Statistics
| Branch: | Revision:

ffmpeg / libavcodec / i386 / idct_mmx_xvid.c @ 5509bffa

History | View | Annotate | Download (23.8 KB)

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 program is free software; you can redistribute it and/or modify it
9
// *  under the terms of the GNU General Public License as published by
10
// *  the Free Software Foundation; either version 2 of the License, or
11
// *  (at your option) any later version.
12
// *
13
// *  This program is distributed in the hope that it will be useful,
14
// *  but WITHOUT ANY WARRANTY; without even the implied warranty of
15
// *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16
// *  GNU General Public License for more details.
17
// *
18
// *  You should have received a copy of the GNU General Public License
19
// *  along with this program; if not, write to the Free Software Foundation,
20
// *  Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21
//
22
// *
23
// * $Id$
24
// *
25
// ***************************************************************************/
26

    
27
// ****************************************************************************
28
//
29
// Originally provided by Intel at AP-922
30
// http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
31
// (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
32
// but in a limited edition.
33
// New macro implements a column part for precise iDCT
34
// The routine precision now satisfies IEEE standard 1180-1990.
35
//
36
// Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
37
// Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org>
38
//
39
// http://www.elecard.com/peter/idct.html
40
// http://www.linuxvideo.org/mpeg2dec/
41
//
42
// ***************************************************************************/
43
//
44
// These examples contain code fragments for first stage iDCT 8x8
45
// (for rows) and first stage DCT 8x8 (for columns)
46
//
47

    
48
// conversion to gcc syntax by michael niedermayer
49

    
50

    
51
#include <inttypes.h>
52
#include "../avcodec.h"
53

    
54
//=============================================================================
55
// Macros and other preprocessor constants
56
//=============================================================================
57

    
58
#define BITS_INV_ACC    5                              // 4 or 5 for IEEE
59
#define SHIFT_INV_ROW   (16 - BITS_INV_ACC) //11
60
#define SHIFT_INV_COL   (1 + BITS_INV_ACC) //6
61
#define RND_INV_ROW     (1024 * (6 - BITS_INV_ACC))
62
#define RND_INV_COL     (16 * (BITS_INV_ACC - 3))
63
#define RND_INV_CORR    (RND_INV_COL - 1)
64

    
65
#define BITS_FRW_ACC    3                              // 2 or 3 for accuracy
66
#define SHIFT_FRW_COL   BITS_FRW_ACC
67
#define SHIFT_FRW_ROW   (BITS_FRW_ACC + 17)
68
#define RND_FRW_ROW     (262144*(BITS_FRW_ACC - 1))
69

    
70

    
71
//-----------------------------------------------------------------------------
72
// Various memory constants (trigonometric values or rounding values)
73
//-----------------------------------------------------------------------------
74

    
75

    
76
static const int16_t tg_1_16[4*4] attribute_used __attribute__ ((aligned(8))) = {
77
  13036,13036,13036,13036,        // tg * (2<<16) + 0.5
78
  27146,27146,27146,27146,        // tg * (2<<16) + 0.5
79
  -21746,-21746,-21746,-21746,    // tg * (2<<16) + 0.5
80
  23170,23170,23170,23170};       // cos * (2<<15) + 0.5
81

    
82
static const int32_t rounder_0[2*8] attribute_used __attribute__ ((aligned(8))) = {
83
  65536,65536,
84
  3597,3597,
85
  2260,2260,
86
  1203,1203,
87
  0,0,
88
  120,120,
89
  512,512,
90
  512,512};
91

    
92
//-----------------------------------------------------------------------------
93
//
94
// The first stage iDCT 8x8 - inverse DCTs of rows
95
//
96
//-----------------------------------------------------------------------------
97
// The 8-point inverse DCT direct algorithm
98
//-----------------------------------------------------------------------------
99
//
100
// static const short w[32] = {
101
//       FIX(cos_4_16),  FIX(cos_2_16),  FIX(cos_4_16),  FIX(cos_6_16),
102
//       FIX(cos_4_16),  FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_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_2_16),  FIX(cos_4_16), -FIX(cos_6_16),
105
//       FIX(cos_1_16),  FIX(cos_3_16),  FIX(cos_5_16),  FIX(cos_7_16),
106
//       FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16),
107
//       FIX(cos_5_16), -FIX(cos_1_16),  FIX(cos_7_16),  FIX(cos_3_16),
108
//       FIX(cos_7_16), -FIX(cos_5_16),  FIX(cos_3_16), -FIX(cos_1_16) };
109
//
110
// #define DCT_8_INV_ROW(x, y)
111
// {
112
//       int a0, a1, a2, a3, b0, b1, b2, b3;
113
//
114
//       a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3];
115
//       a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7];
116
//       a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11];
117
//       a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15];
118
//       b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19];
119
//       b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23];
120
//       b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27];
121
//       b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31];
122
//
123
//       y[0] = SHIFT_ROUND ( a0 + b0 );
124
//       y[1] = SHIFT_ROUND ( a1 + b1 );
125
//       y[2] = SHIFT_ROUND ( a2 + b2 );
126
//       y[3] = SHIFT_ROUND ( a3 + b3 );
127
//       y[4] = SHIFT_ROUND ( a3 - b3 );
128
//       y[5] = SHIFT_ROUND ( a2 - b2 );
129
//       y[6] = SHIFT_ROUND ( a1 - b1 );
130
//       y[7] = SHIFT_ROUND ( a0 - b0 );
131
// }
132
//
133
//-----------------------------------------------------------------------------
134
//
135
// In this implementation the outputs of the iDCT-1D are multiplied
136
//       for rows 0,4 - by cos_4_16,
137
//       for rows 1,7 - by cos_1_16,
138
//       for rows 2,6 - by cos_2_16,
139
//       for rows 3,5 - by cos_3_16
140
// and are shifted to the left for better accuracy
141
//
142
// For the constants used,
143
//       FIX(float_const) = (short) (float_const * (1<<15) + 0.5)
144
//
145
//-----------------------------------------------------------------------------
146

    
147
//-----------------------------------------------------------------------------
148
// Tables for mmx processors
149
//-----------------------------------------------------------------------------
150

    
151
// Table for rows 0,4 - constants are multiplied by cos_4_16
152
static const int16_t tab_i_04_mmx[32*4] attribute_used __attribute__ ((aligned(8))) = {
153
  16384,16384,16384,-16384,       // movq-> w06 w04 w02 w00
154
  21407,8867,8867,-21407,         // w07 w05 w03 w01
155
  16384,-16384,16384,16384,       // w14 w12 w10 w08
156
  -8867,21407,-21407,-8867,       // w15 w13 w11 w09
157
  22725,12873,19266,-22725,       // w22 w20 w18 w16
158
  19266,4520,-4520,-12873,        // w23 w21 w19 w17
159
  12873,4520,4520,19266,          // w30 w28 w26 w24
160
  -22725,19266,-12873,-22725,     // w31 w29 w27 w25
161
// Table for rows 1,7 - constants are multiplied by cos_1_16
162
  22725,22725,22725,-22725,       // movq-> w06 w04 w02 w00
163
  29692,12299,12299,-29692,       // w07 w05 w03 w01
164
  22725,-22725,22725,22725,       // w14 w12 w10 w08
165
  -12299,29692,-29692,-12299,     // w15 w13 w11 w09
166
  31521,17855,26722,-31521,       // w22 w20 w18 w16
167
  26722,6270,-6270,-17855,        // w23 w21 w19 w17
168
  17855,6270,6270,26722,          // w30 w28 w26 w24
169
  -31521,26722,-17855,-31521,     // w31 w29 w27 w25
170
// Table for rows 2,6 - constants are multiplied by cos_2_16
171
  21407,21407,21407,-21407,       // movq-> w06 w04 w02 w00
172
  27969,11585,11585,-27969,       // w07 w05 w03 w01
173
  21407,-21407,21407,21407,       // w14 w12 w10 w08
174
  -11585,27969,-27969,-11585,     // w15 w13 w11 w09
175
  29692,16819,25172,-29692,       // w22 w20 w18 w16
176
  25172,5906,-5906,-16819,        // w23 w21 w19 w17
177
  16819,5906,5906,25172,          // w30 w28 w26 w24
178
  -29692,25172,-16819,-29692,     // w31 w29 w27 w25
179
// Table for rows 3,5 - constants are multiplied by cos_3_16
180
  19266,19266,19266,-19266,       // movq-> w06 w04 w02 w00
181
  25172,10426,10426,-25172,       // w07 w05 w03 w01
182
  19266,-19266,19266,19266,       // w14 w12 w10 w08
183
  -10426,25172,-25172,-10426,     // w15 w13 w11 w09
184
  26722,15137,22654,-26722,       // w22 w20 w18 w16
185
  22654,5315,-5315,-15137,        // w23 w21 w19 w17
186
  15137,5315,5315,22654,          // w30 w28 w26 w24
187
  -26722,22654,-15137,-26722,     // w31 w29 w27 w25
188
};
189
//-----------------------------------------------------------------------------
190
// Tables for xmm processors
191
//-----------------------------------------------------------------------------
192

    
193
// %3 for rows 0,4 - constants are multiplied by cos_4_16
194
static const int16_t tab_i_04_xmm[32*4] attribute_used __attribute__ ((aligned(8))) = {
195
  16384,21407,16384,8867,      // movq-> w05 w04 w01 w00
196
  16384,8867,-16384,-21407,    // w07 w06 w03 w02
197
  16384,-8867,16384,-21407,    // w13 w12 w09 w08
198
  -16384,21407,16384,-8867,    // w15 w14 w11 w10
199
  22725,19266,19266,-4520,     // w21 w20 w17 w16
200
  12873,4520,-22725,-12873,    // w23 w22 w19 w18
201
  12873,-22725,4520,-12873,    // w29 w28 w25 w24
202
  4520,19266,19266,-22725,     // w31 w30 w27 w26
203
// %3 for rows 1,7 - constants are multiplied by cos_1_16
204
  22725,29692,22725,12299,     // movq-> w05 w04 w01 w00
205
  22725,12299,-22725,-29692,   // w07 w06 w03 w02
206
  22725,-12299,22725,-29692,   // w13 w12 w09 w08
207
  -22725,29692,22725,-12299,   // w15 w14 w11 w10
208
  31521,26722,26722,-6270,     // w21 w20 w17 w16
209
  17855,6270,-31521,-17855,    // w23 w22 w19 w18
210
  17855,-31521,6270,-17855,    // w29 w28 w25 w24
211
  6270,26722,26722,-31521,     // w31 w30 w27 w26
212
// %3 for rows 2,6 - constants are multiplied by cos_2_16
213
  21407,27969,21407,11585,     // movq-> w05 w04 w01 w00
214
  21407,11585,-21407,-27969,   // w07 w06 w03 w02
215
  21407,-11585,21407,-27969,   // w13 w12 w09 w08
216
  -21407,27969,21407,-11585,   // w15 w14 w11 w10
217
  29692,25172,25172,-5906,     // w21 w20 w17 w16
218
  16819,5906,-29692,-16819,    // w23 w22 w19 w18
219
  16819,-29692,5906,-16819,    // w29 w28 w25 w24
220
  5906,25172,25172,-29692,     // w31 w30 w27 w26
221
// %3 for rows 3,5 - constants are multiplied by cos_3_16
222
  19266,25172,19266,10426,     // movq-> w05 w04 w01 w00
223
  19266,10426,-19266,-25172,   // w07 w06 w03 w02
224
  19266,-10426,19266,-25172,   // w13 w12 w09 w08
225
  -19266,25172,19266,-10426,   // w15 w14 w11 w10
226
  26722,22654,22654,-5315,     // w21 w20 w17 w16
227
  15137,5315,-26722,-15137,    // w23 w22 w19 w18
228
  15137,-26722,5315,-15137,    // w29 w28 w25 w24
229
  5315,22654,22654,-26722,     // w31 w30 w27 w26
230
};
231
//=============================================================================
232
// Helper macros for the code
233
//=============================================================================
234

    
235
//-----------------------------------------------------------------------------
236
// DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER
237
//-----------------------------------------------------------------------------
238

    
239
#define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\
240
  "movq " #A1 ",%%mm0              \n\t"/* 0 ; x3 x2 x1 x0*/\
241
  "movq 8+" #A1 ",%%mm1            \n\t"/* 1 ; x7 x6 x5 x4*/\
242
  "movq %%mm0,%%mm2            \n\t"/* 2  ; x3 x2 x1 x0*/\
243
  "movq " #A3 ",%%mm3              \n\t"/* 3 ; w06 w04 w02 w00*/\
244
  "punpcklwd %%mm1,%%mm0       \n\t"/* x5 x1 x4 x0*/\
245
  "movq %%mm0,%%mm5            \n\t"/* 5 ; x5 x1 x4 x0*/\
246
  "punpckldq %%mm0,%%mm0       \n\t"/* x4 x0 x4 x0*/\
247
  "movq 8+" #A3 ",%%mm4            \n\t"/* 4 ; w07 w05 w03 w01*/\
248
  "punpckhwd %%mm1,%%mm2       \n\t"/* 1 ; x7 x3 x6 x2*/\
249
  "pmaddwd %%mm0,%%mm3         \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\
250
  "movq %%mm2,%%mm6            \n\t"/* 6 ; x7 x3 x6 x2*/\
251
  "movq 32+" #A3 ",%%mm1           \n\t"/* 1 ; w22 w20 w18 w16*/\
252
  "punpckldq %%mm2,%%mm2       \n\t"/* x6 x2 x6 x2*/\
253
  "pmaddwd %%mm2,%%mm4         \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\
254
  "punpckhdq %%mm5,%%mm5       \n\t"/* x5 x1 x5 x1*/\
255
  "pmaddwd 16+" #A3 ",%%mm0        \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\
256
  "punpckhdq %%mm6,%%mm6       \n\t"/* x7 x3 x7 x3*/\
257
  "movq 40+" #A3 ",%%mm7           \n\t"/* 7 ; w23 w21 w19 w17*/\
258
  "pmaddwd %%mm5,%%mm1         \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\
259
  "paddd " #A4 ",%%mm3             \n\t"/* +%4*/\
260
  "pmaddwd %%mm6,%%mm7         \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\
261
  "pmaddwd 24+" #A3 ",%%mm2        \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\
262
  "paddd %%mm4,%%mm3           \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
263
  "pmaddwd 48+" #A3 ",%%mm5        \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\
264
  "movq %%mm3,%%mm4            \n\t"/* 4 ; a1 a0*/\
265
  "pmaddwd 56+" #A3 ",%%mm6        \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\
266
  "paddd %%mm7,%%mm1           \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
267
  "paddd " #A4 ",%%mm0             \n\t"/* +%4*/\
268
  "psubd %%mm1,%%mm3           \n\t"/* a1-b1 a0-b0*/\
269
  "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\
270
  "paddd %%mm4,%%mm1           \n\t"/* 4 ; a1+b1 a0+b0*/\
271
  "paddd %%mm2,%%mm0           \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\
272
  "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\
273
  "paddd %%mm6,%%mm5           \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
274
  "movq %%mm0,%%mm4            \n\t"/* 4 ; a3 a2*/\
275
  "paddd %%mm5,%%mm0           \n\t"/* a3+b3 a2+b2*/\
276
  "psubd %%mm5,%%mm4           \n\t"/* 5 ; a3-b3 a2-b2*/\
277
  "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
278
  "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
279
  "packssdw %%mm0,%%mm1        \n\t"/* 0 ; y3 y2 y1 y0*/\
280
  "packssdw %%mm3,%%mm4        \n\t"/* 3 ; y6 y7 y4 y5*/\
281
  "movq %%mm4,%%mm7            \n\t"/* 7 ; y6 y7 y4 y5*/\
282
  "psrld $16,%%mm4            \n\t"/* 0 y6 0 y4*/\
283
  "pslld $16,%%mm7            \n\t"/* y7 0 y5 0*/\
284
  "movq %%mm1," #A2 "              \n\t"/* 1 ; save y3 y2 y1 y0*/\
285
  "por %%mm4,%%mm7             \n\t"/* 4 ; y7 y6 y5 y4*/\
286
  "movq %%mm7,8            +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
287

    
288

    
289
//-----------------------------------------------------------------------------
290
// DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
291
//-----------------------------------------------------------------------------
292

    
293
#define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
294
  "movq " #A1 ",%%mm0                  \n\t"/* 0     ; x3 x2 x1 x0*/\
295
  "movq 8+" #A1 ",%%mm1                \n\t"/* 1     ; x7 x6 x5 x4*/\
296
  "movq %%mm0,%%mm2                \n\t"/* 2     ; x3 x2 x1 x0*/\
297
  "movq " #A3 ",%%mm3                  \n\t"/* 3     ; w05 w04 w01 w00*/\
298
  "pshufw $0b10001000,%%mm0,%%mm0  \n\t"/* x2 x0 x2 x0*/\
299
  "movq 8+" #A3 ",%%mm4                \n\t"/* 4     ; w07 w06 w03 w02*/\
300
  "movq %%mm1,%%mm5                \n\t"/* 5     ; x7 x6 x5 x4*/\
301
  "pmaddwd %%mm0,%%mm3             \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
302
  "movq 32+" #A3 ",%%mm6               \n\t"/* 6     ; w21 w20 w17 w16*/\
303
  "pshufw $0b10001000,%%mm1,%%mm1  \n\t"/* x6 x4 x6 x4*/\
304
  "pmaddwd %%mm1,%%mm4             \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
305
  "movq 40+" #A3 ",%%mm7               \n\t"/* 7    ; w23 w22 w19 w18*/\
306
  "pshufw $0b11011101,%%mm2,%%mm2  \n\t"/* x3 x1 x3 x1*/\
307
  "pmaddwd %%mm2,%%mm6             \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
308
  "pshufw $0b11011101,%%mm5,%%mm5  \n\t"/* x7 x5 x7 x5*/\
309
  "pmaddwd %%mm5,%%mm7             \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
310
  "paddd " #A4 ",%%mm3                 \n\t"/* +%4*/\
311
  "pmaddwd 16+" #A3 ",%%mm0            \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
312
  "paddd %%mm4,%%mm3               \n\t"/* 4     ; a1=sum(even1) a0=sum(even0)*/\
313
  "pmaddwd 24+" #A3 ",%%mm1            \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
314
  "movq %%mm3,%%mm4                \n\t"/* 4     ; a1 a0*/\
315
  "pmaddwd 48+" #A3 ",%%mm2            \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
316
  "paddd %%mm7,%%mm6               \n\t"/* 7     ; b1=sum(odd1) b0=sum(odd0)*/\
317
  "pmaddwd 56+" #A3 ",%%mm5            \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
318
  "paddd %%mm6,%%mm3               \n\t"/* a1+b1 a0+b0*/\
319
  "paddd " #A4 ",%%mm0                 \n\t"/* +%4*/\
320
  "psrad $11,%%mm3     \n\t"/* y1=a1+b1 y0=a0+b0*/\
321
  "paddd %%mm1,%%mm0               \n\t"/* 1     ; a3=sum(even3) a2=sum(even2)*/\
322
  "psubd %%mm6,%%mm4               \n\t"/* 6     ; a1-b1 a0-b0*/\
323
  "movq %%mm0,%%mm7                \n\t"/* 7     ; a3 a2*/\
324
  "paddd %%mm5,%%mm2               \n\t"/* 5     ; b3=sum(odd3) b2=sum(odd2)*/\
325
  "paddd %%mm2,%%mm0               \n\t"/* a3+b3 a2+b2*/\
326
  "psrad $11,%%mm4     \n\t"/* y6=a1-b1 y7=a0-b0*/\
327
  "psubd %%mm2,%%mm7               \n\t"/* 2     ; a3-b3 a2-b2*/\
328
  "psrad $11,%%mm0     \n\t"/* y3=a3+b3 y2=a2+b2*/\
329
  "psrad $11,%%mm7     \n\t"/* y4=a3-b3 y5=a2-b2*/\
330
  "packssdw %%mm0,%%mm3            \n\t"/* 0     ; y3 y2 y1 y0*/\
331
  "packssdw %%mm4,%%mm7            \n\t"/* 4     ; y6 y7 y4 y5*/\
332
  "movq %%mm3, " #A2 "                  \n\t"/* 3     ; save y3 y2 y1 y0*/\
333
  "pshufw $0b10110001,%%mm7,%%mm7  \n\t"/* y7 y6 y5 y4*/\
334
  "movq %%mm7,8                +" #A2 "\n\t"/* 7     ; save y7 y6 y5 y4*/\
335

    
336

    
337
//-----------------------------------------------------------------------------
338
//
339
// The first stage DCT 8x8 - forward DCTs of columns
340
//
341
// The %2puts are multiplied
342
// for rows 0,4 - on cos_4_16,
343
// for rows 1,7 - on cos_1_16,
344
// for rows 2,6 - on cos_2_16,
345
// for rows 3,5 - on cos_3_16
346
// and are shifted to the left for rise of accuracy
347
//
348
//-----------------------------------------------------------------------------
349
//
350
// The 8-point scaled forward DCT algorithm (26a8m)
351
//
352
//-----------------------------------------------------------------------------
353
//
354
// #define DCT_8_FRW_COL(x, y)
355
//{
356
// short t0, t1, t2, t3, t4, t5, t6, t7;
357
// short tp03, tm03, tp12, tm12, tp65, tm65;
358
// short tp465, tm465, tp765, tm765;
359
//
360
// t0 = LEFT_SHIFT ( x[0] + x[7] );
361
// t1 = LEFT_SHIFT ( x[1] + x[6] );
362
// t2 = LEFT_SHIFT ( x[2] + x[5] );
363
// t3 = LEFT_SHIFT ( x[3] + x[4] );
364
// t4 = LEFT_SHIFT ( x[3] - x[4] );
365
// t5 = LEFT_SHIFT ( x[2] - x[5] );
366
// t6 = LEFT_SHIFT ( x[1] - x[6] );
367
// t7 = LEFT_SHIFT ( x[0] - x[7] );
368
//
369
// tp03 = t0 + t3;
370
// tm03 = t0 - t3;
371
// tp12 = t1 + t2;
372
// tm12 = t1 - t2;
373
//
374
// y[0] = tp03 + tp12;
375
// y[4] = tp03 - tp12;
376
//
377
// y[2] = tm03 + tm12 * tg_2_16;
378
// y[6] = tm03 * tg_2_16 - tm12;
379
//
380
// tp65 =(t6 +t5 )*cos_4_16;
381
// tm65 =(t6 -t5 )*cos_4_16;
382
//
383
// tp765 = t7 + tp65;
384
// tm765 = t7 - tp65;
385
// tp465 = t4 + tm65;
386
// tm465 = t4 - tm65;
387
//
388
// y[1] = tp765 + tp465 * tg_1_16;
389
// y[7] = tp765 * tg_1_16 - tp465;
390
// y[5] = tm765 * tg_3_16 + tm465;
391
// y[3] = tm765 - tm465 * tg_3_16;
392
//}
393
//
394
//-----------------------------------------------------------------------------
395

    
396
//-----------------------------------------------------------------------------
397
// DCT_8_INV_COL_4  INP,OUT
398
//-----------------------------------------------------------------------------
399

    
400
#define DCT_8_INV_COL(A1,A2)\
401
  "movq 2*8(%3),%%mm0\n\t"\
402
  "movq 16*3+" #A1 ",%%mm3\n\t"\
403
  "movq %%mm0,%%mm1            \n\t"/* tg_3_16*/\
404
  "movq 16*5+" #A1 ",%%mm5\n\t"\
405
  "pmulhw %%mm3,%%mm0          \n\t"/* x3*(tg_3_16-1)*/\
406
  "movq (%3),%%mm4\n\t"\
407
  "pmulhw %%mm5,%%mm1          \n\t"/* x5*(tg_3_16-1)*/\
408
  "movq 16*7+" #A1 ",%%mm7\n\t"\
409
  "movq %%mm4,%%mm2            \n\t"/* tg_1_16*/\
410
  "movq 16*1+" #A1 ",%%mm6\n\t"\
411
  "pmulhw %%mm7,%%mm4          \n\t"/* x7*tg_1_16*/\
412
  "paddsw %%mm3,%%mm0          \n\t"/* x3*tg_3_16*/\
413
  "pmulhw %%mm6,%%mm2          \n\t"/* x1*tg_1_16*/\
414
  "paddsw %%mm3,%%mm1          \n\t"/* x3+x5*(tg_3_16-1)*/\
415
  "psubsw %%mm5,%%mm0          \n\t"/* x3*tg_3_16-x5 = tm35*/\
416
  "movq 3*8(%3),%%mm3\n\t"\
417
  "paddsw %%mm5,%%mm1          \n\t"/* x3+x5*tg_3_16 = tp35*/\
418
  "paddsw %%mm6,%%mm4          \n\t"/* x1+tg_1_16*x7 = tp17*/\
419
  "psubsw %%mm7,%%mm2          \n\t"/* x1*tg_1_16-x7 = tm17*/\
420
  "movq %%mm4,%%mm5            \n\t"/* tp17*/\
421
  "movq %%mm2,%%mm6            \n\t"/* tm17*/\
422
  "paddsw %%mm1,%%mm5          \n\t"/* tp17+tp35 = b0*/\
423
  "psubsw %%mm0,%%mm6          \n\t"/* tm17-tm35 = b3*/\
424
  "psubsw %%mm1,%%mm4          \n\t"/* tp17-tp35 = t1*/\
425
  "paddsw %%mm0,%%mm2          \n\t"/* tm17+tm35 = t2*/\
426
  "movq 1*8(%3),%%mm7\n\t"\
427
  "movq %%mm4,%%mm1            \n\t"/* t1*/\
428
  "movq %%mm5,3*16         +" #A2 "\n\t"/* save b0*/\
429
  "paddsw %%mm2,%%mm1          \n\t"/* t1+t2*/\
430
  "movq %%mm6,5*16         +" #A2 "\n\t"/* save b3*/\
431
  "psubsw %%mm2,%%mm4          \n\t"/* t1-t2*/\
432
  "movq 2*16+" #A1 ",%%mm5\n\t"\
433
  "movq %%mm7,%%mm0            \n\t"/* tg_2_16*/\
434
  "movq 6*16+" #A1 ",%%mm6\n\t"\
435
  "pmulhw %%mm5,%%mm0          \n\t"/* x2*tg_2_16*/\
436
  "pmulhw %%mm6,%%mm7          \n\t"/* x6*tg_2_16*/\
437
  "pmulhw %%mm3,%%mm1          \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
438
  "movq 0*16+" #A1 ",%%mm2\n\t"\
439
  "pmulhw %%mm3,%%mm4          \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
440
  "psubsw %%mm6,%%mm0          \n\t"/* t2*tg_2_16-x6 = tm26*/\
441
  "movq %%mm2,%%mm3            \n\t"/* x0*/\
442
  "movq 4*16+" #A1 ",%%mm6\n\t"\
443
  "paddsw %%mm5,%%mm7          \n\t"/* x2+x6*tg_2_16 = tp26*/\
444
  "paddsw %%mm6,%%mm2          \n\t"/* x0+x4 = tp04*/\
445
  "psubsw %%mm6,%%mm3          \n\t"/* x0-x4 = tm04*/\
446
  "movq %%mm2,%%mm5            \n\t"/* tp04*/\
447
  "movq %%mm3,%%mm6            \n\t"/* tm04*/\
448
  "psubsw %%mm7,%%mm2          \n\t"/* tp04-tp26 = a3*/\
449
  "paddsw %%mm0,%%mm3          \n\t"/* tm04+tm26 = a1*/\
450
  "paddsw %%mm1,%%mm1          \n\t"/* b1*/\
451
  "paddsw %%mm4,%%mm4          \n\t"/* b2*/\
452
  "paddsw %%mm7,%%mm5          \n\t"/* tp04+tp26 = a0*/\
453
  "psubsw %%mm0,%%mm6          \n\t"/* tm04-tm26 = a2*/\
454
  "movq %%mm3,%%mm7            \n\t"/* a1*/\
455
  "movq %%mm6,%%mm0            \n\t"/* a2*/\
456
  "paddsw %%mm1,%%mm3          \n\t"/* a1+b1*/\
457
  "paddsw %%mm4,%%mm6          \n\t"/* a2+b2*/\
458
  "psraw $6,%%mm3 \n\t"/* dst1*/\
459
  "psubsw %%mm1,%%mm7          \n\t"/* a1-b1*/\
460
  "psraw $6,%%mm6 \n\t"/* dst2*/\
461
  "psubsw %%mm4,%%mm0          \n\t"/* a2-b2*/\
462
  "movq 3*16+" #A2 ",%%mm1         \n\t"/* load b0*/\
463
  "psraw $6,%%mm7 \n\t"/* dst6*/\
464
  "movq %%mm5,%%mm4            \n\t"/* a0*/\
465
  "psraw $6,%%mm0 \n\t"/* dst5*/\
466
  "movq %%mm3,1*16+" #A2 "\n\t"\
467
  "paddsw %%mm1,%%mm5          \n\t"/* a0+b0*/\
468
  "movq %%mm6,2*16+" #A2 "\n\t"\
469
  "psubsw %%mm1,%%mm4          \n\t"/* a0-b0*/\
470
  "movq 5*16+" #A2 ",%%mm3         \n\t"/* load b3*/\
471
  "psraw $6,%%mm5 \n\t"/* dst0*/\
472
  "movq %%mm2,%%mm6            \n\t"/* a3*/\
473
  "psraw $6,%%mm4 \n\t"/* dst7*/\
474
  "movq %%mm0,5*16+" #A2 "\n\t"\
475
  "paddsw %%mm3,%%mm2          \n\t"/* a3+b3*/\
476
  "movq %%mm7,6*16+" #A2 "\n\t"\
477
  "psubsw %%mm3,%%mm6          \n\t"/* a3-b3*/\
478
  "movq %%mm5,0*16+" #A2 "\n\t"\
479
  "psraw $6,%%mm2 \n\t"/* dst3*/\
480
  "movq %%mm4,7*16+" #A2 "\n\t"\
481
  "psraw $6,%%mm6 \n\t"/* dst4*/\
482
  "movq %%mm2,3*16+" #A2 "\n\t"\
483
  "movq %%mm6,4*16+" #A2 "\n\t"
484

    
485
//=============================================================================
486
// Code
487
//=============================================================================
488

    
489
//-----------------------------------------------------------------------------
490
// void idct_mmx(uint16_t block[64]);
491
//-----------------------------------------------------------------------------
492

    
493

    
494
void ff_idct_xvid_mmx(short *block){
495
asm volatile(
496
            //# Process each row
497
    DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
498
    DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
499
    DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
500
    DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
501
    DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
502
    DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
503
    DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
504
    DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
505

    
506
            //# Process the columns (4 at a time)
507
    DCT_8_INV_COL(0(%0), 0(%0))
508
    DCT_8_INV_COL(8(%0), 8(%0))
509
    :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
510
}
511

    
512
//-----------------------------------------------------------------------------
513
// void idct_xmm(uint16_t block[64]);
514
//-----------------------------------------------------------------------------
515

    
516

    
517
void ff_idct_xvid_mmx2(short *block){
518
asm volatile(
519
            //# Process each row
520
    DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
521
    DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
522
    DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
523
    DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
524
    DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
525
    DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
526
    DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
527
    DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
528

    
529
            //# Process the columns (4 at a time)
530
    DCT_8_INV_COL(0(%0), 0(%0))
531
    DCT_8_INV_COL(8(%0), 8(%0))
532
    :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
533
}
534