FFmpeg: libavcodec/x86/idct_mmx

00001 /*
00002  * XVID MPEG-4 VIDEO CODEC
00003  * - MMX and XMM forward discrete cosine transform -
00004  *
00005  * Copyright(C) 2001 Peter Ross <pross@xvid.org>
00006  *
00007  * Originally provided by Intel at AP-922
00008  * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
00009  * (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
00010  * but in a limited edition.
00011  * New macro implements a column part for precise iDCT
00012  * The routine precision now satisfies IEEE standard 1180-1990.
00013  *
00014  * Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
00015  * Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org>
00016  *
00017  * http://www.elecard.com/peter/idct.html
00018  * http://www.linuxvideo.org/mpeg2dec/
00019  *
00020  * These examples contain code fragments for first stage iDCT 8x8
00021  * (for rows) and first stage DCT 8x8 (for columns)
00022  *
00023  * conversion to gcc syntax by Michael Niedermayer
00024  *
00025  * This file is part of FFmpeg.
00026  *
00027  * FFmpeg is free software; you can redistribute it and/or
00028  * modify it under the terms of the GNU Lesser General Public
00029  * License as published by the Free Software Foundation; either
00030  * version 2.1 of the License, or (at your option) any later version.
00031  *
00032  * FFmpeg is distributed in the hope that it will be useful,
00033  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00034  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00035  * Lesser General Public License for more details.
00036  *
00037  * You should have received a copy of the GNU Lesser General Public License
00038  * along with FFmpeg; if not, write to the Free Software Foundation,
00039  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00040  */
00041 
00042 #include <inttypes.h>
00043 
00044 #include "config.h"
00045 #include "libavcodec/avcodec.h"
00046 #include "libavutil/mem.h"
00047 #include "dsputil_mmx.h"
00048 #include "idct_xvid.h"
00049 
00050 #if HAVE_INLINE_ASM
00051 
00052 //=============================================================================
00053 // Macros and other preprocessor constants
00054 //=============================================================================
00055 
00056 #define BITS_INV_ACC    5                              // 4 or 5 for IEEE
00057 #define SHIFT_INV_ROW   (16 - BITS_INV_ACC) //11
00058 #define SHIFT_INV_COL   (1 + BITS_INV_ACC) //6
00059 #define RND_INV_ROW     (1024 * (6 - BITS_INV_ACC))
00060 #define RND_INV_COL     (16 * (BITS_INV_ACC - 3))
00061 #define RND_INV_CORR    (RND_INV_COL - 1)
00062 
00063 #define BITS_FRW_ACC    3                              // 2 or 3 for accuracy
00064 #define SHIFT_FRW_COL   BITS_FRW_ACC
00065 #define SHIFT_FRW_ROW   (BITS_FRW_ACC + 17)
00066 #define RND_FRW_ROW     (262144*(BITS_FRW_ACC - 1))
00067 
00068 
00069 //-----------------------------------------------------------------------------
00070 // Various memory constants (trigonometric values or rounding values)
00071 //-----------------------------------------------------------------------------
00072 
00073 
00074 DECLARE_ALIGNED(8, static const int16_t, tg_1_16)[4*4] = {
00075   13036,13036,13036,13036,        // tg * (2<<16) + 0.5
00076   27146,27146,27146,27146,        // tg * (2<<16) + 0.5
00077   -21746,-21746,-21746,-21746,    // tg * (2<<16) + 0.5
00078   23170,23170,23170,23170};       // cos * (2<<15) + 0.5
00079 
00080 DECLARE_ALIGNED(8, static const int32_t, rounder_0)[2*8] = {
00081   65536,65536,
00082   3597,3597,
00083   2260,2260,
00084   1203,1203,
00085   0,0,
00086   120,120,
00087   512,512,
00088   512,512};
00089 
00090 //-----------------------------------------------------------------------------
00091 //
00092 // The first stage iDCT 8x8 - inverse DCTs of rows
00093 //
00094 //-----------------------------------------------------------------------------
00095 // The 8-point inverse DCT direct algorithm
00096 //-----------------------------------------------------------------------------
00097 //
00098 // static const short w[32] = {
00099 //       FIX(cos_4_16),  FIX(cos_2_16),  FIX(cos_4_16),  FIX(cos_6_16),
00100 //       FIX(cos_4_16),  FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16),
00101 //       FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16),  FIX(cos_2_16),
00102 //       FIX(cos_4_16), -FIX(cos_2_16),  FIX(cos_4_16), -FIX(cos_6_16),
00103 //       FIX(cos_1_16),  FIX(cos_3_16),  FIX(cos_5_16),  FIX(cos_7_16),
00104 //       FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16),
00105 //       FIX(cos_5_16), -FIX(cos_1_16),  FIX(cos_7_16),  FIX(cos_3_16),
00106 //       FIX(cos_7_16), -FIX(cos_5_16),  FIX(cos_3_16), -FIX(cos_1_16) };
00107 //
00108 // #define DCT_8_INV_ROW(x, y)
00109 // {
00110 //       int a0, a1, a2, a3, b0, b1, b2, b3;
00111 //
00112 //       a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3];
00113 //       a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7];
00114 //       a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11];
00115 //       a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15];
00116 //       b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19];
00117 //       b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23];
00118 //       b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27];
00119 //       b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31];
00120 //
00121 //       y[0] = SHIFT_ROUND ( a0 + b0 );
00122 //       y[1] = SHIFT_ROUND ( a1 + b1 );
00123 //       y[2] = SHIFT_ROUND ( a2 + b2 );
00124 //       y[3] = SHIFT_ROUND ( a3 + b3 );
00125 //       y[4] = SHIFT_ROUND ( a3 - b3 );
00126 //       y[5] = SHIFT_ROUND ( a2 - b2 );
00127 //       y[6] = SHIFT_ROUND ( a1 - b1 );
00128 //       y[7] = SHIFT_ROUND ( a0 - b0 );
00129 // }
00130 //
00131 //-----------------------------------------------------------------------------
00132 //
00133 // In this implementation the outputs of the iDCT-1D are multiplied
00134 //       for rows 0,4 - by cos_4_16,
00135 //       for rows 1,7 - by cos_1_16,
00136 //       for rows 2,6 - by cos_2_16,
00137 //       for rows 3,5 - by cos_3_16
00138 // and are shifted to the left for better accuracy
00139 //
00140 // For the constants used,
00141 //       FIX(float_const) = (short) (float_const * (1<<15) + 0.5)
00142 //
00143 //-----------------------------------------------------------------------------
00144 
00145 //-----------------------------------------------------------------------------
00146 // Tables for mmx processors
00147 //-----------------------------------------------------------------------------
00148 
00149 // Table for rows 0,4 - constants are multiplied by cos_4_16
00150 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_mmx)[32*4] = {
00151   16384,16384,16384,-16384,       // movq-> w06 w04 w02 w00
00152   21407,8867,8867,-21407,         // w07 w05 w03 w01
00153   16384,-16384,16384,16384,       // w14 w12 w10 w08
00154   -8867,21407,-21407,-8867,       // w15 w13 w11 w09
00155   22725,12873,19266,-22725,       // w22 w20 w18 w16
00156   19266,4520,-4520,-12873,        // w23 w21 w19 w17
00157   12873,4520,4520,19266,          // w30 w28 w26 w24
00158   -22725,19266,-12873,-22725,     // w31 w29 w27 w25
00159 // Table for rows 1,7 - constants are multiplied by cos_1_16
00160   22725,22725,22725,-22725,       // movq-> w06 w04 w02 w00
00161   29692,12299,12299,-29692,       // w07 w05 w03 w01
00162   22725,-22725,22725,22725,       // w14 w12 w10 w08
00163   -12299,29692,-29692,-12299,     // w15 w13 w11 w09
00164   31521,17855,26722,-31521,       // w22 w20 w18 w16
00165   26722,6270,-6270,-17855,        // w23 w21 w19 w17
00166   17855,6270,6270,26722,          // w30 w28 w26 w24
00167   -31521,26722,-17855,-31521,     // w31 w29 w27 w25
00168 // Table for rows 2,6 - constants are multiplied by cos_2_16
00169   21407,21407,21407,-21407,       // movq-> w06 w04 w02 w00
00170   27969,11585,11585,-27969,       // w07 w05 w03 w01
00171   21407,-21407,21407,21407,       // w14 w12 w10 w08
00172   -11585,27969,-27969,-11585,     // w15 w13 w11 w09
00173   29692,16819,25172,-29692,       // w22 w20 w18 w16
00174   25172,5906,-5906,-16819,        // w23 w21 w19 w17
00175   16819,5906,5906,25172,          // w30 w28 w26 w24
00176   -29692,25172,-16819,-29692,     // w31 w29 w27 w25
00177 // Table for rows 3,5 - constants are multiplied by cos_3_16
00178   19266,19266,19266,-19266,       // movq-> w06 w04 w02 w00
00179   25172,10426,10426,-25172,       // w07 w05 w03 w01
00180   19266,-19266,19266,19266,       // w14 w12 w10 w08
00181   -10426,25172,-25172,-10426,     // w15 w13 w11 w09
00182   26722,15137,22654,-26722,       // w22 w20 w18 w16
00183   22654,5315,-5315,-15137,        // w23 w21 w19 w17
00184   15137,5315,5315,22654,          // w30 w28 w26 w24
00185   -26722,22654,-15137,-26722,     // w31 w29 w27 w25
00186 };
00187 //-----------------------------------------------------------------------------
00188 // Tables for xmm processors
00189 //-----------------------------------------------------------------------------
00190 
00191 // %3 for rows 0,4 - constants are multiplied by cos_4_16
00192 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_xmm)[32*4] = {
00193   16384,21407,16384,8867,      // movq-> w05 w04 w01 w00
00194   16384,8867,-16384,-21407,    // w07 w06 w03 w02
00195   16384,-8867,16384,-21407,    // w13 w12 w09 w08
00196   -16384,21407,16384,-8867,    // w15 w14 w11 w10
00197   22725,19266,19266,-4520,     // w21 w20 w17 w16
00198   12873,4520,-22725,-12873,    // w23 w22 w19 w18
00199   12873,-22725,4520,-12873,    // w29 w28 w25 w24
00200   4520,19266,19266,-22725,     // w31 w30 w27 w26
00201 // %3 for rows 1,7 - constants are multiplied by cos_1_16
00202   22725,29692,22725,12299,     // movq-> w05 w04 w01 w00
00203   22725,12299,-22725,-29692,   // w07 w06 w03 w02
00204   22725,-12299,22725,-29692,   // w13 w12 w09 w08
00205   -22725,29692,22725,-12299,   // w15 w14 w11 w10
00206   31521,26722,26722,-6270,     // w21 w20 w17 w16
00207   17855,6270,-31521,-17855,    // w23 w22 w19 w18
00208   17855,-31521,6270,-17855,    // w29 w28 w25 w24
00209   6270,26722,26722,-31521,     // w31 w30 w27 w26
00210 // %3 for rows 2,6 - constants are multiplied by cos_2_16
00211   21407,27969,21407,11585,     // movq-> w05 w04 w01 w00
00212   21407,11585,-21407,-27969,   // w07 w06 w03 w02
00213   21407,-11585,21407,-27969,   // w13 w12 w09 w08
00214   -21407,27969,21407,-11585,   // w15 w14 w11 w10
00215   29692,25172,25172,-5906,     // w21 w20 w17 w16
00216   16819,5906,-29692,-16819,    // w23 w22 w19 w18
00217   16819,-29692,5906,-16819,    // w29 w28 w25 w24
00218   5906,25172,25172,-29692,     // w31 w30 w27 w26
00219 // %3 for rows 3,5 - constants are multiplied by cos_3_16
00220   19266,25172,19266,10426,     // movq-> w05 w04 w01 w00
00221   19266,10426,-19266,-25172,   // w07 w06 w03 w02
00222   19266,-10426,19266,-25172,   // w13 w12 w09 w08
00223   -19266,25172,19266,-10426,   // w15 w14 w11 w10
00224   26722,22654,22654,-5315,     // w21 w20 w17 w16
00225   15137,5315,-26722,-15137,    // w23 w22 w19 w18
00226   15137,-26722,5315,-15137,    // w29 w28 w25 w24
00227   5315,22654,22654,-26722,     // w31 w30 w27 w26
00228 };
00229 //=============================================================================
00230 // Helper macros for the code
00231 //=============================================================================
00232 
00233 //-----------------------------------------------------------------------------
00234 // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER
00235 //-----------------------------------------------------------------------------
00236 
00237 #define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\
00238   "movq " #A1 ",%%mm0              \n\t"/* 0 ; x3 x2 x1 x0*/\
00239   "movq 8+" #A1 ",%%mm1            \n\t"/* 1 ; x7 x6 x5 x4*/\
00240   "movq %%mm0,%%mm2            \n\t"/* 2  ; x3 x2 x1 x0*/\
00241   "movq " #A3 ",%%mm3              \n\t"/* 3 ; w06 w04 w02 w00*/\
00242   "punpcklwd %%mm1,%%mm0       \n\t"/* x5 x1 x4 x0*/\
00243   "movq %%mm0,%%mm5            \n\t"/* 5 ; x5 x1 x4 x0*/\
00244   "punpckldq %%mm0,%%mm0       \n\t"/* x4 x0 x4 x0*/\
00245   "movq 8+" #A3 ",%%mm4            \n\t"/* 4 ; w07 w05 w03 w01*/\
00246   "punpckhwd %%mm1,%%mm2       \n\t"/* 1 ; x7 x3 x6 x2*/\
00247   "pmaddwd %%mm0,%%mm3         \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\
00248   "movq %%mm2,%%mm6            \n\t"/* 6 ; x7 x3 x6 x2*/\
00249   "movq 32+" #A3 ",%%mm1           \n\t"/* 1 ; w22 w20 w18 w16*/\
00250   "punpckldq %%mm2,%%mm2       \n\t"/* x6 x2 x6 x2*/\
00251   "pmaddwd %%mm2,%%mm4         \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\
00252   "punpckhdq %%mm5,%%mm5       \n\t"/* x5 x1 x5 x1*/\
00253   "pmaddwd 16+" #A3 ",%%mm0        \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\
00254   "punpckhdq %%mm6,%%mm6       \n\t"/* x7 x3 x7 x3*/\
00255   "movq 40+" #A3 ",%%mm7           \n\t"/* 7 ; w23 w21 w19 w17*/\
00256   "pmaddwd %%mm5,%%mm1         \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\
00257   "paddd " #A4 ",%%mm3             \n\t"/* +%4*/\
00258   "pmaddwd %%mm6,%%mm7         \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\
00259   "pmaddwd 24+" #A3 ",%%mm2        \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\
00260   "paddd %%mm4,%%mm3           \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
00261   "pmaddwd 48+" #A3 ",%%mm5        \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\
00262   "movq %%mm3,%%mm4            \n\t"/* 4 ; a1 a0*/\
00263   "pmaddwd 56+" #A3 ",%%mm6        \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\
00264   "paddd %%mm7,%%mm1           \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
00265   "paddd " #A4 ",%%mm0             \n\t"/* +%4*/\
00266   "psubd %%mm1,%%mm3           \n\t"/* a1-b1 a0-b0*/\
00267   "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\
00268   "paddd %%mm4,%%mm1           \n\t"/* 4 ; a1+b1 a0+b0*/\
00269   "paddd %%mm2,%%mm0           \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\
00270   "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\
00271   "paddd %%mm6,%%mm5           \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
00272   "movq %%mm0,%%mm4            \n\t"/* 4 ; a3 a2*/\
00273   "paddd %%mm5,%%mm0           \n\t"/* a3+b3 a2+b2*/\
00274   "psubd %%mm5,%%mm4           \n\t"/* 5 ; a3-b3 a2-b2*/\
00275   "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
00276   "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
00277   "packssdw %%mm0,%%mm1        \n\t"/* 0 ; y3 y2 y1 y0*/\
00278   "packssdw %%mm3,%%mm4        \n\t"/* 3 ; y6 y7 y4 y5*/\
00279   "movq %%mm4,%%mm7            \n\t"/* 7 ; y6 y7 y4 y5*/\
00280   "psrld $16,%%mm4            \n\t"/* 0 y6 0 y4*/\
00281   "pslld $16,%%mm7            \n\t"/* y7 0 y5 0*/\
00282   "movq %%mm1," #A2 "              \n\t"/* 1 ; save y3 y2 y1 y0*/\
00283   "por %%mm4,%%mm7             \n\t"/* 4 ; y7 y6 y5 y4*/\
00284   "movq %%mm7,8            +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
00285 
00286 
00287 //-----------------------------------------------------------------------------
00288 // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
00289 //-----------------------------------------------------------------------------
00290 
00291 #define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
00292   "movq " #A1 ",%%mm0                  \n\t"/* 0     ; x3 x2 x1 x0*/\
00293   "movq 8+" #A1 ",%%mm1                \n\t"/* 1     ; x7 x6 x5 x4*/\
00294   "movq %%mm0,%%mm2                \n\t"/* 2     ; x3 x2 x1 x0*/\
00295   "movq " #A3 ",%%mm3                  \n\t"/* 3     ; w05 w04 w01 w00*/\
00296   "pshufw $0x88,%%mm0,%%mm0        \n\t"/* x2 x0 x2 x0*/\
00297   "movq 8+" #A3 ",%%mm4                \n\t"/* 4     ; w07 w06 w03 w02*/\
00298   "movq %%mm1,%%mm5                \n\t"/* 5     ; x7 x6 x5 x4*/\
00299   "pmaddwd %%mm0,%%mm3             \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
00300   "movq 32+" #A3 ",%%mm6               \n\t"/* 6     ; w21 w20 w17 w16*/\
00301   "pshufw $0x88,%%mm1,%%mm1        \n\t"/* x6 x4 x6 x4*/\
00302   "pmaddwd %%mm1,%%mm4             \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
00303   "movq 40+" #A3 ",%%mm7               \n\t"/* 7    ; w23 w22 w19 w18*/\
00304   "pshufw $0xdd,%%mm2,%%mm2        \n\t"/* x3 x1 x3 x1*/\
00305   "pmaddwd %%mm2,%%mm6             \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
00306   "pshufw $0xdd,%%mm5,%%mm5        \n\t"/* x7 x5 x7 x5*/\
00307   "pmaddwd %%mm5,%%mm7             \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
00308   "paddd " #A4 ",%%mm3                 \n\t"/* +%4*/\
00309   "pmaddwd 16+" #A3 ",%%mm0            \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
00310   "paddd %%mm4,%%mm3               \n\t"/* 4     ; a1=sum(even1) a0=sum(even0)*/\
00311   "pmaddwd 24+" #A3 ",%%mm1            \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
00312   "movq %%mm3,%%mm4                \n\t"/* 4     ; a1 a0*/\
00313   "pmaddwd 48+" #A3 ",%%mm2            \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
00314   "paddd %%mm7,%%mm6               \n\t"/* 7     ; b1=sum(odd1) b0=sum(odd0)*/\
00315   "pmaddwd 56+" #A3 ",%%mm5            \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
00316   "paddd %%mm6,%%mm3               \n\t"/* a1+b1 a0+b0*/\
00317   "paddd " #A4 ",%%mm0                 \n\t"/* +%4*/\
00318   "psrad $11,%%mm3     \n\t"/* y1=a1+b1 y0=a0+b0*/\
00319   "paddd %%mm1,%%mm0               \n\t"/* 1     ; a3=sum(even3) a2=sum(even2)*/\
00320   "psubd %%mm6,%%mm4               \n\t"/* 6     ; a1-b1 a0-b0*/\
00321   "movq %%mm0,%%mm7                \n\t"/* 7     ; a3 a2*/\
00322   "paddd %%mm5,%%mm2               \n\t"/* 5     ; b3=sum(odd3) b2=sum(odd2)*/\
00323   "paddd %%mm2,%%mm0               \n\t"/* a3+b3 a2+b2*/\
00324   "psrad $11,%%mm4     \n\t"/* y6=a1-b1 y7=a0-b0*/\
00325   "psubd %%mm2,%%mm7               \n\t"/* 2     ; a3-b3 a2-b2*/\
00326   "psrad $11,%%mm0     \n\t"/* y3=a3+b3 y2=a2+b2*/\
00327   "psrad $11,%%mm7     \n\t"/* y4=a3-b3 y5=a2-b2*/\
00328   "packssdw %%mm0,%%mm3            \n\t"/* 0     ; y3 y2 y1 y0*/\
00329   "packssdw %%mm4,%%mm7            \n\t"/* 4     ; y6 y7 y4 y5*/\
00330   "movq %%mm3, " #A2 "                  \n\t"/* 3     ; save y3 y2 y1 y0*/\
00331   "pshufw $0xb1,%%mm7,%%mm7        \n\t"/* y7 y6 y5 y4*/\
00332   "movq %%mm7,8                +" #A2 "\n\t"/* 7     ; save y7 y6 y5 y4*/\
00333 
00334 
00335 //-----------------------------------------------------------------------------
00336 //
00337 // The first stage DCT 8x8 - forward DCTs of columns
00338 //
00339 // The %2puts are multiplied
00340 // for rows 0,4 - on cos_4_16,
00341 // for rows 1,7 - on cos_1_16,
00342 // for rows 2,6 - on cos_2_16,
00343 // for rows 3,5 - on cos_3_16
00344 // and are shifted to the left for rise of accuracy
00345 //
00346 //-----------------------------------------------------------------------------
00347 //
00348 // The 8-point scaled forward DCT algorithm (26a8m)
00349 //
00350 //-----------------------------------------------------------------------------
00351 //
00352 // #define DCT_8_FRW_COL(x, y)
00353 //{
00354 // short t0, t1, t2, t3, t4, t5, t6, t7;
00355 // short tp03, tm03, tp12, tm12, tp65, tm65;
00356 // short tp465, tm465, tp765, tm765;
00357 //
00358 // t0 = LEFT_SHIFT ( x[0] + x[7] );
00359 // t1 = LEFT_SHIFT ( x[1] + x[6] );
00360 // t2 = LEFT_SHIFT ( x[2] + x[5] );
00361 // t3 = LEFT_SHIFT ( x[3] + x[4] );
00362 // t4 = LEFT_SHIFT ( x[3] - x[4] );
00363 // t5 = LEFT_SHIFT ( x[2] - x[5] );
00364 // t6 = LEFT_SHIFT ( x[1] - x[6] );
00365 // t7 = LEFT_SHIFT ( x[0] - x[7] );
00366 //
00367 // tp03 = t0 + t3;
00368 // tm03 = t0 - t3;
00369 // tp12 = t1 + t2;
00370 // tm12 = t1 - t2;
00371 //
00372 // y[0] = tp03 + tp12;
00373 // y[4] = tp03 - tp12;
00374 //
00375 // y[2] = tm03 + tm12 * tg_2_16;
00376 // y[6] = tm03 * tg_2_16 - tm12;
00377 //
00378 // tp65 =(t6 +t5 )*cos_4_16;
00379 // tm65 =(t6 -t5 )*cos_4_16;
00380 //
00381 // tp765 = t7 + tp65;
00382 // tm765 = t7 - tp65;
00383 // tp465 = t4 + tm65;
00384 // tm465 = t4 - tm65;
00385 //
00386 // y[1] = tp765 + tp465 * tg_1_16;
00387 // y[7] = tp765 * tg_1_16 - tp465;
00388 // y[5] = tm765 * tg_3_16 + tm465;
00389 // y[3] = tm765 - tm465 * tg_3_16;
00390 //}
00391 //
00392 //-----------------------------------------------------------------------------
00393 
00394 //-----------------------------------------------------------------------------
00395 // DCT_8_INV_COL_4  INP,OUT
00396 //-----------------------------------------------------------------------------
00397 
00398 #define DCT_8_INV_COL(A1,A2)\
00399   "movq 2*8(%3),%%mm0\n\t"\
00400   "movq 16*3+" #A1 ",%%mm3\n\t"\
00401   "movq %%mm0,%%mm1            \n\t"/* tg_3_16*/\
00402   "movq 16*5+" #A1 ",%%mm5\n\t"\
00403   "pmulhw %%mm3,%%mm0          \n\t"/* x3*(tg_3_16-1)*/\
00404   "movq (%3),%%mm4\n\t"\
00405   "pmulhw %%mm5,%%mm1          \n\t"/* x5*(tg_3_16-1)*/\
00406   "movq 16*7+" #A1 ",%%mm7\n\t"\
00407   "movq %%mm4,%%mm2            \n\t"/* tg_1_16*/\
00408   "movq 16*1+" #A1 ",%%mm6\n\t"\
00409   "pmulhw %%mm7,%%mm4          \n\t"/* x7*tg_1_16*/\
00410   "paddsw %%mm3,%%mm0          \n\t"/* x3*tg_3_16*/\
00411   "pmulhw %%mm6,%%mm2          \n\t"/* x1*tg_1_16*/\
00412   "paddsw %%mm3,%%mm1          \n\t"/* x3+x5*(tg_3_16-1)*/\
00413   "psubsw %%mm5,%%mm0          \n\t"/* x3*tg_3_16-x5 = tm35*/\
00414   "movq 3*8(%3),%%mm3\n\t"\
00415   "paddsw %%mm5,%%mm1          \n\t"/* x3+x5*tg_3_16 = tp35*/\
00416   "paddsw %%mm6,%%mm4          \n\t"/* x1+tg_1_16*x7 = tp17*/\
00417   "psubsw %%mm7,%%mm2          \n\t"/* x1*tg_1_16-x7 = tm17*/\
00418   "movq %%mm4,%%mm5            \n\t"/* tp17*/\
00419   "movq %%mm2,%%mm6            \n\t"/* tm17*/\
00420   "paddsw %%mm1,%%mm5          \n\t"/* tp17+tp35 = b0*/\
00421   "psubsw %%mm0,%%mm6          \n\t"/* tm17-tm35 = b3*/\
00422   "psubsw %%mm1,%%mm4          \n\t"/* tp17-tp35 = t1*/\
00423   "paddsw %%mm0,%%mm2          \n\t"/* tm17+tm35 = t2*/\
00424   "movq 1*8(%3),%%mm7\n\t"\
00425   "movq %%mm4,%%mm1            \n\t"/* t1*/\
00426   "movq %%mm5,3*16         +" #A2 "\n\t"/* save b0*/\
00427   "paddsw %%mm2,%%mm1          \n\t"/* t1+t2*/\
00428   "movq %%mm6,5*16         +" #A2 "\n\t"/* save b3*/\
00429   "psubsw %%mm2,%%mm4          \n\t"/* t1-t2*/\
00430   "movq 2*16+" #A1 ",%%mm5\n\t"\
00431   "movq %%mm7,%%mm0            \n\t"/* tg_2_16*/\
00432   "movq 6*16+" #A1 ",%%mm6\n\t"\
00433   "pmulhw %%mm5,%%mm0          \n\t"/* x2*tg_2_16*/\
00434   "pmulhw %%mm6,%%mm7          \n\t"/* x6*tg_2_16*/\
00435   "pmulhw %%mm3,%%mm1          \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
00436   "movq 0*16+" #A1 ",%%mm2\n\t"\
00437   "pmulhw %%mm3,%%mm4          \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
00438   "psubsw %%mm6,%%mm0          \n\t"/* t2*tg_2_16-x6 = tm26*/\
00439   "movq %%mm2,%%mm3            \n\t"/* x0*/\
00440   "movq 4*16+" #A1 ",%%mm6\n\t"\
00441   "paddsw %%mm5,%%mm7          \n\t"/* x2+x6*tg_2_16 = tp26*/\
00442   "paddsw %%mm6,%%mm2          \n\t"/* x0+x4 = tp04*/\
00443   "psubsw %%mm6,%%mm3          \n\t"/* x0-x4 = tm04*/\
00444   "movq %%mm2,%%mm5            \n\t"/* tp04*/\
00445   "movq %%mm3,%%mm6            \n\t"/* tm04*/\
00446   "psubsw %%mm7,%%mm2          \n\t"/* tp04-tp26 = a3*/\
00447   "paddsw %%mm0,%%mm3          \n\t"/* tm04+tm26 = a1*/\
00448   "paddsw %%mm1,%%mm1          \n\t"/* b1*/\
00449   "paddsw %%mm4,%%mm4          \n\t"/* b2*/\
00450   "paddsw %%mm7,%%mm5          \n\t"/* tp04+tp26 = a0*/\
00451   "psubsw %%mm0,%%mm6          \n\t"/* tm04-tm26 = a2*/\
00452   "movq %%mm3,%%mm7            \n\t"/* a1*/\
00453   "movq %%mm6,%%mm0            \n\t"/* a2*/\
00454   "paddsw %%mm1,%%mm3          \n\t"/* a1+b1*/\
00455   "paddsw %%mm4,%%mm6          \n\t"/* a2+b2*/\
00456   "psraw $6,%%mm3 \n\t"/* dst1*/\
00457   "psubsw %%mm1,%%mm7          \n\t"/* a1-b1*/\
00458   "psraw $6,%%mm6 \n\t"/* dst2*/\
00459   "psubsw %%mm4,%%mm0          \n\t"/* a2-b2*/\
00460   "movq 3*16+" #A2 ",%%mm1         \n\t"/* load b0*/\
00461   "psraw $6,%%mm7 \n\t"/* dst6*/\
00462   "movq %%mm5,%%mm4            \n\t"/* a0*/\
00463   "psraw $6,%%mm0 \n\t"/* dst5*/\
00464   "movq %%mm3,1*16+" #A2 "\n\t"\
00465   "paddsw %%mm1,%%mm5          \n\t"/* a0+b0*/\
00466   "movq %%mm6,2*16+" #A2 "\n\t"\
00467   "psubsw %%mm1,%%mm4          \n\t"/* a0-b0*/\
00468   "movq 5*16+" #A2 ",%%mm3         \n\t"/* load b3*/\
00469   "psraw $6,%%mm5 \n\t"/* dst0*/\
00470   "movq %%mm2,%%mm6            \n\t"/* a3*/\
00471   "psraw $6,%%mm4 \n\t"/* dst7*/\
00472   "movq %%mm0,5*16+" #A2 "\n\t"\
00473   "paddsw %%mm3,%%mm2          \n\t"/* a3+b3*/\
00474   "movq %%mm7,6*16+" #A2 "\n\t"\
00475   "psubsw %%mm3,%%mm6          \n\t"/* a3-b3*/\
00476   "movq %%mm5,0*16+" #A2 "\n\t"\
00477   "psraw $6,%%mm2 \n\t"/* dst3*/\
00478   "movq %%mm4,7*16+" #A2 "\n\t"\
00479   "psraw $6,%%mm6 \n\t"/* dst4*/\
00480   "movq %%mm2,3*16+" #A2 "\n\t"\
00481   "movq %%mm6,4*16+" #A2 "\n\t"
00482 
00483 //=============================================================================
00484 // Code
00485 //=============================================================================
00486 
00487 //-----------------------------------------------------------------------------
00488 // void idct_mmx(uint16_t block[64]);
00489 //-----------------------------------------------------------------------------
00490 
00491 
00492 void ff_idct_xvid_mmx(short *block){
00493 __asm__ volatile(
00494             //# Process each row
00495     DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
00496     DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
00497     DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
00498     DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
00499     DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
00500     DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
00501     DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
00502     DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
00503 
00504             //# Process the columns (4 at a time)
00505     DCT_8_INV_COL(0(%0), 0(%0))
00506     DCT_8_INV_COL(8(%0), 8(%0))
00507     :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
00508 }
00509 
00510 //-----------------------------------------------------------------------------
00511 // void idct_xmm(uint16_t block[64]);
00512 //-----------------------------------------------------------------------------
00513 
00514 
00515 void ff_idct_xvid_mmx2(short *block){
00516 __asm__ volatile(
00517             //# Process each row
00518     DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
00519     DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
00520     DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
00521     DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
00522     DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
00523     DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
00524     DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
00525     DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
00526 
00527             //# Process the columns (4 at a time)
00528     DCT_8_INV_COL(0(%0), 0(%0))
00529     DCT_8_INV_COL(8(%0), 8(%0))
00530     :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
00531 }
00532 
00533 void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, DCTELEM *block)
00534 {
00535     ff_idct_xvid_mmx(block);
00536     ff_put_pixels_clamped_mmx(block, dest, line_size);
00537 }
00538 
00539 void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, DCTELEM *block)
00540 {
00541     ff_idct_xvid_mmx(block);
00542     ff_add_pixels_clamped_mmx(block, dest, line_size);
00543 }
00544 
00545 void ff_idct_xvid_mmx2_put(uint8_t *dest, int line_size, DCTELEM *block)
00546 {
00547     ff_idct_xvid_mmx2(block);
00548     ff_put_pixels_clamped_mmx(block, dest, line_size);
00549 }
00550 
00551 void ff_idct_xvid_mmx2_add(uint8_t *dest, int line_size, DCTELEM *block)
00552 {
00553     ff_idct_xvid_mmx2(block);
00554     ff_add_pixels_clamped_mmx(block, dest, line_size);
00555 }
00556 
00557 #endif /* HAVE_INLINE_ASM */
libavcodec/x86/idct_mmx_xvid.c
