Parent Directory | Revision Log
Revision 1.6 - (view) (download)
1 : | edgomez | 1.2 | ;/* |
2 : | ; * Simple IDCT MMX | ||
3 : | ; * | ||
4 : | ; * Copyright (c) 2001, 2002 Michael Niedermayer <michaelni@gmx.at> | ||
5 : | ; * | ||
6 : | ; * This library is free software; you can redistribute it and/or | ||
7 : | ; * modify it under the terms of the GNU Lesser General Public | ||
8 : | ; * License as published by the Free Software Foundation; either | ||
9 : | ; * version 2 of the License, or (at your option) any later version. | ||
10 : | ; * | ||
11 : | ; * This library is distributed in the hope that it will be useful, | ||
12 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
13 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
14 : | ; * Lesser General Public License for more details. | ||
15 : | ; * | ||
16 : | ; * You should have received a copy of the GNU Lesser General Public | ||
17 : | ; * License along with this library; if not, write to the Free Software | ||
18 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
19 : | ; * | ||
20 : | ; * Ported to nasm by Peter Ross <pross@xvid.org> | ||
21 : | ; */ | ||
22 : | |||
23 : | edgomez | 1.4 | BITS 32 |
24 : | edgomez | 1.2 | |
25 : | edgomez | 1.4 | ;============================================================================= |
26 : | ; Macros and other preprocessor constants | ||
27 : | ;============================================================================= | ||
28 : | edgomez | 1.2 | |
29 : | edgomez | 1.4 | %macro cglobal 1 |
30 : | %ifdef PREFIX | ||
31 : | edgomez | 1.6 | %ifdef MARK_FUNCS |
32 : | global _%1:function | ||
33 : | %define %1 _%1:function | ||
34 : | %else | ||
35 : | global _%1 | ||
36 : | %define %1 _%1 | ||
37 : | %endif | ||
38 : | edgomez | 1.4 | %else |
39 : | edgomez | 1.6 | %ifdef MARK_FUNCS |
40 : | global %1:function | ||
41 : | %else | ||
42 : | global %1 | ||
43 : | %endif | ||
44 : | edgomez | 1.4 | %endif |
45 : | %endmacro | ||
46 : | edgomez | 1.2 | |
47 : | %define ROW_SHIFT 11 | ||
48 : | %define COL_SHIFT 20 | ||
49 : | %define C0 23170 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 23170.475006 | ||
50 : | %define C1 22725 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 22725.260826 | ||
51 : | %define C2 21407 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 21406.727617 | ||
52 : | %define C3 19266 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 19265.545870 | ||
53 : | %define C4 16383 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5 = 16384.000000 | ||
54 : | %define C5 12873 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 12872.826198 | ||
55 : | %define C6 8867 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 8866.956905 | ||
56 : | %define C7 4520 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 4520.335430 | ||
57 : | |||
58 : | edgomez | 1.4 | ;=========================================================================== |
59 : | ; Data (Read Only) | ||
60 : | ;=========================================================================== | ||
61 : | |||
62 : | %ifdef FORMAT_COFF | ||
63 : | edgomez | 1.5 | SECTION .rodata |
64 : | edgomez | 1.4 | %else |
65 : | edgomez | 1.5 | SECTION .rodata align=16 |
66 : | edgomez | 1.4 | %endif |
67 : | |||
68 : | ;----------------------------------------------------------------------------- | ||
69 : | ; Trigonometric Tables | ||
70 : | ;----------------------------------------------------------------------------- | ||
71 : | |||
72 : | ALIGN 16 | ||
73 : | wm1010: | ||
74 : | dw 0, 0xffff, 0, 0xffff | ||
75 : | |||
76 : | ALIGN 16 | ||
77 : | d40000: | ||
78 : | dd 0x40000, 0 | ||
79 : | |||
80 : | ALIGN 16 | ||
81 : | coeffs: | ||
82 : | dw 1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0, ; 0 | ||
83 : | dw 1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0, ; 8 | ||
84 : | |||
85 : | dw C4, C4, C4, C4 ; 16 | ||
86 : | dw C4, -C4, C4, -C4 ; 24 | ||
87 : | |||
88 : | dw C2, C6, C2, C6 ; 32 | ||
89 : | dw C6, -C2, C6, -C2 ; 40 | ||
90 : | |||
91 : | dw C1, C3, C1, C3 ; 48 | ||
92 : | dw C5, C7, C5, C7 ; 56 | ||
93 : | |||
94 : | dw C3, -C7, C3, -C7 ; 64 | ||
95 : | dw -C1, -C5, -C1, -C5 ; 72 | ||
96 : | |||
97 : | dw C5, -C1, C5, -C1 ; 80 | ||
98 : | dw C7, C3, C7, C3 ; 88 | ||
99 : | |||
100 : | dw C7, -C5, C7, -C5 ; 96 | ||
101 : | dw C3, -C1, C3, -C1 ; 104 | ||
102 : | edgomez | 1.2 | |
103 : | |||
104 : | ;=========================================================================== | ||
105 : | edgomez | 1.4 | ; Helper macros |
106 : | edgomez | 1.2 | ;=========================================================================== |
107 : | edgomez | 1.4 | |
108 : | ;--------------------------------------------------------------------------- | ||
109 : | ; DC_COND_IDCT | ||
110 : | ;--------------------------------------------------------------------------- | ||
111 : | edgomez | 1.2 | |
112 : | %macro DC_COND_IDCT 8 | ||
113 : | %define src0 %1 | ||
114 : | %define src4 %2 | ||
115 : | %define src1 %3 | ||
116 : | %define src5 %4 | ||
117 : | %define dst %5 | ||
118 : | %define rounder_op %6 | ||
119 : | %define rounder_arg %7 | ||
120 : | %define shift %8 | ||
121 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
122 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
123 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
124 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
125 : | movq mm4,[wm1010] | ||
126 : | pand mm4,mm0 | ||
127 : | por mm4,mm1 | ||
128 : | por mm4,mm2 | ||
129 : | por mm4,mm3 | ||
130 : | packssdw mm4,mm4 | ||
131 : | movd eax,mm4 | ||
132 : | or eax,eax | ||
133 : | jz near .skip1 | ||
134 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
135 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
136 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
137 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
138 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
139 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
140 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
141 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
142 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
143 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
144 : | rounder_op mm4, rounder_arg | ||
145 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
146 : | paddd mm4,mm5 ; A0 a0 | ||
147 : | psubd mm6,mm5 ; A3 a3 | ||
148 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
149 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
150 : | rounder_op mm0, rounder_arg | ||
151 : | paddd mm1,mm0 ; A1 a1 | ||
152 : | paddd mm0,mm0 | ||
153 : | psubd mm0,mm1 ; A2 a2 | ||
154 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
155 : | paddd mm7,mm5 ; B0 b0 | ||
156 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
157 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
158 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
159 : | paddd mm4,mm4 ; 2A0 2a0 | ||
160 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
161 : | paddd mm5,mm2 ; B1 b1 | ||
162 : | psrad mm7,shift | ||
163 : | psrad mm4,shift | ||
164 : | movq mm2,mm1 ; A1 a1 | ||
165 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
166 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
167 : | psrad mm1,shift | ||
168 : | psrad mm2,shift | ||
169 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
170 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
171 : | movq [dst],mm7 | ||
172 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
173 : | movq mm4,[coeffs+80] ;-C1 C5 -C1 C5 | ||
174 : | movq [dst + 24],mm2 | ||
175 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
176 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
177 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
178 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
179 : | movq mm2,mm0 ; A2 a2 | ||
180 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
181 : | paddd mm4,mm7 ; B2 b2 | ||
182 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
183 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
184 : | psrad mm2,shift | ||
185 : | psrad mm0,shift | ||
186 : | movq mm4,mm6 ; A3 a3 | ||
187 : | paddd mm3,mm1 ; B3 b3 | ||
188 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
189 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
190 : | psrad mm6,shift | ||
191 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
192 : | movq [ dst + 8],mm2 | ||
193 : | psrad mm4,shift | ||
194 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
195 : | movq [ dst + 16],mm4 | ||
196 : | jmp short .skip2 | ||
197 : | edgomez | 1.2 | .skip1 |
198 : | edgomez | 1.4 | pslld mm0,16 |
199 : | paddd mm0,[d40000] | ||
200 : | psrad mm0,13 | ||
201 : | packssdw mm0,mm0 | ||
202 : | movq [ dst ],mm0 | ||
203 : | movq [ dst + 8],mm0 | ||
204 : | movq [ dst + 16],mm0 | ||
205 : | movq [ dst + 24],mm0 | ||
206 : | edgomez | 1.2 | .skip2 |
207 : | edgomez | 1.4 | %undef src0 |
208 : | %undef src4 | ||
209 : | %undef src1 | ||
210 : | %undef src5 | ||
211 : | %undef dst | ||
212 : | %undef rounder_op | ||
213 : | %undef rounder_arg | ||
214 : | %undef shift | ||
215 : | edgomez | 1.2 | %endmacro |
216 : | |||
217 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
218 : | ; Z_COND_IDCT | ||
219 : | ;--------------------------------------------------------------------------- | ||
220 : | edgomez | 1.2 | |
221 : | %macro Z_COND_IDCT 9 | ||
222 : | edgomez | 1.4 | %define src0 %1 |
223 : | %define src4 %2 | ||
224 : | %define src1 %3 | ||
225 : | %define src5 %4 | ||
226 : | %define dst %5 | ||
227 : | %define rounder_op %6 | ||
228 : | %define rounder_arg %7 | ||
229 : | %define shift %8 | ||
230 : | %define bt %9 | ||
231 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
232 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
233 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
234 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
235 : | movq mm4,mm0 | ||
236 : | por mm4,mm1 | ||
237 : | por mm4,mm2 | ||
238 : | por mm4,mm3 | ||
239 : | packssdw mm4,mm4 | ||
240 : | movd eax,mm4 | ||
241 : | or eax,eax | ||
242 : | jz near bt | ||
243 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
244 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
245 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
246 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
247 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
248 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
249 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
250 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
251 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
252 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
253 : | rounder_op mm4, rounder_arg | ||
254 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
255 : | paddd mm4,mm5 ; A0 a0 | ||
256 : | psubd mm6,mm5 ; A3 a3 | ||
257 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
258 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
259 : | rounder_op mm0, rounder_arg | ||
260 : | paddd mm1,mm0 ; A1 a1 | ||
261 : | paddd mm0,mm0 | ||
262 : | psubd mm0,mm1 ; A2 a2 | ||
263 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
264 : | paddd mm7,mm5 ; B0 b0 | ||
265 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
266 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
267 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
268 : | paddd mm4,mm4 ; 2A0 2a0 | ||
269 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
270 : | paddd mm5,mm2 ; B1 b1 | ||
271 : | psrad mm7,shift | ||
272 : | psrad mm4,shift | ||
273 : | movq mm2,mm1 ; A1 a1 | ||
274 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
275 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
276 : | psrad mm1,shift | ||
277 : | psrad mm2,shift | ||
278 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
279 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
280 : | movq [ dst ],mm7 | ||
281 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
282 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
283 : | movq [ dst + 24 ],mm2 | ||
284 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
285 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
286 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
287 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
288 : | movq mm2,mm0 ; A2 a2 | ||
289 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
290 : | paddd mm4,mm7 ; B2 b2 | ||
291 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
292 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
293 : | psrad mm2,shift | ||
294 : | psrad mm0,shift | ||
295 : | movq mm4,mm6 ; A3 a3 | ||
296 : | paddd mm3,mm1 ; B3 b3 | ||
297 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
298 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
299 : | psrad mm6,shift | ||
300 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
301 : | movq [ dst + 8],mm2 | ||
302 : | psrad mm4,shift | ||
303 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
304 : | movq [dst + 16],mm4 | ||
305 : | edgomez | 1.2 | %undef src0 |
306 : | %undef src4 | ||
307 : | %undef src1 | ||
308 : | %undef src5 | ||
309 : | %undef dst | ||
310 : | %undef rounder_op | ||
311 : | %undef rounder_arg | ||
312 : | %undef shift | ||
313 : | %undef bt | ||
314 : | %endmacro | ||
315 : | |||
316 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
317 : | ; IDCT0 | ||
318 : | ;--------------------------------------------------------------------------- | ||
319 : | edgomez | 1.2 | |
320 : | %macro IDCT0 8 | ||
321 : | %define src0 %1 | ||
322 : | edgomez | 1.4 | %define src4 %2 |
323 : | %define src1 %3 | ||
324 : | %define src5 %4 | ||
325 : | %define dst %5 | ||
326 : | %define rounder_op %6 | ||
327 : | %define rounder_arg %7 | ||
328 : | %define shift %8 | ||
329 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
330 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
331 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
332 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
333 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
334 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
335 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
336 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
337 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
338 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
339 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
340 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
341 : | ; rounder_op mm4, rounder_arg | ||
342 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
343 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
344 : | ; rounder_op mm0, rounder_arg | ||
345 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
346 : | paddd mm4,mm5 ; A0 a0 | ||
347 : | psubd mm6,mm5 ; A3 a3 | ||
348 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
349 : | paddd mm0,mm1 ; A1 a1 | ||
350 : | psubd mm5,mm1 ; A2 a2 | ||
351 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
352 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
353 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
354 : | paddd mm7,mm1 ; B0 b0 | ||
355 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
356 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
357 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
358 : | paddd mm4,mm4 ; 2A0 2a0 | ||
359 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
360 : | paddd mm1,mm2 ; B1 b1 | ||
361 : | psrad mm7,shift | ||
362 : | psrad mm4,shift | ||
363 : | movq mm2,mm0 ; A1 a1 | ||
364 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
365 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
366 : | psrad mm0,shift | ||
367 : | psrad mm2,shift | ||
368 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
369 : | movd [ dst ],mm7 | ||
370 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
371 : | movd [ dst + 16],mm0 | ||
372 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
373 : | movd [ dst + 96 ],mm2 | ||
374 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
375 : | movd [ dst + 112],mm4 | ||
376 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
377 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
378 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
379 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
380 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
381 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
382 : | movq mm2,mm5 ; A2 a2 | ||
383 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
384 : | paddd mm4,mm7 ; B2 b2 | ||
385 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
386 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
387 : | psrad mm2,shift | ||
388 : | psrad mm5,shift | ||
389 : | movq mm4,mm6 ; A3 a3 | ||
390 : | paddd mm3,mm0 ; B3 b3 | ||
391 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
392 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
393 : | psrad mm6,shift | ||
394 : | psrad mm4,shift | ||
395 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
396 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
397 : | movd [ dst + 32 ],mm2 | ||
398 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
399 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
400 : | movd [ dst + 48 ],mm6 | ||
401 : | movd [ dst + 64 ],mm4 | ||
402 : | movd [ dst + 80 ],mm5 | ||
403 : | edgomez | 1.2 | %undef src0 |
404 : | %undef src4 | ||
405 : | %undef src1 | ||
406 : | %undef src5 | ||
407 : | %undef dst | ||
408 : | %undef rounder_op | ||
409 : | %undef rounder_arg | ||
410 : | %undef shift | ||
411 : | %endmacro | ||
412 : | |||
413 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
414 : | ; IDCT4 | ||
415 : | ;--------------------------------------------------------------------------- | ||
416 : | edgomez | 1.2 | |
417 : | %macro IDCT4 8 | ||
418 : | %define src0 %1 | ||
419 : | %define src4 %2 | ||
420 : | %define src1 %3 | ||
421 : | %define src5 %4 | ||
422 : | %define dst %5 | ||
423 : | %define rounder_op %6 | ||
424 : | %define rounder_arg %7 | ||
425 : | %define shift %8 | ||
426 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
427 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
428 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
429 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
430 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
431 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
432 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
433 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
434 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
435 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
436 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
437 : | ; rounder_op mm4, rounder_arg | ||
438 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
439 : | ; rounder_op mm0, rounder_arg | ||
440 : | paddd mm4,mm5 ; A0 a0 | ||
441 : | psubd mm6,mm5 ; A3 a3 | ||
442 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
443 : | paddd mm0,mm1 ; A1 a1 | ||
444 : | psubd mm5,mm1 ; A2 a2 | ||
445 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
446 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
447 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
448 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
449 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
450 : | paddd mm4,mm4 ; 2A0 2a0 | ||
451 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
452 : | psrad mm1,shift | ||
453 : | psrad mm4,shift | ||
454 : | movq mm2,mm0 ; A1 a1 | ||
455 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
456 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
457 : | psrad mm0,shift | ||
458 : | psrad mm2,shift | ||
459 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
460 : | movd [ dst ],mm1 | ||
461 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
462 : | movd [ dst + 16 ],mm0 | ||
463 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
464 : | movd [ dst + 96 ],mm2 | ||
465 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
466 : | movd [ dst + 112 ],mm4 | ||
467 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
468 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
469 : | movq mm2,mm5 ; A2 a2 | ||
470 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
471 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
472 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
473 : | psrad mm2,shift | ||
474 : | psrad mm5,shift | ||
475 : | movq mm1,mm6 ; A3 a3 | ||
476 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
477 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
478 : | psrad mm6,shift | ||
479 : | psrad mm1,shift | ||
480 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
481 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
482 : | movd [dst + 32],mm2 | ||
483 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
484 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
485 : | movd [dst + 48],mm6 | ||
486 : | movd [dst + 64],mm1 | ||
487 : | movd [dst + 80],mm5 | ||
488 : | edgomez | 1.2 | %undef src0 |
489 : | %undef src4 | ||
490 : | %undef src1 | ||
491 : | %undef src5 | ||
492 : | %undef dst | ||
493 : | %undef rounder_op | ||
494 : | %undef rounder_arg | ||
495 : | %undef shift | ||
496 : | %endmacro | ||
497 : | |||
498 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
499 : | ; IDCT6 | ||
500 : | ;--------------------------------------------------------------------------- | ||
501 : | edgomez | 1.2 | |
502 : | %macro IDCT6 8 | ||
503 : | %define src0 %1 | ||
504 : | %define src4 %2 | ||
505 : | %define src1 %3 | ||
506 : | %define src5 %4 | ||
507 : | %define dst %5 | ||
508 : | %define rounder_op %6 | ||
509 : | %define rounder_arg %7 | ||
510 : | %define shift %8 | ||
511 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
512 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
513 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
514 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
515 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
516 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
517 : | ; rounder_op mm4, rounder_arg | ||
518 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
519 : | ; rounder_op mm0, rounder_arg | ||
520 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
521 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
522 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
523 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
524 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
525 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
526 : | paddd mm4,mm4 ; 2A0 2a0 | ||
527 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
528 : | psrad mm1,shift | ||
529 : | psrad mm4,shift | ||
530 : | movq mm2,mm0 ; A1 a1 | ||
531 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
532 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
533 : | psrad mm0,shift | ||
534 : | psrad mm2,shift | ||
535 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
536 : | movd [ dst ],mm1 | ||
537 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
538 : | movd [ dst + 16 ],mm0 | ||
539 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
540 : | movd [ dst + 96 ],mm2 | ||
541 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
542 : | movd [ dst + 112 ],mm4 | ||
543 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
544 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
545 : | movq mm2,mm5 ; A2 a2 | ||
546 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
547 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
548 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
549 : | psrad mm2,shift | ||
550 : | psrad mm5,shift | ||
551 : | movq mm1,mm6 ; A3 a3 | ||
552 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
553 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
554 : | psrad mm6,shift | ||
555 : | psrad mm1,shift | ||
556 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
557 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
558 : | movd [dst + 32],mm2 | ||
559 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
560 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
561 : | movd [dst + 48],mm6 | ||
562 : | movd [dst + 64],mm1 | ||
563 : | movd [dst + 80],mm5 | ||
564 : | %undef src0 | ||
565 : | %undef src4 | ||
566 : | %undef src1 | ||
567 : | %undef src5 | ||
568 : | %undef dst | ||
569 : | %undef rounder_op | ||
570 : | %undef rounder_arg | ||
571 : | edgomez | 1.2 | %undef shift |
572 : | %endmacro | ||
573 : | |||
574 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
575 : | ; IDCT2 | ||
576 : | ;--------------------------------------------------------------------------- | ||
577 : | edgomez | 1.2 | |
578 : | %macro IDCT2 8 | ||
579 : | %define src0 %1 | ||
580 : | %define src4 %2 | ||
581 : | %define src1 %3 | ||
582 : | %define src5 %4 | ||
583 : | %define dst %5 | ||
584 : | %define rounder_op %6 | ||
585 : | %define rounder_arg %7 | ||
586 : | edgomez | 1.4 | %define shift %8 |
587 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
588 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
589 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
590 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
591 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
592 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
593 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
594 : | ; rounder_op mm4, rounder_arg | ||
595 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
596 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
597 : | ; rounder_op mm0, rounder_arg | ||
598 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
599 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
600 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
601 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
602 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
603 : | paddd mm7,mm1 ; B0 b0 | ||
604 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
605 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
606 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
607 : | paddd mm4,mm4 ; 2A0 2a0 | ||
608 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
609 : | paddd mm1,mm2 ; B1 b1 | ||
610 : | psrad mm7,shift | ||
611 : | psrad mm4,shift | ||
612 : | movq mm2,mm0 ; A1 a1 | ||
613 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
614 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
615 : | psrad mm0,shift | ||
616 : | psrad mm2,shift | ||
617 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
618 : | movd [dst],mm7 | ||
619 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
620 : | movd [dst + 16],mm0 | ||
621 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
622 : | movd [dst + 96],mm2 | ||
623 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
624 : | movd [dst + 112],mm4 | ||
625 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
626 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
627 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
628 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
629 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
630 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
631 : | movq mm2,mm5 ; A2 a2 | ||
632 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
633 : | paddd mm4,mm7 ; B2 b2 | ||
634 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
635 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
636 : | psrad mm2,shift | ||
637 : | psrad mm5,shift | ||
638 : | movq mm4,mm6 ; A3 a3 | ||
639 : | paddd mm3,mm0 ; B3 b3 | ||
640 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
641 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
642 : | psrad mm6,shift | ||
643 : | psrad mm4,shift | ||
644 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
645 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
646 : | movd [dst + 32],mm2 | ||
647 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
648 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
649 : | movd [dst + 48],mm6 | ||
650 : | movd [dst + 64],mm4 | ||
651 : | movd [dst + 80],mm5 | ||
652 : | %undef src0 | ||
653 : | %undef src4 | ||
654 : | %undef src1 | ||
655 : | %undef src5 | ||
656 : | %undef dst | ||
657 : | %undef rounder_op | ||
658 : | %undef rounder_arg | ||
659 : | %undef shift | ||
660 : | %endmacro | ||
661 : | |||
662 : | ;--------------------------------------------------------------------------- | ||
663 : | ; IDCT3 | ||
664 : | ;--------------------------------------------------------------------------- | ||
665 : | |||
666 : | %macro IDCT3 8 | ||
667 : | %define src0 %1 | ||
668 : | %define src4 %2 | ||
669 : | %define src1 %3 | ||
670 : | %define src5 %4 | ||
671 : | %define dst %5 | ||
672 : | %define rounder_op %6 | ||
673 : | %define rounder_arg %7 | ||
674 : | %define shift %8 | ||
675 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
676 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
677 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
678 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
679 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
680 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
681 : | ; rounder_op mm4, rounder_arg | ||
682 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
683 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
684 : | ; rounder_op mm0, rounder_arg | ||
685 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
686 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
687 : | movq mm3,[coeffs+64] | ||
688 : | pmaddwd mm3,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
689 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
690 : | paddd mm4,mm4 ; 2A0 2a0 | ||
691 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
692 : | psrad mm7,shift | ||
693 : | psrad mm4,shift | ||
694 : | movq mm1,mm0 ; A1 a1 | ||
695 : | paddd mm0,mm3 ; A1+B1 a1+b1 | ||
696 : | psubd mm1,mm3 ; A1-B1 a1-b1 | ||
697 : | psrad mm0,shift | ||
698 : | psrad mm1,shift | ||
699 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
700 : | movd [dst],mm7 | ||
701 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
702 : | movd [dst + 16],mm0 | ||
703 : | packssdw mm1,mm1 ; A1-B1 a1-b1 | ||
704 : | movd [dst + 96],mm1 | ||
705 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
706 : | movd [dst + 112],mm4 | ||
707 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
708 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
709 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
710 : | movq mm1,mm5 ; A2 a2 | ||
711 : | paddd mm1,mm4 ; A2+B2 a2+b2 | ||
712 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
713 : | psrad mm1,shift | ||
714 : | psrad mm5,shift | ||
715 : | movq mm4,mm6 ; A3 a3 | ||
716 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
717 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
718 : | psrad mm6,shift | ||
719 : | psrad mm4,shift | ||
720 : | packssdw mm1,mm1 ; A2+B2 a2+b2 | ||
721 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
722 : | movd [dst + 32],mm1 | ||
723 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
724 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
725 : | movd [dst + 48],mm6 | ||
726 : | movd [dst + 64],mm4 | ||
727 : | movd [dst + 80],mm5 | ||
728 : | %undef src0 | ||
729 : | edgomez | 1.2 | %undef src4 |
730 : | %undef src1 | ||
731 : | %undef src5 | ||
732 : | %undef dst | ||
733 : | %undef rounder_op | ||
734 : | %undef rounder_arg | ||
735 : | %undef shift | ||
736 : | %endmacro | ||
737 : | |||
738 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
739 : | ; IDCT5 | ||
740 : | ;--------------------------------------------------------------------------- | ||
741 : | edgomez | 1.2 | |
742 : | edgomez | 1.4 | %macro IDCT5 8 |
743 : | edgomez | 1.2 | %define src0 %1 |
744 : | %define src4 %2 | ||
745 : | %define src1 %3 | ||
746 : | %define src5 %4 | ||
747 : | %define dst %5 | ||
748 : | %define rounder_op %6 | ||
749 : | %define rounder_arg %7 | ||
750 : | %define shift %8 | ||
751 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
752 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
753 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
754 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
755 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
756 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
757 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
758 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
759 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
760 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
761 : | ; rounder_op mm4, rounder_arg | ||
762 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
763 : | paddd mm4,mm5 ; A0 a0 | ||
764 : | ; rounder_op mm0, rounder_arg | ||
765 : | psubd mm6,mm5 ; A3 a3 | ||
766 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
767 : | paddd mm0,mm1 ; A1 a1 | ||
768 : | psubd mm5,mm1 ; A2 a2 | ||
769 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
770 : | movq mm3,[src4 + 8] ; R6 R2 r6 r2 | ||
771 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
772 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
773 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
774 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
775 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
776 : | pmaddwd mm7,mm3 ; C6R6+C2R2 C6r6+C2r2 | ||
777 : | pmaddwd mm3,[coeffs+40] ; -C2R6+C6R2 -C2r6+C6r2 | ||
778 : | ; rounder_op mm1, rounder_arg | ||
779 : | paddd mm7,mm1 ; A0 a0 | ||
780 : | paddd mm1,mm1 ; 2C0 2c0 | ||
781 : | ; rounder_op mm2, rounder_arg | ||
782 : | psubd mm1,mm7 ; A3 a3 | ||
783 : | paddd mm3,mm2 ; A1 a1 | ||
784 : | paddd mm2,mm2 ; 2C1 2c1 | ||
785 : | psubd mm2,mm3 ; A2 a2 | ||
786 : | psrad mm4,shift | ||
787 : | psrad mm7,shift | ||
788 : | psrad mm3,shift | ||
789 : | packssdw mm4,mm7 ; A0 a0 | ||
790 : | movq [dst],mm4 | ||
791 : | psrad mm0,shift | ||
792 : | packssdw mm0,mm3 ; A1 a1 | ||
793 : | movq [dst + 16],mm0 | ||
794 : | movq [dst + 96],mm0 | ||
795 : | movq [dst + 112],mm4 | ||
796 : | psrad mm5,shift | ||
797 : | psrad mm6,shift | ||
798 : | psrad mm2,shift | ||
799 : | packssdw mm5,mm2 ; A2-B2 a2-b2 | ||
800 : | movq [dst + 32],mm5 | ||
801 : | psrad mm1,shift | ||
802 : | packssdw mm6,mm1 ; A3+B3 a3+b3 | ||
803 : | movq [dst + 48],mm6 | ||
804 : | movq [dst + 64],mm6 | ||
805 : | movq [dst + 80],mm5 | ||
806 : | %undef src0 | ||
807 : | %undef src4 | ||
808 : | %undef src1 | ||
809 : | %undef src5 | ||
810 : | %undef dst | ||
811 : | %undef rounder_op | ||
812 : | edgomez | 1.2 | %undef rounder_arg |
813 : | %undef shift | ||
814 : | %endmacro | ||
815 : | |||
816 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
817 : | ; IDCT1 | ||
818 : | ;--------------------------------------------------------------------------- | ||
819 : | edgomez | 1.2 | |
820 : | edgomez | 1.4 | %macro IDCT1 8 |
821 : | edgomez | 1.2 | %define src0 %1 |
822 : | %define src4 %2 | ||
823 : | %define src1 %3 | ||
824 : | %define src5 %4 | ||
825 : | edgomez | 1.4 | %define dst %5 |
826 : | %define rounder_op %6 | ||
827 : | %define rounder_arg %7 | ||
828 : | %define shift %8 | ||
829 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
830 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
831 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
832 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
833 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
834 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
835 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
836 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
837 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
838 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
839 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
840 : | ; rounder_op mm4, rounder_arg | ||
841 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
842 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
843 : | ; rounder_op mm0, rounder_arg | ||
844 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
845 : | paddd mm4,mm5 ; A0 a0 | ||
846 : | psubd mm6,mm5 ; A3 a3 | ||
847 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
848 : | paddd mm0,mm1 ; A1 a1 | ||
849 : | psubd mm5,mm1 ; A2 a2 | ||
850 : | movq mm1,[coeffs+64] | ||
851 : | pmaddwd mm1,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
852 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
853 : | paddd mm4,mm4 ; 2A0 2a0 | ||
854 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
855 : | psrad mm7,shift | ||
856 : | psrad mm4,shift | ||
857 : | movq mm3,mm0 ; A1 a1 | ||
858 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
859 : | psubd mm3,mm1 ; A1-B1 a1-b1 | ||
860 : | psrad mm0,shift | ||
861 : | psrad mm3,shift | ||
862 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
863 : | movd [dst],mm7 | ||
864 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
865 : | movd [dst + 16],mm0 | ||
866 : | packssdw mm3,mm3 ; A1-B1 a1-b1 | ||
867 : | movd [dst + 96],mm3 | ||
868 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
869 : | movd [dst + 112],mm4 | ||
870 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
871 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
872 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
873 : | movq mm3,mm5 ; A2 a2 | ||
874 : | paddd mm3,mm4 ; A2+B2 a2+b2 | ||
875 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
876 : | psrad mm3,shift | ||
877 : | psrad mm5,shift | ||
878 : | movq mm4,mm6 ; A3 a3 | ||
879 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
880 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
881 : | psrad mm6,shift | ||
882 : | packssdw mm3,mm3 ; A2+B2 a2+b2 | ||
883 : | movd [dst + 32],mm3 | ||
884 : | psrad mm4,shift | ||
885 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
886 : | movd [dst + 48],mm6 | ||
887 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
888 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
889 : | movd [dst + 64],mm4 | ||
890 : | movd [dst + 80],mm5 | ||
891 : | %undef src0 | ||
892 : | %undef src4 | ||
893 : | %undef src1 | ||
894 : | %undef src5 | ||
895 : | %undef dst | ||
896 : | %undef rounder_op | ||
897 : | %undef rounder_arg | ||
898 : | %undef shift | ||
899 : | edgomez | 1.2 | %endmacro |
900 : | |||
901 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
902 : | ; IDCT7 | ||
903 : | ;--------------------------------------------------------------------------- | ||
904 : | edgomez | 1.2 | |
905 : | edgomez | 1.4 | %macro IDCT7 8 |
906 : | edgomez | 1.2 | %define src0 %1 |
907 : | %define src4 %2 | ||
908 : | %define src1 %3 | ||
909 : | %define src5 %4 | ||
910 : | %define dst %5 | ||
911 : | %define rounder_op %6 | ||
912 : | %define rounder_arg %7 | ||
913 : | %define shift %8 | ||
914 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
915 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
916 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
917 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
918 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
919 : | ; rounder_op mm4, rounder_arg | ||
920 : | ; rounder_op mm0, rounder_arg | ||
921 : | psrad mm4,shift | ||
922 : | psrad mm0,shift | ||
923 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
924 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
925 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
926 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
927 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
928 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
929 : | ; rounder_op mm1, rounder_arg | ||
930 : | ; rounder_op mm2, rounder_arg | ||
931 : | psrad mm1,shift | ||
932 : | packssdw mm4,mm1 ; A0 a0 | ||
933 : | movq [dst],mm4 | ||
934 : | psrad mm2,shift | ||
935 : | packssdw mm0,mm2 ; A1 a1 | ||
936 : | movq [dst + 16],mm0 | ||
937 : | movq [dst + 96],mm0 | ||
938 : | movq [dst + 112],mm4 | ||
939 : | movq [dst + 32],mm0 | ||
940 : | movq [dst + 48],mm4 | ||
941 : | movq [dst + 64],mm4 | ||
942 : | movq [dst + 80],mm0 | ||
943 : | edgomez | 1.2 | %undef src0 |
944 : | %undef src4 | ||
945 : | %undef src1 | ||
946 : | %undef src5 | ||
947 : | %undef dst | ||
948 : | %undef rounder_op | ||
949 : | %undef rounder_arg | ||
950 : | %undef shift | ||
951 : | %endmacro | ||
952 : | |||
953 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
954 : | ; Permutation helpers | ||
955 : | ;--------------------------------------------------------------------------- | ||
956 : | |||
957 : | %macro XLODA 2 | ||
958 : | mov bx, [srcP+2*%2] ; get src contents | ||
959 : | mov ax, [srcP+2*%1] ; get dest contents | ||
960 : | mov [srcP+2*%1], bx ; store new dest val | ||
961 : | %endmacro | ||
962 : | edgomez | 1.2 | |
963 : | edgomez | 1.4 | %macro XCHGA 2 |
964 : | mov ax, [srcP+2*%1] ; get dest contents | ||
965 : | mov [srcP+2*%1], bx ; store new dest val | ||
966 : | %endmacro | ||
967 : | edgomez | 1.2 | |
968 : | edgomez | 1.4 | %macro XCHGB 2 |
969 : | mov bx, [srcP+2*%1] ; get dest contents | ||
970 : | mov [srcP+2*%1], ax ; store new dest val | ||
971 : | %endmacro | ||
972 : | |||
973 : | %macro XSTRA 2 | ||
974 : | mov [srcP+2*%1], bx ; store dest val | ||
975 : | %endmacro | ||
976 : | edgomez | 1.2 | |
977 : | edgomez | 1.4 | %macro XSTRB 2 |
978 : | mov [srcP+2*%1], ax ; store dest val | ||
979 : | edgomez | 1.2 | %endmacro |
980 : | |||
981 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
982 : | ; Permutation macro | ||
983 : | ;--------------------------------------------------------------------------- | ||
984 : | |||
985 : | %macro PERMUTEP 1 | ||
986 : | %define srcP %1 | ||
987 : | push ebx | ||
988 : | |||
989 : | ; XCHGA 0x00, 0x00 ; nothing to do | ||
990 : | |||
991 : | XLODA 0x08, 0x01 | ||
992 : | XCHGB 0x10, 0x08 | ||
993 : | XCHGA 0x20, 0x10 | ||
994 : | XCHGB 0x02, 0x20 | ||
995 : | XCHGA 0x04, 0x02 | ||
996 : | XSTRB 0x01, 0x04 | ||
997 : | |||
998 : | XLODA 0x09, 0x03 | ||
999 : | XCHGB 0x18, 0x09 | ||
1000 : | XCHGA 0x12, 0x18 | ||
1001 : | XCHGB 0x24, 0x12 | ||
1002 : | XSTRA 0x03, 0x24 | ||
1003 : | |||
1004 : | XLODA 0x0C, 0x05 | ||
1005 : | XCHGB 0x11, 0x0C | ||
1006 : | XCHGA 0x28, 0x11 | ||
1007 : | XCHGB 0x30, 0x28 | ||
1008 : | XCHGA 0x22, 0x30 | ||
1009 : | XCHGB 0x06, 0x22 | ||
1010 : | XSTRA 0x05, 0x06 | ||
1011 : | |||
1012 : | XLODA 0x0D, 0x07 | ||
1013 : | XCHGB 0x1C, 0x0D | ||
1014 : | XCHGA 0x13, 0x1C | ||
1015 : | XCHGB 0x29, 0x13 | ||
1016 : | XCHGA 0x38, 0x29 | ||
1017 : | XCHGB 0x32, 0x38 | ||
1018 : | XCHGA 0x26, 0x32 | ||
1019 : | XSTRB 0x07, 0x26 | ||
1020 : | |||
1021 : | XLODA 0x14, 0x0A | ||
1022 : | XCHGB 0x21, 0x14 | ||
1023 : | XSTRA 0x0A, 0x21 | ||
1024 : | |||
1025 : | XLODA 0x19, 0x0B | ||
1026 : | XCHGB 0x1A, 0x19 | ||
1027 : | XCHGA 0x16, 0x1A | ||
1028 : | XCHGB 0x25, 0x16 | ||
1029 : | XCHGA 0x0E, 0x25 | ||
1030 : | XCHGB 0x15, 0x0E | ||
1031 : | XCHGA 0x2C, 0x15 | ||
1032 : | XCHGB 0x31, 0x2C | ||
1033 : | XCHGA 0x2A, 0x31 | ||
1034 : | XCHGB 0x34, 0x2A | ||
1035 : | XCHGA 0x23, 0x34 | ||
1036 : | XSTRB 0x0B, 0x23 | ||
1037 : | |||
1038 : | XLODA 0x1D, 0x0F | ||
1039 : | XCHGB 0x1E, 0x1D | ||
1040 : | XCHGA 0x17, 0x1E | ||
1041 : | XCHGB 0x2D, 0x17 | ||
1042 : | XCHGA 0x3C, 0x2D | ||
1043 : | XCHGB 0x33, 0x3C | ||
1044 : | XCHGA 0x2B, 0x33 | ||
1045 : | XCHGB 0x39, 0x2B | ||
1046 : | XCHGA 0x3A, 0x39 | ||
1047 : | XCHGB 0x36, 0x3A | ||
1048 : | XCHGA 0x27, 0x36 | ||
1049 : | XSTRB 0x0F, 0x27 | ||
1050 : | |||
1051 : | ; XCHGA 0x1B, 0x1B | ||
1052 : | |||
1053 : | ; XCHGA 0x1F, 0x1F | ||
1054 : | |||
1055 : | XLODA 0x35, 0x2E | ||
1056 : | XSTRB 0x2E, 0x35 | ||
1057 : | |||
1058 : | XLODA 0x3D, 0x2F | ||
1059 : | XCHGB 0x3E, 0x3D | ||
1060 : | XCHGA 0x37, 0x3E | ||
1061 : | XSTRB 0x2F, 0x37 | ||
1062 : | edgomez | 1.2 | |
1063 : | edgomez | 1.4 | ; XCHGA 0x3B, 0x3B |
1064 : | edgomez | 1.2 | |
1065 : | edgomez | 1.4 | ; XCHGA 0x3F, 0x3F |
1066 : | pop ebx | ||
1067 : | %undef srcP | ||
1068 : | edgomez | 1.2 | %endmacro |
1069 : | |||
1070 : | edgomez | 1.4 | ;============================================================================= |
1071 : | ; Code | ||
1072 : | ;============================================================================= | ||
1073 : | |||
1074 : | SECTION .text | ||
1075 : | |||
1076 : | cglobal simple_idct_mmx_P | ||
1077 : | cglobal simple_idct_mmx | ||
1078 : | edgomez | 1.2 | |
1079 : | edgomez | 1.4 | ;----------------------------------------------------------------------------- |
1080 : | ; void simple_idct_mmx_P(int16_t * const block) | ||
1081 : | Isibaar | 1.3 | ; expects input data to be permutated |
1082 : | edgomez | 1.4 | ;----------------------------------------------------------------------------- |
1083 : | |||
1084 : | ALIGN 16 | ||
1085 : | simple_idct_mmx_P: | ||
1086 : | sub esp, 128 | ||
1087 : | mov edx, [esp+128+4] | ||
1088 : | |||
1089 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
1090 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
1091 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .four | ||
1092 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .two | ||
1093 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .one | ||
1094 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1095 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1096 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1097 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1098 : | jmp .ret | ||
1099 : | edgomez | 1.2 | |
1100 : | edgomez | 1.4 | ALIGN 16 |
1101 : | edgomez | 1.2 | .four |
1102 : | edgomez | 1.4 | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .six |
1103 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .five | ||
1104 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1105 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1106 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1107 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1108 : | jmp .ret | ||
1109 : | edgomez | 1.2 | |
1110 : | edgomez | 1.4 | ALIGN 16 |
1111 : | edgomez | 1.2 | .six |
1112 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .seven |
1113 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1114 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1115 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1116 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1117 : | jmp .ret | ||
1118 : | edgomez | 1.2 | |
1119 : | edgomez | 1.4 | ALIGN 16 |
1120 : | edgomez | 1.2 | .two |
1121 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .three |
1122 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1123 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1124 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1125 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1126 : | jmp .ret | ||
1127 : | edgomez | 1.2 | |
1128 : | edgomez | 1.4 | ALIGN 16 |
1129 : | edgomez | 1.2 | .three |
1130 : | edgomez | 1.4 | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1131 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1132 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1133 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1134 : | jmp .ret | ||
1135 : | edgomez | 1.2 | |
1136 : | edgomez | 1.4 | ALIGN 16 |
1137 : | edgomez | 1.2 | .five |
1138 : | edgomez | 1.4 | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1139 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1140 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1141 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1142 : | jmp .ret | ||
1143 : | edgomez | 1.2 | |
1144 : | edgomez | 1.4 | ALIGN 16 |
1145 : | edgomez | 1.2 | .one |
1146 : | edgomez | 1.4 | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1147 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1148 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1149 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1150 : | jmp .ret | ||
1151 : | edgomez | 1.2 | |
1152 : | edgomez | 1.4 | ALIGN 16 |
1153 : | edgomez | 1.2 | .seven |
1154 : | edgomez | 1.4 | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1155 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1156 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1157 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1158 : | edgomez | 1.2 | |
1159 : | .ret | ||
1160 : | edgomez | 1.4 | add esp, 128 |
1161 : | Isibaar | 1.3 | |
1162 : | edgomez | 1.4 | ret |
1163 : | Isibaar | 1.3 | |
1164 : | edgomez | 1.4 | |
1165 : | ;----------------------------------------------------------------------------- | ||
1166 : | ; void simple_idct_mmx(int16_t * const block) | ||
1167 : | Isibaar | 1.3 | ; |
1168 : | edgomez | 1.4 | ; simple_idct_mmx is the same function as simple_idct_mmx_P above except that |
1169 : | ; on entry it will do a fast in-line and in-place permutation on the iDCT parm | ||
1170 : | ; list. This means that same parm list will also not have to be copied on the | ||
1171 : | ; way out. - trbarry 6/2003 | ||
1172 : | ;----------------------------------------------------------------------------- | ||
1173 : | |||
1174 : | ALIGN 16 | ||
1175 : | simple_idct_mmx: | ||
1176 : | sub esp, 128 | ||
1177 : | mov edx, [esp+128+4] | ||
1178 : | PERMUTEP edx ; permute parm list in place | ||
1179 : | |||
1180 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
1181 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
1182 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .fourP | ||
1183 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .twoP | ||
1184 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .oneP | ||
1185 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1186 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1187 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1188 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1189 : | jmp .retP | ||
1190 : | Isibaar | 1.3 | |
1191 : | edgomez | 1.4 | ALIGN 16 |
1192 : | Isibaar | 1.3 | .fourP |
1193 : | edgomez | 1.4 | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .sixP |
1194 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .fiveP | ||
1195 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1196 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1197 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1198 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1199 : | jmp .retP | ||
1200 : | Isibaar | 1.3 | |
1201 : | edgomez | 1.4 | ALIGN 16 |
1202 : | Isibaar | 1.3 | .sixP |
1203 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .sevenP |
1204 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1205 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1206 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1207 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1208 : | jmp .retP | ||
1209 : | Isibaar | 1.3 | |
1210 : | edgomez | 1.4 | ALIGN 16 |
1211 : | Isibaar | 1.3 | .twoP |
1212 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .threeP |
1213 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1214 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1215 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1216 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1217 : | jmp .retP | ||
1218 : | Isibaar | 1.3 | |
1219 : | edgomez | 1.4 | ALIGN 16 |
1220 : | Isibaar | 1.3 | .threeP |
1221 : | edgomez | 1.4 | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1222 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1223 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1224 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1225 : | jmp .retP | ||
1226 : | Isibaar | 1.3 | |
1227 : | edgomez | 1.4 | ALIGN 16 |
1228 : | Isibaar | 1.3 | .fiveP |
1229 : | edgomez | 1.4 | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1230 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1231 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1232 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1233 : | jmp .retP | ||
1234 : | Isibaar | 1.3 | |
1235 : | edgomez | 1.4 | ALIGN 16 |
1236 : | Isibaar | 1.3 | .oneP |
1237 : | edgomez | 1.4 | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1238 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1239 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1240 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1241 : | jmp .retP | ||
1242 : | Isibaar | 1.3 | |
1243 : | edgomez | 1.4 | ALIGN 16 |
1244 : | Isibaar | 1.3 | .sevenP |
1245 : | edgomez | 1.4 | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1246 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1247 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1248 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1249 : | Isibaar | 1.3 | |
1250 : | .retP | ||
1251 : | edgomez | 1.4 | add esp, 128 |
1252 : | edgomez | 1.2 | |
1253 : | edgomez | 1.4 | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |