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/asm.h"
00026 #include "dsputil_mmx.h"
00027 
00028 #if HAVE_INLINE_ASM
00029 
00030 #define ROW_SHIFT 11
00031 #define COL_SHIFT 6
00032 
00033 #define round(bias) ((int)(((bias)+0.5) * (1<<ROW_SHIFT)))
00034 #define rounder(bias) {round (bias), round (bias)}
00035 
00036 
00037 #if 0
00038 /* C row IDCT - it is just here to document the MMXEXT and MMX versions */
00039 static inline void idct_row (int16_t * row, int offset,
00040                              int16_t * table, int32_t * rounder)
00041 {
00042     int C1, C2, C3, C4, C5, C6, C7;
00043     int a0, a1, a2, a3, b0, b1, b2, b3;
00044 
00045     row += offset;
00046 
00047     C1 = table[1];
00048     C2 = table[2];
00049     C3 = table[3];
00050     C4 = table[4];
00051     C5 = table[5];
00052     C6 = table[6];
00053     C7 = table[7];
00054 
00055     a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + *rounder;
00056     a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + *rounder;
00057     a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + *rounder;
00058     a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + *rounder;
00059 
00060     b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
00061     b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
00062     b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
00063     b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
00064 
00065     row[0] = (a0 + b0) >> ROW_SHIFT;
00066     row[1] = (a1 + b1) >> ROW_SHIFT;
00067     row[2] = (a2 + b2) >> ROW_SHIFT;
00068     row[3] = (a3 + b3) >> ROW_SHIFT;
00069     row[4] = (a3 - b3) >> ROW_SHIFT;
00070     row[5] = (a2 - b2) >> ROW_SHIFT;
00071     row[6] = (a1 - b1) >> ROW_SHIFT;
00072     row[7] = (a0 - b0) >> ROW_SHIFT;
00073 }
00074 #endif
00075 
00076 
00077 /* MMXEXT row IDCT */
00078 
00079 #define mmxext_table(c1,c2,c3,c4,c5,c6,c7)      {  c4,  c2, -c4, -c2,   \
00080                                                    c4,  c6,  c4,  c6,   \
00081                                                    c1,  c3, -c1, -c5,   \
00082                                                    c5,  c7,  c3, -c7,   \
00083                                                    c4, -c6,  c4, -c6,   \
00084                                                   -c4,  c2,  c4, -c2,   \
00085                                                    c5, -c1,  c3, -c1,   \
00086                                                    c7,  c3,  c7, -c5 }
00087 
00088 static inline void mmxext_row_head (int16_t * const row, const int offset,
00089                                     const int16_t * const table)
00090 {
00091     __asm__ volatile(
00092         "movq     (%0), %%mm2        \n\t"  /* mm2 = x6 x4 x2 x0 */
00093 
00094         "movq    8(%0), %%mm5        \n\t"  /* mm5 = x7 x5 x3 x1 */
00095         "movq    %%mm2, %%mm0        \n\t"  /* mm0 = x6 x4 x2 x0 */
00096 
00097         "movq     (%1), %%mm3        \n\t"  /* mm3 = -C2 -C4 C2 C4 */
00098         "movq    %%mm5, %%mm6        \n\t"  /* mm6 = x7 x5 x3 x1 */
00099 
00100         "movq    8(%1), %%mm4        \n\t"  /* mm4 = C6 C4 C6 C4 */
00101         "pmaddwd %%mm0, %%mm3        \n\t"  /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00102 
00103         "pshufw  $0x4e, %%mm2, %%mm2 \n\t"  /* mm2 = x2 x0 x6 x4 */
00104         :: "r" ((row+offset)), "r" (table)
00105     );
00106 }
00107 
00108 static inline void mmxext_row (const int16_t * const table,
00109                                const int32_t * const rounder)
00110 {
00111     __asm__ volatile (
00112         "movq    16(%0), %%mm1         \n\t" /* mm1 = -C5 -C1 C3 C1 */
00113         "pmaddwd  %%mm2, %%mm4         \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
00114 
00115         "pmaddwd 32(%0), %%mm0         \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
00116         "pshufw   $0x4e, %%mm6, %%mm6  \n\t" /* mm6 = x3 x1 x7 x5 */
00117 
00118         "movq    24(%0), %%mm7         \n\t" /* mm7 = -C7 C3 C7 C5 */
00119         "pmaddwd  %%mm5, %%mm1         \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */
00120 
00121         "paddd     (%1), %%mm3         \n\t" /* mm3 += rounder */
00122         "pmaddwd  %%mm6, %%mm7         \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
00123 
00124         "pmaddwd 40(%0), %%mm2         \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */
00125         "paddd    %%mm4, %%mm3         \n\t" /* mm3 = a1 a0 + rounder */
00126 
00127         "pmaddwd 48(%0), %%mm5         \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
00128         "movq     %%mm3, %%mm4         \n\t" /* mm4 = a1 a0 + rounder */
00129 
00130         "pmaddwd 56(%0), %%mm6         \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
00131         "paddd    %%mm7, %%mm1         \n\t" /* mm1 = b1 b0 */
00132 
00133         "paddd     (%1), %%mm0         \n\t" /* mm0 += rounder */
00134         "psubd    %%mm1, %%mm3         \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
00135 
00136         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3         \n\t" /* mm3 = y6 y7 */
00137         "paddd    %%mm4, %%mm1         \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
00138 
00139         "paddd    %%mm2, %%mm0         \n\t" /* mm0 = a3 a2 + rounder */
00140         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1         \n\t" /* mm1 = y1 y0 */
00141 
00142         "paddd    %%mm6, %%mm5         \n\t" /* mm5 = b3 b2 */
00143         "movq     %%mm0, %%mm4         \n\t" /* mm4 = a3 a2 + rounder */
00144 
00145         "paddd    %%mm5, %%mm0         \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
00146         "psubd    %%mm5, %%mm4         \n\t" /* mm4 = a3-b3 a2-b2 + rounder */
00147         : : "r" (table), "r" (rounder));
00148 }
00149 
00150 static inline void mmxext_row_tail (int16_t * const row, const int store)
00151 {
00152     __asm__ volatile (
00153         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0        \n\t"  /* mm0 = y3 y2 */
00154 
00155         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4  \n\t"  /* mm4 = y4 y5 */
00156 
00157         "packssdw %%mm0, %%mm1        \n\t"  /* mm1 = y3 y2 y1 y0 */
00158 
00159         "packssdw %%mm3, %%mm4        \n\t"  /* mm4 = y6 y7 y4 y5 */
00160 
00161         "movq     %%mm1, (%0)         \n\t"  /* save y3 y2 y1 y0 */
00162         "pshufw   $0xb1, %%mm4, %%mm4 \n\t"  /* mm4 = y7 y6 y5 y4 */
00163 
00164         /* slot */
00165 
00166         "movq     %%mm4, 8(%0)        \n\t"  /* save y7 y6 y5 y4 */
00167         :: "r" (row+store)
00168         );
00169 }
00170 
00171 static inline void mmxext_row_mid (int16_t * const row, const int store,
00172                                    const int offset,
00173                                    const int16_t * const table)
00174 {
00175     __asm__ volatile (
00176         "movq     (%0,%1), %%mm2       \n\t" /* mm2 = x6 x4 x2 x0 */
00177         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t"   /* mm0 = y3 y2 */
00178 
00179         "movq    8(%0,%1), %%mm5       \n\t" /* mm5 = x7 x5 x3 x1 */
00180         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
00181 
00182         "packssdw   %%mm0, %%mm1       \n\t" /* mm1 = y3 y2 y1 y0 */
00183         "movq       %%mm5, %%mm6       \n\t" /* mm6 = x7 x5 x3 x1 */
00184 
00185         "packssdw   %%mm3, %%mm4       \n\t" /* mm4 = y6 y7 y4 y5 */
00186         "movq       %%mm2, %%mm0       \n\t" /* mm0 = x6 x4 x2 x0 */
00187 
00188         "movq       %%mm1, (%0,%2)     \n\t" /* save y3 y2 y1 y0 */
00189         "pshufw     $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
00190 
00191         "movq        (%3), %%mm3       \n\t" /* mm3 = -C2 -C4 C2 C4 */
00192         "movq       %%mm4, 8(%0,%2)    \n\t" /* save y7 y6 y5 y4 */
00193 
00194         "pmaddwd    %%mm0, %%mm3       \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
00195 
00196         "movq       8(%3), %%mm4       \n\t" /* mm4 = C6 C4 C6 C4 */
00197         "pshufw     $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */
00198         :: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00199         );
00200 }
00201 
00202 
00203 /* MMX row IDCT */
00204 
00205 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00206                                            c4,  c6, -c4, -c2,   \
00207                                            c1,  c3,  c3, -c7,   \
00208                                            c5,  c7, -c1, -c5,   \
00209                                            c4, -c6,  c4, -c2,   \
00210                                           -c4,  c2,  c4, -c6,   \
00211                                            c5, -c1,  c7, -c5,   \
00212                                            c7,  c3,  c3, -c1 }
00213 
00214 static inline void mmx_row_head (int16_t * const row, const int offset,
00215                                  const int16_t * const table)
00216 {
00217     __asm__ volatile (
00218         "movq (%0), %%mm2       \n\t"    /* mm2 = x6 x4 x2 x0 */
00219 
00220         "movq 8(%0), %%mm5      \n\t"    /* mm5 = x7 x5 x3 x1 */
00221         "movq %%mm2, %%mm0      \n\t"    /* mm0 = x6 x4 x2 x0 */
00222 
00223         "movq (%1), %%mm3       \n\t"    /* mm3 = C6 C4 C2 C4 */
00224         "movq %%mm5, %%mm6      \n\t"    /* mm6 = x7 x5 x3 x1 */
00225 
00226         "punpckldq %%mm0, %%mm0 \n\t"    /* mm0 = x2 x0 x2 x0 */
00227 
00228         "movq 8(%1), %%mm4      \n\t"    /* mm4 = -C2 -C4 C6 C4 */
00229         "pmaddwd %%mm0, %%mm3   \n\t"    /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00230 
00231         "movq 16(%1), %%mm1     \n\t"    /* mm1 = -C7 C3 C3 C1 */
00232         "punpckhdq %%mm2, %%mm2 \n\t"    /* mm2 = x6 x4 x6 x4 */
00233         :: "r" ((row+offset)), "r" (table)
00234         );
00235 }
00236 
00237 static inline void mmx_row (const int16_t * const table,
00238                             const int32_t * const rounder)
00239 {
00240     __asm__ volatile (
00241         "pmaddwd   %%mm2, %%mm4    \n\t"  /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
00242         "punpckldq %%mm5, %%mm5    \n\t"  /* mm5 = x3 x1 x3 x1 */
00243 
00244         "pmaddwd  32(%0), %%mm0    \n\t"  /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
00245         "punpckhdq %%mm6, %%mm6    \n\t"  /* mm6 = x7 x5 x7 x5 */
00246 
00247         "movq     24(%0), %%mm7    \n\t"  /* mm7 = -C5 -C1 C7 C5 */
00248         "pmaddwd   %%mm5, %%mm1    \n\t"  /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
00249 
00250         "paddd      (%1), %%mm3    \n\t"  /* mm3 += rounder */
00251         "pmaddwd   %%mm6, %%mm7    \n\t"  /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
00252 
00253         "pmaddwd  40(%0), %%mm2    \n\t"  /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
00254         "paddd     %%mm4, %%mm3    \n\t"  /* mm3 = a1 a0 + rounder */
00255 
00256         "pmaddwd  48(%0), %%mm5    \n\t"  /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
00257         "movq      %%mm3, %%mm4    \n\t"  /* mm4 = a1 a0 + rounder */
00258 
00259         "pmaddwd  56(%0), %%mm6    \n\t"  /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
00260         "paddd     %%mm7, %%mm1    \n\t"  /* mm1 = b1 b0 */
00261 
00262         "paddd      (%1), %%mm0    \n\t"  /* mm0 += rounder */
00263         "psubd     %%mm1, %%mm3    \n\t"  /* mm3 = a1-b1 a0-b0 + rounder */
00264 
00265         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3    \n\t"  /* mm3 = y6 y7 */
00266         "paddd     %%mm4, %%mm1    \n\t"  /* mm1 = a1+b1 a0+b0 + rounder */
00267 
00268         "paddd     %%mm2, %%mm0    \n\t"  /* mm0 = a3 a2 + rounder */
00269         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1    \n\t"  /* mm1 = y1 y0 */
00270 
00271         "paddd     %%mm6, %%mm5    \n\t"  /* mm5 = b3 b2 */
00272         "movq      %%mm0, %%mm7    \n\t"  /* mm7 = a3 a2 + rounder */
00273 
00274         "paddd     %%mm5, %%mm0    \n\t"  /* mm0 = a3+b3 a2+b2 + rounder */
00275         "psubd     %%mm5, %%mm7    \n\t"  /* mm7 = a3-b3 a2-b2 + rounder */
00276         :: "r" (table), "r" (rounder)
00277         );
00278 }
00279 
00280 static inline void mmx_row_tail (int16_t * const row, const int store)
00281 {
00282     __asm__ volatile (
00283         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0      \n\t" /* mm0 = y3 y2 */
00284 
00285         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7      \n\t" /* mm7 = y4 y5 */
00286 
00287         "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
00288 
00289         "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
00290 
00291         "movq %%mm1, (%0)      \n\t" /* save y3 y2 y1 y0 */
00292         "movq %%mm7, %%mm4     \n\t" /* mm4 = y6 y7 y4 y5 */
00293 
00294         "pslld $16, %%mm7      \n\t" /* mm7 = y7 0 y5 0 */
00295 
00296         "psrld $16, %%mm4      \n\t" /* mm4 = 0 y6 0 y4 */
00297 
00298         "por %%mm4, %%mm7      \n\t" /* mm7 = y7 y6 y5 y4 */
00299 
00300         /* slot */
00301 
00302         "movq %%mm7, 8(%0)     \n\t" /* save y7 y6 y5 y4 */
00303         :: "r" (row+store)
00304         );
00305 }
00306 
00307 static inline void mmx_row_mid (int16_t * const row, const int store,
00308                                 const int offset, const int16_t * const table)
00309 {
00310 
00311     __asm__ volatile (
00312         "movq    (%0,%1), %%mm2    \n\t" /* mm2 = x6 x4 x2 x0 */
00313         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
00314 
00315         "movq   8(%0,%1), %%mm5    \n\t" /* mm5 = x7 x5 x3 x1 */
00316         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
00317 
00318         "packssdw  %%mm0, %%mm1    \n\t" /* mm1 = y3 y2 y1 y0 */
00319         "movq      %%mm5, %%mm6    \n\t" /* mm6 = x7 x5 x3 x1 */
00320 
00321         "packssdw  %%mm3, %%mm7    \n\t" /* mm7 = y6 y7 y4 y5 */
00322         "movq      %%mm2, %%mm0    \n\t" /* mm0 = x6 x4 x2 x0 */
00323 
00324         "movq      %%mm1, (%0,%2)  \n\t" /* save y3 y2 y1 y0 */
00325         "movq      %%mm7, %%mm1    \n\t" /* mm1 = y6 y7 y4 y5 */
00326 
00327         "punpckldq %%mm0, %%mm0    \n\t" /* mm0 = x2 x0 x2 x0 */
00328         "psrld       $16, %%mm7    \n\t" /* mm7 = 0 y6 0 y4 */
00329 
00330         "movq       (%3), %%mm3    \n\t" /* mm3 = C6 C4 C2 C4 */
00331         "pslld       $16, %%mm1    \n\t" /* mm1 = y7 0 y5 0 */
00332 
00333         "movq      8(%3), %%mm4    \n\t" /* mm4 = -C2 -C4 C6 C4 */
00334         "por       %%mm1, %%mm7    \n\t" /* mm7 = y7 y6 y5 y4 */
00335 
00336         "movq     16(%3), %%mm1    \n\t" /* mm1 = -C7 C3 C3 C1 */
00337         "punpckhdq %%mm2, %%mm2    \n\t" /* mm2 = x6 x4 x6 x4 */
00338 
00339         "movq      %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
00340         "pmaddwd   %%mm0, %%mm3    \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00341         : : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00342         );
00343 }
00344 
00345 
00346 #if 0
00347 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
00348 static inline void idct_col (int16_t * col, int offset)
00349 {
00350 /* multiplication - as implemented on mmx */
00351 #define F(c,x) (((c) * (x)) >> 16)
00352 
00353 /* saturation - it helps us handle torture test cases */
00354 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00355 
00356     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00357     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00358     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00359     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00360 
00361     col += offset;
00362 
00363     x0 = col[0*8];
00364     x1 = col[1*8];
00365     x2 = col[2*8];
00366     x3 = col[3*8];
00367     x4 = col[4*8];
00368     x5 = col[5*8];
00369     x6 = col[6*8];
00370     x7 = col[7*8];
00371 
00372     u04 = S (x0 + x4);
00373     v04 = S (x0 - x4);
00374     u26 = S (F (T2, x6) + x2);
00375     v26 = S (F (T2, x2) - x6);
00376 
00377     a0 = S (u04 + u26);
00378     a1 = S (v04 + v26);
00379     a2 = S (v04 - v26);
00380     a3 = S (u04 - u26);
00381 
00382     u17 = S (F (T1, x7) + x1);
00383     v17 = S (F (T1, x1) - x7);
00384     u35 = S (F (T3, x5) + x3);
00385     v35 = S (F (T3, x3) - x5);
00386 
00387     b0 = S (u17 + u35);
00388     b3 = S (v17 - v35);
00389     u12 = S (u17 - u35);
00390     v12 = S (v17 + v35);
00391     u12 = S (2 * F (C4, u12));
00392     v12 = S (2 * F (C4, v12));
00393     b1 = S (u12 + v12);
00394     b2 = S (u12 - v12);
00395 
00396     y0 = S (a0 + b0) >> COL_SHIFT;
00397     y1 = S (a1 + b1) >> COL_SHIFT;
00398     y2 = S (a2 + b2) >> COL_SHIFT;
00399     y3 = S (a3 + b3) >> COL_SHIFT;
00400 
00401     y4 = S (a3 - b3) >> COL_SHIFT;
00402     y5 = S (a2 - b2) >> COL_SHIFT;
00403     y6 = S (a1 - b1) >> COL_SHIFT;
00404     y7 = S (a0 - b0) >> COL_SHIFT;
00405 
00406     col[0*8] = y0;
00407     col[1*8] = y1;
00408     col[2*8] = y2;
00409     col[3*8] = y3;
00410     col[4*8] = y4;
00411     col[5*8] = y5;
00412     col[6*8] = y6;
00413     col[7*8] = y7;
00414 }
00415 #endif
00416 
00417 
00418 /* MMX column IDCT */
00419 static inline void idct_col (int16_t * const col, const int offset)
00420 {
00421 #define T1 13036
00422 #define T2 27146
00423 #define T3 43790
00424 #define C4 23170
00425 
00426     DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
00427         T1,T1,T1,T1,
00428         T2,T2,T2,T2,
00429         T3,T3,T3,T3,
00430         C4,C4,C4,C4
00431     };
00432 
00433     /* column code adapted from Peter Gubanov */
00434     /* http://www.elecard.com/peter/idct.shtml */
00435 
00436     __asm__ volatile (
00437         "movq      (%0), %%mm0    \n\t" /* mm0 = T1 */
00438 
00439         "movq   2*8(%1), %%mm1    \n\t" /* mm1 = x1 */
00440         "movq     %%mm0, %%mm2    \n\t" /* mm2 = T1 */
00441 
00442         "movq 7*2*8(%1), %%mm4    \n\t" /* mm4 = x7 */
00443         "pmulhw   %%mm1, %%mm0    \n\t" /* mm0 = T1*x1 */
00444 
00445         "movq    16(%0), %%mm5    \n\t" /* mm5 = T3 */
00446         "pmulhw   %%mm4, %%mm2    \n\t" /* mm2 = T1*x7 */
00447 
00448         "movq 2*5*8(%1), %%mm6    \n\t" /* mm6 = x5 */
00449         "movq     %%mm5, %%mm7    \n\t" /* mm7 = T3-1 */
00450 
00451         "movq 3*8*2(%1), %%mm3    \n\t" /* mm3 = x3 */
00452         "psubsw   %%mm4, %%mm0    \n\t" /* mm0 = v17 */
00453 
00454         "movq     8(%0), %%mm4    \n\t" /* mm4 = T2 */
00455         "pmulhw   %%mm3, %%mm5    \n\t" /* mm5 = (T3-1)*x3 */
00456 
00457         "paddsw   %%mm2, %%mm1    \n\t" /* mm1 = u17 */
00458         "pmulhw   %%mm6, %%mm7    \n\t" /* mm7 = (T3-1)*x5 */
00459 
00460         /* slot */
00461 
00462         "movq     %%mm4, %%mm2    \n\t" /* mm2 = T2 */
00463         "paddsw   %%mm3, %%mm5    \n\t" /* mm5 = T3*x3 */
00464 
00465         "pmulhw 2*8*2(%1), %%mm4  \n\t" /* mm4 = T2*x2 */
00466         "paddsw   %%mm6, %%mm7    \n\t" /* mm7 = T3*x5 */
00467 
00468         "psubsw   %%mm6, %%mm5    \n\t" /* mm5 = v35 */
00469         "paddsw   %%mm3, %%mm7    \n\t" /* mm7 = u35 */
00470 
00471         "movq 6*8*2(%1), %%mm3    \n\t" /* mm3 = x6 */
00472         "movq     %%mm0, %%mm6    \n\t" /* mm6 = v17 */
00473 
00474         "pmulhw   %%mm3, %%mm2    \n\t" /* mm2 = T2*x6 */
00475         "psubsw   %%mm5, %%mm0    \n\t" /* mm0 = b3 */
00476 
00477         "psubsw   %%mm3, %%mm4    \n\t" /* mm4 = v26 */
00478         "paddsw   %%mm6, %%mm5    \n\t" /* mm5 = v12 */
00479 
00480         "movq     %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
00481         "movq     %%mm1, %%mm6    \n\t" /* mm6 = u17 */
00482 
00483         "paddsw 2*8*2(%1), %%mm2  \n\t" /* mm2 = u26 */
00484         "paddsw   %%mm7, %%mm6    \n\t" /* mm6 = b0 */
00485 
00486         "psubsw   %%mm7, %%mm1    \n\t" /* mm1 = u12 */
00487         "movq     %%mm1, %%mm7    \n\t" /* mm7 = u12 */
00488 
00489         "movq   0*8(%1), %%mm3    \n\t" /* mm3 = x0 */
00490         "paddsw   %%mm5, %%mm1    \n\t" /* mm1 = u12+v12 */
00491 
00492         "movq    24(%0), %%mm0    \n\t" /* mm0 = C4/2 */
00493         "psubsw   %%mm5, %%mm7    \n\t" /* mm7 = u12-v12 */
00494 
00495         "movq     %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
00496         "pmulhw   %%mm0, %%mm1    \n\t" /* mm1 = b1/2 */
00497 
00498         "movq     %%mm4, %%mm6    \n\t" /* mm6 = v26 */
00499         "pmulhw   %%mm0, %%mm7    \n\t" /* mm7 = b2/2 */
00500 
00501         "movq 4*8*2(%1), %%mm5    \n\t" /* mm5 = x4 */
00502         "movq     %%mm3, %%mm0    \n\t" /* mm0 = x0 */
00503 
00504         "psubsw   %%mm5, %%mm3    \n\t" /* mm3 = v04 */
00505         "paddsw   %%mm5, %%mm0    \n\t" /* mm0 = u04 */
00506 
00507         "paddsw   %%mm3, %%mm4    \n\t" /* mm4 = a1 */
00508         "movq     %%mm0, %%mm5    \n\t" /* mm5 = u04 */
00509 
00510         "psubsw   %%mm6, %%mm3    \n\t" /* mm3 = a2 */
00511         "paddsw   %%mm2, %%mm5    \n\t" /* mm5 = a0 */
00512 
00513         "paddsw   %%mm1, %%mm1    \n\t" /* mm1 = b1 */
00514         "psubsw   %%mm2, %%mm0    \n\t" /* mm0 = a3 */
00515 
00516         "paddsw   %%mm7, %%mm7    \n\t" /* mm7 = b2 */
00517         "movq     %%mm3, %%mm2    \n\t" /* mm2 = a2 */
00518 
00519         "movq     %%mm4, %%mm6    \n\t" /* mm6 = a1 */
00520         "paddsw   %%mm7, %%mm3    \n\t" /* mm3 = a2+b2 */
00521 
00522         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
00523         "paddsw   %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
00524 
00525         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
00526         "psubsw   %%mm1, %%mm6    \n\t" /* mm6 = a1-b1 */
00527 
00528         "movq 5*8*2(%1), %%mm1    \n\t" /* mm1 = b0 */
00529         "psubsw   %%mm7, %%mm2    \n\t" /* mm2 = a2-b2 */
00530 
00531         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
00532         "movq     %%mm5, %%mm7    \n\t" /* mm7 = a0 */
00533 
00534         "movq     %%mm4, 1*8*2(%1)\n\t" /* save y1 */
00535         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
00536 
00537         "movq     %%mm3, 2*8*2(%1)\n\t" /* save y2 */
00538         "paddsw   %%mm1, %%mm5    \n\t" /* mm5 = a0+b0 */
00539 
00540         "movq 3*8*2(%1), %%mm4    \n\t" /* mm4 = b3 */
00541         "psubsw   %%mm1, %%mm7    \n\t" /* mm7 = a0-b0 */
00542 
00543         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
00544         "movq     %%mm0, %%mm3    \n\t" /* mm3 = a3 */
00545 
00546         "movq     %%mm2, 5*8*2(%1)\n\t" /* save y5 */
00547         "psubsw   %%mm4, %%mm3    \n\t" /* mm3 = a3-b3 */
00548 
00549         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
00550         "paddsw   %%mm0, %%mm4    \n\t" /* mm4 = a3+b3 */
00551 
00552         "movq     %%mm5, 0*8*2(%1)\n\t" /* save y0 */
00553         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
00554 
00555         "movq     %%mm6, 6*8*2(%1)\n\t" /* save y6 */
00556         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
00557 
00558         "movq     %%mm7, 7*8*2(%1)\n\t" /* save y7 */
00559 
00560         "movq     %%mm3, 4*8*2(%1)\n\t" /* save y4 */
00561 
00562         "movq     %%mm4, 3*8*2(%1)\n\t" /* save y3 */
00563         :: "r" (t1_vector), "r" (col+offset)
00564         );
00565 
00566 #undef T1
00567 #undef T2
00568 #undef T3
00569 #undef C4
00570 }
00571 
00572 
00573 DECLARE_ALIGNED(8, static const int32_t, rounder0)[] =
00574     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
00575 DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0);
00576 DECLARE_ALIGNED(8, static const int32_t, rounder1)[] =
00577     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
00578 DECLARE_ALIGNED(8, static const int32_t, rounder7)[] =
00579     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
00580 DECLARE_ALIGNED(8, static const int32_t, rounder2)[] =
00581     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
00582 DECLARE_ALIGNED(8, static const int32_t, rounder6)[] =
00583     rounder (-0.25);                /* C2 * (C6-C2)/2 */
00584 DECLARE_ALIGNED(8, static const int32_t, rounder3)[] =
00585     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
00586 DECLARE_ALIGNED(8, static const int32_t, rounder5)[] =
00587     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
00588 
00589 #undef COL_SHIFT
00590 #undef ROW_SHIFT
00591 
00592 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
00593 void idct (int16_t * const block)                                       \
00594 {                                                                       \
00595     DECLARE_ALIGNED(16, static const int16_t, table04)[] =              \
00596         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00597     DECLARE_ALIGNED(16, static const int16_t, table17)[] =              \
00598         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00599     DECLARE_ALIGNED(16, static const int16_t, table26)[] =              \
00600         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00601     DECLARE_ALIGNED(16, static const int16_t, table35)[] =              \
00602         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00603                                                                         \
00604     idct_row_head (block, 0*8, table04);                                \
00605     idct_row (table04, rounder0);                                       \
00606     idct_row_mid (block, 0*8, 4*8, table04);                            \
00607     idct_row (table04, rounder4);                                       \
00608     idct_row_mid (block, 4*8, 1*8, table17);                            \
00609     idct_row (table17, rounder1);                                       \
00610     idct_row_mid (block, 1*8, 7*8, table17);                            \
00611     idct_row (table17, rounder7);                                       \
00612     idct_row_mid (block, 7*8, 2*8, table26);                            \
00613     idct_row (table26, rounder2);                                       \
00614     idct_row_mid (block, 2*8, 6*8, table26);                            \
00615     idct_row (table26, rounder6);                                       \
00616     idct_row_mid (block, 6*8, 3*8, table35);                            \
00617     idct_row (table35, rounder3);                                       \
00618     idct_row_mid (block, 3*8, 5*8, table35);                            \
00619     idct_row (table35, rounder5);                                       \
00620     idct_row_tail (block, 5*8);                                         \
00621                                                                         \
00622     idct_col (block, 0);                                                \
00623     idct_col (block, 4);                                                \
00624 }
00625 
00626 declare_idct (ff_mmxext_idct, mmxext_table,
00627               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00628 
00629 declare_idct (ff_mmx_idct, mmx_table,
00630               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00631 
00632 #endif /* HAVE_INLINE_ASM */
libavcodec/x86/idct_mmx.c
