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