FFmpeg
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
idct_mmx_xvid.c
Go to the documentation of this file.
1 /*
2  * XVID MPEG-4 VIDEO CODEC
3  * - MMX and XMM forward discrete cosine transform -
4  *
5  * Copyright(C) 2001 Peter Ross <pross@xvid.org>
6  *
7  * Originally provided by Intel at AP-922
8  * http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
9  * (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
10  * but in a limited edition.
11  * New macro implements a column part for precise iDCT
12  * The routine precision now satisfies IEEE standard 1180-1990.
13  *
14  * Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
15  * Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org>
16  *
17  * http://www.elecard.com/peter/idct.html
18  * http://www.linuxvideo.org/mpeg2dec/
19  *
20  * These examples contain code fragments for first stage iDCT 8x8
21  * (for rows) and first stage DCT 8x8 (for columns)
22  *
23  * conversion to gcc syntax by Michael Niedermayer
24  *
25  * This file is part of FFmpeg.
26  *
27  * FFmpeg is free software; you can redistribute it and/or
28  * modify it under the terms of the GNU Lesser General Public
29  * License as published by the Free Software Foundation; either
30  * version 2.1 of the License, or (at your option) any later version.
31  *
32  * FFmpeg is distributed in the hope that it will be useful,
33  * but WITHOUT ANY WARRANTY; without even the implied warranty of
34  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
35  * Lesser General Public License for more details.
36  *
37  * You should have received a copy of the GNU Lesser General Public License
38  * along with FFmpeg; if not, write to the Free Software Foundation,
39  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
40  */
41 
42 #include <inttypes.h>
43 
44 #include "config.h"
45 #include "libavcodec/avcodec.h"
46 #include "libavutil/mem.h"
47 #include "dsputil_x86.h"
48 #include "idct_xvid.h"
49 
50 #if HAVE_INLINE_ASM
51 
52 //=============================================================================
53 // Macros and other preprocessor constants
54 //=============================================================================
55 
56 #define BITS_INV_ACC 5 // 4 or 5 for IEEE
57 #define SHIFT_INV_ROW (16 - BITS_INV_ACC) //11
58 #define SHIFT_INV_COL (1 + BITS_INV_ACC) //6
59 #define RND_INV_ROW (1024 * (6 - BITS_INV_ACC))
60 #define RND_INV_COL (16 * (BITS_INV_ACC - 3))
61 #define RND_INV_CORR (RND_INV_COL - 1)
62 
63 #define BITS_FRW_ACC 3 // 2 or 3 for accuracy
64 #define SHIFT_FRW_COL BITS_FRW_ACC
65 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17)
66 #define RND_FRW_ROW (262144*(BITS_FRW_ACC - 1))
67 
68 
69 //-----------------------------------------------------------------------------
70 // Various memory constants (trigonometric values or rounding values)
71 //-----------------------------------------------------------------------------
72 
73 
74 DECLARE_ALIGNED(8, static const int16_t, tg_1_16)[4*4] = {
75  13036,13036,13036,13036, // tg * (2<<16) + 0.5
76  27146,27146,27146,27146, // tg * (2<<16) + 0.5
77  -21746,-21746,-21746,-21746, // tg * (2<<16) + 0.5
78  23170,23170,23170,23170}; // cos * (2<<15) + 0.5
79 
80 DECLARE_ALIGNED(8, static const int32_t, rounder_0)[2*8] = {
81  65536,65536,
82  3597,3597,
83  2260,2260,
84  1203,1203,
85  0,0,
86  120,120,
87  512,512,
88  512,512};
89 
90 //-----------------------------------------------------------------------------
91 //
92 // The first stage iDCT 8x8 - inverse DCTs of rows
93 //
94 //-----------------------------------------------------------------------------
95 // The 8-point inverse DCT direct algorithm
96 //-----------------------------------------------------------------------------
97 //
98 // static const short w[32] = {
99 // FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16),
100 // FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16),
101 // FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16),
102 // FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16),
103 // FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16),
104 // FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16),
105 // FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16),
106 // FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) };
107 //
108 // #define DCT_8_INV_ROW(x, y)
109 // {
110 // int a0, a1, a2, a3, b0, b1, b2, b3;
111 //
112 // a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3];
113 // a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7];
114 // a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11];
115 // a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15];
116 // b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19];
117 // b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23];
118 // b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27];
119 // b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31];
120 //
121 // y[0] = SHIFT_ROUND ( a0 + b0 );
122 // y[1] = SHIFT_ROUND ( a1 + b1 );
123 // y[2] = SHIFT_ROUND ( a2 + b2 );
124 // y[3] = SHIFT_ROUND ( a3 + b3 );
125 // y[4] = SHIFT_ROUND ( a3 - b3 );
126 // y[5] = SHIFT_ROUND ( a2 - b2 );
127 // y[6] = SHIFT_ROUND ( a1 - b1 );
128 // y[7] = SHIFT_ROUND ( a0 - b0 );
129 // }
130 //
131 //-----------------------------------------------------------------------------
132 //
133 // In this implementation the outputs of the iDCT-1D are multiplied
134 // for rows 0,4 - by cos_4_16,
135 // for rows 1,7 - by cos_1_16,
136 // for rows 2,6 - by cos_2_16,
137 // for rows 3,5 - by cos_3_16
138 // and are shifted to the left for better accuracy
139 //
140 // For the constants used,
141 // FIX(float_const) = (short) (float_const * (1<<15) + 0.5)
142 //
143 //-----------------------------------------------------------------------------
144 
145 //-----------------------------------------------------------------------------
146 // Tables for mmx processors
147 //-----------------------------------------------------------------------------
148 
149 // Table for rows 0,4 - constants are multiplied by cos_4_16
150 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_mmx)[32*4] = {
151  16384,16384,16384,-16384, // movq-> w06 w04 w02 w00
152  21407,8867,8867,-21407, // w07 w05 w03 w01
153  16384,-16384,16384,16384, // w14 w12 w10 w08
154  -8867,21407,-21407,-8867, // w15 w13 w11 w09
155  22725,12873,19266,-22725, // w22 w20 w18 w16
156  19266,4520,-4520,-12873, // w23 w21 w19 w17
157  12873,4520,4520,19266, // w30 w28 w26 w24
158  -22725,19266,-12873,-22725, // w31 w29 w27 w25
159 // Table for rows 1,7 - constants are multiplied by cos_1_16
160  22725,22725,22725,-22725, // movq-> w06 w04 w02 w00
161  29692,12299,12299,-29692, // w07 w05 w03 w01
162  22725,-22725,22725,22725, // w14 w12 w10 w08
163  -12299,29692,-29692,-12299, // w15 w13 w11 w09
164  31521,17855,26722,-31521, // w22 w20 w18 w16
165  26722,6270,-6270,-17855, // w23 w21 w19 w17
166  17855,6270,6270,26722, // w30 w28 w26 w24
167  -31521,26722,-17855,-31521, // w31 w29 w27 w25
168 // Table for rows 2,6 - constants are multiplied by cos_2_16
169  21407,21407,21407,-21407, // movq-> w06 w04 w02 w00
170  27969,11585,11585,-27969, // w07 w05 w03 w01
171  21407,-21407,21407,21407, // w14 w12 w10 w08
172  -11585,27969,-27969,-11585, // w15 w13 w11 w09
173  29692,16819,25172,-29692, // w22 w20 w18 w16
174  25172,5906,-5906,-16819, // w23 w21 w19 w17
175  16819,5906,5906,25172, // w30 w28 w26 w24
176  -29692,25172,-16819,-29692, // w31 w29 w27 w25
177 // Table for rows 3,5 - constants are multiplied by cos_3_16
178  19266,19266,19266,-19266, // movq-> w06 w04 w02 w00
179  25172,10426,10426,-25172, // w07 w05 w03 w01
180  19266,-19266,19266,19266, // w14 w12 w10 w08
181  -10426,25172,-25172,-10426, // w15 w13 w11 w09
182  26722,15137,22654,-26722, // w22 w20 w18 w16
183  22654,5315,-5315,-15137, // w23 w21 w19 w17
184  15137,5315,5315,22654, // w30 w28 w26 w24
185  -26722,22654,-15137,-26722, // w31 w29 w27 w25
186 };
187 //-----------------------------------------------------------------------------
188 // Tables for xmm processors
189 //-----------------------------------------------------------------------------
190 
191 // %3 for rows 0,4 - constants are multiplied by cos_4_16
192 DECLARE_ALIGNED(8, static const int16_t, tab_i_04_xmm)[32*4] = {
193  16384,21407,16384,8867, // movq-> w05 w04 w01 w00
194  16384,8867,-16384,-21407, // w07 w06 w03 w02
195  16384,-8867,16384,-21407, // w13 w12 w09 w08
196  -16384,21407,16384,-8867, // w15 w14 w11 w10
197  22725,19266,19266,-4520, // w21 w20 w17 w16
198  12873,4520,-22725,-12873, // w23 w22 w19 w18
199  12873,-22725,4520,-12873, // w29 w28 w25 w24
200  4520,19266,19266,-22725, // w31 w30 w27 w26
201 // %3 for rows 1,7 - constants are multiplied by cos_1_16
202  22725,29692,22725,12299, // movq-> w05 w04 w01 w00
203  22725,12299,-22725,-29692, // w07 w06 w03 w02
204  22725,-12299,22725,-29692, // w13 w12 w09 w08
205  -22725,29692,22725,-12299, // w15 w14 w11 w10
206  31521,26722,26722,-6270, // w21 w20 w17 w16
207  17855,6270,-31521,-17855, // w23 w22 w19 w18
208  17855,-31521,6270,-17855, // w29 w28 w25 w24
209  6270,26722,26722,-31521, // w31 w30 w27 w26
210 // %3 for rows 2,6 - constants are multiplied by cos_2_16
211  21407,27969,21407,11585, // movq-> w05 w04 w01 w00
212  21407,11585,-21407,-27969, // w07 w06 w03 w02
213  21407,-11585,21407,-27969, // w13 w12 w09 w08
214  -21407,27969,21407,-11585, // w15 w14 w11 w10
215  29692,25172,25172,-5906, // w21 w20 w17 w16
216  16819,5906,-29692,-16819, // w23 w22 w19 w18
217  16819,-29692,5906,-16819, // w29 w28 w25 w24
218  5906,25172,25172,-29692, // w31 w30 w27 w26
219 // %3 for rows 3,5 - constants are multiplied by cos_3_16
220  19266,25172,19266,10426, // movq-> w05 w04 w01 w00
221  19266,10426,-19266,-25172, // w07 w06 w03 w02
222  19266,-10426,19266,-25172, // w13 w12 w09 w08
223  -19266,25172,19266,-10426, // w15 w14 w11 w10
224  26722,22654,22654,-5315, // w21 w20 w17 w16
225  15137,5315,-26722,-15137, // w23 w22 w19 w18
226  15137,-26722,5315,-15137, // w29 w28 w25 w24
227  5315,22654,22654,-26722, // w31 w30 w27 w26
228 };
229 //=============================================================================
230 // Helper macros for the code
231 //=============================================================================
232 
233 //-----------------------------------------------------------------------------
234 // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER
235 //-----------------------------------------------------------------------------
236 
237 #define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\
238  "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
239  "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
240  "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
241  "movq " #A3 ",%%mm3 \n\t"/* 3 ; w06 w04 w02 w00*/\
242  "punpcklwd %%mm1,%%mm0 \n\t"/* x5 x1 x4 x0*/\
243  "movq %%mm0,%%mm5 \n\t"/* 5 ; x5 x1 x4 x0*/\
244  "punpckldq %%mm0,%%mm0 \n\t"/* x4 x0 x4 x0*/\
245  "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w05 w03 w01*/\
246  "punpckhwd %%mm1,%%mm2 \n\t"/* 1 ; x7 x3 x6 x2*/\
247  "pmaddwd %%mm0,%%mm3 \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\
248  "movq %%mm2,%%mm6 \n\t"/* 6 ; x7 x3 x6 x2*/\
249  "movq 32+" #A3 ",%%mm1 \n\t"/* 1 ; w22 w20 w18 w16*/\
250  "punpckldq %%mm2,%%mm2 \n\t"/* x6 x2 x6 x2*/\
251  "pmaddwd %%mm2,%%mm4 \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\
252  "punpckhdq %%mm5,%%mm5 \n\t"/* x5 x1 x5 x1*/\
253  "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\
254  "punpckhdq %%mm6,%%mm6 \n\t"/* x7 x3 x7 x3*/\
255  "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w21 w19 w17*/\
256  "pmaddwd %%mm5,%%mm1 \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\
257  "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
258  "pmaddwd %%mm6,%%mm7 \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\
259  "pmaddwd 24+" #A3 ",%%mm2 \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\
260  "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
261  "pmaddwd 48+" #A3 ",%%mm5 \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\
262  "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
263  "pmaddwd 56+" #A3 ",%%mm6 \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\
264  "paddd %%mm7,%%mm1 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
265  "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
266  "psubd %%mm1,%%mm3 \n\t"/* a1-b1 a0-b0*/\
267  "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\
268  "paddd %%mm4,%%mm1 \n\t"/* 4 ; a1+b1 a0+b0*/\
269  "paddd %%mm2,%%mm0 \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\
270  "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\
271  "paddd %%mm6,%%mm5 \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
272  "movq %%mm0,%%mm4 \n\t"/* 4 ; a3 a2*/\
273  "paddd %%mm5,%%mm0 \n\t"/* a3+b3 a2+b2*/\
274  "psubd %%mm5,%%mm4 \n\t"/* 5 ; a3-b3 a2-b2*/\
275  "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
276  "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
277  "packssdw %%mm0,%%mm1 \n\t"/* 0 ; y3 y2 y1 y0*/\
278  "packssdw %%mm3,%%mm4 \n\t"/* 3 ; y6 y7 y4 y5*/\
279  "movq %%mm4,%%mm7 \n\t"/* 7 ; y6 y7 y4 y5*/\
280  "psrld $16,%%mm4 \n\t"/* 0 y6 0 y4*/\
281  "pslld $16,%%mm7 \n\t"/* y7 0 y5 0*/\
282  "movq %%mm1," #A2 " \n\t"/* 1 ; save y3 y2 y1 y0*/\
283  "por %%mm4,%%mm7 \n\t"/* 4 ; y7 y6 y5 y4*/\
284  "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
285 
286 
287 //-----------------------------------------------------------------------------
288 // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
289 //-----------------------------------------------------------------------------
290 
291 #define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
292  "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
293  "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
294  "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
295  "movq " #A3 ",%%mm3 \n\t"/* 3 ; w05 w04 w01 w00*/\
296  "pshufw $0x88,%%mm0,%%mm0 \n\t"/* x2 x0 x2 x0*/\
297  "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w06 w03 w02*/\
298  "movq %%mm1,%%mm5 \n\t"/* 5 ; x7 x6 x5 x4*/\
299  "pmaddwd %%mm0,%%mm3 \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
300  "movq 32+" #A3 ",%%mm6 \n\t"/* 6 ; w21 w20 w17 w16*/\
301  "pshufw $0x88,%%mm1,%%mm1 \n\t"/* x6 x4 x6 x4*/\
302  "pmaddwd %%mm1,%%mm4 \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
303  "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w22 w19 w18*/\
304  "pshufw $0xdd,%%mm2,%%mm2 \n\t"/* x3 x1 x3 x1*/\
305  "pmaddwd %%mm2,%%mm6 \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
306  "pshufw $0xdd,%%mm5,%%mm5 \n\t"/* x7 x5 x7 x5*/\
307  "pmaddwd %%mm5,%%mm7 \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
308  "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
309  "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
310  "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
311  "pmaddwd 24+" #A3 ",%%mm1 \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
312  "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
313  "pmaddwd 48+" #A3 ",%%mm2 \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
314  "paddd %%mm7,%%mm6 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
315  "pmaddwd 56+" #A3 ",%%mm5 \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
316  "paddd %%mm6,%%mm3 \n\t"/* a1+b1 a0+b0*/\
317  "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
318  "psrad $11,%%mm3 \n\t"/* y1=a1+b1 y0=a0+b0*/\
319  "paddd %%mm1,%%mm0 \n\t"/* 1 ; a3=sum(even3) a2=sum(even2)*/\
320  "psubd %%mm6,%%mm4 \n\t"/* 6 ; a1-b1 a0-b0*/\
321  "movq %%mm0,%%mm7 \n\t"/* 7 ; a3 a2*/\
322  "paddd %%mm5,%%mm2 \n\t"/* 5 ; b3=sum(odd3) b2=sum(odd2)*/\
323  "paddd %%mm2,%%mm0 \n\t"/* a3+b3 a2+b2*/\
324  "psrad $11,%%mm4 \n\t"/* y6=a1-b1 y7=a0-b0*/\
325  "psubd %%mm2,%%mm7 \n\t"/* 2 ; a3-b3 a2-b2*/\
326  "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
327  "psrad $11,%%mm7 \n\t"/* y4=a3-b3 y5=a2-b2*/\
328  "packssdw %%mm0,%%mm3 \n\t"/* 0 ; y3 y2 y1 y0*/\
329  "packssdw %%mm4,%%mm7 \n\t"/* 4 ; y6 y7 y4 y5*/\
330  "movq %%mm3, " #A2 " \n\t"/* 3 ; save y3 y2 y1 y0*/\
331  "pshufw $0xb1,%%mm7,%%mm7 \n\t"/* y7 y6 y5 y4*/\
332  "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
333 
334 
335 //-----------------------------------------------------------------------------
336 //
337 // The first stage DCT 8x8 - forward DCTs of columns
338 //
339 // The %2puts are multiplied
340 // for rows 0,4 - on cos_4_16,
341 // for rows 1,7 - on cos_1_16,
342 // for rows 2,6 - on cos_2_16,
343 // for rows 3,5 - on cos_3_16
344 // and are shifted to the left for rise of accuracy
345 //
346 //-----------------------------------------------------------------------------
347 //
348 // The 8-point scaled forward DCT algorithm (26a8m)
349 //
350 //-----------------------------------------------------------------------------
351 //
352 // #define DCT_8_FRW_COL(x, y)
353 //{
354 // short t0, t1, t2, t3, t4, t5, t6, t7;
355 // short tp03, tm03, tp12, tm12, tp65, tm65;
356 // short tp465, tm465, tp765, tm765;
357 //
358 // t0 = LEFT_SHIFT ( x[0] + x[7] );
359 // t1 = LEFT_SHIFT ( x[1] + x[6] );
360 // t2 = LEFT_SHIFT ( x[2] + x[5] );
361 // t3 = LEFT_SHIFT ( x[3] + x[4] );
362 // t4 = LEFT_SHIFT ( x[3] - x[4] );
363 // t5 = LEFT_SHIFT ( x[2] - x[5] );
364 // t6 = LEFT_SHIFT ( x[1] - x[6] );
365 // t7 = LEFT_SHIFT ( x[0] - x[7] );
366 //
367 // tp03 = t0 + t3;
368 // tm03 = t0 - t3;
369 // tp12 = t1 + t2;
370 // tm12 = t1 - t2;
371 //
372 // y[0] = tp03 + tp12;
373 // y[4] = tp03 - tp12;
374 //
375 // y[2] = tm03 + tm12 * tg_2_16;
376 // y[6] = tm03 * tg_2_16 - tm12;
377 //
378 // tp65 =(t6 +t5 )*cos_4_16;
379 // tm65 =(t6 -t5 )*cos_4_16;
380 //
381 // tp765 = t7 + tp65;
382 // tm765 = t7 - tp65;
383 // tp465 = t4 + tm65;
384 // tm465 = t4 - tm65;
385 //
386 // y[1] = tp765 + tp465 * tg_1_16;
387 // y[7] = tp765 * tg_1_16 - tp465;
388 // y[5] = tm765 * tg_3_16 + tm465;
389 // y[3] = tm765 - tm465 * tg_3_16;
390 //}
391 //
392 //-----------------------------------------------------------------------------
393 
394 //-----------------------------------------------------------------------------
395 // DCT_8_INV_COL_4 INP,OUT
396 //-----------------------------------------------------------------------------
397 
398 #define DCT_8_INV_COL(A1,A2)\
399  "movq 2*8(%3),%%mm0\n\t"\
400  "movq 16*3+" #A1 ",%%mm3\n\t"\
401  "movq %%mm0,%%mm1 \n\t"/* tg_3_16*/\
402  "movq 16*5+" #A1 ",%%mm5\n\t"\
403  "pmulhw %%mm3,%%mm0 \n\t"/* x3*(tg_3_16-1)*/\
404  "movq (%3),%%mm4\n\t"\
405  "pmulhw %%mm5,%%mm1 \n\t"/* x5*(tg_3_16-1)*/\
406  "movq 16*7+" #A1 ",%%mm7\n\t"\
407  "movq %%mm4,%%mm2 \n\t"/* tg_1_16*/\
408  "movq 16*1+" #A1 ",%%mm6\n\t"\
409  "pmulhw %%mm7,%%mm4 \n\t"/* x7*tg_1_16*/\
410  "paddsw %%mm3,%%mm0 \n\t"/* x3*tg_3_16*/\
411  "pmulhw %%mm6,%%mm2 \n\t"/* x1*tg_1_16*/\
412  "paddsw %%mm3,%%mm1 \n\t"/* x3+x5*(tg_3_16-1)*/\
413  "psubsw %%mm5,%%mm0 \n\t"/* x3*tg_3_16-x5 = tm35*/\
414  "movq 3*8(%3),%%mm3\n\t"\
415  "paddsw %%mm5,%%mm1 \n\t"/* x3+x5*tg_3_16 = tp35*/\
416  "paddsw %%mm6,%%mm4 \n\t"/* x1+tg_1_16*x7 = tp17*/\
417  "psubsw %%mm7,%%mm2 \n\t"/* x1*tg_1_16-x7 = tm17*/\
418  "movq %%mm4,%%mm5 \n\t"/* tp17*/\
419  "movq %%mm2,%%mm6 \n\t"/* tm17*/\
420  "paddsw %%mm1,%%mm5 \n\t"/* tp17+tp35 = b0*/\
421  "psubsw %%mm0,%%mm6 \n\t"/* tm17-tm35 = b3*/\
422  "psubsw %%mm1,%%mm4 \n\t"/* tp17-tp35 = t1*/\
423  "paddsw %%mm0,%%mm2 \n\t"/* tm17+tm35 = t2*/\
424  "movq 1*8(%3),%%mm7\n\t"\
425  "movq %%mm4,%%mm1 \n\t"/* t1*/\
426  "movq %%mm5,3*16 +" #A2 "\n\t"/* save b0*/\
427  "paddsw %%mm2,%%mm1 \n\t"/* t1+t2*/\
428  "movq %%mm6,5*16 +" #A2 "\n\t"/* save b3*/\
429  "psubsw %%mm2,%%mm4 \n\t"/* t1-t2*/\
430  "movq 2*16+" #A1 ",%%mm5\n\t"\
431  "movq %%mm7,%%mm0 \n\t"/* tg_2_16*/\
432  "movq 6*16+" #A1 ",%%mm6\n\t"\
433  "pmulhw %%mm5,%%mm0 \n\t"/* x2*tg_2_16*/\
434  "pmulhw %%mm6,%%mm7 \n\t"/* x6*tg_2_16*/\
435  "pmulhw %%mm3,%%mm1 \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
436  "movq 0*16+" #A1 ",%%mm2\n\t"\
437  "pmulhw %%mm3,%%mm4 \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
438  "psubsw %%mm6,%%mm0 \n\t"/* t2*tg_2_16-x6 = tm26*/\
439  "movq %%mm2,%%mm3 \n\t"/* x0*/\
440  "movq 4*16+" #A1 ",%%mm6\n\t"\
441  "paddsw %%mm5,%%mm7 \n\t"/* x2+x6*tg_2_16 = tp26*/\
442  "paddsw %%mm6,%%mm2 \n\t"/* x0+x4 = tp04*/\
443  "psubsw %%mm6,%%mm3 \n\t"/* x0-x4 = tm04*/\
444  "movq %%mm2,%%mm5 \n\t"/* tp04*/\
445  "movq %%mm3,%%mm6 \n\t"/* tm04*/\
446  "psubsw %%mm7,%%mm2 \n\t"/* tp04-tp26 = a3*/\
447  "paddsw %%mm0,%%mm3 \n\t"/* tm04+tm26 = a1*/\
448  "paddsw %%mm1,%%mm1 \n\t"/* b1*/\
449  "paddsw %%mm4,%%mm4 \n\t"/* b2*/\
450  "paddsw %%mm7,%%mm5 \n\t"/* tp04+tp26 = a0*/\
451  "psubsw %%mm0,%%mm6 \n\t"/* tm04-tm26 = a2*/\
452  "movq %%mm3,%%mm7 \n\t"/* a1*/\
453  "movq %%mm6,%%mm0 \n\t"/* a2*/\
454  "paddsw %%mm1,%%mm3 \n\t"/* a1+b1*/\
455  "paddsw %%mm4,%%mm6 \n\t"/* a2+b2*/\
456  "psraw $6,%%mm3 \n\t"/* dst1*/\
457  "psubsw %%mm1,%%mm7 \n\t"/* a1-b1*/\
458  "psraw $6,%%mm6 \n\t"/* dst2*/\
459  "psubsw %%mm4,%%mm0 \n\t"/* a2-b2*/\
460  "movq 3*16+" #A2 ",%%mm1 \n\t"/* load b0*/\
461  "psraw $6,%%mm7 \n\t"/* dst6*/\
462  "movq %%mm5,%%mm4 \n\t"/* a0*/\
463  "psraw $6,%%mm0 \n\t"/* dst5*/\
464  "movq %%mm3,1*16+" #A2 "\n\t"\
465  "paddsw %%mm1,%%mm5 \n\t"/* a0+b0*/\
466  "movq %%mm6,2*16+" #A2 "\n\t"\
467  "psubsw %%mm1,%%mm4 \n\t"/* a0-b0*/\
468  "movq 5*16+" #A2 ",%%mm3 \n\t"/* load b3*/\
469  "psraw $6,%%mm5 \n\t"/* dst0*/\
470  "movq %%mm2,%%mm6 \n\t"/* a3*/\
471  "psraw $6,%%mm4 \n\t"/* dst7*/\
472  "movq %%mm0,5*16+" #A2 "\n\t"\
473  "paddsw %%mm3,%%mm2 \n\t"/* a3+b3*/\
474  "movq %%mm7,6*16+" #A2 "\n\t"\
475  "psubsw %%mm3,%%mm6 \n\t"/* a3-b3*/\
476  "movq %%mm5,0*16+" #A2 "\n\t"\
477  "psraw $6,%%mm2 \n\t"/* dst3*/\
478  "movq %%mm4,7*16+" #A2 "\n\t"\
479  "psraw $6,%%mm6 \n\t"/* dst4*/\
480  "movq %%mm2,3*16+" #A2 "\n\t"\
481  "movq %%mm6,4*16+" #A2 "\n\t"
482 
483 //=============================================================================
484 // Code
485 //=============================================================================
486 
487 //-----------------------------------------------------------------------------
488 // void idct_mmx(uint16_t block[64]);
489 //-----------------------------------------------------------------------------
490 
491 
492 void ff_idct_xvid_mmx(short *block){
493 __asm__ volatile(
494  //# Process each row
495  DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
496  DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
497  DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
498  DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
499  DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
500  DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
501  DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
502  DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
503 
504  //# Process the columns (4 at a time)
505  DCT_8_INV_COL(0(%0), 0(%0))
506  DCT_8_INV_COL(8(%0), 8(%0))
507  :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
508 }
509 
510 //-----------------------------------------------------------------------------
511 // void idct_xmm(uint16_t block[64]);
512 //-----------------------------------------------------------------------------
513 
514 
515 void ff_idct_xvid_mmxext(short *block)
516 {
517 __asm__ volatile(
518  //# Process each row
519  DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
520  DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
521  DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
522  DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
523  DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
524  DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
525  DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
526  DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
527 
528  //# Process the columns (4 at a time)
529  DCT_8_INV_COL(0(%0), 0(%0))
530  DCT_8_INV_COL(8(%0), 8(%0))
531  :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
532 }
533 
534 void ff_idct_xvid_mmx_put(uint8_t *dest, int line_size, int16_t *block)
535 {
536  ff_idct_xvid_mmx(block);
537  ff_put_pixels_clamped_mmx(block, dest, line_size);
538 }
539 
540 void ff_idct_xvid_mmx_add(uint8_t *dest, int line_size, int16_t *block)
541 {
542  ff_idct_xvid_mmx(block);
543  ff_add_pixels_clamped_mmx(block, dest, line_size);
544 }
545 
546 void ff_idct_xvid_mmxext_put(uint8_t *dest, int line_size, int16_t *block)
547 {
548  ff_idct_xvid_mmxext(block);
549  ff_put_pixels_clamped_mmx(block, dest, line_size);
550 }
551 
552 void ff_idct_xvid_mmxext_add(uint8_t *dest, int line_size, int16_t *block)
553 {
554  ff_idct_xvid_mmxext(block);
555  ff_add_pixels_clamped_mmx(block, dest, line_size);
556 }
557 
558 #endif /* HAVE_INLINE_ASM */