Cinelerra: hvirtual/quicktime/ffmpeg/libavcodec/i386/idct

00001 /*
00002  * Note: For libavcodec, this code can also be used under the LGPL license
00003  */
00004 /*
00005  * idct_mmx.c
00006  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
00007  *
00008  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
00009  *
00010  * mpeg2dec is free software; you can redistribute it and/or modify
00011  * it under the terms of the GNU General Public License as published by
00012  * the Free Software Foundation; either version 2 of the License, or
00013  * (at your option) any later version.
00014  *
00015  * mpeg2dec is distributed in the hope that it will be useful,
00016  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00017  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
00018  * GNU General Public License for more details.
00019  *
00020  * You should have received a copy of the GNU General Public License
00021  * along with this program; if not, write to the Free Software
00022  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
00023  */
00024 
00025 #include "../common.h"
00026 #include "../dsputil.h"
00027 
00028 #include "mmx.h"
00029 
00030 #define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
00031 
00032 #define ROW_SHIFT 11
00033 #define COL_SHIFT 6
00034 
00035 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
00036 #define rounder(bias) {round (bias), round (bias)}
00037 
00038 #if 0
00039 /* C row IDCT - its just here to document the MMXEXT and MMX versions */
00040 static inline void idct_row (int16_t * row, int offset,
00041                              int16_t * table, int32_t * rounder)
00042 {
00043     int C1, C2, C3, C4, C5, C6, C7;
00044     int a0, a1, a2, a3, b0, b1, b2, b3;
00045 
00046     row += offset;
00047 
00048     C1 = table[1];
00049     C2 = table[2];
00050     C3 = table[3];
00051     C4 = table[4];
00052     C5 = table[5];
00053     C6 = table[6];
00054     C7 = table[7];
00055 
00056     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
00057     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
00058     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
00059     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
00060 
00061     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
00062     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
00063     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
00064     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
00065 
00066     row[0] = (a0 + b0) >> ROW_SHIFT;
00067     row[1] = (a1 + b1) >> ROW_SHIFT;
00068     row[2] = (a2 + b2) >> ROW_SHIFT;
00069     row[3] = (a3 + b3) >> ROW_SHIFT;
00070     row[4] = (a3 - b3) >> ROW_SHIFT;
00071     row[5] = (a2 - b2) >> ROW_SHIFT;
00072     row[6] = (a1 - b1) >> ROW_SHIFT;
00073     row[7] = (a0 - b0) >> ROW_SHIFT;
00074 }
00075 #endif
00076 
00077 
00078 /* MMXEXT row IDCT */
00079 
00080 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
00081                                                    c4,  c6,  c4,  c6,   \
00082                                                    c1,  c3, -c1, -c5,   \
00083                                                    c5,  c7,  c3, -c7,   \
00084                                                    c4, -c6,  c4, -c6,   \
00085                                                   -c4,  c2,  c4, -c2,   \
00086                                                    c5, -c1,  c3, -c1,   \
00087                                                    c7,  c3,  c7, -c5 }
00088 
00089 static inline void mmxext_row_head (int16_t * row, int offset, int16_t * 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 (int16_t * table, int32_t * 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 * row, 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 * row, int store,
