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 "libavutil/x86_cpu.h"
00027 #include "dsputil_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     __asm__ volatile(
00091         "movq     (%0), %%mm2        \n\t"  /* mm2 = x6 x4 x2 x0 */
00092 
00093         "movq    8(%0), %%mm5        \n\t"  /* mm5 = x7 x5 x3 x1 */
00094         "movq    %%mm2, %%mm0        \n\t"  /* mm0 = x6 x4 x2 x0 */
00095 
00096         "movq     (%1), %%mm3        \n\t"  /* mm3 = -C2 -C4 C2 C4 */
00097         "movq    %%mm5, %%mm6        \n\t"  /* mm6 = x7 x5 x3 x1 */
00098 
00099         "movq    8(%1), %%mm4        \n\t"  /* mm4 = C6 C4 C6 C4 */
00100         "pmaddwd %%mm0, %%mm3        \n\t"  /* mm3 = -C4*x4-C2*x6 C4*x0+C2*x2 */
00101 
00102         "pshufw  $0x4e, %%mm2, %%mm2 \n\t"  /* mm2 = x2 x0 x6 x4 */
00103         :: "r" ((row+offset)), "r" (table)
00104     );
00105 }
00106 
00107 static inline void mmxext_row (const int16_t * const table,
00108                                const int32_t * const rounder)
00109 {
00110     __asm__ volatile (
00111         "movq    16(%0), %%mm1         \n\t" /* mm1 = -C5 -C1 C3 C1 */
00112         "pmaddwd  %%mm2, %%mm4         \n\t" /* mm4 = C4*x0+C6*x2 C4*x4+C6*x6 */
00113 
00114         "pmaddwd 32(%0), %%mm0         \n\t" /* mm0 = C4*x4-C6*x6 C4*x0-C6*x2 */
00115         "pshufw   $0x4e, %%mm6, %%mm6  \n\t" /* mm6 = x3 x1 x7 x5 */
00116 
00117         "movq    24(%0), %%mm7         \n\t" /* mm7 = -C7 C3 C7 C5 */
00118         "pmaddwd  %%mm5, %%mm1         \n\t" /* mm1= -C1*x5-C5*x7 C1*x1+C3*x3 */
00119 
00120         "paddd     (%1), %%mm3         \n\t" /* mm3 += rounder */
00121         "pmaddwd  %%mm6, %%mm7         \n\t" /* mm7 = C3*x1-C7*x3 C5*x5+C7*x7 */
00122 
00123         "pmaddwd 40(%0), %%mm2         \n\t" /* mm2= C4*x0-C2*x2 -C4*x4+C2*x6 */
00124         "paddd    %%mm4, %%mm3         \n\t" /* mm3 = a1 a0 + rounder */
00125 
00126         "pmaddwd 48(%0), %%mm5         \n\t" /* mm5 = C3*x5-C1*x7 C5*x1-C1*x3 */
00127         "movq     %%mm3, %%mm4         \n\t" /* mm4 = a1 a0 + rounder */
00128 
00129         "pmaddwd 56(%0), %%mm6         \n\t" /* mm6 = C7*x1-C5*x3 C7*x5+C3*x7 */
00130         "paddd    %%mm7, %%mm1         \n\t" /* mm1 = b1 b0 */
00131 
00132         "paddd     (%1), %%mm0         \n\t" /* mm0 += rounder */
00133         "psubd    %%mm1, %%mm3         \n\t" /* mm3 = a1-b1 a0-b0 + rounder */
00134 
00135         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3         \n\t" /* mm3 = y6 y7 */
00136         "paddd    %%mm4, %%mm1         \n\t" /* mm1 = a1+b1 a0+b0 + rounder */
00137 
00138         "paddd    %%mm2, %%mm0         \n\t" /* mm0 = a3 a2 + rounder */
00139         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1         \n\t" /* mm1 = y1 y0 */
00140 
00141         "paddd    %%mm6, %%mm5         \n\t" /* mm5 = b3 b2 */
00142         "movq     %%mm0, %%mm4         \n\t" /* mm4 = a3 a2 + rounder */
00143 
00144         "paddd    %%mm5, %%mm0         \n\t" /* mm0 = a3+b3 a2+b2 + rounder */
00145         "psubd    %%mm5, %%mm4         \n\t" /* mm4 = a3-b3 a2-b2 + rounder */
