FFmpeg: libavcodec/x86/idct

00001 /*
00002  * Copyright (C) 1999-2001 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
00003  *
00004  * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
00005  * See http://libmpeg2.sourceforge.net/ for updates.
00006  *
00007  * mpeg2dec is free software; you can redistribute it and/or modify
00008  * it under the terms of the GNU General Public License as published by
00009  * the Free Software Foundation; either version 2 of the License, or
00010  * (at your option) any later version.
00011  *
00012  * mpeg2dec is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
00015  * GNU General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU General Public License
00018  * along with mpeg2dec; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00022 #include "libavutil/common.h"
00023 #include "libavcodec/dsputil.h"
00024 
00025 #include "libavutil/x86_cpu.h"
00026 #include "dsputil_mmx.h"
00027 
00028 #define ROW_SHIFT 11
00029 #define COL_SHIFT 6
00030 
00031 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
00032 #define rounder(bias) {round (bias), round (bias)}
00033 
00034 
00035 #if 0
00036 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
00037 static inline void idct_row (int16_t * row, int offset,
00038                              int16_t * table, int32_t * rounder)
00039 {
00040     int C1, C2, C3, C4, C5, C6, C7;
00041     int a0, a1, a2, a3, b0, b1, b2, b3;
00042 
00043     row += offset;
00044 
00045     C1 = table[1];
00046     C2 = table[2];
00047     C3 = table[3];
00048     C4 = table[4];
00049     C5 = table[5];
00050     C6 = table[6];
00051     C7 = table[7];
00052 
00053     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
00054     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
00055     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
00056     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
00057 
00058     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
00059     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
00060     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
00061     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
00062 
00063     row[0] = (a0 + b0) >> ROW_SHIFT;
00064     row[1] = (a1 + b1) >> ROW_SHIFT;
00065     row[2] = (a2 + b2) >> ROW_SHIFT;
00066     row[3] = (a3 + b3) >> ROW_SHIFT;
00067     row[4] = (a3 - b3) >> ROW_SHIFT;
00068     row[5] = (a2 - b2) >> ROW_SHIFT;
00069     row[6] = (a1 - b1) >> ROW_SHIFT;
00070     row[7] = (a0 - b0) >> ROW_SHIFT;
00071 }
00072 #endif
00073 
00074 
00075 /* MMXEXT row IDCT */
00076 
00077 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
00078                                                    c4,  c6,  c4,  c6,   \
00079                                                    c1,  c3, -c1, -c5,   \
00080                                                    c5,  c7,  c3, -c7,   \
00081                                                    c4, -c6,  c4, -c6,   \
00082                                                   -c4,  c2,  c4, -c2,   \
00083                                                    c5, -c1,  c3, -c1,   \
00084                                                    c7,  c3,  c7, -c5 }
00085 
00086 static inline void mmxext_row_head (int16_t * const row, const int offset,
00087                                     const int16_t * const table)
00088 {
00089     __asm__ volatile(
00090         "movq     (%0), %%mm2        \n\t"  /* mm2 = x6 x4 x2 x0 */
00091 
00092         "movq    8(%0), %%mm5        \n\t"  /* mm5 = x7 x5 x3 x1 */
00093         "movq    %%mm2, %%mm0        \n\t"  /* mm0 = x6 x4 x2 x0 */
00094 
00095         "movq     (%1), %%mm3        \n\t"  /* mm3 = -C2 -C4 C2 C4 */
00096         "movq    %%mm5, %%mm6        \n\t"  /* mm6 = x7 x5 x3 x1 */
00097 
00098         "movq    8(%1), %%mm4        \n\t"  /* mm4 = C6 C4 C6 C4 */
00099         "pmaddwd %%mm0, %%mm3        \n\t"  /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00100 
00101         "pshufw  $0x4e, %%mm2, %%mm2 \n\t"  /* mm2 = x2 x0 x6 x4 */
00102         :: "r" ((row+offset)), "r" (table)
00103     );
00104 }
00105 
00106 static inline void mmxext_row (const int16_t * const table,
00107                                const int32_t * const rounder)
00108 {
00109     __asm__ volatile (
00110         "movq    16(%0), %%mm1         \n\t" /* mm1 = -C5 -C1 C3 C1 */
00111         "pmaddwd  %%mm2, %%mm4         \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
00112 
00113         "pmaddwd 32(%0), %%mm0         \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
00114         "pshufw   $0x4e, %%mm6, %%mm6  \n\t" /* mm6 = x3 x1 x7 x5 */
00115 
00116         "movq    24(%0), %%mm7         \n\t" /* mm7 = -C7 C3 C7 C5 */
00117         "pmaddwd  %%mm5, %%mm1         \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */
00118 
00119         "paddd     (%1), %%mm3         \n\t" /* mm3 += rounder */
00120         "pmaddwd  %%mm6, %%mm7         \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
00121 
00122         "pmaddwd 40(%0), %%mm2         \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */
00123         "paddd    %%mm4, %%mm3         \n\t" /* mm3 = a1 a0 + rounder */
00124 
00125         "pmaddwd 48(%0), %%mm5         \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
00126         "movq     %%mm3, %%mm4         \n\t" /* mm4 = a1 a0 + rounder */
00127 
00128         "pmaddwd 56(%0), %%mm6         \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
00129         "paddd    %%mm7, %%mm1         \n\t" /* mm1 = b1 b0 */
00130 
00131         "paddd     (%1), %%mm0         \n\t" /* mm0 += rounder */
00132         "psubd    %%mm1, %%mm3         \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
00133 
00134         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3         \n\t" /* mm3 = y6 y7 */
00135         "paddd    %%mm4, %%mm1         \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
00136 
00137         "paddd    %%mm2, %%mm0         \n\t" /* mm0 = a3 a2 + rounder */
00138         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1         \n\t" /* mm1 = y1 y0 */
00139 
00140         "paddd    %%mm6, %%mm5         \n\t" /* mm5 = b3 b2 */
00141         "movq     %%mm0, %%mm4         \n\t" /* mm4 = a3 a2 + rounder */
00142 
00143         "paddd    %%mm5, %%mm0         \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
00144         "psubd    %%mm5, %%mm4         \n\t" /* mm4 = a3-b3 a2-b2 + rounder */
00145         : : "r" (table), "r" (rounder));
00146 }
00147 
00148 static inline void mmxext_row_tail (int16_t * const row, const int store)
00149 {
00150     __asm__ volatile (
00151         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0        \n\t"  /* mm0 = y3 y2 */
00152 
00153         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4  \n\t"  /* mm4 = y4 y5 */
00154 
00155         "packssdw %%mm0, %%mm1        \n\t"  /* mm1 = y3 y2 y1 y0 */
00156 
00157         "packssdw %%mm3, %%mm4        \n\t"  /* mm4 = y6 y7 y4 y5 */
00158 
00159         "movq     %%mm1, (%0)         \n\t"  /* save y3 y2 y1 y0 */
00160         "pshufw   $0xb1, %%mm4, %%mm4 \n\t"  /* mm4 = y7 y6 y5 y4 */
00161 
00162         /* slot */
00163 
00164         "movq     %%mm4, 8(%0)        \n\t"  /* save y7 y6 y5 y4 */
00165         :: "r" (row+store)
00166         );
00167 }
00168 
00169 static inline void mmxext_row_mid (int16_t * const row, const int store,
00170                                    const int offset,
00171                                    const int16_t * const table)
00172 {
00173     __asm__ volatile (
00174         "movq     (%0,%1), %%mm2       \n\t" /* mm2 = x6 x4 x2 x0 */
00175         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t"   /* mm0 = y3 y2 */
00176 
00177         "movq    8(%0,%1), %%mm5       \n\t" /* mm5 = x7 x5 x3 x1 */
00178         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
00179 
00180         "packssdw   %%mm0, %%mm1       \n\t" /* mm1 = y3 y2 y1 y0 */
00181         "movq       %%mm5, %%mm6       \n\t" /* mm6 = x7 x5 x3 x1 */
00182 
00183         "packssdw   %%mm3, %%mm4       \n\t" /* mm4 = y6 y7 y4 y5 */
00184         "movq       %%mm2, %%mm0       \n\t" /* mm0 = x6 x4 x2 x0 */
00185 
00186         "movq       %%mm1, (%0,%2)     \n\t" /* save y3 y2 y1 y0 */
00187         "pshufw     $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
00188 
00189         "movq        (%3), %%mm3       \n\t" /* mm3 = -C2 -C4 C2 C4 */
00190         "movq       %%mm4, 8(%0,%2)    \n\t" /* save y7 y6 y5 y4 */
00191 
00192         "pmaddwd    %%mm0, %%mm3       \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
00193 
00194         "movq       8(%3), %%mm4       \n\t" /* mm4 = C6 C4 C6 C4 */
00195         "pshufw     $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */
00196         :: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00197         );
00198 }
00199 
00200 
00201 /* MMX row IDCT */
00202 
00203 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00204                                            c4,  c6, -c4, -c2,   \
00205                                            c1,  c3,  c3, -c7,   \
00206                                            c5,  c7, -c1, -c5,   \
00207                                            c4, -c6,  c4, -c2,   \
00208                                           -c4,  c2,  c4, -c6,   \
00209                                            c5, -c1,  c7, -c5,   \
00210                                            c7,  c3,  c3, -c1 }
00211 
00212 static inline void mmx_row_head (int16_t * const row, const int offset,
00213                                  const int16_t * const table)
00214 {
00215     __asm__ volatile (
00216         "movq (%0), %%mm2       \n\t"    /* mm2 = x6 x4 x2 x0 */
00217 
00218         "movq 8(%0), %%mm5      \n\t"    /* mm5 = x7 x5 x3 x1 */
00219         "movq %%mm2, %%mm0      \n\t"    /* mm0 = x6 x4 x2 x0 */
00220 
00221         "movq (%1), %%mm3       \n\t"    /* mm3 = C6 C4 C2 C4 */
00222         "movq %%mm5, %%mm6      \n\t"    /* mm6 = x7 x5 x3 x1 */
00223 
00224         "punpckldq %%mm0, %%mm0 \n\t"    /* mm0 = x2 x0 x2 x0 */
00225 
00226         "movq 8(%1), %%mm4      \n\t"    /* mm4 = -C2 -C4 C6 C4 */
00227         "pmaddwd %%mm0, %%mm3   \n\t"    /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00228 
00229         "movq 16(%1), %%mm1     \n\t"    /* mm1 = -C7 C3 C3 C1 */
00230         "punpckhdq %%mm2, %%mm2 \n\t"    /* mm2 = x6 x4 x6 x4 */
00231         :: "r" ((row+offset)), "r" (table)
00232         );
00233 }
00234 
00235 static inline void mmx_row (const int16_t * const table,
00236                             const int32_t * const rounder)
00237 {
00238     __asm__ volatile (
00239         "pmaddwd   %%mm2, %%mm4    \n\t"  /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
00240         "punpckldq %%mm5, %%mm5    \n\t"  /* mm5 = x3 x1 x3 x1 */
00241 
00242         "pmaddwd  32(%0), %%mm0    \n\t"  /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
00243         "punpckhdq %%mm6, %%mm6    \n\t"  /* mm6 = x7 x5 x7 x5 */
00244 
00245         "movq     24(%0), %%mm7    \n\t"  /* mm7 = -C5 -C1 C7 C5 */
00246         "pmaddwd   %%mm5, %%mm1    \n\t"  /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
00247 
00248         "paddd      (%1), %%mm3    \n\t"  /* mm3 += rounder */
00249         "pmaddwd   %%mm6, %%mm7    \n\t"  /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
00250 
00251         "pmaddwd  40(%0), %%mm2    \n\t"  /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
00252         "paddd     %%mm4, %%mm3    \n\t"  /* mm3 = a1 a0 + rounder */
00253 
00254         "pmaddwd  48(%0), %%mm5    \n\t"  /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
00255         "movq      %%mm3, %%mm4    \n\t"  /* mm4 = a1 a0 + rounder */
00256 
00257         "pmaddwd  56(%0), %%mm6    \n\t"  /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
00258         "paddd     %%mm7, %%mm1    \n\t"  /* mm1 = b1 b0 */
00259 
00260         "paddd      (%1), %%mm0    \n\t"  /* mm0 += rounder */
00261         "psubd     %%mm1, %%mm3    \n\t"  /* mm3 = a1-b1 a0-b0 + rounder */
00262 
00263         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3    \n\t"  /* mm3 = y6 y7 */
00264         "paddd     %%mm4, %%mm1    \n\t"  /* mm1 = a1+b1 a0+b0 + rounder */
00265 
00266         "paddd     %%mm2, %%mm0    \n\t"  /* mm0 = a3 a2 + rounder */
00267         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1    \n\t"  /* mm1 = y1 y0 */
00268 
00269         "paddd     %%mm6, %%mm5    \n\t"  /* mm5 = b3 b2 */
00270         "movq      %%mm0, %%mm7    \n\t"  /* mm7 = a3 a2 + rounder */
00271 
00272         "paddd     %%mm5, %%mm0    \n\t"  /* mm0 = a3+b3 a2+b2 + rounder */
00273         "psubd     %%mm5, %%mm7    \n\t"  /* mm7 = a3-b3 a2-b2 + rounder */
00274         :: "r" (table), "r" (rounder)
00275         );
00276 }
00277 
00278 static inline void mmx_row_tail (int16_t * const row, const int store)
00279 {
00280     __asm__ volatile (
00281         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0      \n\t" /* mm0 = y3 y2 */
00282 
00283         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7      \n\t" /* mm7 = y4 y5 */
00284 
00285         "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
00286 
00287         "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
00288 
00289         "movq %%mm1, (%0)      \n\t" /* save y3 y2 y1 y0 */
00290         "movq %%mm7, %%mm4     \n\t" /* mm4 = y6 y7 y4 y5 */
00291 
00292         "pslld $16, %%mm7      \n\t" /* mm7 = y7 0 y5 0 */
00293 
00294         "psrld $16, %%mm4      \n\t" /* mm4 = 0 y6 0 y4 */
00295 
00296         "por %%mm4, %%mm7      \n\t" /* mm7 = y7 y6 y5 y4 */
00297 
00298         /* slot */
00299 
00300         "movq %%mm7, 8(%0)     \n\t" /* save y7 y6 y5 y4 */
00301         :: "r" (row+store)
00302         );
00303 }
00304 
00305 static inline void mmx_row_mid (int16_t * const row, const int store,
00306                                 const int offset, const int16_t * const table)
00307 {
00308 
00309     __asm__ volatile (
00310         "movq    (%0,%1), %%mm2    \n\t" /* mm2 = x6 x4 x2 x0 */
00311         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
00312 
00313         "movq   8(%0,%1), %%mm5    \n\t" /* mm5 = x7 x5 x3 x1 */
00314         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
00315 
00316         "packssdw  %%mm0, %%mm1    \n\t" /* mm1 = y3 y2 y1 y0 */
00317         "movq      %%mm5, %%mm6    \n\t" /* mm6 = x7 x5 x3 x1 */
00318 
00319         "packssdw  %%mm3, %%mm7    \n\t" /* mm7 = y6 y7 y4 y5 */
00320         "movq      %%mm2, %%mm0    \n\t" /* mm0 = x6 x4 x2 x0 */
00321 
00322         "movq      %%mm1, (%0,%2)  \n\t" /* save y3 y2 y1 y0 */
00323         "movq      %%mm7, %%mm1    \n\t" /* mm1 = y6 y7 y4 y5 */
00324 
00325         "punpckldq %%mm0, %%mm0    \n\t" /* mm0 = x2 x0 x2 x0 */
00326         "psrld       $16, %%mm7    \n\t" /* mm7 = 0 y6 0 y4 */
00327 
00328         "movq       (%3), %%mm3    \n\t" /* mm3 = C6 C4 C2 C4 */
00329         "pslld       $16, %%mm1    \n\t" /* mm1 = y7 0 y5 0 */
00330 
00331         "movq      8(%3), %%mm4    \n\t" /* mm4 = -C2 -C4 C6 C4 */
00332         "por       %%mm1, %%mm7    \n\t" /* mm7 = y7 y6 y5 y4 */
00333 
00334         "movq     16(%3), %%mm1    \n\t" /* mm1 = -C7 C3 C3 C1 */
00335         "punpckhdq %%mm2, %%mm2    \n\t" /* mm2 = x6 x4 x6 x4 */
00336 
00337         "movq      %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
00338         "pmaddwd   %%mm0, %%mm3    \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00339         : : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00340         );
00341 }
00342 
00343 
00344 #if 0
00345 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
00346 static inline void idct_col (int16_t * col, int offset)
00347 {
00348 /* multiplication - as implemented on mmx */
00349 #define F(c,x) (((c) * (x)) >> 16)
00350 
00351 /* saturation - it helps us handle torture test cases */
00352 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00353 
00354     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00355     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00356     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00357     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00358 
00359     col += offset;
00360 
00361     x0 = col[0*8];
00362     x1 = col[1*8];
00363     x2 = col[2*8];
00364     x3 = col[3*8];
00365     x4 = col[4*8];
00366     x5 = col[5*8];
00367     x6 = col[6*8];
00368     x7 = col[7*8];
00369 
00370     u04 = S (x0 + x4);
00371     v04 = S (x0 - x4);
00372     u26 = S (F (T2, x6) + x2);
00373     v26 = S (F (T2, x2) - x6);
00374 
00375     a0 = S (u04 + u26);
00376     a1 = S (v04 + v26);
00377     a2 = S (v04 - v26);
00378     a3 = S (u04 - u26);
00379 
00380     u17 = S (F (T1, x7) + x1);
00381     v17 = S (F (T1, x1) - x7);
00382     u35 = S (F (T3, x5) + x3);
00383     v35 = S (F (T3, x3) - x5);
00384 
00385     b0 = S (u17 + u35);
00386     b3 = S (v17 - v35);
00387     u12 = S (u17 - u35);
00388     v12 = S (v17 + v35);
00389     u12 = S (2 * F (C4, u12));
00390     v12 = S (2 * F (C4, v12));
00391     b1 = S (u12 + v12);
00392     b2 = S (u12 - v12);
00393 
00394     y0 = S (a0 + b0) >> COL_SHIFT;
00395     y1 = S (a1 + b1) >> COL_SHIFT;
00396     y2 = S (a2 + b2) >> COL_SHIFT;
00397     y3 = S (a3 + b3) >> COL_SHIFT;
00398 
00399     y4 = S (a3 - b3) >> COL_SHIFT;
00400     y5 = S (a2 - b2) >> COL_SHIFT;
00401     y6 = S (a1 - b1) >> COL_SHIFT;
00402     y7 = S (a0 - b0) >> COL_SHIFT;
00403 
00404     col[0*8] = y0;
00405     col[1*8] = y1;
00406     col[2*8] = y2;
00407     col[3*8] = y3;
00408     col[4*8] = y4;
00409     col[5*8] = y5;
00410     col[6*8] = y6;
00411     col[7*8] = y7;
00412 }
00413 #endif
00414 
00415 
00416 /* MMX column IDCT */
00417 static inline void idct_col (int16_t * const col, const int offset)
00418 {
00419 #define T1 13036
00420 #define T2 27146
00421 #define T3 43790
00422 #define C4 23170
00423 
00424     DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
00425         T1,T1,T1,T1,
00426         T2,T2,T2,T2,
00427         T3,T3,T3,T3,
00428         C4,C4,C4,C4
00429     };
00430 
00431     /* column code adapted from Peter Gubanov */
00432     /* http://www.elecard.com/peter/idct.shtml */
00433 
00434     __asm__ volatile (
00435         "movq      (%0), %%mm0    \n\t" /* mm0 = T1 */
00436 
00437         "movq   2*8(%1), %%mm1    \n\t" /* mm1 = x1 */
00438         "movq     %%mm0, %%mm2    \n\t" /* mm2 = T1 */
00439 
00440         "movq 7*2*8(%1), %%mm4    \n\t" /* mm4 = x7 */
00441         "pmulhw   %%mm1, %%mm0    \n\t" /* mm0 = T1*x1 */
00442 
00443         "movq    16(%0), %%mm5    \n\t" /* mm5 = T3 */
00444         "pmulhw   %%mm4, %%mm2    \n\t" /* mm2 = T1*x7 */
00445 
00446         "movq 2*5*8(%1), %%mm6    \n\t" /* mm6 = x5 */
00447         "movq     %%mm5, %%mm7    \n\t" /* mm7 = T3-1 */
00448 
00449         "movq 3*8*2(%1), %%mm3    \n\t" /* mm3 = x3 */
00450         "psubsw   %%mm4, %%mm0    \n\t" /* mm0 = v17 */
00451 
00452         "movq     8(%0), %%mm4    \n\t" /* mm4 = T2 */
00453         "pmulhw   %%mm3, %%mm5    \n\t" /* mm5 = (T3-1)*x3 */
00454 
00455         "paddsw   %%mm2, %%mm1    \n\t" /* mm1 = u17 */
00456         "pmulhw   %%mm6, %%mm7    \n\t" /* mm7 = (T3-1)*x5 */
00457 
00458         /* slot */
00459 
00460         "movq     %%mm4, %%mm2    \n\t" /* mm2 = T2 */
00461         "paddsw   %%mm3, %%mm5    \n\t" /* mm5 = T3*x3 */
00462 
00463         "pmulhw 2*8*2(%1), %%mm4  \n\t" /* mm4 = T2*x2 */
00464         "paddsw   %%mm6, %%mm7    \n\t" /* mm7 = T3*x5 */
00465 
00466         "psubsw   %%mm6, %%mm5    \n\t" /* mm5 = v35 */
00467         "paddsw   %%mm3, %%mm7    \n\t" /* mm7 = u35 */
00468 
00469         "movq 6*8*2(%1), %%mm3    \n\t" /* mm3 = x6 */
00470         "movq     %%mm0, %%mm6    \n\t" /* mm6 = v17 */
00471 
00472         "pmulhw   %%mm3, %%mm2    \n\t" /* mm2 = T2*x6 */
00473         "psubsw   %%mm5, %%mm0    \n\t" /* mm0 = b3 */
00474 
00475         "psubsw   %%mm3, %%mm4    \n\t" /* mm4 = v26 */
00476         "paddsw   %%mm6, %%mm5    \n\t" /* mm5 = v12 */
00477 
00478         "movq     %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
00479         "movq     %%mm1, %%mm6    \n\t" /* mm6 = u17 */
00480 
00481         "paddsw 2*8*2(%1), %%mm2  \n\t" /* mm2 = u26 */
00482         "paddsw   %%mm7, %%mm6    \n\t" /* mm6 = b0 */
00483 
00484         "psubsw   %%mm7, %%mm1    \n\t" /* mm1 = u12 */
00485         "movq     %%mm1, %%mm7    \n\t" /* mm7 = u12 */
00486 
00487         "movq   0*8(%1), %%mm3    \n\t" /* mm3 = x0 */
00488         "paddsw   %%mm5, %%mm1    \n\t" /* mm1 = u12+v12 */
00489 
00490         "movq    24(%0), %%mm0    \n\t" /* mm0 = C4/2 */
00491         "psubsw   %%mm5, %%mm7    \n\t" /* mm7 = u12-v12 */
00492 
00493         "movq     %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
00494         "pmulhw   %%mm0, %%mm1    \n\t" /* mm1 = b1/2 */
00495 
00496         "movq     %%mm4, %%mm6    \n\t" /* mm6 = v26 */
00497         "pmulhw   %%mm0, %%mm7    \n\t" /* mm7 = b2/2 */
00498 
00499         "movq 4*8*2(%1), %%mm5    \n\t" /* mm5 = x4 */
00500         "movq     %%mm3, %%mm0    \n\t" /* mm0 = x0 */
00501 
00502         "psubsw   %%mm5, %%mm3    \n\t" /* mm3 = v04 */
00503         "paddsw   %%mm5, %%mm0    \n\t" /* mm0 = u04 */
00504 
00505         "paddsw   %%mm3, %%mm4    \n\t" /* mm4 = a1 */
00506         "movq     %%mm0, %%mm5    \n\t" /* mm5 = u04 */
00507 
00508         "psubsw   %%mm6, %%mm3    \n\t" /* mm3 = a2 */
00509         "paddsw   %%mm2, %%mm5    \n\t" /* mm5 = a0 */
00510 
00511         "paddsw   %%mm1, %%mm1    \n\t" /* mm1 = b1 */
00512         "psubsw   %%mm2, %%mm0    \n\t" /* mm0 = a3 */
00513 
00514         "paddsw   %%mm7, %%mm7    \n\t" /* mm7 = b2 */
00515         "movq     %%mm3, %%mm2    \n\t" /* mm2 = a2 */
00516 
00517         "movq     %%mm4, %%mm6    \n\t" /* mm6 = a1 */
00518         "paddsw   %%mm7, %%mm3    \n\t" /* mm3 = a2+b2 */
00519 
00520         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
00521         "paddsw   %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
00522 
00523         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
00524         "psubsw   %%mm1, %%mm6    \n\t" /* mm6 = a1-b1 */
00525 
00526         "movq 5*8*2(%1), %%mm1    \n\t" /* mm1 = b0 */
00527         "psubsw   %%mm7, %%mm2    \n\t" /* mm2 = a2-b2 */
00528 
00529         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
00530         "movq     %%mm5, %%mm7    \n\t" /* mm7 = a0 */
00531 
00532         "movq     %%mm4, 1*8*2(%1)\n\t" /* save y1 */
00533         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
00534 
00535         "movq     %%mm3, 2*8*2(%1)\n\t" /* save y2 */
00536         "paddsw   %%mm1, %%mm5    \n\t" /* mm5 = a0+b0 */
00537 
00538         "movq 3*8*2(%1), %%mm4    \n\t" /* mm4 = b3 */
00539         "psubsw   %%mm1, %%mm7    \n\t" /* mm7 = a0-b0 */
00540 
00541         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
00542         "movq     %%mm0, %%mm3    \n\t" /* mm3 = a3 */
00543 
00544         "movq     %%mm2, 5*8*2(%1)\n\t" /* save y5 */
00545         "psubsw   %%mm4, %%mm3    \n\t" /* mm3 = a3-b3 */
00546 
00547         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
00548         "paddsw   %%mm0, %%mm4    \n\t" /* mm4 = a3+b3 */
00549 
00550         "movq     %%mm5, 0*8*2(%1)\n\t" /* save y0 */
00551         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
00552 
00553         "movq     %%mm6, 6*8*2(%1)\n\t" /* save y6 */
00554         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
00555 
00556         "movq     %%mm7, 7*8*2(%1)\n\t" /* save y7 */
00557 
00558         "movq     %%mm3, 4*8*2(%1)\n\t" /* save y4 */
00559 
00560         "movq     %%mm4, 3*8*2(%1)\n\t" /* save y3 */
00561         :: "r" (t1_vector), "r" (col+offset)
00562         );
00563 
00564 #undef T1
00565 #undef T2
00566 #undef T3
00567 #undef C4
00568 }
00569 
00570 
00571 DECLARE_ALIGNED(8, static const int32_t, rounder0)[] =
00572     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
00573 DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0);
00574 DECLARE_ALIGNED(8, static const int32_t, rounder1)[] =
00575     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
00576 DECLARE_ALIGNED(8, static const int32_t, rounder7)[] =
00577     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
00578 DECLARE_ALIGNED(8, static const int32_t, rounder2)[] =
00579     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
00580 DECLARE_ALIGNED(8, static const int32_t, rounder6)[] =
00581     rounder (-0.25);                /* C2 * (C6-C2)/2 */
00582 DECLARE_ALIGNED(8, static const int32_t, rounder3)[] =
00583     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
00584 DECLARE_ALIGNED(8, static const int32_t, rounder5)[] =
00585     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
00586 
00587 #undef COL_SHIFT
00588 #undef ROW_SHIFT
00589 
00590 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
00591 void idct (int16_t * const block)                                       \
00592 {                                                                       \
00593     DECLARE_ALIGNED(16, static const int16_t, table04)[] =              \
00594         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00595     DECLARE_ALIGNED(16, static const int16_t, table17)[] =              \
00596         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00597     DECLARE_ALIGNED(16, static const int16_t, table26)[] =              \
00598         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00599     DECLARE_ALIGNED(16, static const int16_t, table35)[] =              \
00600         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00601                                                                         \
00602     idct_row_head (block, 0*8, table04);                                \
00603     idct_row (table04, rounder0);                                       \
00604     idct_row_mid (block, 0*8, 4*8, table04);                            \
00605     idct_row (table04, rounder4);                                       \
00606     idct_row_mid (block, 4*8, 1*8, table17);                            \
00607     idct_row (table17, rounder1);                                       \
00608     idct_row_mid (block, 1*8, 7*8, table17);                            \
00609     idct_row (table17, rounder7);                                       \
00610     idct_row_mid (block, 7*8, 2*8, table26);                            \
00611     idct_row (table26, rounder2);                                       \
00612     idct_row_mid (block, 2*8, 6*8, table26);                            \
00613     idct_row (table26, rounder6);                                       \
00614     idct_row_mid (block, 6*8, 3*8, table35);                            \
00615     idct_row (table35, rounder3);                                       \
00616     idct_row_mid (block, 3*8, 5*8, table35);                            \
00617     idct_row (table35, rounder5);                                       \
00618     idct_row_tail (block, 5*8);                                         \
00619                                                                         \
00620     idct_col (block, 0);                                                \
00621     idct_col (block, 4);                                                \
00622 }
00623 
00624 declare_idct (ff_mmxext_idct, mmxext_table,
00625               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00626 
00627 declare_idct (ff_mmx_idct, mmx_table,
00628               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00629
libavcodec/x86/idct_mmx.c
