Parent Directory | Revision Log
Revision 1.4 - (view) (download)
1 : | edgomez | 1.2 | ;/**************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - SSE2 inverse discrete cosine transform - | ||
5 : | ; * | ||
6 : | ; * Copyright(C) 2002 Dmitry Rozhdestvensky | ||
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 | ||
20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | ; * | ||
22 : | ; * $Id$ | ||
23 : | ; * | ||
24 : | ; ***************************************************************************/ | ||
25 : | |||
26 : | BITS 32 | ||
27 : | |||
28 : | ;============================================================================= | ||
29 : | ; Macros and other preprocessor constants | ||
30 : | ;============================================================================= | ||
31 : | |||
32 : | %macro cglobal 1 | ||
33 : | %ifdef PREFIX | ||
34 : | edgomez | 1.4 | %ifdef MARK_FUNCS |
35 : | global _%1:function | ||
36 : | %define %1 _%1:function | ||
37 : | %else | ||
38 : | global _%1 | ||
39 : | %define %1 _%1 | ||
40 : | %endif | ||
41 : | edgomez | 1.2 | %else |
42 : | edgomez | 1.4 | %ifdef MARK_FUNCS |
43 : | global %1:function | ||
44 : | %else | ||
45 : | global %1 | ||
46 : | %endif | ||
47 : | edgomez | 1.2 | %endif |
48 : | %endmacro | ||
49 : | |||
50 : | %define BITS_INV_ACC 5 ; 4 or 5 for IEEE | ||
51 : | %define SHIFT_INV_ROW 16 - BITS_INV_ACC | ||
52 : | %define SHIFT_INV_COL 1 + BITS_INV_ACC | ||
53 : | %define RND_INV_ROW 1024 * (6 - BITS_INV_ACC) ; 1 << (SHIFT_INV_ROW-1) | ||
54 : | %define RND_INV_COL 16 * (BITS_INV_ACC - 3) ; 1 << (SHIFT_INV_COL-1) | ||
55 : | %define RND_INV_CORR RND_INV_COL - 1 ; correction -1.0 and round | ||
56 : | |||
57 : | %define BITS_FRW_ACC 3 ; 2 or 3 for accuracy | ||
58 : | %define SHIFT_FRW_COL BITS_FRW_ACC | ||
59 : | %define SHIFT_FRW_ROW BITS_FRW_ACC + 17 | ||
60 : | %define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) | ||
61 : | |||
62 : | ;============================================================================= | ||
63 : | ; Local Data (Read Only) | ||
64 : | ;============================================================================= | ||
65 : | |||
66 : | %ifdef FORMAT_COFF | ||
67 : | edgomez | 1.3 | SECTION .rodata |
68 : | edgomez | 1.2 | %else |
69 : | edgomez | 1.3 | SECTION .rodata align=16 |
70 : | edgomez | 1.2 | %endif |
71 : | |||
72 : | ALIGN 16 | ||
73 : | tab_i_04: | ||
74 : | dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 | ||
75 : | dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 | ||
76 : | dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 | ||
77 : | dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 | ||
78 : | dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 | ||
79 : | dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 | ||
80 : | dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 | ||
81 : | dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 | ||
82 : | |||
83 : | ; Table for rows 1,7 - constants are multiplied by cos_1_16 | ||
84 : | tab_i_17: | ||
85 : | dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 | ||
86 : | dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 | ||
87 : | dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 | ||
88 : | dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 | ||
89 : | dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 | ||
90 : | dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 | ||
91 : | dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 | ||
92 : | dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 | ||
93 : | |||
94 : | ; Table for rows 2,6 - constants are multiplied by cos_2_16 | ||
95 : | tab_i_26: | ||
96 : | dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 | ||
97 : | dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 | ||
98 : | dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 | ||
99 : | dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 | ||
100 : | dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 | ||
101 : | dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 | ||
102 : | dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 | ||
103 : | dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 | ||
104 : | |||
105 : | ; Table for rows 3,5 - constants are multiplied by cos_3_16 | ||
106 : | tab_i_35: | ||
107 : | dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 | ||
108 : | dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 | ||
109 : | dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 | ||
110 : | dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 | ||
111 : | dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 | ||
112 : | dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 | ||
113 : | dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 | ||
114 : | dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 | ||
115 : | |||
116 : | %if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 | ||
117 : | rounder_2_0: dd 65536, 65536 | ||
118 : | dd 65536, 65536 | ||
119 : | rounder_2_4: dd 0, 0 | ||
120 : | dd 0, 0 | ||
121 : | rounder_2_1: dd 7195, 7195 | ||
122 : | dd 7195, 7195 | ||
123 : | rounder_2_7: dd 1024, 1024 | ||
124 : | dd 1024, 1024 | ||
125 : | rounder_2_2: dd 4520, 4520 | ||
126 : | dd 4520, 4520 | ||
127 : | rounder_2_6: dd 1024, 1024 | ||
128 : | dd 1024, 1024 | ||
129 : | rounder_2_3: dd 2407, 2407 | ||
130 : | dd 2407, 2407 | ||
131 : | rounder_2_5: dd 240, 240 | ||
132 : | dd 240, 240 | ||
133 : | |||
134 : | %elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 | ||
135 : | rounder_2_0: dd 65536, 65536 | ||
136 : | dd 65536, 65536 | ||
137 : | rounder_2_4: dd 0, 0 | ||
138 : | dd 0, 0 | ||
139 : | rounder_2_1: dd 3597, 3597 | ||
140 : | dd 3597, 3597 | ||
141 : | rounder_2_7: dd 512, 512 | ||
142 : | dd 512, 512 | ||
143 : | rounder_2_2: dd 2260, 2260 | ||
144 : | dd 2260, 2260 | ||
145 : | rounder_2_6: dd 512, 512 | ||
146 : | dd 512, 512 | ||
147 : | rounder_2_3: dd 1203, 1203 | ||
148 : | dd 1203, 1203 | ||
149 : | rounder_2_5: dd 120, 120 | ||
150 : | dd 120, 120 | ||
151 : | %else | ||
152 : | |||
153 : | %error Invalid SHIFT_INV_ROW specified | ||
154 : | |||
155 : | %endif | ||
156 : | |||
157 : | tg_1_16: dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 | ||
158 : | dw 13036, 13036, 13036, 13036 | ||
159 : | tg_2_16: dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 | ||
160 : | dw 27146, 27146, 27146, 27146 | ||
161 : | tg_3_16: dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 | ||
162 : | dw -21746, -21746, -21746, -21746 | ||
163 : | ocos_4_16: dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 | ||
164 : | dw 23170, 23170, 23170, 23170 | ||
165 : | |||
166 : | ;============================================================================= | ||
167 : | ; Code | ||
168 : | ;============================================================================= | ||
169 : | |||
170 : | SECTION .text | ||
171 : | |||
172 : | cglobal idct_sse2_dmitry | ||
173 : | |||
174 : | ;----------------------------------------------------------------------------- | ||
175 : | ; Helper macro - ROW iDCT | ||
176 : | ;----------------------------------------------------------------------------- | ||
177 : | |||
178 : | %macro DCT_8_INV_ROW_1_SSE2 4 | ||
179 : | pshufhw xmm1, [%1], 11011000b ;x 75643210 | ||
180 : | pshuflw xmm1, xmm1, 11011000b ;x 75643120 | ||
181 : | pshufd xmm0, xmm1, 00000000b ;x 20202020 | ||
182 : | pmaddwd xmm0, [%3] ;w 13 12 9 8 5410 | ||
183 : | |||
184 : | ;a 3210 first part | ||
185 : | pshufd xmm2, xmm1, 10101010b ;x 64646464 | ||
186 : | pmaddwd xmm2, [%3+16] ;w 15 14 11 10 7632 | ||
187 : | |||
188 : | ;a 3210 second part | ||
189 : | paddd xmm2, xmm0 ;a 3210 ready | ||
190 : | paddd xmm2, [%4] ;must be 4 dwords long, not 2 as for sse1 | ||
191 : | movdqa xmm5, xmm2 | ||
192 : | |||
193 : | pshufd xmm3, xmm1, 01010101b ;x 31313131 | ||
194 : | pmaddwd xmm3, [%3+32] ;w 29 28 25 24 21 20 17 16 | ||
195 : | |||
196 : | ;b 3210 first part | ||
197 : | pshufd xmm4, xmm1, 11111111b ;x 75757575 | ||
198 : | pmaddwd xmm4, [%3+48] ;w 31 30 27 26 23 22 19 18 | ||
199 : | |||
200 : | ;b 3210 second part | ||
201 : | paddd xmm3,xmm4 ;b 3210 ready | ||
202 : | |||
203 : | paddd xmm2, xmm3 ;will be y 3210 | ||
204 : | psubd xmm5, xmm3 ;will be y 4567 | ||
205 : | psrad xmm2, SHIFT_INV_ROW | ||
206 : | psrad xmm5, SHIFT_INV_ROW | ||
207 : | packssdw xmm2, xmm5 ;y 45673210 | ||
208 : | pshufhw xmm6, xmm2, 00011011b ;y 76543210 | ||
209 : | movdqa [%2], xmm6 | ||
210 : | %endmacro | ||
211 : | |||
212 : | ;----------------------------------------------------------------------------- | ||
213 : | ; Helper macro - Columns iDCT | ||
214 : | ;----------------------------------------------------------------------------- | ||
215 : | |||
216 : | %macro DCT_8_INV_COL_4_SSE2 2 | ||
217 : | movdqa xmm0, [%1+16*0] ;x0 (all columns) | ||
218 : | movdqa xmm2, [%1+16*4] ;x4 | ||
219 : | movdqa xmm1, xmm0 | ||
220 : | |||
221 : | movdqa xmm4, [%1+16*2] ;x2 | ||
222 : | movdqa xmm5, [%1+16*6] ;x6 | ||
223 : | movdqa xmm6, [tg_2_16] | ||
224 : | movdqa xmm7, xmm6 | ||
225 : | |||
226 : | paddsw xmm0, xmm2 ;u04=x0+x4 | ||
227 : | psubsw xmm1, xmm2 ;v04=x0-x4 | ||
228 : | movdqa xmm3, xmm0 | ||
229 : | movdqa xmm2, xmm1 | ||
230 : | |||
231 : | pmulhw xmm6, xmm4 | ||
232 : | pmulhw xmm7, xmm5 | ||
233 : | psubsw xmm6, xmm5 ;v26=x2*T2-x6 | ||
234 : | paddsw xmm7, xmm4 ;u26=x6*T2+x2 | ||
235 : | |||
236 : | paddsw xmm1, xmm6 ;a1=v04+v26 | ||
237 : | paddsw xmm0, xmm7 ;a0=u04+u26 | ||
238 : | psubsw xmm2, xmm6 ;a2=v04-v26 | ||
239 : | psubsw xmm3, xmm7 ;a3=u04-u26 | ||
240 : | |||
241 : | movdqa [%2+16*0], xmm0 ;store a3-a0 to | ||
242 : | movdqa [%2+16*6], xmm1 ;free registers | ||
243 : | movdqa [%2+16*2], xmm2 | ||
244 : | movdqa [%2+16*4], xmm3 | ||
245 : | |||
246 : | movdqa xmm0, [%1+16*1] ;x1 | ||
247 : | movdqa xmm1, [%1+16*7] ;x7 | ||
248 : | movdqa xmm2, [tg_1_16] | ||
249 : | movdqa xmm3, xmm2 | ||
250 : | |||
251 : | movdqa xmm4, [%1+16*3] ;x3 | ||
252 : | movdqa xmm5, [%1+16*5] ;x5 | ||
253 : | movdqa xmm6, [tg_3_16] | ||