00163                                    int offset, int16_t * table)
00164 {
00165     movq_m2r (*(row+offset), mm2);      // mm2 = x6 x4 x2 x0
00166     psrad_i2r (ROW_SHIFT, mm0);         // mm0 = y3 y2
00167 
00168     movq_m2r (*(row+offset+4), mm5);    // mm5 = x7 x5 x3 x1
00169     psrad_i2r (ROW_SHIFT, mm4);         // mm4 = y4 y5
00170 
00171     packssdw_r2r (mm0, mm1);            // mm1 = y3 y2 y1 y0
00172     movq_r2r (mm5, mm6);                // mm6 = x7 x5 x3 x1
00173 
00174     packssdw_r2r (mm3, mm4);            // mm4 = y6 y7 y4 y5
00175     movq_r2r (mm2, mm0);                // mm0 = x6 x4 x2 x0
00176 
00177     movq_r2m (mm1, *(row+store));       // save y3 y2 y1 y0
00178     pshufw_r2r (mm4, mm4, 0xb1);        // mm4 = y7 y6 y5 y4
00179 
00180     movq_m2r (*table, mm3);             // mm3 = -C2 -C4 C2 C4
00181     movq_r2m (mm4, *(row+store+4));     // save y7 y6 y5 y4
00182 
00183     pmaddwd_r2r (mm0, mm3);             // mm3 = -C4*x4-C2*x6 C4*x0+C2*x2
00184 
00185     movq_m2r (*(table+4), mm4);         // mm4 = C6 C4 C6 C4
00186     pshufw_r2r (mm2, mm2, 0x4e);        // mm2 = x2 x0 x6 x4
00187 }
00188 
00189 
00190 /* MMX row IDCT */
00191 
00192 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00193                                            c4,  c6, -c4, -c2,   \
00194                                            c1,  c3,  c3, -c7,   \
00195                                            c5,  c7, -c1, -c5,   \
00196                                            c4, -c6,  c4, -c2,   \
00197                                           -c4,  c2,  c4, -c6,   \
00198                                            c5, -c1,  c7, -c5,   \
00199                                            c7,  c3,  c3, -c1 }
00200 
00201 static inline void mmx_row_head (int16_t * row, int offset, int16_t * table)
00202 {
00203     movq_m2r (*(row+offset), mm2);      // mm2 = x6 x4 x2 x0
00204 
00205     movq_m2r (*(row+offset+4), mm5);    // mm5 = x7 x5 x3 x1
00206     movq_r2r (mm2, mm0);                // mm0 = x6 x4 x2 x0
00207 
00208     movq_m2r (*table, mm3);             // mm3 = C6 C4 C2 C4
00209     movq_r2r (mm5, mm6);                // mm6 = x7 x5 x3 x1
00210 
00211     punpckldq_r2r (mm0, mm0);           // mm0 = x2 x0 x2 x0
00212 
00213     movq_m2r (*(table+4), mm4);         // mm4 = -C2 -C4 C6 C4
00214     pmaddwd_r2r (mm0, mm3);             // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
00215 
00216     movq_m2r (*(table+8), mm1);         // mm1 = -C7 C3 C3 C1
00217     punpckhdq_r2r (mm2, mm2);           // mm2 = x6 x4 x6 x4
00218 }
00219 
00220 static inline void mmx_row (int16_t * table, int32_t * rounder)
00221 {
00222     pmaddwd_r2r (mm2, mm4);             // mm4 = -C4*x4-C2*x6 C4*x4+C6*x6
00223     punpckldq_r2r (mm5, mm5);           // mm5 = x3 x1 x3 x1
00224 
00225     pmaddwd_m2r (*(table+16), mm0);     // mm0 = C4*x0-C2*x2 C4*x0-C6*x2
00226     punpckhdq_r2r (mm6, mm6);           // mm6 = x7 x5 x7 x5
00227 
00228     movq_m2r (*(table+12), mm7);        // mm7 = -C5 -C1 C7 C5
00229     pmaddwd_r2r (mm5, mm1);             // mm1 = C3*x1-C7*x3 C1*x1+C3*x3
00230 
00231     paddd_m2r (*rounder, mm3);          // mm3 += rounder
00232     pmaddwd_r2r (mm6, mm7);             // mm7 = -C1*x5-C5*x7 C5*x5+C7*x7
00233 
00234     pmaddwd_m2r (*(table+20), mm2);     // mm2 = C4*x4-C6*x6 -C4*x4+C2*x6
00235     paddd_r2r (mm4, mm3);               // mm3 = a1 a0 + rounder
00236 
00237     pmaddwd_m2r (*(table+24), mm5);     // mm5 = C7*x1-C5*x3 C5*x1-C1*x3
00238     movq_r2r (mm3, mm4);                // mm4 = a1 a0 + rounder
00239 
00240     pmaddwd_m2r (*(table+28), mm6);     // mm6 = C3*x5-C1*x7 C7*x5+C3*x7
00241     paddd_r2r (mm7, mm1);               // mm1 = b1 b0
00242 
00243     paddd_m2r (*rounder, mm0);          // mm0 += rounder
00244     psubd_r2r (mm1, mm3);               // mm3 = a1-b1 a0-b0 + rounder
00245 
00246     psrad_i2r (ROW_SHIFT, mm3);         // mm3 = y6 y7
00247     paddd_r2r (mm4, mm1);               // mm1 = a1+b1 a0+b0 + rounder
00248 
00249     paddd_r2r (mm2, mm0);               // mm0 = a3 a2 + rounder
00250     psrad_i2r (ROW_SHIFT, mm1);         // mm1 = y1 y0
00251 
00252     paddd_r2r (mm6, mm5);               // mm5 = b3 b2
00253     movq_r2r (mm0, mm7);                // mm7 = a3 a2 + rounder
00254 
00255     paddd_r2r (mm5, mm0);               // mm0 = a3+b3 a2+b2 + rounder
00256     psubd_r2r (mm5, mm7);               // mm7 = a3-b3 a2-b2 + rounder
00257 }
00258 
00259 static inline void mmx_row_tail (int16_t * row, int store)
00260 {
00261     psrad_i2r (ROW_SHIFT, mm0);         // mm0 = y3 y2
00262 
00263     psrad_i2r (ROW_SHIFT, mm7);         // mm7 = y4 y5
00264 
00265     packssdw_r2r (mm0, mm1);            // mm1 = y3 y2 y1 y0
00266 
00267     packssdw_r2r (mm3, mm7);            // mm7 = y6 y7 y4 y5
00268 
00269     movq_r2m (mm1, *(row+store));       // save y3 y2 y1 y0
00270     movq_r2r (mm7, mm4);                // mm4 = y6 y7 y4 y5
00271 
00272     pslld_i2r (16, mm7);                // mm7 = y7 0 y5 0
00273 
00274     psrld_i2r (16, mm4);                // mm4 = 0 y6 0 y4
00275 
00276     por_r2r (mm4, mm7);                 // mm7 = y7 y6 y5 y4
00277 
00278     /* slot */
00279 
00280     movq_r2m (mm7, *(row+store+4));     // save y7 y6 y5 y4
00281 }
00282 
00283 static inline void mmx_row_mid (int16_t * row, int store,
00284                                 int offset, int16_t * table)
00285 {
00286     movq_m2r (*(row+offset), mm2);      // mm2 = x6 x4 x2 x0
00287     psrad_i2r (ROW_SHIFT, mm0);         // mm0 = y3 y2
00288 
00289     movq_m2r (*(row+offset+4), mm5);    // mm5 = x7 x5 x3 x1
00290     psrad_i2r (ROW_SHIFT, mm7);         // mm7 = y4 y5
00291 
00292     packssdw_r2r (mm0, mm1);            // mm1 = y3 y2 y1 y0
00293     movq_r2r (mm5, mm6);                // mm6 = x7 x5 x3 x1
00294 
00295     packssdw_r2r (mm3, mm7);            // mm7 = y6 y7 y4 y5
00296     movq_r2r (mm2, mm0);                // mm0 = x6 x4 x2 x0
00297 
00298     movq_r2m (mm1, *(row+store));       // save y3 y2 y1 y0
00299     movq_r2r (mm7, mm1);                // mm1 = y6 y7 y4 y5
00300 
00301     punpckldq_r2r (mm0, mm0);           // mm0 = x2 x0 x2 x0
00302     psrld_i2r (16, mm7);                // mm7 = 0 y6 0 y4
00303 
00304     movq_m2r (*table, mm3);             // mm3 = C6 C4 C2 C4
00305     pslld_i2r (16, mm1);                // mm1 = y7 0 y5 0
00306 
00307     movq_m2r (*(table+4), mm4);         // mm4 = -C2 -C4 C6 C4
00308     por_r2r (mm1, mm7);                 // mm7 = y7 y6 y5 y4
00309 
00310     movq_m2r (*(table+8), mm1);         // mm1 = -C7 C3 C3 C1
00311     punpckhdq_r2r (mm2, mm2);           // mm2 = x6 x4 x6 x4
00312 
00313     movq_r2m (mm7, *(row+store+4));     // save y7 y6 y5 y4
00314     pmaddwd_r2r (mm0, mm3);             // mm3 = C4*x0+C6*x2 C4*x0+C2*x2
00315 }
00316 
00317 
00318 #if 0
00319 // C column IDCT - its just here to document the MMXEXT and MMX versions
00320 static inline void idct_col (int16_t * col, int offset)
00321 {
00322 /* multiplication - as implemented on mmx */
00323 #define F(c,x) (((c) * (x)) >> 16)
00324 
00325 /* saturation - it helps us handle torture test cases */
00326 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00327 
00328     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00329     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00330     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00331     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00332 
00333     col += offset;
00334 
00335     x0 = col[0*8];
00336     x1 = col[1*8];
00337     x2 = col[2*8];
00338     x3 = col[3*8];
00339     x4 = col[4*8];
00340     x5 = col[5*8];
00341     x6 = col[6*8];
00342     x7 = col[7*8];
00343 
00344     u04 = S (x0 + x4);
00345     v04 = S (x0 - x4);
00346     u26 = S (F (T2, x6) + x2);
00347     v26 = S (F (T2, x2) - x6);
00348 
00349     a0 = S (u04 + u26);
00350     a1 = S (v04 + v26);
00351     a2 = S (v04 - v26);
00352     a3 = S (u04 - u26);
00353 
00354     u17 = S (F (T1, x7) + x1);
00355     v17 = S (F (T1, x1) - x7);
00356     u35 = S (F (T3, x5) + x3);
00357     v35 = S (F (T3, x3) - x5);
00358 
00359     b0 = S (u17 + u35);
00360     b3 = S (v17 - v35);
00361     u12 = S (u17 - u35);
00362     v12 = S (v17 + v35);
00363     u12 = S (2 * F (C4, u12));
00364     v12 = S (2 * F (C4, v12));
00365     b1 = S (u12 + v12);
00366     b2 = S (u12 - v12);
00367 
00368     y0 = S (a0 + b0) >> COL_SHIFT;
00369     y1 = S (a1 + b1) >> COL_SHIFT;
00370     y2 = S (a2 + b2) >> COL_SHIFT;
00371     y3 = S (a3 + b3) >> COL_SHIFT;
00372 
00373     y4 = S (a3 - b3) >> COL_SHIFT;
00374     y5 = S (a2 - b2) >> COL_SHIFT;
00375     y6 = S (a1 - b1) >> COL_SHIFT;
00376     y7 = S (a0 - b0) >> COL_SHIFT;
00377 
00378     col[0*8] = y0;
00379     col[1*8] = y1;
00380     col[2*8] = y2;
00381     col[3*8] = y3;
00382     col[4*8] = y4;
00383     col[5*8] = y5;
00384     col[6*8] = y6;
00385     col[7*8] = y7;
00386 }
00387 #endif
00388 
00389 
00390 // MMX column IDCT
00391 static inline void idct_col (int16_t * col, int offset)
00392 {
00393 #define T1 13036
00394 #define T2 27146
00395 #define T3 43790
00396 #define C4 23170
00397 
00398     static short _T1[] ATTR_ALIGN(8) = {T1,T1,T1,T1};
00399     static short _T2[] ATTR_ALIGN(8) = {T2,T2,T2,T2};
00400     static short _T3[] ATTR_ALIGN(8) = {T3,T3,T3,T3};
00401     static short _C4[] ATTR_ALIGN(8) = {C4,C4,C4,C4};
00402 
00403     /* column code adapted from peter gubanov */
00404     /* http://www.elecard.com/peter/idct.shtml */
00405 
00406     movq_m2r (*_T1, mm0);               // mm0 = T1
00407 
00408     movq_m2r (*(col+offset+1*8), mm1);  // mm1 = x1
00409     movq_r2r (mm0, mm2);                // mm2 = T1
00410 
00411     movq_m2r (*(col+offset+7*8), mm4);  // mm4 = x7
00412     pmulhw_r2r (mm1, mm0);              // mm0 = T1*x1
00413 
00414     movq_m2r (*_T3, mm5);               // mm5 = T3
00415     pmulhw_r2r (mm4, mm2);              // mm2 = T1*x7
00416 
00417     movq_m2r (*(col+offset+5*8), mm6);  // mm6 = x5
00418     movq_r2r (mm5, mm7);                // mm7 = T3-1
00419 
00420     movq_m2r (*(col+offset+3*8), mm3);  // mm3 = x3
00421     psubsw_r2r (mm4, mm0);              // mm0 = v17
00422 
00423     movq_m2r (*_T2, mm4);               // mm4 = T2
00424     pmulhw_r2r (mm3, mm5);              // mm5 = (T3-1)*x3
00425 
00426     paddsw_r2r (mm2, mm1);              // mm1 = u17
00427     pmulhw_r2r (mm6, mm7);              // mm7 = (T3-1)*x5
00428 
00429     /* slot */
00430 
00431     movq_r2r (mm4, mm2);                // mm2 = T2
00432     paddsw_r2r (mm3, mm5);              // mm5 = T3*x3
00433 
00434     pmulhw_m2r (*(col+offset+2*8), mm4);// mm4 = T2*x2
00435     paddsw_r2r (mm6, mm7);              // mm7 = T3*x5
00436 
00437     psubsw_r2r (mm6, mm5);              // mm5 = v35
00438     paddsw_r2r (mm3, mm7);              // mm7 = u35
00439 
00440     movq_m2r (*(col+offset+6*8), mm3);  // mm3 = x6
00441     movq_r2r (mm0, mm6);                // mm6 = v17
00442 
00443     pmulhw_r2r (mm3, mm2);              // mm2 = T2*x6
00444     psubsw_r2r (mm5, mm0);              // mm0 = b3
00445 
00446     psubsw_r2r (mm3, mm4);              // mm4 = v26
00447     paddsw_r2r (mm6, mm5);              // mm5 = v12
00448 
00449     movq_r2m (mm0, *(col+offset+3*8));  // save b3 in scratch0
00450     movq_r2r (mm1, mm6);                // mm6 = u17
00451 
00452     paddsw_m2r (*(col+offset+2*8), mm2);// mm2 = u26
00453     paddsw_r2r (mm7, mm6);              // mm6 = b0
00454 
00455     psubsw_r2r (mm7, mm1);              // mm1 = u12
00456     movq_r2r (mm1, mm7);                // mm7 = u12
00457 
00458     movq_m2r (*(col+offset+0*8), mm3);  // mm3 = x0
00459     paddsw_r2r (mm5, mm1);              // mm1 = u12+v12
00460 
00461     movq_m2r (*_C4, mm0);               // mm0 = C4/2
00462     psubsw_r2r (mm5, mm7);              // mm7 = u12-v12
00463 
00464     movq_r2m (mm6, *(col+offset+5*8));  // save b0 in scratch1
00465     pmulhw_r2r (mm0, mm1);              // mm1 = b1/2
00466 
00467     movq_r2r (mm4, mm6);                // mm6 = v26
00468     pmulhw_r2r (mm0, mm7);              // mm7 = b2/2
00469 
00470     movq_m2r (*(col+offset+4*8), mm5);  // mm5 = x4
00471     movq_r2r (mm3, mm0);                // mm0 = x0
00472 
00473     psubsw_r2r (mm5, mm3);              // mm3 = v04
00474     paddsw_r2r (mm5, mm0);              // mm0 = u04
00475 
00476     paddsw_r2r (mm3, mm4);              // mm4 = a1
00477     movq_r2r (mm0, mm5);                // mm5 = u04
00478 
00479     psubsw_r2r (mm6, mm3);              // mm3 = a2
00480     paddsw_r2r (mm2, mm5);              // mm5 = a0
00481 
00482     paddsw_r2r (mm1, mm1);              // mm1 = b1
00483     psubsw_r2r (mm2, mm0);              // mm0 = a3
00484 
00485     paddsw_r2r (mm7, mm7);              // mm7 = b2
00486     movq_r2r (mm3, mm2);                // mm2 = a2
00487 
00488     movq_r2r (mm4, mm6);                // mm6 = a1
00489     paddsw_r2r (mm7, mm3);              // mm3 = a2+b2
00490 
00491     psraw_i2r (COL_SHIFT, mm3);         // mm3 = y2
00492     paddsw_r2r (mm1, mm4);              // mm4 = a1+b1
00493 
00494     psraw_i2r (COL_SHIFT, mm4);         // mm4 = y1
00495     psubsw_r2r (mm1, mm6);              // mm6 = a1-b1
00496 
00497     movq_m2r (*(col+offset+5*8), mm1);  // mm1 = b0
00498     psubsw_r2r (mm7, mm2);              // mm2 = a2-b2
00499 
00500     psraw_i2r (COL_SHIFT, mm6);         // mm6 = y6
00501     movq_r2r (mm5, mm7);                // mm7 = a0
00502 
00503     movq_r2m (mm4, *(col+offset+1*8));  // save y1
00504     psraw_i2r (COL_SHIFT, mm2);         // mm2 = y5
00505 
00506     movq_r2m (mm3, *(col+offset+2*8));  // save y2
00507     paddsw_r2r (mm1, mm5);              // mm5 = a0+b0
00508 
00509     movq_m2r (*(col+offset+3*8), mm4);  // mm4 = b3
00510     psubsw_r2r (mm1, mm7);              // mm7 = a0-b0
00511 
00512     psraw_i2r (COL_SHIFT, mm5);         // mm5 = y0
00513     movq_r2r (mm0, mm3);                // mm3 = a3
00514 
00515     movq_r2m (mm2, *(col+offset+5*8));  // save y5
00516     psubsw_r2r (mm4, mm3);              // mm3 = a3-b3
00517 
00518     psraw_i2r (COL_SHIFT, mm7);         // mm7 = y7
00519     paddsw_r2r (mm0, mm4);              // mm4 = a3+b3
00520 
00521     movq_r2m (mm5, *(col+offset+0*8));  // save y0
00522     psraw_i2r (COL_SHIFT, mm3);         // mm3 = y4
00523 
00524     movq_r2m (mm6, *(col+offset+6*8));  // save y6
00525     psraw_i2r (COL_SHIFT, mm4);         // mm4 = y3
00526 
00527     movq_r2m (mm7, *(col+offset+7*8));  // save y7
00528 
00529     movq_r2m (mm3, *(col+offset+4*8));  // save y4
00530 
00531     movq_r2m (mm4, *(col+offset+3*8));  // save y3
00532 
00533 #undef T1
00534 #undef T2
00535 #undef T3
00536 #undef C4
00537 }
00538 
00539 static int32_t rounder0[] ATTR_ALIGN(8) =
00540     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
00541 static int32_t rounder4[] ATTR_ALIGN(8) = rounder (0);
00542 static int32_t rounder1[] ATTR_ALIGN(8) =
00543     rounder (1.25683487303);    /* C1*(C1/C4+C1+C7)/2 */
00544 static int32_t rounder7[] ATTR_ALIGN(8) =
00545     rounder (-0.25);            /* C1*(C7/C4+C7-C1)/2 */
00546 static int32_t rounder2[] ATTR_ALIGN(8) =
00547     rounder (0.60355339059);    /* C2 * (C6+C2)/2 */
00548 static int32_t rounder6[] ATTR_ALIGN(8) =
00549     rounder (-0.25);            /* C2 * (C6-C2)/2 */
00550 static int32_t rounder3[] ATTR_ALIGN(8) =
00551     rounder (0.087788325588);   /* C3*(-C3/C4+C3+C5)/2 */
00552 static int32_t rounder5[] ATTR_ALIGN(8) =
00553     rounder (-0.441341716183);  /* C3*(-C5/C4+C5-C3)/2 */
00554 
00555 #undef COL_SHIFT
00556 #undef ROW_SHIFT
00557 
00558 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid)      \
00559 void idct (int16_t * block)                                     \
00560 {                                                                       \
00561     static int16_t table04[] ATTR_ALIGN(16) =                           \
00562         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00563     static int16_t table17[] ATTR_ALIGN(16) =                           \
00564         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00565     static int16_t table26[] ATTR_ALIGN(16) =                           \
00566         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00567     static int16_t table35[] ATTR_ALIGN(16) =                           \
00568         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00569                                                                         \
00570     idct_row_head (block, 0*8, table04);                                \
00571     idct_row (table04, rounder0);                                       \
00572     idct_row_mid (block, 0*8, 4*8, table04);                            \
00573     idct_row (table04, rounder4);                                       \
00574     idct_row_mid (block, 4*8, 1*8, table17);                            \
00575     idct_row (table17, rounder1);                                       \
00576     idct_row_mid (block, 1*8, 7*8, table17);                            \
00577     idct_row (table17, rounder7);                                       \
00578     idct_row_mid (block, 7*8, 2*8, table26);                            \
00579     idct_row (table26, rounder2);                                       \
00580     idct_row_mid (block, 2*8, 6*8, table26);                            \
00581     idct_row (table26, rounder6);                                       \
00582     idct_row_mid (block, 6*8, 3*8, table35);                            \
00583     idct_row (table35, rounder3);                                       \
00584     idct_row_mid (block, 3*8, 5*8, table35);                            \
00585     idct_row (table35, rounder5);                                       \
00586     idct_row_tail (block, 5*8);                                         \
00587                                                                         \
00588     idct_col (block, 0);                                                \
00589     idct_col (block, 4);                                                \
00590 }
00591 
00592 void ff_mmx_idct(DCTELEM *block);
00593 void ff_mmxext_idct(DCTELEM *block);
00594 
00595 declare_idct (ff_mmxext_idct, mmxext_table,
00596               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00597 
00598 declare_idct (ff_mmx_idct, mmx_table,
00599               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00600
hvirtual/quicktime/ffmpeg/libavcodec/i386/idct_mmx.c
