Mercurial > audlegacy-plugins

comparison src/ffmpeg/libavcodec/i386/idct_mmx_xvid.c @ 808:e8776388b02a trunk

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
[svn] - add ffmpeg
author nenolod
date Mon, 12 Mar 2007 11:18:54 -0700
parents
children
comparison
equal deleted inserted replaced
807:0f9c8d4d3ac4 808:e8776388b02a
1 ///****************************************************************************
2 // *
3 // * XVID MPEG-4 VIDEO CODEC
4 // * - MMX and XMM forward discrete cosine transform -
5 // *
6 // * Copyright(C) 2001 Peter Ross <pross@xvid.org>
7 // *
8 // * This file is part of FFmpeg.
9 // *
10 // * FFmpeg is free software; you can redistribute it and/or
11 // * modify it under the terms of the GNU Lesser General Public
12 // * License as published by the Free Software Foundation; either
13 // * version 2.1 of the License, or (at your option) any later version.
14 // *
15 // * FFmpeg is distributed in the hope that it will be useful,
16 // * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 // * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 // * Lesser General Public License for more details.
19 // *
20 // * You should have received a copy of the GNU Lesser General Public License
21 // * along with FFmpeg; if not, write to the Free Software Foundation,
22 // * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 // *
24 // * $Id: idct_mmx_xvid.c 6577 2006-10-07 15:30:46Z diego $
25 // *
26 // ***************************************************************************/
27
28 // ****************************************************************************
29 //
30 // Originally provided by Intel at AP-922
31 // http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
32 // (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
33 // but in a limited edition.
34 // New macro implements a column part for precise iDCT
35 // The routine precision now satisfies IEEE standard 1180-1990.
36 //
37 // Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
38 // Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org>
39 //
40 // http://www.elecard.com/peter/idct.html
41 // http://www.linuxvideo.org/mpeg2dec/
42 //
43 // ***************************************************************************/
44 //
45 // These examples contain code fragments for first stage iDCT 8x8
46 // (for rows) and first stage DCT 8x8 (for columns)
47 //
48
49 // conversion to gcc syntax by michael niedermayer
50
51
52 #include <inttypes.h>
53 #include "../avcodec.h"
54
55 //=============================================================================
56 // Macros and other preprocessor constants
57 //=============================================================================
58
59 #define BITS_INV_ACC 5 // 4 or 5 for IEEE
60 #define SHIFT_INV_ROW (16 - BITS_INV_ACC) //11
61 #define SHIFT_INV_COL (1 + BITS_INV_ACC) //6
62 #define RND_INV_ROW (1024 * (6 - BITS_INV_ACC))
63 #define RND_INV_COL (16 * (BITS_INV_ACC - 3))
64 #define RND_INV_CORR (RND_INV_COL - 1)
65
66 #define BITS_FRW_ACC 3 // 2 or 3 for accuracy
67 #define SHIFT_FRW_COL BITS_FRW_ACC
68 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17)
69 #define RND_FRW_ROW (262144*(BITS_FRW_ACC - 1))
70
71
72 //-----------------------------------------------------------------------------
73 // Various memory constants (trigonometric values or rounding values)
74 //-----------------------------------------------------------------------------
75
76
77 static const int16_t tg_1_16[4*4] attribute_used __attribute__ ((aligned(8))) = {
78 13036,13036,13036,13036, // tg * (2<<16) + 0.5
79 27146,27146,27146,27146, // tg * (2<<16) + 0.5
80 -21746,-21746,-21746,-21746, // tg * (2<<16) + 0.5
81 23170,23170,23170,23170}; // cos * (2<<15) + 0.5
82
83 static const int32_t rounder_0[2*8] attribute_used __attribute__ ((aligned(8))) = {
84 65536,65536,
85 3597,3597,
86 2260,2260,
87 1203,1203,
88 0,0,
89 120,120,
90 512,512,
91 512,512};
92
93 //-----------------------------------------------------------------------------
94 //
95 // The first stage iDCT 8x8 - inverse DCTs of rows
96 //
97 //-----------------------------------------------------------------------------
98 // The 8-point inverse DCT direct algorithm
99 //-----------------------------------------------------------------------------
100 //
101 // static const short w[32] = {
102 // FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16),
103 // FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16),
104 // FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16),
105 // FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16),
106 // FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16),
107 // FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16),
108 // FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16),
109 // FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) };
110 //
111 // #define DCT_8_INV_ROW(x, y)
112 // {
113 // int a0, a1, a2, a3, b0, b1, b2, b3;
114 //
115 // a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3];
116 // a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7];
117 // a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11];
118 // a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15];
119 // b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19];
120 // b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23];
121 // b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27];
122 // b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31];
123 //
124 // y[0] = SHIFT_ROUND ( a0 + b0 );
125 // y[1] = SHIFT_ROUND ( a1 + b1 );
126 // y[2] = SHIFT_ROUND ( a2 + b2 );
127 // y[3] = SHIFT_ROUND ( a3 + b3 );
128 // y[4] = SHIFT_ROUND ( a3 - b3 );
129 // y[5] = SHIFT_ROUND ( a2 - b2 );
130 // y[6] = SHIFT_ROUND ( a1 - b1 );
131 // y[7] = SHIFT_ROUND ( a0 - b0 );
132 // }
133 //
134 //-----------------------------------------------------------------------------
135 //
136 // In this implementation the outputs of the iDCT-1D are multiplied
137 // for rows 0,4 - by cos_4_16,
138 // for rows 1,7 - by cos_1_16,
139 // for rows 2,6 - by cos_2_16,
140 // for rows 3,5 - by cos_3_16
141 // and are shifted to the left for better accuracy
142 //
143 // For the constants used,
144 // FIX(float_const) = (short) (float_const * (1<<15) + 0.5)
145 //
146 //-----------------------------------------------------------------------------
147
148 //-----------------------------------------------------------------------------
149 // Tables for mmx processors
150 //-----------------------------------------------------------------------------
151
152 // Table for rows 0,4 - constants are multiplied by cos_4_16
153 static const int16_t tab_i_04_mmx[32*4] attribute_used __attribute__ ((aligned(8))) = {
154 16384,16384,16384,-16384, // movq-> w06 w04 w02 w00
155 21407,8867,8867,-21407, // w07 w05 w03 w01
156 16384,-16384,16384,16384, // w14 w12 w10 w08
157 -8867,21407,-21407,-8867, // w15 w13 w11 w09
158 22725,12873,19266,-22725, // w22 w20 w18 w16
159 19266,4520,-4520,-12873, // w23 w21 w19 w17
160 12873,4520,4520,19266, // w30 w28 w26 w24
161 -22725,19266,-12873,-22725, // w31 w29 w27 w25
162 // Table for rows 1,7 - constants are multiplied by cos_1_16
163 22725,22725,22725,-22725, // movq-> w06 w04 w02 w00
164 29692,12299,12299,-29692, // w07 w05 w03 w01
165 22725,-22725,22725,22725, // w14 w12 w10 w08
166 -12299,29692,-29692,-12299, // w15 w13 w11 w09
167 31521,17855,26722,-31521, // w22 w20 w18 w16
168 26722,6270,-6270,-17855, // w23 w21 w19 w17
169 17855,6270,6270,26722, // w30 w28 w26 w24
170 -31521,26722,-17855,-31521, // w31 w29 w27 w25
171 // Table for rows 2,6 - constants are multiplied by cos_2_16
172 21407,21407,21407,-21407, // movq-> w06 w04 w02 w00
173 27969,11585,11585,-27969, // w07 w05 w03 w01
174 21407,-21407,21407,21407, // w14 w12 w10 w08
175 -11585,27969,-27969,-11585, // w15 w13 w11 w09
176 29692,16819,25172,-29692, // w22 w20 w18 w16
177 25172,5906,-5906,-16819, // w23 w21 w19 w17
178 16819,5906,5906,25172, // w30 w28 w26 w24
179 -29692,25172,-16819,-29692, // w31 w29 w27 w25
180 // Table for rows 3,5 - constants are multiplied by cos_3_16
181 19266,19266,19266,-19266, // movq-> w06 w04 w02 w00
182 25172,10426,10426,-25172, // w07 w05 w03 w01
183 19266,-19266,19266,19266, // w14 w12 w10 w08
184 -10426,25172,-25172,-10426, // w15 w13 w11 w09
185 26722,15137,22654,-26722, // w22 w20 w18 w16
186 22654,5315,-5315,-15137, // w23 w21 w19 w17
187 15137,5315,5315,22654, // w30 w28 w26 w24
188 -26722,22654,-15137,-26722, // w31 w29 w27 w25
189 };
190 //-----------------------------------------------------------------------------
191 // Tables for xmm processors
192 //-----------------------------------------------------------------------------
193
194 // %3 for rows 0,4 - constants are multiplied by cos_4_16
195 static const int16_t tab_i_04_xmm[32*4] attribute_used __attribute__ ((aligned(8))) = {
196 16384,21407,16384,8867, // movq-> w05 w04 w01 w00
197 16384,8867,-16384,-21407, // w07 w06 w03 w02
198 16384,-8867,16384,-21407, // w13 w12 w09 w08
199 -16384,21407,16384,-8867, // w15 w14 w11 w10
200 22725,19266,19266,-4520, // w21 w20 w17 w16
201 12873,4520,-22725,-12873, // w23 w22 w19 w18
202 12873,-22725,4520,-12873, // w29 w28 w25 w24
203 4520,19266,19266,-22725, // w31 w30 w27 w26
204 // %3 for rows 1,7 - constants are multiplied by cos_1_16
205 22725,29692,22725,12299, // movq-> w05 w04 w01 w00
206 22725,12299,-22725,-29692, // w07 w06 w03 w02
207 22725,-12299,22725,-29692, // w13 w12 w09 w08
208 -22725,29692,22725,-12299, // w15 w14 w11 w10
209 31521,26722,26722,-6270, // w21 w20 w17 w16
210 17855,6270,-31521,-17855, // w23 w22 w19 w18
211 17855,-31521,6270,-17855, // w29 w28 w25 w24
212 6270,26722,26722,-31521, // w31 w30 w27 w26
213 // %3 for rows 2,6 - constants are multiplied by cos_2_16
214 21407,27969,21407,11585, // movq-> w05 w04 w01 w00
215 21407,11585,-21407,-27969, // w07 w06 w03 w02
216 21407,-11585,21407,-27969, // w13 w12 w09 w08
217 -21407,27969,21407,-11585, // w15 w14 w11 w10
218 29692,25172,25172,-5906, // w21 w20 w17 w16
219 16819,5906,-29692,-16819, // w23 w22 w19 w18
220 16819,-29692,5906,-16819, // w29 w28 w25 w24
221 5906,25172,25172,-29692, // w31 w30 w27 w26
222 // %3 for rows 3,5 - constants are multiplied by cos_3_16
223 19266,25172,19266,10426, // movq-> w05 w04 w01 w00
224 19266,10426,-19266,-25172, // w07 w06 w03 w02
225 19266,-10426,19266,-25172, // w13 w12 w09 w08
226 -19266,25172,19266,-10426, // w15 w14 w11 w10
227 26722,22654,22654,-5315, // w21 w20 w17 w16
228 15137,5315,-26722,-15137, // w23 w22 w19 w18
229 15137,-26722,5315,-15137, // w29 w28 w25 w24
230 5315,22654,22654,-26722, // w31 w30 w27 w26
231 };
232 //=============================================================================
233 // Helper macros for the code
234 //=============================================================================
235
236 //-----------------------------------------------------------------------------
237 // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER
238 //-----------------------------------------------------------------------------
239
240 #define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\
241 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
242 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
243 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
244 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w06 w04 w02 w00*/\
245 "punpcklwd %%mm1,%%mm0 \n\t"/* x5 x1 x4 x0*/\
246 "movq %%mm0,%%mm5 \n\t"/* 5 ; x5 x1 x4 x0*/\
247 "punpckldq %%mm0,%%mm0 \n\t"/* x4 x0 x4 x0*/\
248 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w05 w03 w01*/\
249 "punpckhwd %%mm1,%%mm2 \n\t"/* 1 ; x7 x3 x6 x2*/\
250 "pmaddwd %%mm0,%%mm3 \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\
251 "movq %%mm2,%%mm6 \n\t"/* 6 ; x7 x3 x6 x2*/\
252 "movq 32+" #A3 ",%%mm1 \n\t"/* 1 ; w22 w20 w18 w16*/\
253 "punpckldq %%mm2,%%mm2 \n\t"/* x6 x2 x6 x2*/\
254 "pmaddwd %%mm2,%%mm4 \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\
255 "punpckhdq %%mm5,%%mm5 \n\t"/* x5 x1 x5 x1*/\
256 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\
257 "punpckhdq %%mm6,%%mm6 \n\t"/* x7 x3 x7 x3*/\
258 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w21 w19 w17*/\
259 "pmaddwd %%mm5,%%mm1 \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\
260 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
261 "pmaddwd %%mm6,%%mm7 \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\
262 "pmaddwd 24+" #A3 ",%%mm2 \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\
263 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
264 "pmaddwd 48+" #A3 ",%%mm5 \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\
265 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
266 "pmaddwd 56+" #A3 ",%%mm6 \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\
267 "paddd %%mm7,%%mm1 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
268 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
269 "psubd %%mm1,%%mm3 \n\t"/* a1-b1 a0-b0*/\
270 "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\
271 "paddd %%mm4,%%mm1 \n\t"/* 4 ; a1+b1 a0+b0*/\
272 "paddd %%mm2,%%mm0 \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\
273 "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\
274 "paddd %%mm6,%%mm5 \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
275 "movq %%mm0,%%mm4 \n\t"/* 4 ; a3 a2*/\
276 "paddd %%mm5,%%mm0 \n\t"/* a3+b3 a2+b2*/\
277 "psubd %%mm5,%%mm4 \n\t"/* 5 ; a3-b3 a2-b2*/\
278 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
279 "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
280 "packssdw %%mm0,%%mm1 \n\t"/* 0 ; y3 y2 y1 y0*/\
281 "packssdw %%mm3,%%mm4 \n\t"/* 3 ; y6 y7 y4 y5*/\
282 "movq %%mm4,%%mm7 \n\t"/* 7 ; y6 y7 y4 y5*/\
283 "psrld $16,%%mm4 \n\t"/* 0 y6 0 y4*/\
284 "pslld $16,%%mm7 \n\t"/* y7 0 y5 0*/\
285 "movq %%mm1," #A2 " \n\t"/* 1 ; save y3 y2 y1 y0*/\
286 "por %%mm4,%%mm7 \n\t"/* 4 ; y7 y6 y5 y4*/\
287 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