00146         : : "r" (table), "r" (rounder));
00147 }
00148 
00149 static inline void mmxext_row_tail (int16_t * const row, const int store)
00150 {
00151     __asm__ volatile (
00152         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0        \n\t"  /* mm0 = y3 y2 */
00153 
00154         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4  \n\t"  /* mm4 = y4 y5 */
00155 
00156         "packssdw %%mm0, %%mm1        \n\t"  /* mm1 = y3 y2 y1 y0 */
00157 
00158         "packssdw %%mm3, %%mm4        \n\t"  /* mm4 = y6 y7 y4 y5 */
00159 
00160         "movq     %%mm1, (%0)         \n\t"  /* save y3 y2 y1 y0 */
00161         "pshufw   $0xb1, %%mm4, %%mm4 \n\t"  /* mm4 = y7 y6 y5 y4 */
00162 
00163         /* slot */
00164 
00165         "movq     %%mm4, 8(%0)        \n\t"  /* save y7 y6 y5 y4 */
00166         :: "r" (row+store)
00167         );
00168 }
00169 
00170 static inline void mmxext_row_mid (int16_t * const row, const int store,
00171                                    const int offset,
00172                                    const int16_t * const table)
00173 {
00174     __asm__ volatile (
00175         "movq     (%0,%1), %%mm2       \n\t" /* mm2 = x6 x4 x2 x0 */
00176         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t"   /* mm0 = y3 y2 */
00177 
00178         "movq    8(%0,%1), %%mm5       \n\t" /* mm5 = x7 x5 x3 x1 */
00179         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm4 \n\t" /* mm4 = y4 y5 */
00180 
00181         "packssdw   %%mm0, %%mm1       \n\t" /* mm1 = y3 y2 y1 y0 */
00182         "movq       %%mm5, %%mm6       \n\t" /* mm6 = x7 x5 x3 x1 */
00183 
00184         "packssdw   %%mm3, %%mm4       \n\t" /* mm4 = y6 y7 y4 y5 */
00185         "movq       %%mm2, %%mm0       \n\t" /* mm0 = x6 x4 x2 x0 */
00186 
00187         "movq       %%mm1, (%0,%2)     \n\t" /* save y3 y2 y1 y0 */
00188         "pshufw     $0xb1, %%mm4, %%mm4\n\t" /* mm4 = y7 y6 y5 y4 */
00189 
00190         "movq        (%3), %%mm3       \n\t" /* mm3 = -C2 -C4 C2 C4 */
00191         "movq       %%mm4, 8(%0,%2)    \n\t" /* save y7 y6 y5 y4 */
00192 
00193         "pmaddwd    %%mm0, %%mm3       \n\t" /* mm3= -C4*x4-C2*x6 C4*x0+C2*x2 */
00194 
00195         "movq       8(%3), %%mm4       \n\t" /* mm4 = C6 C4 C6 C4 */
00196         "pshufw     $0x4e, %%mm2, %%mm2\n\t" /* mm2 = x2 x0 x6 x4 */
00197         :: "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00198         );
00199 }
00200 
00201 
00202 /* MMX row IDCT */
00203 
00204 #define mmx_table(c1,c2,c3,c4,c5,c6,c7) {  c4,  c2,  c4,  c6,   \
00205                                            c4,  c6, -c4, -c2,   \
00206                                            c1,  c3,  c3, -c7,   \
00207                                            c5,  c7, -c1, -c5,   \
00208                                            c4, -c6,  c4, -c2,   \
00209                                           -c4,  c2,  c4, -c6,   \
00210                                            c5, -c1,  c7, -c5,   \
00211                                            c7,  c3,  c3, -c1 }
00212 
00213 static inline void mmx_row_head (int16_t * const row, const int offset,
00214                                  const int16_t * const table)
00215 {
00216     __asm__ volatile (
00217         "movq (%0), %%mm2       \n\t"    /* mm2 = x6 x4 x2 x0 */
00218 
00219         "movq 8(%0), %%mm5      \n\t"    /* mm5 = x7 x5 x3 x1 */
00220         "movq %%mm2, %%mm0      \n\t"    /* mm0 = x6 x4 x2 x0 */
00221 
00222         "movq (%1), %%mm3       \n\t"    /* mm3 = C6 C4 C2 C4 */
00223         "movq %%mm5, %%mm6      \n\t"    /* mm6 = x7 x5 x3 x1 */
00224 
00225         "punpckldq %%mm0, %%mm0 \n\t"    /* mm0 = x2 x0 x2 x0 */
00226 
00227         "movq 8(%1), %%mm4      \n\t"    /* mm4 = -C2 -C4 C6 C4 */
00228         "pmaddwd %%mm0, %%mm3   \n\t"    /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00229 
00230         "movq 16(%1), %%mm1     \n\t"    /* mm1 = -C7 C3 C3 C1 */
00231         "punpckhdq %%mm2, %%mm2 \n\t"    /* mm2 = x6 x4 x6 x4 */
00232         :: "r" ((row+offset)), "r" (table)
00233         );
00234 }
00235 
00236 static inline void mmx_row (const int16_t * const table,
00237                             const int32_t * const rounder)
00238 {
00239     __asm__ volatile (
00240         "pmaddwd   %%mm2, %%mm4    \n\t"  /* mm4 = -C4*x4-C2*x6 C4*x4+C6*x6 */
00241         "punpckldq %%mm5, %%mm5    \n\t"  /* mm5 = x3 x1 x3 x1 */
00242 
00243         "pmaddwd  32(%0), %%mm0    \n\t"  /* mm0 = C4*x0-C2*x2 C4*x0-C6*x2 */
00244         "punpckhdq %%mm6, %%mm6    \n\t"  /* mm6 = x7 x5 x7 x5 */
00245 
00246         "movq     24(%0), %%mm7    \n\t"  /* mm7 = -C5 -C1 C7 C5 */
00247         "pmaddwd   %%mm5, %%mm1    \n\t"  /* mm1 = C3*x1-C7*x3 C1*x1+C3*x3 */
00248 
00249         "paddd      (%1), %%mm3    \n\t"  /* mm3 += rounder */
00250         "pmaddwd   %%mm6, %%mm7    \n\t"  /* mm7 = -C1*x5-C5*x7 C5*x5+C7*x7 */
00251 
00252         "pmaddwd  40(%0), %%mm2    \n\t"  /* mm2 = C4*x4-C6*x6 -C4*x4+C2*x6 */
00253         "paddd     %%mm4, %%mm3    \n\t"  /* mm3 = a1 a0 + rounder */
00254 
00255         "pmaddwd  48(%0), %%mm5    \n\t"  /* mm5 = C7*x1-C5*x3 C5*x1-C1*x3 */
00256         "movq      %%mm3, %%mm4    \n\t"  /* mm4 = a1 a0 + rounder */
00257 
00258         "pmaddwd  56(%0), %%mm6    \n\t"  /* mm6 = C3*x5-C1*x7 C7*x5+C3*x7 */
00259         "paddd     %%mm7, %%mm1    \n\t"  /* mm1 = b1 b0 */
00260 
00261         "paddd      (%1), %%mm0    \n\t"  /* mm0 += rounder */
00262         "psubd     %%mm1, %%mm3    \n\t"  /* mm3 = a1-b1 a0-b0 + rounder */
00263 
00264         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm3    \n\t"  /* mm3 = y6 y7 */
00265         "paddd     %%mm4, %%mm1    \n\t"  /* mm1 = a1+b1 a0+b0 + rounder */
00266 
00267         "paddd     %%mm2, %%mm0    \n\t"  /* mm0 = a3 a2 + rounder */
00268         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm1    \n\t"  /* mm1 = y1 y0 */
00269 
00270         "paddd     %%mm6, %%mm5    \n\t"  /* mm5 = b3 b2 */
00271         "movq      %%mm0, %%mm7    \n\t"  /* mm7 = a3 a2 + rounder */
00272 
00273         "paddd     %%mm5, %%mm0    \n\t"  /* mm0 = a3+b3 a2+b2 + rounder */
00274         "psubd     %%mm5, %%mm7    \n\t"  /* mm7 = a3-b3 a2-b2 + rounder */
00275         :: "r" (table), "r" (rounder)
00276         );
00277 }
00278 
00279 static inline void mmx_row_tail (int16_t * const row, const int store)
00280 {
00281     __asm__ volatile (
00282         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0      \n\t" /* mm0 = y3 y2 */
00283 
00284         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7      \n\t" /* mm7 = y4 y5 */
00285 
00286         "packssdw %%mm0, %%mm1 \n\t" /* mm1 = y3 y2 y1 y0 */
00287 
00288         "packssdw %%mm3, %%mm7 \n\t" /* mm7 = y6 y7 y4 y5 */
00289 
00290         "movq %%mm1, (%0)      \n\t" /* save y3 y2 y1 y0 */
00291         "movq %%mm7, %%mm4     \n\t" /* mm4 = y6 y7 y4 y5 */
00292 
00293         "pslld $16, %%mm7      \n\t" /* mm7 = y7 0 y5 0 */
00294 
00295         "psrld $16, %%mm4      \n\t" /* mm4 = 0 y6 0 y4 */
00296 
00297         "por %%mm4, %%mm7      \n\t" /* mm7 = y7 y6 y5 y4 */
00298 
00299         /* slot */
00300 
00301         "movq %%mm7, 8(%0)     \n\t" /* save y7 y6 y5 y4 */
00302         :: "r" (row+store)
00303         );
00304 }
00305 
00306 static inline void mmx_row_mid (int16_t * const row, const int store,
00307                                 const int offset, const int16_t * const table)
00308 {
00309 
00310     __asm__ volatile (
00311         "movq    (%0,%1), %%mm2    \n\t" /* mm2 = x6 x4 x2 x0 */
00312         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm0 \n\t" /* mm0 = y3 y2 */
00313 
00314         "movq   8(%0,%1), %%mm5    \n\t" /* mm5 = x7 x5 x3 x1 */
00315         "psrad $" AV_STRINGIFY(ROW_SHIFT) ", %%mm7 \n\t" /* mm7 = y4 y5 */
00316 
00317         "packssdw  %%mm0, %%mm1    \n\t" /* mm1 = y3 y2 y1 y0 */
00318         "movq      %%mm5, %%mm6    \n\t" /* mm6 = x7 x5 x3 x1 */
00319 
00320         "packssdw  %%mm3, %%mm7    \n\t" /* mm7 = y6 y7 y4 y5 */
00321         "movq      %%mm2, %%mm0    \n\t" /* mm0 = x6 x4 x2 x0 */
00322 
00323         "movq      %%mm1, (%0,%2)  \n\t" /* save y3 y2 y1 y0 */
00324         "movq      %%mm7, %%mm1    \n\t" /* mm1 = y6 y7 y4 y5 */
00325 
00326         "punpckldq %%mm0, %%mm0    \n\t" /* mm0 = x2 x0 x2 x0 */
00327         "psrld       $16, %%mm7    \n\t" /* mm7 = 0 y6 0 y4 */
00328 
00329         "movq       (%3), %%mm3    \n\t" /* mm3 = C6 C4 C2 C4 */
00330         "pslld       $16, %%mm1    \n\t" /* mm1 = y7 0 y5 0 */
00331 
00332         "movq      8(%3), %%mm4    \n\t" /* mm4 = -C2 -C4 C6 C4 */
00333         "por       %%mm1, %%mm7    \n\t" /* mm7 = y7 y6 y5 y4 */
00334 
00335         "movq     16(%3), %%mm1    \n\t" /* mm1 = -C7 C3 C3 C1 */
00336         "punpckhdq %%mm2, %%mm2    \n\t" /* mm2 = x6 x4 x6 x4 */
00337 
00338         "movq      %%mm7, 8(%0,%2) \n\t" /* save y7 y6 y5 y4 */
00339         "pmaddwd   %%mm0, %%mm3    \n\t" /* mm3 = C4*x0+C6*x2 C4*x0+C2*x2 */
00340         : : "r" (row), "r" ((x86_reg) (2*offset)), "r" ((x86_reg) (2*store)), "r" (table)
00341         );
00342 }
00343 
00344 
00345 #if 0
00346 /* C column IDCT - it is just here to document the MMXEXT and MMX versions */
00347 static inline void idct_col (int16_t * col, int offset)
00348 {
00349 /* multiplication - as implemented on mmx */
00350 #define F(c,x) (((c) * (x)) >> 16)
00351 
00352 /* saturation - it helps us handle torture test cases */
00353 #define S(x) (((x)>32767) ? 32767 : ((x)<-32768) ? -32768 : (x))
00354 
00355     int16_t x0, x1, x2, x3, x4, x5, x6, x7;
00356     int16_t y0, y1, y2, y3, y4, y5, y6, y7;
00357     int16_t a0, a1, a2, a3, b0, b1, b2, b3;
00358     int16_t u04, v04, u26, v26, u17, v17, u35, v35, u12, v12;
00359 
00360     col += offset;
00361 
00362     x0 = col[0*8];
00363     x1 = col[1*8];
00364     x2 = col[2*8];
00365     x3 = col[3*8];
00366     x4 = col[4*8];
00367     x5 = col[5*8];
00368     x6 = col[6*8];
00369     x7 = col[7*8];
00370 
00371     u04 = S (x0 + x4);
00372     v04 = S (x0 - x4);
00373     u26 = S (F (T2, x6) + x2);
00374     v26 = S (F (T2, x2) - x6);
00375 
00376     a0 = S (u04 + u26);
00377     a1 = S (v04 + v26);
00378     a2 = S (v04 - v26);
00379     a3 = S (u04 - u26);
00380 
00381     u17 = S (F (T1, x7) + x1);
00382     v17 = S (F (T1, x1) - x7);
00383     u35 = S (F (T3, x5) + x3);
00384     v35 = S (F (T3, x3) - x5);
00385 
00386     b0 = S (u17 + u35);
00387     b3 = S (v17 - v35);
00388     u12 = S (u17 - u35);
00389     v12 = S (v17 + v35);
00390     u12 = S (2 * F (C4, u12));
00391     v12 = S (2 * F (C4, v12));
00392     b1 = S (u12 + v12);
00393     b2 = S (u12 - v12);
00394 
00395     y0 = S (a0 + b0) >> COL_SHIFT;
00396     y1 = S (a1 + b1) >> COL_SHIFT;
00397     y2 = S (a2 + b2) >> COL_SHIFT;
00398     y3 = S (a3 + b3) >> COL_SHIFT;
00399 
00400     y4 = S (a3 - b3) >> COL_SHIFT;
00401     y5 = S (a2 - b2) >> COL_SHIFT;
00402     y6 = S (a1 - b1) >> COL_SHIFT;
00403     y7 = S (a0 - b0) >> COL_SHIFT;
00404 
00405     col[0*8] = y0;
00406     col[1*8] = y1;
00407     col[2*8] = y2;
00408     col[3*8] = y3;
00409     col[4*8] = y4;
00410     col[5*8] = y5;
00411     col[6*8] = y6;
00412     col[7*8] = y7;
00413 }
00414 #endif
00415 
00416 
00417 /* MMX column IDCT */
00418 static inline void idct_col (int16_t * const col, const int offset)
00419 {
00420 #define T1 13036
00421 #define T2 27146
00422 #define T3 43790
00423 #define C4 23170
00424 
00425     DECLARE_ALIGNED(8, static const short, t1_vector)[] = {
00426         T1,T1,T1,T1,
00427         T2,T2,T2,T2,
00428         T3,T3,T3,T3,
00429         C4,C4,C4,C4
00430     };
00431 
00432     /* column code adapted from Peter Gubanov */
00433     /* http://www.elecard.com/peter/idct.shtml */
00434 
00435     __asm__ volatile (
00436         "movq      (%0), %%mm0    \n\t" /* mm0 = T1 */
00437 
00438         "movq   2*8(%1), %%mm1    \n\t" /* mm1 = x1 */
00439         "movq     %%mm0, %%mm2    \n\t" /* mm2 = T1 */
00440 
00441         "movq 7*2*8(%1), %%mm4    \n\t" /* mm4 = x7 */
00442         "pmulhw   %%mm1, %%mm0    \n\t" /* mm0 = T1*x1 */
00443 
00444         "movq    16(%0), %%mm5    \n\t" /* mm5 = T3 */
00445         "pmulhw   %%mm4, %%mm2    \n\t" /* mm2 = T1*x7 */
00446 
00447         "movq 2*5*8(%1), %%mm6    \n\t" /* mm6 = x5 */
00448         "movq     %%mm5, %%mm7    \n\t" /* mm7 = T3-1 */
00449 
00450         "movq 3*8*2(%1), %%mm3    \n\t" /* mm3 = x3 */
00451         "psubsw   %%mm4, %%mm0    \n\t" /* mm0 = v17 */
00452 
00453         "movq     8(%0), %%mm4    \n\t" /* mm4 = T2 */
00454         "pmulhw   %%mm3, %%mm5    \n\t" /* mm5 = (T3-1)*x3 */
00455 
00456         "paddsw   %%mm2, %%mm1    \n\t" /* mm1 = u17 */
00457         "pmulhw   %%mm6, %%mm7    \n\t" /* mm7 = (T3-1)*x5 */
00458 
00459         /* slot */
00460 
00461         "movq     %%mm4, %%mm2    \n\t" /* mm2 = T2 */
00462         "paddsw   %%mm3, %%mm5    \n\t" /* mm5 = T3*x3 */
00463 
00464         "pmulhw 2*8*2(%1), %%mm4  \n\t" /* mm4 = T2*x2 */
00465         "paddsw   %%mm6, %%mm7    \n\t" /* mm7 = T3*x5 */
00466 
00467         "psubsw   %%mm6, %%mm5    \n\t" /* mm5 = v35 */
00468         "paddsw   %%mm3, %%mm7    \n\t" /* mm7 = u35 */
00469 
00470         "movq 6*8*2(%1), %%mm3    \n\t" /* mm3 = x6 */
00471         "movq     %%mm0, %%mm6    \n\t" /* mm6 = v17 */
00472 
00473         "pmulhw   %%mm3, %%mm2    \n\t" /* mm2 = T2*x6 */
00474         "psubsw   %%mm5, %%mm0    \n\t" /* mm0 = b3 */
00475 
00476         "psubsw   %%mm3, %%mm4    \n\t" /* mm4 = v26 */
00477         "paddsw   %%mm6, %%mm5    \n\t" /* mm5 = v12 */
00478 
00479         "movq     %%mm0, 3*8*2(%1)\n\t" /* save b3 in scratch0 */
00480         "movq     %%mm1, %%mm6    \n\t" /* mm6 = u17 */
00481 
00482         "paddsw 2*8*2(%1), %%mm2  \n\t" /* mm2 = u26 */
00483         "paddsw   %%mm7, %%mm6    \n\t" /* mm6 = b0 */
00484 
00485         "psubsw   %%mm7, %%mm1    \n\t" /* mm1 = u12 */
00486         "movq     %%mm1, %%mm7    \n\t" /* mm7 = u12 */
00487 
00488         "movq   0*8(%1), %%mm3    \n\t" /* mm3 = x0 */
00489         "paddsw   %%mm5, %%mm1    \n\t" /* mm1 = u12+v12 */
00490 
00491         "movq    24(%0), %%mm0    \n\t" /* mm0 = C4/2 */
00492         "psubsw   %%mm5, %%mm7    \n\t" /* mm7 = u12-v12 */
00493 
00494         "movq     %%mm6, 5*8*2(%1)\n\t" /* save b0 in scratch1 */
00495         "pmulhw   %%mm0, %%mm1    \n\t" /* mm1 = b1/2 */
00496 
00497         "movq     %%mm4, %%mm6    \n\t" /* mm6 = v26 */
00498         "pmulhw   %%mm0, %%mm7    \n\t" /* mm7 = b2/2 */
00499 
00500         "movq 4*8*2(%1), %%mm5    \n\t" /* mm5 = x4 */
00501         "movq     %%mm3, %%mm0    \n\t" /* mm0 = x0 */
00502 
00503         "psubsw   %%mm5, %%mm3    \n\t" /* mm3 = v04 */
00504         "paddsw   %%mm5, %%mm0    \n\t" /* mm0 = u04 */
00505 
00506         "paddsw   %%mm3, %%mm4    \n\t" /* mm4 = a1 */
00507         "movq     %%mm0, %%mm5    \n\t" /* mm5 = u04 */
00508 
00509         "psubsw   %%mm6, %%mm3    \n\t" /* mm3 = a2 */
00510         "paddsw   %%mm2, %%mm5    \n\t" /* mm5 = a0 */
00511 
00512         "paddsw   %%mm1, %%mm1    \n\t" /* mm1 = b1 */
00513         "psubsw   %%mm2, %%mm0    \n\t" /* mm0 = a3 */
00514 
00515         "paddsw   %%mm7, %%mm7    \n\t" /* mm7 = b2 */
00516         "movq     %%mm3, %%mm2    \n\t" /* mm2 = a2 */
00517 
00518         "movq     %%mm4, %%mm6    \n\t" /* mm6 = a1 */
00519         "paddsw   %%mm7, %%mm3    \n\t" /* mm3 = a2+b2 */
00520 
00521         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y2 */
00522         "paddsw   %%mm1, %%mm4\n\t" /* mm4 = a1+b1 */
00523 
00524         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y1 */
00525         "psubsw   %%mm1, %%mm6    \n\t" /* mm6 = a1-b1 */
00526 
00527         "movq 5*8*2(%1), %%mm1    \n\t" /* mm1 = b0 */
00528         "psubsw   %%mm7, %%mm2    \n\t" /* mm2 = a2-b2 */
00529 
00530         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm6\n\t" /* mm6 = y6 */
00531         "movq     %%mm5, %%mm7    \n\t" /* mm7 = a0 */
00532 
00533         "movq     %%mm4, 1*8*2(%1)\n\t" /* save y1 */
00534         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm2\n\t" /* mm2 = y5 */
00535 
00536         "movq     %%mm3, 2*8*2(%1)\n\t" /* save y2 */
00537         "paddsw   %%mm1, %%mm5    \n\t" /* mm5 = a0+b0 */
00538 
00539         "movq 3*8*2(%1), %%mm4    \n\t" /* mm4 = b3 */
00540         "psubsw   %%mm1, %%mm7    \n\t" /* mm7 = a0-b0 */
00541 
00542         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm5\n\t" /* mm5 = y0 */
00543         "movq     %%mm0, %%mm3    \n\t" /* mm3 = a3 */
00544 
00545         "movq     %%mm2, 5*8*2(%1)\n\t" /* save y5 */
00546         "psubsw   %%mm4, %%mm3    \n\t" /* mm3 = a3-b3 */
00547 
00548         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm7\n\t" /* mm7 = y7 */
00549         "paddsw   %%mm0, %%mm4    \n\t" /* mm4 = a3+b3 */
00550 
00551         "movq     %%mm5, 0*8*2(%1)\n\t" /* save y0 */
00552         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm3\n\t" /* mm3 = y4 */
00553 
00554         "movq     %%mm6, 6*8*2(%1)\n\t" /* save y6 */
00555         "psraw $" AV_STRINGIFY(COL_SHIFT) ", %%mm4\n\t" /* mm4 = y3 */
00556 
00557         "movq     %%mm7, 7*8*2(%1)\n\t" /* save y7 */
00558 
00559         "movq     %%mm3, 4*8*2(%1)\n\t" /* save y4 */
00560 
00561         "movq     %%mm4, 3*8*2(%1)\n\t" /* save y3 */
00562         :: "r" (t1_vector), "r" (col+offset)
00563         );
00564 
00565 #undef T1
00566 #undef T2
00567 #undef T3
00568 #undef C4
00569 }
00570 
00571 
00572 DECLARE_ALIGNED(8, static const int32_t, rounder0)[] =
00573     rounder ((1 << (COL_SHIFT - 1)) - 0.5);
00574 DECLARE_ALIGNED(8, static const int32_t, rounder4)[] = rounder (0);
00575 DECLARE_ALIGNED(8, static const int32_t, rounder1)[] =
00576     rounder (1.25683487303);        /* C1*(C1/C4+C1+C7)/2 */
00577 DECLARE_ALIGNED(8, static const int32_t, rounder7)[] =
00578     rounder (-0.25);                /* C1*(C7/C4+C7-C1)/2 */
00579 DECLARE_ALIGNED(8, static const int32_t, rounder2)[] =
00580     rounder (0.60355339059);        /* C2 * (C6+C2)/2 */
00581 DECLARE_ALIGNED(8, static const int32_t, rounder6)[] =
00582     rounder (-0.25);                /* C2 * (C6-C2)/2 */
00583 DECLARE_ALIGNED(8, static const int32_t, rounder3)[] =
00584     rounder (0.087788325588);       /* C3*(-C3/C4+C3+C5)/2 */
00585 DECLARE_ALIGNED(8, static const int32_t, rounder5)[] =
00586     rounder (-0.441341716183);      /* C3*(-C5/C4+C5-C3)/2 */
00587 
00588 #undef COL_SHIFT
00589 #undef ROW_SHIFT
00590 
00591 #define declare_idct(idct,table,idct_row_head,idct_row,idct_row_tail,idct_row_mid) \
00592 void idct (int16_t * const block)                                       \
00593 {                                                                       \
00594     DECLARE_ALIGNED(16, static const int16_t, table04)[] =              \
00595         table (22725, 21407, 19266, 16384, 12873,  8867, 4520);         \
00596     DECLARE_ALIGNED(16, static const int16_t, table17)[] =              \
00597         table (31521, 29692, 26722, 22725, 17855, 12299, 6270);         \
00598     DECLARE_ALIGNED(16, static const int16_t, table26)[] =              \
00599         table (29692, 27969, 25172, 21407, 16819, 11585, 5906);         \
00600     DECLARE_ALIGNED(16, static const int16_t, table35)[] =              \
00601         table (26722, 25172, 22654, 19266, 15137, 10426, 5315);         \
00602                                                                         \
00603     idct_row_head (block, 0*8, table04);                                \
00604     idct_row (table04, rounder0);                                       \
00605     idct_row_mid (block, 0*8, 4*8, table04);                            \
00606     idct_row (table04, rounder4);                                       \
00607     idct_row_mid (block, 4*8, 1*8, table17);                            \
00608     idct_row (table17, rounder1);                                       \
00609     idct_row_mid (block, 1*8, 7*8, table17);                            \
00610     idct_row (table17, rounder7);                                       \
00611     idct_row_mid (block, 7*8, 2*8, table26);                            \
00612     idct_row (table26, rounder2);                                       \
00613     idct_row_mid (block, 2*8, 6*8, table26);                            \
00614     idct_row (table26, rounder6);                                       \
00615     idct_row_mid (block, 6*8, 3*8, table35);                            \
00616     idct_row (table35, rounder3);                                       \
00617     idct_row_mid (block, 3*8, 5*8, table35);                            \
00618     idct_row (table35, rounder5);                                       \
00619     idct_row_tail (block, 5*8);                                         \
00620                                                                         \
00621     idct_col (block, 0);                                                \
00622     idct_col (block, 4);                                                \
00623 }
00624 
00625 declare_idct (ff_mmxext_idct, mmxext_table,
00626               mmxext_row_head, mmxext_row, mmxext_row_tail, mmxext_row_mid)
00627 
00628 declare_idct (ff_mmx_idct, mmx_table,
00629               mmx_row_head, mmx_row, mmx_row_tail, mmx_row_mid)
00630
libavcodec/x86/idct_mmx.c
