_col (block, 4);                                                \
00593 }
00594 
00595 
00596 #define COPY_MMX(offset,r0,r1,r2)       \
00597 do {                                    \
00598     movq_m2r (*(block+offset), r0);     \
00599     dest += stride;                     \
00600     movq_m2r (*(block+offset+4), r1);   \
00601     movq_r2m (r2, *dest);               \
00602     packuswb_r2r (r1, r0);              \
00603 } while (0)
00604 
00605 static inline void block_copy (int16_t * const block, uint8_t * dest,
00606                                const int stride)
00607 {
00608     movq_m2r (*(block+0*8), mm0);
00609     movq_m2r (*(block+0*8+4), mm1);
00610     movq_m2r (*(block+1*8), mm2);
00611     packuswb_r2r (mm1, mm0);
00612     movq_m2r (*(block+1*8+4), mm3);
00613     movq_r2m (mm0, *dest);
00614     packuswb_r2r (mm3, mm2);
00615     COPY_MMX (2*8, mm0, mm1, mm2);
00616     COPY_MMX (3*8, mm2, mm3, mm0);
00617     COPY_MMX (4*8, mm0, mm1, mm2);
00618     COPY_MMX (5*8, mm2, mm3, mm0);
00619     COPY_MMX (6*8, mm0, mm1, mm2);
00620     COPY_MMX (7*8, mm2, mm3, mm0);
00621     movq_r2m (mm2, *(dest+stride));
00622 }
00623 
00624 
00625 #define ADD_MMX(offset,r1,r2,r3,r4)     \
00626 do {                                    \
00627     movq_m2r (*(dest+2*stride), r1);    \
00628     packuswb_r2r (r4, r3);              \
00629     movq_r2r (r1, r2);                  \
00630     dest += stride;                     \
00631     movq_r2m (r3, *dest);               \
00632     punpcklbw_r2r (mm0, r1);            \
00633     paddsw_m2r (*(block+offset), r1);   \
00634     punpckhbw_r2r (mm0, r2);            \
00635     paddsw_m2r (*(block+offset+4), r2); \
00636 } while (0)
00637 
00638 static inline void block_add (int16_t * const block, uint8_t * dest,
00639                               const int stride)
00640 {
00641     movq_m2r (*dest, mm1);
00642     pxor_r2r (mm0, mm0);
00643     movq_m2r (*(dest+stride), mm3);
00644     movq_r2r (mm1, mm2);
00645     punpcklbw_r2r (mm0, mm1);
00646     movq_r2r (mm3, mm4);
00647     paddsw_m2r (*(block+0*8), mm1);
00648     punpckhbw_r2r (mm0, mm2);
00649     paddsw_m2r (*(block+0*8+4), mm2);
00650     punpcklbw_r2r (mm0, mm3);
00651     paddsw_m2r (*(block+1*8), mm3);
00652     packuswb_r2r (mm2, mm1);
00653     punpckhbw_r2r (mm0, mm4);
00654     movq_r2m (mm1, *dest);
00655     paddsw_m2r (*(block+1*8+4), mm4);
00656     ADD_MMX (2*8, mm1, mm2, mm3, mm4);
00657     ADD_MMX (3*8, mm3, mm4, mm1, mm2);
00658     ADD_MMX (4*8, mm1, mm2, mm3, mm4);
00659     ADD_MMX (5*8, mm3, mm4, mm1, mm2);
00660     ADD_MMX (6*8, mm1, mm2, mm3, mm4);
00661     ADD_MMX (7*8, mm3, mm4, mm1, mm2);
00662     packuswb_r2r (mm4, mm3);
00663     movq_r2m (mm3, *(dest+stride));
00664 }
00665 
00666 
00667 static inline void block_zero (int16_t * const block)
00668 {
00669     pxor_r2r (mm0, mm0);
00670     movq_r2m (mm0, *(block+0*4));
00671     movq_r2m (mm0, *(block+1*4));
00672     movq_r2m (mm0, *(block+2*4));
00673     movq_r2m (mm0, *(block+3*4));
00674     movq_r2m (mm0, *(block+4*4));
00675     movq_r2m (mm0, *(block+5*4));
00676     movq_r2m (mm0, *(block+6*4));
00677     movq_r2m (mm0, *(block+7*4));
00678     movq_r2m (mm0, *(block+8*4));
00679     movq_r2m (mm0, *(block+9*4));
00680     movq_r2m (mm0, *(block+10*4));
00681     movq_r2m (mm0, *(block+11*4));
00682     movq_r2m (mm0, *(block+12*4));
00683     movq_r2m (mm0, *(block+13*4));
00684     movq_r2m (mm0, *(block+14*4));
00685     movq_r2m (mm0, *(block+15*4));
00686 }
00687 
00688 
00689 #define CPU_MMXEXT 0
00690 #define CPU_MMX 1
00691 
00692 #define dup4(reg)                       \
00693 do {                                    \
00694     if (cpu != CPU_MMXEXT) {            \
00695         punpcklwd_r2r (reg, reg);       \
00696         punpckldq_r2r (reg, reg);       \
00697     } else                              \
00698         pshufw_r2r (reg, reg, 0x00);    \
00699 } while (0)
00700 
00701 static inline void block_add_DC (int16_t * const block, uint8_t * dest,
00702                                  const int stride, const int cpu)
00703 {
00704     movd_v2r ((block[0] + 4) >> 3, mm0);
00705     pxor_r2r (mm1, mm1);
00706     movq_m2r (*dest, mm2);
00707     dup4 (mm0);
00708     psubsw_r2r (mm0, mm1);
00709     packuswb_r2r (mm0, mm0);
00710     paddusb_r2r (mm0, mm2);
00711     packuswb_r2r (mm1, mm1);
00712     movq_m2r (*(dest + stride), mm3);
00713     psubusb_r2r (mm1, mm2);
00714     block[0] = 0;
00715     paddusb_r2r (mm0, mm3);
00716     movq_r2m (mm2, *dest);
00717     psubusb_r2r (mm1, mm3);
00718     movq_m2r (*(dest + 2*stride), mm2);
00719     dest += stride;
00720     movq_r2m (mm3, *dest);
00721     paddusb_r2r (mm0, mm2);
00722     movq_m2r (*(dest + 2*stride), mm3);
00723     psubusb_r2r (mm1, mm2);
00724     dest += stride;
00725     paddusb_r2r (mm0, mm3);
00726     movq_r2m (mm2, *dest);
00727     psubusb_r2r (mm1, mm3);
00728     movq_m2r (*(dest + 2*stride), mm2);
00729     dest += stride;
00730     movq_r2m (mm3, *dest);
00731     paddusb_r2r (mm0, mm2);
00732     movq_m2r (*(dest + 2*stride), mm3);
00733     psubusb_r2r (mm1, mm2);
00734     dest += stride;
00735     paddusb_r2r (mm0, mm3);
00736     movq_r2m (mm2, *dest);
00737     psubusb_r2r (mm1, mm3);
00738     movq_m2r (*(dest + 2*stride), mm2);
00739     dest += stride;
00740     movq_r2m (mm3, *dest);
00741     paddusb_r2r (mm0, mm2);
00742     movq_m2r (*(dest + 2*stride), mm3);
00743     psubusb_r2r (mm1, mm2);
00744     block[63] = 0;
00745     paddusb_r2r (mm0, mm3);
00746     movq_r2m (mm2, *(dest + stride));
00747     psubusb_r2r (mm1, mm3);
00748     movq_r2m (mm3, *(dest + 2*stride));
00749 }
00750 
00751 
00752 declare_idct (mmxext_idct, mmxext_table,
00753               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00754 
00755 void mpeg2_idct_copy_mmxext (int16_t * const block, uint8_t * const dest,
00756                              const int stride)
00757 {
00758     mmxext_idct (block);
00759     block_copy (block, dest, stride);
00760     block_zero (block);
00761 }
00762 
00763 void mpeg2_idct_add_mmxext (const int last, int16_t * const block,
00764                             uint8_t * const dest, const int stride)
00765 {
00766     if (last != 129 || (block[0] & 7) == 4) {
00767         mmxext_idct (block);
00768         block_add (block, dest, stride);
00769         block_zero (block);
00770     } else
00771         block_add_DC (block, dest, stride, CPU_MMXEXT);
00772 }
00773 
00774 
00775 declare_idct (mmx_idct, mmx_table,
00776               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00777 
00778 void mpeg2_idct_copy_mmx (int16_t * const block, uint8_t * const dest,
00779                           const int stride)
00780 {
00781     mmx_idct (block);
00782     block_copy (block, dest, stride);
00783     block_zero (block);
00784 }
00785 
00786 void mpeg2_idct_add_mmx (const int last, int16_t * const block,
00787                          uint8_t * const dest, const int stride)
00788 {
00789     if (last != 129 || (block[0] & 7) == 4) {
00790         mmx_idct (block);
00791         block_add (block, dest, stride);
00792         block_zero (block);
00793     } else
00794         block_add_DC (block, dest, stride, CPU_MMX);
00795 }
00796 
00797 
00798 void mpeg2_idct_mmx_init (void)
00799 {
00800     extern uint8_t mpeg2_scan_norm[64];
00801     extern uint8_t mpeg2_scan_alt[64];
00802     int i, j;
00803 
00804     /* the mmx/mmxext idct uses a reordered input, so we patch scan tables */
00805 
00806     for (i = 0; i < 64; i++) {
00807         j = mpeg2_scan_norm[i];
00808         mpeg2_scan_norm[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
00809         j = mpeg2_scan_alt[i];
00810         mpeg2_scan_alt[i] = (j & 0x38) | ((j & 6) >> 1) | ((j & 1) << 2);
00811     }
00812 }
00813 
00814 #endif