288
289
290 //-----------------------------------------------------------------------------
291 // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
292 //-----------------------------------------------------------------------------
293
294 #define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
295 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
296 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
297 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
298 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w05 w04 w01 w00*/\
299 "pshufw $0x88,%%mm0,%%mm0 \n\t"/* x2 x0 x2 x0*/\
300 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w06 w03 w02*/\
301 "movq %%mm1,%%mm5 \n\t"/* 5 ; x7 x6 x5 x4*/\
302 "pmaddwd %%mm0,%%mm3 \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
303 "movq 32+" #A3 ",%%mm6 \n\t"/* 6 ; w21 w20 w17 w16*/\
304 "pshufw $0x88,%%mm1,%%mm1 \n\t"/* x6 x4 x6 x4*/\
305 "pmaddwd %%mm1,%%mm4 \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
306 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w22 w19 w18*/\
307 "pshufw $0xdd,%%mm2,%%mm2 \n\t"/* x3 x1 x3 x1*/\
308 "pmaddwd %%mm2,%%mm6 \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
309 "pshufw $0xdd,%%mm5,%%mm5 \n\t"/* x7 x5 x7 x5*/\
310 "pmaddwd %%mm5,%%mm7 \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
311 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
312 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
313 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
314 "pmaddwd 24+" #A3 ",%%mm1 \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
315 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
316 "pmaddwd 48+" #A3 ",%%mm2 \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
317 "paddd %%mm7,%%mm6 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
318 "pmaddwd 56+" #A3 ",%%mm5 \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
319 "paddd %%mm6,%%mm3 \n\t"/* a1+b1 a0+b0*/\
320 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
321 "psrad $11,%%mm3 \n\t"/* y1=a1+b1 y0=a0+b0*/\
322 "paddd %%mm1,%%mm0 \n\t"/* 1 ; a3=sum(even3) a2=sum(even2)*/\
323 "psubd %%mm6,%%mm4 \n\t"/* 6 ; a1-b1 a0-b0*/\
324 "movq %%mm0,%%mm7 \n\t"/* 7 ; a3 a2*/\
325 "paddd %%mm5,%%mm2 \n\t"/* 5 ; b3=sum(odd3) b2=sum(odd2)*/\
326 "paddd %%mm2,%%mm0 \n\t"/* a3+b3 a2+b2*/\
327 "psrad $11,%%mm4 \n\t"/* y6=a1-b1 y7=a0-b0*/\
328 "psubd %%mm2,%%mm7 \n\t"/* 2 ; a3-b3 a2-b2*/\
329 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
330 "psrad $11,%%mm7 \n\t"/* y4=a3-b3 y5=a2-b2*/\
331 "packssdw %%mm0,%%mm3 \n\t"/* 0 ; y3 y2 y1 y0*/\
332 "packssdw %%mm4,%%mm7 \n\t"/* 4 ; y6 y7 y4 y5*/\
333 "movq %%mm3, " #A2 " \n\t"/* 3 ; save y3 y2 y1 y0*/\
334 "pshufw $0xb1,%%mm7,%%mm7 \n\t"/* y7 y6 y5 y4*/\
335 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
336
337
338 //-----------------------------------------------------------------------------
339 //
340 // The first stage DCT 8x8 - forward DCTs of columns
341 //
342 // The %2puts are multiplied
343 // for rows 0,4 - on cos_4_16,
344 // for rows 1,7 - on cos_1_16,
345 // for rows 2,6 - on cos_2_16,
346 // for rows 3,5 - on cos_3_16
347 // and are shifted to the left for rise of accuracy
348 //
349 //-----------------------------------------------------------------------------
350 //
351 // The 8-point scaled forward DCT algorithm (26a8m)
352 //
353 //-----------------------------------------------------------------------------
354 //
355 // #define DCT_8_FRW_COL(x, y)
356 //{
357 // short t0, t1, t2, t3, t4, t5, t6, t7;
358 // short tp03, tm03, tp12, tm12, tp65, tm65;
359 // short tp465, tm465, tp765, tm765;
360 //
361 // t0 = LEFT_SHIFT ( x[0] + x[7] );
362 // t1 = LEFT_SHIFT ( x[1] + x[6] );
363 // t2 = LEFT_SHIFT ( x[2] + x[5] );
364 // t3 = LEFT_SHIFT ( x[3] + x[4] );
365 // t4 = LEFT_SHIFT ( x[3] - x[4] );
366 // t5 = LEFT_SHIFT ( x[2] - x[5] );
367 // t6 = LEFT_SHIFT ( x[1] - x[6] );
368 // t7 = LEFT_SHIFT ( x[0] - x[7] );
369 //
370 // tp03 = t0 + t3;
371 // tm03 = t0 - t3;
372 // tp12 = t1 + t2;
373 // tm12 = t1 - t2;
374 //
375 // y[0] = tp03 + tp12;
376 // y[4] = tp03 - tp12;
377 //
378 // y[2] = tm03 + tm12 * tg_2_16;
379 // y[6] = tm03 * tg_2_16 - tm12;
380 //
381 // tp65 =(t6 +t5 )*cos_4_16;
382 // tm65 =(t6 -t5 )*cos_4_16;
383 //
384 // tp765 = t7 + tp65;
385 // tm765 = t7 - tp65;
386 // tp465 = t4 + tm65;
387 // tm465 = t4 - tm65;
388 //
389 // y[1] = tp765 + tp465 * tg_1_16;
390 // y[7] = tp765 * tg_1_16 - tp465;
391 // y[5] = tm765 * tg_3_16 + tm465;
392 // y[3] = tm765 - tm465 * tg_3_16;
393 //}
394 //
395 //-----------------------------------------------------------------------------
396
397 //-----------------------------------------------------------------------------
398 // DCT_8_INV_COL_4 INP,OUT
399 //-----------------------------------------------------------------------------
400
401 #define DCT_8_INV_COL(A1,A2)\
402 "movq 2*8(%3),%%mm0\n\t"\
403 "movq 16*3+" #A1 ",%%mm3\n\t"\
404 "movq %%mm0,%%mm1 \n\t"/* tg_3_16*/\
405 "movq 16*5+" #A1 ",%%mm5\n\t"\
406 "pmulhw %%mm3,%%mm0 \n\t"/* x3*(tg_3_16-1)*/\
407 "movq (%3),%%mm4\n\t"\
408 "pmulhw %%mm5,%%mm1 \n\t"/* x5*(tg_3_16-1)*/\
409 "movq 16*7+" #A1 ",%%mm7\n\t"\
410 "movq %%mm4,%%mm2 \n\t"/* tg_1_16*/\
411 "movq 16*1+" #A1 ",%%mm6\n\t"\
412 "pmulhw %%mm7,%%mm4 \n\t"/* x7*tg_1_16*/\
413 "paddsw %%mm3,%%mm0 \n\t"/* x3*tg_3_16*/\
414 "pmulhw %%mm6,%%mm2 \n\t"/* x1*tg_1_16*/\
415 "paddsw %%mm3,%%mm1 \n\t"/* x3+x5*(tg_3_16-1)*/\
416 "psubsw %%mm5,%%mm0 \n\t"/* x3*tg_3_16-x5 = tm35*/\
417 "movq 3*8(%3),%%mm3\n\t"\
418 "paddsw %%mm5,%%mm1 \n\t"/* x3+x5*tg_3_16 = tp35*/\
419 "paddsw %%mm6,%%mm4 \n\t"/* x1+tg_1_16*x7 = tp17*/\
420 "psubsw %%mm7,%%mm2 \n\t"/* x1*tg_1_16-x7 = tm17*/\
421 "movq %%mm4,%%mm5 \n\t"/* tp17*/\
422 "movq %%mm2,%%mm6 \n\t"/* tm17*/\
423 "paddsw %%mm1,%%mm5 \n\t"/* tp17+tp35 = b0*/\
424 "psubsw %%mm0,%%mm6 \n\t"/* tm17-tm35 = b3*/\
425 "psubsw %%mm1,%%mm4 \n\t"/* tp17-tp35 = t1*/\
426 "paddsw %%mm0,%%mm2 \n\t"/* tm17+tm35 = t2*/\
427 "movq 1*8(%3),%%mm7\n\t"\
428 "movq %%mm4,%%mm1 \n\t"/* t1*/\
429 "movq %%mm5,3*16 +" #A2 "\n\t"/* save b0*/\
430 "paddsw %%mm2,%%mm1 \n\t"/* t1+t2*/\
431 "movq %%mm6,5*16 +" #A2 "\n\t"/* save b3*/\
432 "psubsw %%mm2,%%mm4 \n\t"/* t1-t2*/\
433 "movq 2*16+" #A1 ",%%mm5\n\t"\
434 "movq %%mm7,%%mm0 \n\t"/* tg_2_16*/\
435 "movq 6*16+" #A1 ",%%mm6\n\t"\
436 "pmulhw %%mm5,%%mm0 \n\t"/* x2*tg_2_16*/\
437 "pmulhw %%mm6,%%mm7 \n\t"/* x6*tg_2_16*/\
438 "pmulhw %%mm3,%%mm1 \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
439 "movq 0*16+" #A1 ",%%mm2\n\t"\
440 "pmulhw %%mm3,%%mm4 \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
441 "psubsw %%mm6,%%mm0 \n\t"/* t2*tg_2_16-x6 = tm26*/\
442 "movq %%mm2,%%mm3 \n\t"/* x0*/\
443 "movq 4*16+" #A1 ",%%mm6\n\t"\
444 "paddsw %%mm5,%%mm7 \n\t"/* x2+x6*tg_2_16 = tp26*/\
445 "paddsw %%mm6,%%mm2 \n\t"/* x0+x4 = tp04*/\
446 "psubsw %%mm6,%%mm3 \n\t"/* x0-x4 = tm04*/\
447 "movq %%mm2,%%mm5 \n\t"/* tp04*/\
448 "movq %%mm3,%%mm6 \n\t"/* tm04*/\
449 "psubsw %%mm7,%%mm2 \n\t"/* tp04-tp26 = a3*/\
450 "paddsw %%mm0,%%mm3 \n\t"/* tm04+tm26 = a1*/\
451 "paddsw %%mm1,%%mm1 \n\t"/* b1*/\
452 "paddsw %%mm4,%%mm4 \n\t"/* b2*/\
453 "paddsw %%mm7,%%mm5 \n\t"/* tp04+tp26 = a0*/\
454 "psubsw %%mm0,%%mm6 \n\t"/* tm04-tm26 = a2*/\
455 "movq %%mm3,%%mm7 \n\t"/* a1*/\
456 "movq %%mm6,%%mm0 \n\t"/* a2*/\
457 "paddsw %%mm1,%%mm3 \n\t"/* a1+b1*/\
458 "paddsw %%mm4,%%mm6 \n\t"/* a2+b2*/\
459 "psraw $6,%%mm3 \n\t"/* dst1*/\
460 "psubsw %%mm1,%%mm7 \n\t"/* a1-b1*/\
461 "psraw $6,%%mm6 \n\t"/* dst2*/\
462 "psubsw %%mm4,%%mm0 \n\t"/* a2-b2*/\
463 "movq 3*16+" #A2 ",%%mm1 \n\t"/* load b0*/\
464 "psraw $6,%%mm7 \n\t"/* dst6*/\
465 "movq %%mm5,%%mm4 \n\t"/* a0*/\
466 "psraw $6,%%mm0 \n\t"/* dst5*/\
467 "movq %%mm3,1*16+" #A2 "\n\t"\
468 "paddsw %%mm1,%%mm5 \n\t"/* a0+b0*/\
469 "movq %%mm6,2*16+" #A2 "\n\t"\
470 "psubsw %%mm1,%%mm4 \n\t"/* a0-b0*/\
471 "movq 5*16+" #A2 ",%%mm3 \n\t"/* load b3*/\
472 "psraw $6,%%mm5 \n\t"/* dst0*/\
473 "movq %%mm2,%%mm6 \n\t"/* a3*/\
474 "psraw $6,%%mm4 \n\t"/* dst7*/\
475 "movq %%mm0,5*16+" #A2 "\n\t"\
476 "paddsw %%mm3,%%mm2 \n\t"/* a3+b3*/\
477 "movq %%mm7,6*16+" #A2 "\n\t"\
478 "psubsw %%mm3,%%mm6 \n\t"/* a3-b3*/\
479 "movq %%mm5,0*16+" #A2 "\n\t"\
480 "psraw $6,%%mm2 \n\t"/* dst3*/\
481 "movq %%mm4,7*16+" #A2 "\n\t"\
482 "psraw $6,%%mm6 \n\t"/* dst4*/\
483 "movq %%mm2,3*16+" #A2 "\n\t"\
484 "movq %%mm6,4*16+" #A2 "\n\t"
485
486 //=============================================================================
487 // Code
488 //=============================================================================
489
490 //-----------------------------------------------------------------------------
491 // void idct_mmx(uint16_t block[64]);
492 //-----------------------------------------------------------------------------
493
494
495 void ff_idct_xvid_mmx(short *block){
496 asm volatile(
497 //# Process each row
498 DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
499 DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
500 DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
501 DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
502 DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
503 DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
504 DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
505 DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
506
507 //# Process the columns (4 at a time)
508 DCT_8_INV_COL(0(%0), 0(%0))
509 DCT_8_INV_COL(8(%0), 8(%0))
510 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
511 }
512
513 //-----------------------------------------------------------------------------
514 // void idct_xmm(uint16_t block[64]);
515 //-----------------------------------------------------------------------------
516
517
518 void ff_idct_xvid_mmx2(short *block){
519 asm volatile(
520 //# Process each row
521 DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
522 DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
523 DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
524 DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
525 DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
526 DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
527 DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
528 DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
529
530 //# Process the columns (4 at a time)
531 DCT_8_INV_COL(0(%0), 0(%0))
532 DCT_8_INV_COL(8(%0), 8(%0))
533 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
534 }
535