254 : | movdqa xmm7, xmm6 | ||
255 : | |||
256 : | pmulhw xmm2, xmm0 | ||
257 : | pmulhw xmm3, xmm1 | ||
258 : | psubsw xmm2, xmm1 ;v17=x1*T1-x7 | ||
259 : | paddsw xmm3, xmm0 ;u17=x7*T1+x1 | ||
260 : | movdqa xmm0, xmm3 ;u17 | ||
261 : | movdqa xmm1, xmm2 ;v17 | ||
262 : | |||
263 : | pmulhw xmm6, xmm4 ;x3*(t3-1) | ||
264 : | pmulhw xmm7, xmm5 ;x5*(t3-1) | ||
265 : | paddsw xmm6, xmm4 | ||
266 : | paddsw xmm7, xmm5 | ||
267 : | psubsw xmm6, xmm5 ;v35=x3*T3-x5 | ||
268 : | paddsw xmm7, xmm4 ;u35=x5*T3+x3 | ||
269 : | |||
270 : | movdqa xmm4, [ocos_4_16] | ||
271 : | |||
272 : | paddsw xmm0, xmm7 ;b0=u17+u35 | ||
273 : | psubsw xmm1, xmm6 ;b3=v17-v35 | ||
274 : | psubsw xmm3, xmm7 ;u12=u17-v35 | ||
275 : | paddsw xmm2, xmm6 ;v12=v17+v35 | ||
276 : | |||
277 : | movdqa xmm5, xmm3 | ||
278 : | paddsw xmm3, xmm2 ;tb1 | ||
279 : | psubsw xmm5, xmm2 ;tb2 | ||
280 : | pmulhw xmm5, xmm4 | ||
281 : | pmulhw xmm4, xmm3 | ||
282 : | paddsw xmm5, xmm5 | ||
283 : | paddsw xmm4, xmm4 | ||
284 : | |||
285 : | movdqa xmm6, [%2+16*0] ;a0 | ||
286 : | movdqa xmm7, xmm6 | ||
287 : | movdqa xmm2, [%2+16*4] ;a3 | ||
288 : | movdqa xmm3, xmm2 | ||
289 : | |||
290 : | paddsw xmm6, xmm0 | ||
291 : | psubsw xmm7, xmm0 | ||
292 : | psraw xmm6, SHIFT_INV_COL ;y0=a0+b0 | ||
293 : | psraw xmm7, SHIFT_INV_COL ;y7=a0-b0 | ||
294 : | movdqa [%2+16*0], xmm6 | ||
295 : | movdqa [%2+16*7], xmm7 | ||
296 : | |||
297 : | paddsw xmm2, xmm1 | ||
298 : | psubsw xmm3, xmm1 | ||
299 : | psraw xmm2, SHIFT_INV_COL ;y3=a3+b3 | ||
300 : | psraw xmm3, SHIFT_INV_COL ;y4=a3-b3 | ||
301 : | movdqa [%2+16*3], xmm2 | ||
302 : | movdqa [%2+16*4], xmm3 | ||
303 : | |||
304 : | movdqa xmm0, [%2+16*6] ;a1 | ||
305 : | movdqa xmm1, xmm0 | ||
306 : | movdqa xmm6, [%2+16*2] ;a2 | ||
307 : | movdqa xmm7, xmm6 | ||
308 : | |||
309 : | |||
310 : | paddsw xmm0, xmm4 | ||
311 : | psubsw xmm1, xmm4 | ||
312 : | psraw xmm0, SHIFT_INV_COL ;y1=a1+b1 | ||
313 : | psraw xmm1, SHIFT_INV_COL ;y6=a1-b1 | ||
314 : | movdqa [%2+16*1], xmm0 | ||
315 : | movdqa [%2+16*6], xmm1 | ||
316 : | |||
317 : | paddsw xmm6, xmm5 | ||
318 : | psubsw xmm7, xmm5 | ||
319 : | psraw xmm6, SHIFT_INV_COL ;y2=a2+b2 | ||
320 : | psraw xmm7, SHIFT_INV_COL ;y5=a2-b2 | ||
321 : | movdqa [%2+16*2], xmm6 | ||
322 : | movdqa [%2+16*5], xmm7 | ||
323 : | %endmacro | ||
324 : | |||
325 : | ;----------------------------------------------------------------------------- | ||
326 : | ; void idct_sse2_dmitry(int16_t coeff[64]); | ||
327 : | ;----------------------------------------------------------------------------- | ||
328 : | |||
329 : | ALIGN 16 | ||
330 : | idct_sse2_dmitry: | ||
331 : | |||
332 : | mov eax, [esp + 4] | ||
333 : | |||
334 : | DCT_8_INV_ROW_1_SSE2 eax+ 0, eax+ 0, tab_i_04, rounder_2_0 | ||
335 : | DCT_8_INV_ROW_1_SSE2 eax+ 16, eax+ 16, tab_i_17, rounder_2_1 | ||
336 : | DCT_8_INV_ROW_1_SSE2 eax+ 32, eax+ 32, tab_i_26, rounder_2_2 | ||
337 : | DCT_8_INV_ROW_1_SSE2 eax+ 48, eax+ 48, tab_i_35, rounder_2_3 | ||
338 : | DCT_8_INV_ROW_1_SSE2 eax+ 64, eax+ 64, tab_i_04, rounder_2_4 | ||
339 : | DCT_8_INV_ROW_1_SSE2 eax+ 80, eax+ 80, tab_i_35, rounder_2_5 | ||
340 : | DCT_8_INV_ROW_1_SSE2 eax+ 96, eax+ 96, tab_i_26, rounder_2_6 | ||
341 : | DCT_8_INV_ROW_1_SSE2 eax+112, eax+112, tab_i_17, rounder_2_7 | ||
342 : | |||
343 : | DCT_8_INV_COL_4_SSE2 eax, eax | ||
344 : | |||
345 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |