Parent Directory | Revision Log
Revision 1.10 - (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 : | edgomez | 1.7 | global _%1:function %1.endfunc-%1 |
33 : | %define %1 _%1:function %1.endfunc-%1 | ||
34 : | Isibaar | 1.9 | %define ENDFUNC .endfunc |
35 : | edgomez | 1.6 | %else |
36 : | global _%1 | ||
37 : | %define %1 _%1 | ||
38 : | Isibaar | 1.9 | %define ENDFUNC |
39 : | edgomez | 1.6 | %endif |
40 : | edgomez | 1.4 | %else |
41 : | edgomez | 1.6 | %ifdef MARK_FUNCS |
42 : | edgomez | 1.7 | global %1:function %1.endfunc-%1 |
43 : | Isibaar | 1.9 | %define ENDFUNC .endfunc |
44 : | edgomez | 1.6 | %else |
45 : | global %1 | ||
46 : | Isibaar | 1.9 | %define ENDFUNC |
47 : | edgomez | 1.6 | %endif |
48 : | edgomez | 1.4 | %endif |
49 : | %endmacro | ||
50 : | edgomez | 1.2 | |
51 : | %define ROW_SHIFT 11 | ||
52 : | %define COL_SHIFT 20 | ||
53 : | %define C0 23170 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 23170.475006 | ||
54 : | %define C1 22725 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 22725.260826 | ||
55 : | %define C2 21407 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 21406.727617 | ||
56 : | %define C3 19266 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 19265.545870 | ||
57 : | %define C4 16383 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5 = 16384.000000 | ||
58 : | %define C5 12873 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 12872.826198 | ||
59 : | %define C6 8867 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 8866.956905 | ||
60 : | %define C7 4520 ;cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5 = 4520.335430 | ||
61 : | |||
62 : | edgomez | 1.4 | ;=========================================================================== |
63 : | ; Data (Read Only) | ||
64 : | ;=========================================================================== | ||
65 : | |||
66 : | %ifdef FORMAT_COFF | ||
67 : | edgomez | 1.5 | SECTION .rodata |
68 : | edgomez | 1.4 | %else |
69 : | edgomez | 1.5 | SECTION .rodata align=16 |
70 : | edgomez | 1.4 | %endif |
71 : | |||
72 : | ;----------------------------------------------------------------------------- | ||
73 : | ; Trigonometric Tables | ||
74 : | ;----------------------------------------------------------------------------- | ||
75 : | |||
76 : | ALIGN 16 | ||
77 : | wm1010: | ||
78 : | dw 0, 0xffff, 0, 0xffff | ||
79 : | |||
80 : | ALIGN 16 | ||
81 : | d40000: | ||
82 : | dd 0x40000, 0 | ||
83 : | |||
84 : | ALIGN 16 | ||
85 : | coeffs: | ||
86 : | dw 1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0, ; 0 | ||
87 : | dw 1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0, ; 8 | ||
88 : | |||
89 : | dw C4, C4, C4, C4 ; 16 | ||
90 : | dw C4, -C4, C4, -C4 ; 24 | ||
91 : | |||
92 : | dw C2, C6, C2, C6 ; 32 | ||
93 : | dw C6, -C2, C6, -C2 ; 40 | ||
94 : | |||
95 : | dw C1, C3, C1, C3 ; 48 | ||
96 : | dw C5, C7, C5, C7 ; 56 | ||
97 : | |||
98 : | dw C3, -C7, C3, -C7 ; 64 | ||
99 : | dw -C1, -C5, -C1, -C5 ; 72 | ||
100 : | |||
101 : | dw C5, -C1, C5, -C1 ; 80 | ||
102 : | dw C7, C3, C7, C3 ; 88 | ||
103 : | |||
104 : | dw C7, -C5, C7, -C5 ; 96 | ||
105 : | dw C3, -C1, C3, -C1 ; 104 | ||
106 : | edgomez | 1.2 | |
107 : | |||
108 : | ;=========================================================================== | ||
109 : | edgomez | 1.4 | ; Helper macros |
110 : | edgomez | 1.2 | ;=========================================================================== |
111 : | edgomez | 1.4 | |
112 : | ;--------------------------------------------------------------------------- | ||
113 : | ; DC_COND_IDCT | ||
114 : | ;--------------------------------------------------------------------------- | ||
115 : | edgomez | 1.2 | |
116 : | %macro DC_COND_IDCT 8 | ||
117 : | %define src0 %1 | ||
118 : | %define src4 %2 | ||
119 : | %define src1 %3 | ||
120 : | %define src5 %4 | ||
121 : | %define dst %5 | ||
122 : | %define rounder_op %6 | ||
123 : | %define rounder_arg %7 | ||
124 : | %define shift %8 | ||
125 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
126 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
127 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
128 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
129 : | movq mm4,[wm1010] | ||
130 : | pand mm4,mm0 | ||
131 : | por mm4,mm1 | ||
132 : | por mm4,mm2 | ||
133 : | por mm4,mm3 | ||
134 : | packssdw mm4,mm4 | ||
135 : | movd eax,mm4 | ||
136 : | or eax,eax | ||
137 : | jz near .skip1 | ||
138 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
139 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
140 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
141 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
142 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
143 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
144 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
145 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
146 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
147 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
148 : | rounder_op mm4, rounder_arg | ||
149 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
150 : | paddd mm4,mm5 ; A0 a0 | ||
151 : | psubd mm6,mm5 ; A3 a3 | ||
152 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
153 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
154 : | rounder_op mm0, rounder_arg | ||
155 : | paddd mm1,mm0 ; A1 a1 | ||
156 : | paddd mm0,mm0 | ||
157 : | psubd mm0,mm1 ; A2 a2 | ||
158 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
159 : | paddd mm7,mm5 ; B0 b0 | ||
160 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
161 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
162 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
163 : | paddd mm4,mm4 ; 2A0 2a0 | ||
164 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
165 : | paddd mm5,mm2 ; B1 b1 | ||
166 : | psrad mm7,shift | ||
167 : | psrad mm4,shift | ||
168 : | movq mm2,mm1 ; A1 a1 | ||
169 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
170 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
171 : | psrad mm1,shift | ||
172 : | psrad mm2,shift | ||
173 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
174 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
175 : | movq [dst],mm7 | ||
176 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
177 : | movq mm4,[coeffs+80] ;-C1 C5 -C1 C5 | ||
178 : | movq [dst + 24],mm2 | ||
179 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
180 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
181 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
182 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
183 : | movq mm2,mm0 ; A2 a2 | ||
184 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
185 : | paddd mm4,mm7 ; B2 b2 | ||
186 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
187 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
188 : | psrad mm2,shift | ||
189 : | psrad mm0,shift | ||
190 : | movq mm4,mm6 ; A3 a3 | ||
191 : | paddd mm3,mm1 ; B3 b3 | ||
192 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
193 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
194 : | psrad mm6,shift | ||
195 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
196 : | movq [ dst + 8],mm2 | ||
197 : | psrad mm4,shift | ||
198 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
199 : | movq [ dst + 16],mm4 | ||
200 : | jmp short .skip2 | ||
201 : | Isibaar | 1.9 | .skip1: |
202 : | edgomez | 1.4 | pslld mm0,16 |
203 : | paddd mm0,[d40000] | ||
204 : | psrad mm0,13 | ||
205 : | packssdw mm0,mm0 | ||
206 : | movq [ dst ],mm0 | ||
207 : | movq [ dst + 8],mm0 | ||
208 : | movq [ dst + 16],mm0 | ||
209 : | movq [ dst + 24],mm0 | ||
210 : | Isibaar | 1.9 | .skip2: |
211 : | edgomez | 1.4 | %undef src0 |
212 : | %undef src4 | ||
213 : | %undef src1 | ||
214 : | %undef src5 | ||
215 : | %undef dst | ||
216 : | %undef rounder_op | ||
217 : | %undef rounder_arg | ||
218 : | %undef shift | ||
219 : | edgomez | 1.2 | %endmacro |
220 : | |||
221 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
222 : | ; Z_COND_IDCT | ||
223 : | ;--------------------------------------------------------------------------- | ||
224 : | edgomez | 1.2 | |
225 : | %macro Z_COND_IDCT 9 | ||
226 : | edgomez | 1.4 | %define src0 %1 |
227 : | %define src4 %2 | ||
228 : | %define src1 %3 | ||
229 : | %define src5 %4 | ||
230 : | %define dst %5 | ||
231 : | %define rounder_op %6 | ||
232 : | %define rounder_arg %7 | ||
233 : | %define shift %8 | ||
234 : | %define bt %9 | ||
235 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
236 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
237 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
238 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
239 : | movq mm4,mm0 | ||
240 : | por mm4,mm1 | ||
241 : | por mm4,mm2 | ||
242 : | por mm4,mm3 | ||
243 : | packssdw mm4,mm4 | ||
244 : | movd eax,mm4 | ||
245 : | or eax,eax | ||
246 : | jz near bt | ||
247 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
248 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
249 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
250 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
251 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
252 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
253 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
254 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
255 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
256 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
257 : | rounder_op mm4, rounder_arg | ||
258 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
259 : | paddd mm4,mm5 ; A0 a0 | ||
260 : | psubd mm6,mm5 ; A3 a3 | ||
261 : | movq mm5,[coeffs+56] ; C7 C5 C7 C5 | ||
262 : | pmaddwd mm5,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
263 : | rounder_op mm0, rounder_arg | ||
264 : | paddd mm1,mm0 ; A1 a1 | ||
265 : | paddd mm0,mm0 | ||
266 : | psubd mm0,mm1 ; A2 a2 | ||
267 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
268 : | paddd mm7,mm5 ; B0 b0 | ||
269 : | movq mm5,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
270 : | pmaddwd mm5,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
271 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
272 : | paddd mm4,mm4 ; 2A0 2a0 | ||
273 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
274 : | paddd mm5,mm2 ; B1 b1 | ||
275 : | psrad mm7,shift | ||
276 : | psrad mm4,shift | ||
277 : | movq mm2,mm1 ; A1 a1 | ||
278 : | paddd mm1,mm5 ; A1+B1 a1+b1 | ||
279 : | psubd mm2,mm5 ; A1-B1 a1-b1 | ||
280 : | psrad mm1,shift | ||
281 : | psrad mm2,shift | ||
282 : | packssdw mm7,mm1 ; A1+B1 a1+b1 A0+B0 a0+b0 | ||
283 : | packssdw mm2,mm4 ; A0-B0 a0-b0 A1-B1 a1-b1 | ||
284 : | movq [ dst ],mm7 | ||
285 : | movq mm1,[src1] ; R3 R1 r3 r1 | ||
286 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
287 : | movq [ dst + 24 ],mm2 | ||
288 : | pmaddwd mm4,mm1 ; -C1R3+C5R1 -C1r3+C5r1 | ||
289 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
290 : | pmaddwd mm1,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
291 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
292 : | movq mm2,mm0 ; A2 a2 | ||
293 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
294 : | paddd mm4,mm7 ; B2 b2 | ||
295 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
296 : | psubd mm0,mm4 ; a2-B2 a2-b2 | ||
297 : | psrad mm2,shift | ||
298 : | psrad mm0,shift | ||
299 : | movq mm4,mm6 ; A3 a3 | ||
300 : | paddd mm3,mm1 ; B3 b3 | ||
301 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
302 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
303 : | psrad mm6,shift | ||
304 : | packssdw mm2,mm6 ; A3+B3 a3+b3 A2+B2 a2+b2 | ||
305 : | movq [ dst + 8],mm2 | ||
306 : | psrad mm4,shift | ||
307 : | packssdw mm4,mm0 ; A2-B2 a2-b2 A3-B3 a3-b3 | ||
308 : | movq [dst + 16],mm4 | ||
309 : | edgomez | 1.2 | %undef src0 |
310 : | %undef src4 | ||
311 : | %undef src1 | ||
312 : | %undef src5 | ||
313 : | %undef dst | ||
314 : | %undef rounder_op | ||
315 : | %undef rounder_arg | ||
316 : | %undef shift | ||
317 : | %undef bt | ||
318 : | %endmacro | ||
319 : | |||
320 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
321 : | ; IDCT0 | ||
322 : | ;--------------------------------------------------------------------------- | ||
323 : | edgomez | 1.2 | |
324 : | %macro IDCT0 8 | ||
325 : | %define src0 %1 | ||
326 : | edgomez | 1.4 | %define src4 %2 |
327 : | %define src1 %3 | ||
328 : | %define src5 %4 | ||
329 : | %define dst %5 | ||
330 : | %define rounder_op %6 | ||
331 : | %define rounder_arg %7 | ||
332 : | %define shift %8 | ||
333 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
334 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
335 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
336 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
337 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
338 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
339 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
340 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
341 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
342 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
343 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
344 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
345 : | ; rounder_op mm4, rounder_arg | ||
346 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
347 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
348 : | ; rounder_op mm0, rounder_arg | ||
349 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
350 : | paddd mm4,mm5 ; A0 a0 | ||
351 : | psubd mm6,mm5 ; A3 a3 | ||
352 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
353 : | paddd mm0,mm1 ; A1 a1 | ||
354 : | psubd mm5,mm1 ; A2 a2 | ||
355 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
356 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
357 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
358 : | paddd mm7,mm1 ; B0 b0 | ||
359 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
360 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
361 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
362 : | paddd mm4,mm4 ; 2A0 2a0 | ||
363 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
364 : | paddd mm1,mm2 ; B1 b1 | ||
365 : | psrad mm7,shift | ||
366 : | psrad mm4,shift | ||
367 : | movq mm2,mm0 ; A1 a1 | ||
368 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
369 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
370 : | psrad mm0,shift | ||
371 : | psrad mm2,shift | ||
372 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
373 : | movd [ dst ],mm7 | ||
374 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
375 : | movd [ dst + 16],mm0 | ||
376 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
377 : | movd [ dst + 96 ],mm2 | ||
378 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
379 : | movd [ dst + 112],mm4 | ||
380 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
381 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
382 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
383 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
384 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
385 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
386 : | movq mm2,mm5 ; A2 a2 | ||
387 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
388 : | paddd mm4,mm7 ; B2 b2 | ||
389 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
390 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
391 : | psrad mm2,shift | ||
392 : | psrad mm5,shift | ||
393 : | movq mm4,mm6 ; A3 a3 | ||
394 : | paddd mm3,mm0 ; B3 b3 | ||
395 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
396 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
397 : | psrad mm6,shift | ||
398 : | psrad mm4,shift | ||
399 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
400 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
401 : | movd [ dst + 32 ],mm2 | ||
402 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
403 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
404 : | movd [ dst + 48 ],mm6 | ||
405 : | movd [ dst + 64 ],mm4 | ||
406 : | movd [ dst + 80 ],mm5 | ||
407 : | edgomez | 1.2 | %undef src0 |
408 : | %undef src4 | ||
409 : | %undef src1 | ||
410 : | %undef src5 | ||
411 : | %undef dst | ||
412 : | %undef rounder_op | ||
413 : | %undef rounder_arg | ||
414 : | %undef shift | ||
415 : | %endmacro | ||
416 : | |||
417 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
418 : | ; IDCT4 | ||
419 : | ;--------------------------------------------------------------------------- | ||
420 : | edgomez | 1.2 | |
421 : | %macro IDCT4 8 | ||
422 : | %define src0 %1 | ||
423 : | %define src4 %2 | ||
424 : | %define src1 %3 | ||
425 : | %define src5 %4 | ||
426 : | %define dst %5 | ||
427 : | %define rounder_op %6 | ||
428 : | %define rounder_arg %7 | ||
429 : | %define shift %8 | ||
430 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
431 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
432 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
433 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
434 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
435 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
436 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
437 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
438 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
439 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
440 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
441 : | ; rounder_op mm4, rounder_arg | ||
442 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
443 : | ; rounder_op mm0, rounder_arg | ||
444 : | paddd mm4,mm5 ; A0 a0 | ||
445 : | psubd mm6,mm5 ; A3 a3 | ||
446 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
447 : | paddd mm0,mm1 ; A1 a1 | ||
448 : | psubd mm5,mm1 ; A2 a2 | ||
449 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
450 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
451 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
452 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
453 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
454 : | paddd mm4,mm4 ; 2A0 2a0 | ||
455 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
456 : | psrad mm1,shift | ||
457 : | psrad mm4,shift | ||
458 : | movq mm2,mm0 ; A1 a1 | ||
459 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
460 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
461 : | psrad mm0,shift | ||
462 : | psrad mm2,shift | ||
463 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
464 : | movd [ dst ],mm1 | ||
465 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
466 : | movd [ dst + 16 ],mm0 | ||
467 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
468 : | movd [ dst + 96 ],mm2 | ||
469 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
470 : | movd [ dst + 112 ],mm4 | ||
471 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
472 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
473 : | movq mm2,mm5 ; A2 a2 | ||
474 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
475 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
476 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
477 : | psrad mm2,shift | ||
478 : | psrad mm5,shift | ||
479 : | movq mm1,mm6 ; A3 a3 | ||
480 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
481 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
482 : | psrad mm6,shift | ||
483 : | psrad mm1,shift | ||
484 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
485 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
486 : | movd [dst + 32],mm2 | ||
487 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
488 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
489 : | movd [dst + 48],mm6 | ||
490 : | movd [dst + 64],mm1 | ||
491 : | movd [dst + 80],mm5 | ||
492 : | edgomez | 1.2 | %undef src0 |
493 : | %undef src4 | ||
494 : | %undef src1 | ||
495 : | %undef src5 | ||
496 : | %undef dst | ||
497 : | %undef rounder_op | ||
498 : | %undef rounder_arg | ||
499 : | %undef shift | ||
500 : | %endmacro | ||
501 : | |||
502 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
503 : | ; IDCT6 | ||
504 : | ;--------------------------------------------------------------------------- | ||
505 : | edgomez | 1.2 | |
506 : | %macro IDCT6 8 | ||
507 : | %define src0 %1 | ||
508 : | %define src4 %2 | ||
509 : | %define src1 %3 | ||
510 : | %define src5 %4 | ||
511 : | %define dst %5 | ||
512 : | %define rounder_op %6 | ||
513 : | %define rounder_arg %7 | ||
514 : | %define shift %8 | ||
515 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
516 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
517 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
518 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
519 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
520 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
521 : | ; rounder_op mm4, rounder_arg | ||
522 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
523 : | ; rounder_op mm0, rounder_arg | ||
524 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
525 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
526 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
527 : | movq mm7,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
528 : | pmaddwd mm7,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
529 : | paddd mm1,mm4 ; A0+B0 a0+b0 | ||
530 : | paddd mm4,mm4 ; 2A0 2a0 | ||
531 : | psubd mm4,mm1 ; A0-B0 a0-b0 | ||
532 : | psrad mm1,shift | ||
533 : | psrad mm4,shift | ||
534 : | movq mm2,mm0 ; A1 a1 | ||
535 : | paddd mm0,mm7 ; A1+B1 a1+b1 | ||
536 : | psubd mm2,mm7 ; A1-B1 a1-b1 | ||
537 : | psrad mm0,shift | ||
538 : | psrad mm2,shift | ||
539 : | packssdw mm1,mm1 ; A0+B0 a0+b0 | ||
540 : | movd [ dst ],mm1 | ||
541 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
542 : | movd [ dst + 16 ],mm0 | ||
543 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
544 : | movd [ dst + 96 ],mm2 | ||
545 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
546 : | movd [ dst + 112 ],mm4 | ||
547 : | movq mm1,[coeffs+88] ; C3 C7 C3 C7 | ||
548 : | pmaddwd mm1,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
549 : | movq mm2,mm5 ; A2 a2 | ||
550 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
551 : | paddd mm2,mm1 ; A2+B2 a2+b2 | ||
552 : | psubd mm5,mm1 ; a2-B2 a2-b2 | ||
553 : | psrad mm2,shift | ||
554 : | psrad mm5,shift | ||
555 : | movq mm1,mm6 ; A3 a3 | ||
556 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
557 : | psubd mm1,mm3 ; a3-B3 a3-b3 | ||
558 : | psrad mm6,shift | ||
559 : | psrad mm1,shift | ||
560 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
561 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
562 : | movd [dst + 32],mm2 | ||
563 : | packssdw mm1,mm1 ; A3-B3 a3-b3 | ||
564 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
565 : | movd [dst + 48],mm6 | ||
566 : | movd [dst + 64],mm1 | ||
567 : | movd [dst + 80],mm5 | ||
568 : | %undef src0 | ||
569 : | %undef src4 | ||
570 : | %undef src1 | ||
571 : | %undef src5 | ||
572 : | %undef dst | ||
573 : | %undef rounder_op | ||
574 : | %undef rounder_arg | ||
575 : | edgomez | 1.2 | %undef shift |
576 : | %endmacro | ||
577 : | |||
578 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
579 : | ; IDCT2 | ||
580 : | ;--------------------------------------------------------------------------- | ||
581 : | edgomez | 1.2 | |
582 : | %macro IDCT2 8 | ||
583 : | %define src0 %1 | ||
584 : | %define src4 %2 | ||
585 : | %define src1 %3 | ||
586 : | %define src5 %4 | ||
587 : | %define dst %5 | ||
588 : | %define rounder_op %6 | ||
589 : | %define rounder_arg %7 | ||
590 : | edgomez | 1.4 | %define shift %8 |
591 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
592 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
593 : | movq mm3,[src5] ; R7 R5 r7 r5 | ||
594 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
595 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
596 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
597 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
598 : | ; rounder_op mm4, rounder_arg | ||
599 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
600 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
601 : | ; rounder_op mm0, rounder_arg | ||
602 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
603 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
604 : | movq mm1,[coeffs+56] ; C7 C5 C7 C5 | ||
605 : | pmaddwd mm1,mm3 ; C7R7+C5R5 C7r7+C5r5 | ||
606 : | pmaddwd mm2,[coeffs+64] ; -C7R3+C3R1 -C7r3+C3r1 | ||
607 : | paddd mm7,mm1 ; B0 b0 | ||
608 : | movq mm1,[coeffs+72] ; -C5 -C1 -C5 -C1 | ||
609 : | pmaddwd mm1,mm3 ; -C5R7-C1R5 -C5r7-C1r5 | ||
610 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
611 : | paddd mm4,mm4 ; 2A0 2a0 | ||
612 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
613 : | paddd mm1,mm2 ; B1 b1 | ||
614 : | psrad mm7,shift | ||
615 : | psrad mm4,shift | ||
616 : | movq mm2,mm0 ; A1 a1 | ||
617 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
618 : | psubd mm2,mm1 ; A1-B1 a1-b1 | ||
619 : | psrad mm0,shift | ||
620 : | psrad mm2,shift | ||
621 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
622 : | movd [dst],mm7 | ||
623 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
624 : | movd [dst + 16],mm0 | ||
625 : | packssdw mm2,mm2 ; A1-B1 a1-b1 | ||
626 : | movd [dst + 96],mm2 | ||
627 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
628 : | movd [dst + 112],mm4 | ||
629 : | movq mm0,[src1] ; R3 R1 r3 r1 | ||
630 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
631 : | pmaddwd mm4,mm0 ; -C1R3+C5R1 -C1r3+C5r1 | ||
632 : | movq mm7,[coeffs+88] ; C3 C7 C3 C7 | ||
633 : | pmaddwd mm0,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
634 : | pmaddwd mm7,mm3 ; C3R7+C7R5 C3r7+C7r5 | ||
635 : | movq mm2,mm5 ; A2 a2 | ||
636 : | pmaddwd mm3,[coeffs+104] ; -C1R7+C3R5 -C1r7+C3r5 | ||
637 : | paddd mm4,mm7 ; B2 b2 | ||
638 : | paddd mm2,mm4 ; A2+B2 a2+b2 | ||
639 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
640 : | psrad mm2,shift | ||
641 : | psrad mm5,shift | ||
642 : | movq mm4,mm6 ; A3 a3 | ||
643 : | paddd mm3,mm0 ; B3 b3 | ||
644 : | paddd mm6,mm3 ; A3+B3 a3+b3 | ||
645 : | psubd mm4,mm3 ; a3-B3 a3-b3 | ||
646 : | psrad mm6,shift | ||
647 : | psrad mm4,shift | ||
648 : | packssdw mm2,mm2 ; A2+B2 a2+b2 | ||
649 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
650 : | movd [dst + 32],mm2 | ||
651 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
652 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
653 : | movd [dst + 48],mm6 | ||
654 : | movd [dst + 64],mm4 | ||
655 : | movd [dst + 80],mm5 | ||
656 : | %undef src0 | ||
657 : | %undef src4 | ||
658 : | %undef src1 | ||
659 : | %undef src5 | ||
660 : | %undef dst | ||
661 : | %undef rounder_op | ||
662 : | %undef rounder_arg | ||
663 : | %undef shift | ||
664 : | %endmacro | ||
665 : | |||
666 : | ;--------------------------------------------------------------------------- | ||
667 : | ; IDCT3 | ||
668 : | ;--------------------------------------------------------------------------- | ||
669 : | |||
670 : | %macro IDCT3 8 | ||
671 : | %define src0 %1 | ||
672 : | %define src4 %2 | ||
673 : | %define src1 %3 | ||
674 : | %define src5 %4 | ||
675 : | %define dst %5 | ||
676 : | %define rounder_op %6 | ||
677 : | %define rounder_arg %7 | ||
678 : | %define shift %8 | ||
679 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
680 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
681 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
682 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
683 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
684 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
685 : | ; rounder_op mm4, rounder_arg | ||
686 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
687 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
688 : | ; rounder_op mm0, rounder_arg | ||
689 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
690 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
691 : | movq mm3,[coeffs+64] | ||
692 : | pmaddwd mm3,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
693 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
694 : | paddd mm4,mm4 ; 2A0 2a0 | ||
695 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
696 : | psrad mm7,shift | ||
697 : | psrad mm4,shift | ||
698 : | movq mm1,mm0 ; A1 a1 | ||
699 : | paddd mm0,mm3 ; A1+B1 a1+b1 | ||
700 : | psubd mm1,mm3 ; A1-B1 a1-b1 | ||
701 : | psrad mm0,shift | ||
702 : | psrad mm1,shift | ||
703 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
704 : | movd [dst],mm7 | ||
705 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
706 : | movd [dst + 16],mm0 | ||
707 : | packssdw mm1,mm1 ; A1-B1 a1-b1 | ||
708 : | movd [dst + 96],mm1 | ||
709 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
710 : | movd [dst + 112],mm4 | ||
711 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
712 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
713 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
714 : | movq mm1,mm5 ; A2 a2 | ||
715 : | paddd mm1,mm4 ; A2+B2 a2+b2 | ||
716 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
717 : | psrad mm1,shift | ||
718 : | psrad mm5,shift | ||
719 : | movq mm4,mm6 ; A3 a3 | ||
720 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
721 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
722 : | psrad mm6,shift | ||
723 : | psrad mm4,shift | ||
724 : | packssdw mm1,mm1 ; A2+B2 a2+b2 | ||
725 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
726 : | movd [dst + 32],mm1 | ||
727 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
728 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
729 : | movd [dst + 48],mm6 | ||
730 : | movd [dst + 64],mm4 | ||
731 : | movd [dst + 80],mm5 | ||
732 : | %undef src0 | ||
733 : | edgomez | 1.2 | %undef src4 |
734 : | %undef src1 | ||
735 : | %undef src5 | ||
736 : | %undef dst | ||
737 : | %undef rounder_op | ||
738 : | %undef rounder_arg | ||
739 : | %undef shift | ||
740 : | %endmacro | ||
741 : | |||
742 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
743 : | ; IDCT5 | ||
744 : | ;--------------------------------------------------------------------------- | ||
745 : | edgomez | 1.2 | |
746 : | edgomez | 1.4 | %macro IDCT5 8 |
747 : | edgomez | 1.2 | %define src0 %1 |
748 : | %define src4 %2 | ||
749 : | %define src1 %3 | ||
750 : | %define src5 %4 | ||
751 : | %define dst %5 | ||
752 : | %define rounder_op %6 | ||
753 : | %define rounder_arg %7 | ||
754 : | %define shift %8 | ||
755 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
756 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
757 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
758 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
759 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
760 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
761 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
762 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
763 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
764 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
765 : | ; rounder_op mm4, rounder_arg | ||
766 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
767 : | paddd mm4,mm5 ; A0 a0 | ||
768 : | ; rounder_op mm0, rounder_arg | ||
769 : | psubd mm6,mm5 ; A3 a3 | ||
770 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
771 : | paddd mm0,mm1 ; A1 a1 | ||
772 : | psubd mm5,mm1 ; A2 a2 | ||
773 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
774 : | movq mm3,[src4 + 8] ; R6 R2 r6 r2 | ||
775 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
776 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
777 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
778 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
779 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
780 : | pmaddwd mm7,mm3 ; C6R6+C2R2 C6r6+C2r2 | ||
781 : | pmaddwd mm3,[coeffs+40] ; -C2R6+C6R2 -C2r6+C6r2 | ||
782 : | ; rounder_op mm1, rounder_arg | ||
783 : | paddd mm7,mm1 ; A0 a0 | ||
784 : | paddd mm1,mm1 ; 2C0 2c0 | ||
785 : | ; rounder_op mm2, rounder_arg | ||
786 : | psubd mm1,mm7 ; A3 a3 | ||
787 : | paddd mm3,mm2 ; A1 a1 | ||
788 : | paddd mm2,mm2 ; 2C1 2c1 | ||
789 : | psubd mm2,mm3 ; A2 a2 | ||
790 : | psrad mm4,shift | ||
791 : | psrad mm7,shift | ||
792 : | psrad mm3,shift | ||
793 : | packssdw mm4,mm7 ; A0 a0 | ||
794 : | movq [dst],mm4 | ||
795 : | psrad mm0,shift | ||
796 : | packssdw mm0,mm3 ; A1 a1 | ||
797 : | movq [dst + 16],mm0 | ||
798 : | movq [dst + 96],mm0 | ||
799 : | movq [dst + 112],mm4 | ||
800 : | psrad mm5,shift | ||
801 : | psrad mm6,shift | ||
802 : | psrad mm2,shift | ||
803 : | packssdw mm5,mm2 ; A2-B2 a2-b2 | ||
804 : | movq [dst + 32],mm5 | ||
805 : | psrad mm1,shift | ||
806 : | packssdw mm6,mm1 ; A3+B3 a3+b3 | ||
807 : | movq [dst + 48],mm6 | ||
808 : | movq [dst + 64],mm6 | ||
809 : | movq [dst + 80],mm5 | ||
810 : | %undef src0 | ||
811 : | %undef src4 | ||
812 : | %undef src1 | ||
813 : | %undef src5 | ||
814 : | %undef dst | ||
815 : | %undef rounder_op | ||
816 : | edgomez | 1.2 | %undef rounder_arg |
817 : | %undef shift | ||
818 : | %endmacro | ||
819 : | |||
820 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
821 : | ; IDCT1 | ||
822 : | ;--------------------------------------------------------------------------- | ||
823 : | edgomez | 1.2 | |
824 : | edgomez | 1.4 | %macro IDCT1 8 |
825 : | edgomez | 1.2 | %define src0 %1 |
826 : | %define src4 %2 | ||
827 : | %define src1 %3 | ||
828 : | %define src5 %4 | ||
829 : | edgomez | 1.4 | %define dst %5 |
830 : | %define rounder_op %6 | ||
831 : | %define rounder_arg %7 | ||
832 : | %define shift %8 | ||
833 : | movq mm0,[src0] ; R4 R0 r4 r0 | ||
834 : | movq mm1,[src4] ; R6 R2 r6 r2 | ||
835 : | movq mm2,[src1] ; R3 R1 r3 r1 | ||
836 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
837 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
838 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
839 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
840 : | movq mm5,[coeffs+32] ; C6 C2 C6 C2 | ||
841 : | pmaddwd mm5,mm1 ; C6R6+C2R2 C6r6+C2r2 | ||
842 : | movq mm6,[coeffs+40] ; -C2 C6 -C2 C6 | ||
843 : | pmaddwd mm1,mm6 ; -C2R6+C6R2 -C2r6+C6r2 | ||
844 : | ; rounder_op mm4, rounder_arg | ||
845 : | movq mm6,mm4 ; C4R4+C4R0 C4r4+C4r0 | ||
846 : | movq mm7,[coeffs+48] ; C3 C1 C3 C1 | ||
847 : | ; rounder_op mm0, rounder_arg | ||
848 : | pmaddwd mm7,mm2 ; C3R3+C1R1 C3r3+C1r1 | ||
849 : | paddd mm4,mm5 ; A0 a0 | ||
850 : | psubd mm6,mm5 ; A3 a3 | ||
851 : | movq mm5,mm0 ; -C4R4+C4R0 -C4r4+C4r0 | ||
852 : | paddd mm0,mm1 ; A1 a1 | ||
853 : | psubd mm5,mm1 ; A2 a2 | ||
854 : | movq mm1,[coeffs+64] | ||
855 : | pmaddwd mm1,mm2 ; -C7R3+C3R1 -C7r3+C3r1 | ||
856 : | paddd mm7,mm4 ; A0+B0 a0+b0 | ||
857 : | paddd mm4,mm4 ; 2A0 2a0 | ||
858 : | psubd mm4,mm7 ; A0-B0 a0-b0 | ||
859 : | psrad mm7,shift | ||
860 : | psrad mm4,shift | ||
861 : | movq mm3,mm0 ; A1 a1 | ||
862 : | paddd mm0,mm1 ; A1+B1 a1+b1 | ||
863 : | psubd mm3,mm1 ; A1-B1 a1-b1 | ||
864 : | psrad mm0,shift | ||
865 : | psrad mm3,shift | ||
866 : | packssdw mm7,mm7 ; A0+B0 a0+b0 | ||
867 : | movd [dst],mm7 | ||
868 : | packssdw mm0,mm0 ; A1+B1 a1+b1 | ||
869 : | movd [dst + 16],mm0 | ||
870 : | packssdw mm3,mm3 ; A1-B1 a1-b1 | ||
871 : | movd [dst + 96],mm3 | ||
872 : | packssdw mm4,mm4 ; A0-B0 a0-b0 | ||
873 : | movd [dst + 112],mm4 | ||
874 : | movq mm4,[coeffs+80] ; -C1 C5 -C1 C5 | ||
875 : | pmaddwd mm4,mm2 ; -C1R3+C5R1 -C1r3+C5r1 | ||
876 : | pmaddwd mm2,[coeffs+96] ; -C5R3+C7R1 -C5r3+C7r1 | ||
877 : | movq mm3,mm5 ; A2 a2 | ||
878 : | paddd mm3,mm4 ; A2+B2 a2+b2 | ||
879 : | psubd mm5,mm4 ; a2-B2 a2-b2 | ||
880 : | psrad mm3,shift | ||
881 : | psrad mm5,shift | ||
882 : | movq mm4,mm6 ; A3 a3 | ||
883 : | paddd mm6,mm2 ; A3+B3 a3+b3 | ||
884 : | psubd mm4,mm2 ; a3-B3 a3-b3 | ||
885 : | psrad mm6,shift | ||
886 : | packssdw mm3,mm3 ; A2+B2 a2+b2 | ||
887 : | movd [dst + 32],mm3 | ||
888 : | psrad mm4,shift | ||
889 : | packssdw mm6,mm6 ; A3+B3 a3+b3 | ||
890 : | movd [dst + 48],mm6 | ||
891 : | packssdw mm4,mm4 ; A3-B3 a3-b3 | ||
892 : | packssdw mm5,mm5 ; A2-B2 a2-b2 | ||
893 : | movd [dst + 64],mm4 | ||
894 : | movd [dst + 80],mm5 | ||
895 : | %undef src0 | ||
896 : | %undef src4 | ||
897 : | %undef src1 | ||
898 : | %undef src5 | ||
899 : | %undef dst | ||
900 : | %undef rounder_op | ||
901 : | %undef rounder_arg | ||
902 : | %undef shift | ||
903 : | edgomez | 1.2 | %endmacro |
904 : | |||
905 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
906 : | ; IDCT7 | ||
907 : | ;--------------------------------------------------------------------------- | ||
908 : | edgomez | 1.2 | |
909 : | edgomez | 1.4 | %macro IDCT7 8 |
910 : | edgomez | 1.2 | %define src0 %1 |
911 : | %define src4 %2 | ||
912 : | %define src1 %3 | ||
913 : | %define src5 %4 | ||
914 : | %define dst %5 | ||
915 : | %define rounder_op %6 | ||
916 : | %define rounder_arg %7 | ||
917 : | %define shift %8 | ||
918 : | edgomez | 1.4 | movq mm0,[src0] ; R4 R0 r4 r0 |
919 : | movq mm4,[coeffs+16] ; C4 C4 C4 C4 | ||
920 : | pmaddwd mm4,mm0 ; C4R4+C4R0 C4r4+C4r0 | ||
921 : | movq mm5,[coeffs+24] ; -C4 C4 -C4 C4 | ||
922 : | pmaddwd mm0,mm5 ; -C4R4+C4R0 -C4r4+C4r0 | ||
923 : | ; rounder_op mm4, rounder_arg | ||
924 : | ; rounder_op mm0, rounder_arg | ||
925 : | psrad mm4,shift | ||
926 : | psrad mm0,shift | ||
927 : | movq mm2,[src0 + 8] ; R4 R0 r4 r0 | ||
928 : | movq mm1,[coeffs+16] ; C4 C4 C4 C4 | ||
929 : | pmaddwd mm1,mm2 ; C4R4+C4R0 C4r4+C4r0 | ||
930 : | movq mm7,[coeffs+24] ; -C4 C4 -C4 C4 | ||
931 : | pmaddwd mm2,mm7 ; -C4R4+C4R0 -C4r4+C4r0 | ||
932 : | movq mm7,[coeffs+32] ; C6 C2 C6 C2 | ||
933 : | ; rounder_op mm1, rounder_arg | ||
934 : | ; rounder_op mm2, rounder_arg | ||
935 : | psrad mm1,shift | ||
936 : | packssdw mm4,mm1 ; A0 a0 | ||
937 : | movq [dst],mm4 | ||
938 : | psrad mm2,shift | ||
939 : | packssdw mm0,mm2 ; A1 a1 | ||
940 : | movq [dst + 16],mm0 | ||
941 : | movq [dst + 96],mm0 | ||
942 : | movq [dst + 112],mm4 | ||
943 : | movq [dst + 32],mm0 | ||
944 : | movq [dst + 48],mm4 | ||
945 : | movq [dst + 64],mm4 | ||
946 : | movq [dst + 80],mm0 | ||
947 : | edgomez | 1.2 | %undef src0 |
948 : | %undef src4 | ||
949 : | %undef src1 | ||
950 : | %undef src5 | ||
951 : | %undef dst | ||
952 : | %undef rounder_op | ||
953 : | %undef rounder_arg | ||
954 : | %undef shift | ||
955 : | %endmacro | ||
956 : | |||
957 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
958 : | ; Permutation helpers | ||
959 : | ;--------------------------------------------------------------------------- | ||
960 : | |||
961 : | %macro XLODA 2 | ||
962 : | mov bx, [srcP+2*%2] ; get src contents | ||
963 : | mov ax, [srcP+2*%1] ; get dest contents | ||
964 : | mov [srcP+2*%1], bx ; store new dest val | ||
965 : | %endmacro | ||
966 : | edgomez | 1.2 | |
967 : | edgomez | 1.4 | %macro XCHGA 2 |
968 : | mov ax, [srcP+2*%1] ; get dest contents | ||
969 : | mov [srcP+2*%1], bx ; store new dest val | ||
970 : | %endmacro | ||
971 : | edgomez | 1.2 | |
972 : | edgomez | 1.4 | %macro XCHGB 2 |
973 : | mov bx, [srcP+2*%1] ; get dest contents | ||
974 : | mov [srcP+2*%1], ax ; store new dest val | ||
975 : | %endmacro | ||
976 : | |||
977 : | %macro XSTRA 2 | ||
978 : | mov [srcP+2*%1], bx ; store dest val | ||
979 : | %endmacro | ||
980 : | edgomez | 1.2 | |
981 : | edgomez | 1.4 | %macro XSTRB 2 |
982 : | mov [srcP+2*%1], ax ; store dest val | ||
983 : | edgomez | 1.2 | %endmacro |
984 : | |||
985 : | edgomez | 1.4 | ;--------------------------------------------------------------------------- |
986 : | ; Permutation macro | ||
987 : | ;--------------------------------------------------------------------------- | ||
988 : | |||
989 : | %macro PERMUTEP 1 | ||
990 : | %define srcP %1 | ||
991 : | push ebx | ||
992 : | |||
993 : | ; XCHGA 0x00, 0x00 ; nothing to do | ||
994 : | |||
995 : | XLODA 0x08, 0x01 | ||
996 : | XCHGB 0x10, 0x08 | ||
997 : | XCHGA 0x20, 0x10 | ||
998 : | XCHGB 0x02, 0x20 | ||
999 : | XCHGA 0x04, 0x02 | ||
1000 : | XSTRB 0x01, 0x04 | ||
1001 : | |||
1002 : | XLODA 0x09, 0x03 | ||
1003 : | XCHGB 0x18, 0x09 | ||
1004 : | XCHGA 0x12, 0x18 | ||
1005 : | XCHGB 0x24, 0x12 | ||
1006 : | XSTRA 0x03, 0x24 | ||
1007 : | |||
1008 : | XLODA 0x0C, 0x05 | ||
1009 : | XCHGB 0x11, 0x0C | ||
1010 : | XCHGA 0x28, 0x11 | ||
1011 : | XCHGB 0x30, 0x28 | ||
1012 : | XCHGA 0x22, 0x30 | ||
1013 : | XCHGB 0x06, 0x22 | ||
1014 : | XSTRA 0x05, 0x06 | ||
1015 : | |||
1016 : | XLODA 0x0D, 0x07 | ||
1017 : | XCHGB 0x1C, 0x0D | ||
1018 : | XCHGA 0x13, 0x1C | ||
1019 : | XCHGB 0x29, 0x13 | ||
1020 : | XCHGA 0x38, 0x29 | ||
1021 : | XCHGB 0x32, 0x38 | ||
1022 : | XCHGA 0x26, 0x32 | ||
1023 : | XSTRB 0x07, 0x26 | ||
1024 : | |||
1025 : | XLODA 0x14, 0x0A | ||
1026 : | XCHGB 0x21, 0x14 | ||
1027 : | XSTRA 0x0A, 0x21 | ||
1028 : | |||
1029 : | XLODA 0x19, 0x0B | ||
1030 : | XCHGB 0x1A, 0x19 | ||
1031 : | XCHGA 0x16, 0x1A | ||
1032 : | XCHGB 0x25, 0x16 | ||
1033 : | XCHGA 0x0E, 0x25 | ||
1034 : | XCHGB 0x15, 0x0E | ||
1035 : | XCHGA 0x2C, 0x15 | ||
1036 : | XCHGB 0x31, 0x2C | ||
1037 : | XCHGA 0x2A, 0x31 | ||
1038 : | XCHGB 0x34, 0x2A | ||
1039 : | XCHGA 0x23, 0x34 | ||
1040 : | XSTRB 0x0B, 0x23 | ||
1041 : | |||
1042 : | XLODA 0x1D, 0x0F | ||
1043 : | XCHGB 0x1E, 0x1D | ||
1044 : | XCHGA 0x17, 0x1E | ||
1045 : | XCHGB 0x2D, 0x17 | ||
1046 : | XCHGA 0x3C, 0x2D | ||
1047 : | XCHGB 0x33, 0x3C | ||
1048 : | XCHGA 0x2B, 0x33 | ||
1049 : | XCHGB 0x39, 0x2B | ||
1050 : | XCHGA 0x3A, 0x39 | ||
1051 : | XCHGB 0x36, 0x3A | ||
1052 : | XCHGA 0x27, 0x36 | ||
1053 : | XSTRB 0x0F, 0x27 | ||
1054 : | |||
1055 : | ; XCHGA 0x1B, 0x1B | ||
1056 : | |||
1057 : | ; XCHGA 0x1F, 0x1F | ||
1058 : | |||
1059 : | XLODA 0x35, 0x2E | ||
1060 : | XSTRB 0x2E, 0x35 | ||
1061 : | |||
1062 : | XLODA 0x3D, 0x2F | ||
1063 : | XCHGB 0x3E, 0x3D | ||
1064 : | XCHGA 0x37, 0x3E | ||
1065 : | XSTRB 0x2F, 0x37 | ||
1066 : | edgomez | 1.2 | |
1067 : | edgomez | 1.4 | ; XCHGA 0x3B, 0x3B |
1068 : | edgomez | 1.2 | |
1069 : | edgomez | 1.4 | ; XCHGA 0x3F, 0x3F |
1070 : | pop ebx | ||
1071 : | %undef srcP | ||
1072 : | edgomez | 1.2 | %endmacro |
1073 : | |||
1074 : | edgomez | 1.4 | ;============================================================================= |
1075 : | ; Code | ||
1076 : | ;============================================================================= | ||
1077 : | |||
1078 : | SECTION .text | ||
1079 : | |||
1080 : | cglobal simple_idct_mmx_P | ||
1081 : | cglobal simple_idct_mmx | ||
1082 : | edgomez | 1.2 | |
1083 : | edgomez | 1.4 | ;----------------------------------------------------------------------------- |
1084 : | ; void simple_idct_mmx_P(int16_t * const block) | ||
1085 : | Isibaar | 1.3 | ; expects input data to be permutated |
1086 : | edgomez | 1.4 | ;----------------------------------------------------------------------------- |
1087 : | |||
1088 : | ALIGN 16 | ||
1089 : | simple_idct_mmx_P: | ||
1090 : | sub esp, 128 | ||
1091 : | mov edx, [esp+128+4] | ||
1092 : | |||
1093 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
1094 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
1095 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .four | ||
1096 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .two | ||
1097 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .one | ||
1098 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1099 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1100 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1101 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1102 : | jmp .ret | ||
1103 : | edgomez | 1.2 | |
1104 : | edgomez | 1.4 | ALIGN 16 |
1105 : | Isibaar | 1.9 | .four: |
1106 : | edgomez | 1.4 | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .six |
1107 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .five | ||
1108 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1109 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1110 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1111 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1112 : | jmp .ret | ||
1113 : | edgomez | 1.2 | |
1114 : | edgomez | 1.4 | ALIGN 16 |
1115 : | Isibaar | 1.9 | .six: |
1116 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .seven |
1117 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1118 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1119 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1120 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1121 : | jmp .ret | ||
1122 : | edgomez | 1.2 | |
1123 : | edgomez | 1.4 | ALIGN 16 |
1124 : | Isibaar | 1.9 | .two: |
1125 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .three |
1126 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1127 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1128 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1129 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1130 : | jmp .ret | ||
1131 : | edgomez | 1.2 | |
1132 : | edgomez | 1.4 | ALIGN 16 |
1133 : | Isibaar | 1.9 | .three: |
1134 : | edgomez | 1.4 | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1135 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1136 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1137 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1138 : | jmp .ret | ||
1139 : | edgomez | 1.2 | |
1140 : | edgomez | 1.4 | ALIGN 16 |
1141 : | Isibaar | 1.9 | .five: |
1142 : | edgomez | 1.4 | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1143 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1144 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1145 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1146 : | jmp .ret | ||
1147 : | edgomez | 1.2 | |
1148 : | edgomez | 1.4 | ALIGN 16 |
1149 : | Isibaar | 1.9 | .one: |
1150 : | edgomez | 1.4 | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1151 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1152 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1153 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1154 : | jmp .ret | ||
1155 : | edgomez | 1.2 | |
1156 : | edgomez | 1.4 | ALIGN 16 |
1157 : | Isibaar | 1.9 | .seven: |
1158 : | edgomez | 1.4 | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1159 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1160 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1161 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1162 : | edgomez | 1.2 | |
1163 : | Isibaar | 1.9 | .ret: |
1164 : | edgomez | 1.4 | add esp, 128 |
1165 : | Isibaar | 1.3 | |
1166 : | edgomez | 1.4 | ret |
1167 : | Isibaar | 1.9 | ENDFUNC |
1168 : | Isibaar | 1.3 | |
1169 : | edgomez | 1.4 | |
1170 : | ;----------------------------------------------------------------------------- | ||
1171 : | ; void simple_idct_mmx(int16_t * const block) | ||
1172 : | Isibaar | 1.3 | ; |
1173 : | edgomez | 1.4 | ; simple_idct_mmx is the same function as simple_idct_mmx_P above except that |
1174 : | ; on entry it will do a fast in-line and in-place permutation on the iDCT parm | ||
1175 : | ; list. This means that same parm list will also not have to be copied on the | ||
1176 : | ; way out. - trbarry 6/2003 | ||
1177 : | ;----------------------------------------------------------------------------- | ||
1178 : | |||
1179 : | ALIGN 16 | ||
1180 : | simple_idct_mmx: | ||
1181 : | sub esp, 128 | ||
1182 : | mov edx, [esp+128+4] | ||
1183 : | PERMUTEP edx ; permute parm list in place | ||
1184 : | |||
1185 : | ; src0, src4, src1, src5, dst, rndop, rndarg, shift, bt | ||
1186 : | DC_COND_IDCT edx+0, edx+8, edx+16, edx+24, esp, paddd, [coeffs+8], 11 | ||
1187 : | Z_COND_IDCT edx+32, edx+40, edx+48, edx+56, esp+32, paddd, [coeffs], 11, .fourP | ||
1188 : | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .twoP | ||
1189 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .oneP | ||
1190 : | IDCT0 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1191 : | IDCT0 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1192 : | IDCT0 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1193 : | IDCT0 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1194 : | jmp .retP | ||
1195 : | Isibaar | 1.3 | |
1196 : | edgomez | 1.4 | ALIGN 16 |
1197 : | Isibaar | 1.9 | .fourP: |
1198 : | edgomez | 1.4 | Z_COND_IDCT edx+64, edx+72, edx+80, edx+88, esp+64, paddd, [coeffs], 11, .sixP |
1199 : | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .fiveP | ||
1200 : | IDCT4 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1201 : | IDCT4 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1202 : | IDCT4 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1203 : | IDCT4 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1204 : | jmp .retP | ||
1205 : | Isibaar | 1.3 | |
1206 : | edgomez | 1.4 | ALIGN 16 |
1207 : | Isibaar | 1.9 | .sixP: |
1208 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .sevenP |
1209 : | IDCT6 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1210 : | IDCT6 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1211 : | IDCT6 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1212 : | IDCT6 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1213 : | jmp .retP | ||
1214 : | Isibaar | 1.3 | |
1215 : | edgomez | 1.4 | ALIGN 16 |
1216 : | Isibaar | 1.9 | .twoP: |
1217 : | edgomez | 1.4 | Z_COND_IDCT edx+96, edx+104,edx+112,edx+120,esp+96, paddd, [coeffs], 11, .threeP |
1218 : | IDCT2 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 | ||
1219 : | IDCT2 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1220 : | IDCT2 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1221 : | IDCT2 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1222 : | jmp .retP | ||
1223 : | Isibaar | 1.3 | |
1224 : | edgomez | 1.4 | ALIGN 16 |
1225 : | Isibaar | 1.9 | .threeP: |
1226 : | edgomez | 1.4 | IDCT3 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1227 : | IDCT3 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1228 : | IDCT3 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1229 : | IDCT3 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1230 : | jmp .retP | ||
1231 : | Isibaar | 1.3 | |
1232 : | edgomez | 1.4 | ALIGN 16 |
1233 : | Isibaar | 1.9 | .fiveP: |
1234 : | edgomez | 1.4 | IDCT5 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1235 : | ; IDCT5 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1236 : | IDCT5 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1237 : | ; IDCT5 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1238 : | jmp .retP | ||
1239 : | Isibaar | 1.3 | |
1240 : | edgomez | 1.4 | ALIGN 16 |
1241 : | Isibaar | 1.9 | .oneP: |
1242 : | edgomez | 1.4 | IDCT1 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1243 : | IDCT1 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1244 : | IDCT1 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1245 : | IDCT1 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1246 : | jmp .retP | ||
1247 : | Isibaar | 1.3 | |
1248 : | edgomez | 1.4 | ALIGN 16 |
1249 : | Isibaar | 1.9 | .sevenP: |
1250 : | edgomez | 1.4 | IDCT7 esp, esp+64, esp+32, esp+96, edx, nop, 0, 20 |
1251 : | ; IDCT7 esp+8, esp+72, esp+40, esp+104,edx+4, nop, 0, 20 | ||
1252 : | IDCT7 esp+16, esp+80, esp+48, esp+112,edx+8, nop, 0, 20 | ||
1253 : | ; IDCT7 esp+24, esp+88, esp+56, esp+120,edx+12, nop, 0, 20 | ||
1254 : | Isibaar | 1.3 | |
1255 : | Isibaar | 1.9 | .retP: |
1256 : | edgomez | 1.4 | add esp, 128 |
1257 : | edgomez | 1.2 | |
1258 : | edgomez | 1.4 | ret |
1259 : | Isibaar | 1.9 | ENDFUNC |
1260 : | edgomez | 1.7 | |
1261 : | Isibaar | 1.8 | |
1262 : | %ifidn __OUTPUT_FORMAT__,elf | ||
1263 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
1264 : | %endif | ||
1265 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |