Parent Directory | Revision Log
Revision 1176 - (view) (download)
1 : | edgomez | 1174 | ;/************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | edgomez | 1176 | ; * - 3dne Quantization/Dequantization - |
5 : | edgomez | 1174 | ; * |
6 : | edgomez | 1176 | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> |
7 : | ; * 2002 Jaan Kalda | ||
8 : | edgomez | 1174 | ; * |
9 : | ; * This program is free software ; you can redistribute it and/or modify | ||
10 : | ; * it under the terms of the GNU General Public License as published by | ||
11 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
12 : | ; * (at your option) any later version. | ||
13 : | ; * | ||
14 : | ; * This program is distributed in the hope that it will be useful, | ||
15 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
16 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
17 : | ; * GNU General Public License for more details. | ||
18 : | ; * | ||
19 : | ; * You should have received a copy of the GNU General Public License | ||
20 : | ; * along with this program ; if not, write to the Free Software | ||
21 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
22 : | ; * | ||
23 : | edgomez | 1176 | ; * $Id: quantize_mpeg_xmm.asm,v 1.1.2.2 2003-10-09 18:50:22 edgomez Exp $ |
24 : | edgomez | 1174 | ; * |
25 : | ; *************************************************************************/ | ||
26 : | edgomez | 1176 | |
27 : | edgomez | 1174 | ; _3dne functions are compatible with iSSE, but are optimized specifically |
28 : | ; for K7 pipelines | ||
29 : | |||
30 : | ; data/text alignment | ||
31 : | %define ALIGN 8 | ||
32 : | %define SATURATE | ||
33 : | |||
34 : | bits 32 | ||
35 : | |||
36 : | %macro cglobal 1 | ||
37 : | %ifdef PREFIX | ||
38 : | global _%1 | ||
39 : | %define %1 _%1 | ||
40 : | %else | ||
41 : | global %1 | ||
42 : | %endif | ||
43 : | %endmacro | ||
44 : | |||
45 : | %macro cextern 1 | ||
46 : | %ifdef PREFIX | ||
47 : | extern _%1 | ||
48 : | %define %1 _%1 | ||
49 : | %else | ||
50 : | extern %1 | ||
51 : | %endif | ||
52 : | %endmacro | ||
53 : | edgomez | 1176 | |
54 : | ;*************************************************************************** | ||
55 : | ; Local data | ||
56 : | ;*************************************************************************** | ||
57 : | |||
58 : | %ifdef FORMAT_COFF | ||
59 : | SECTION .data data | ||
60 : | %else | ||
61 : | SECTION .data data align=8 | ||
62 : | %endif | ||
63 : | |||
64 : | edgomez | 1174 | align 8 |
65 : | edgomez | 1176 | mmzero: |
66 : | dd 0,0 | ||
67 : | mmx_one: | ||
68 : | times 4 dw 1 | ||
69 : | edgomez | 1174 | |
70 : | ;=========================================================================== | ||
71 : | ; | ||
72 : | ; divide by 2Q table | ||
73 : | ; | ||
74 : | ;=========================================================================== | ||
75 : | |||
76 : | align ALIGN | ||
77 : | edgomez | 1176 | mmx_divs: ;i>2 |
78 : | edgomez | 1174 | %assign i 1 |
79 : | %rep 31 | ||
80 : | times 4 dw ((1 << 15) / i + 1) | ||
81 : | %assign i i+1 | ||
82 : | %endrep | ||
83 : | |||
84 : | align ALIGN | ||
85 : | edgomez | 1176 | mmx_div: ;i>2 |
86 : | edgomez | 1174 | %assign i 1 |
87 : | %rep 31 | ||
88 : | times 4 dw ((1 << 16) / i + 1) | ||
89 : | %assign i i+1 | ||
90 : | %endrep | ||
91 : | |||
92 : | |||
93 : | ;=========================================================================== | ||
94 : | ; | ||
95 : | ; intra matrix | ||
96 : | ; | ||
97 : | ;=========================================================================== | ||
98 : | |||
99 : | %macro FIXX 1 | ||
100 : | dw (1 << 16) / (%1) + 1 | ||
101 : | %endmacro | ||
102 : | |||
103 : | cextern intra_matrix_fixl | ||
104 : | cextern intra_matrix_fix | ||
105 : | cextern intra_matrix1 | ||
106 : | cextern intra_matrix | ||
107 : | |||
108 : | ;=========================================================================== | ||
109 : | ; | ||
110 : | ; inter matrix | ||
111 : | ; | ||
112 : | ;=========================================================================== | ||
113 : | |||
114 : | cextern inter_matrix1 | ||
115 : | cextern inter_matrix | ||
116 : | cextern inter_matrix_fix | ||
117 : | cextern inter_matrix_fixl | ||
118 : | |||
119 : | |||
120 : | edgomez | 1176 | %define nop4 db 08Dh, 074h, 026h,0 |
121 : | %define nop3 add esp, byte 0 | ||
122 : | %define nop2 mov esp, esp | ||
123 : | %define nop7 db 08dh, 02ch, 02dh,0,0,0,0 | ||
124 : | %define nop6 add ebp, dword 0 | ||
125 : | edgomez | 1174 | |
126 : | ;=========================================================================== | ||
127 : | ; | ||
128 : | ; quantd table | ||
129 : | ; | ||
130 : | ;=========================================================================== | ||
131 : | |||
132 : | edgomez | 1176 | %define VM18P 3 |
133 : | %define VM18Q 4 | ||
134 : | edgomez | 1174 | |
135 : | edgomez | 1176 | align 16 |
136 : | quantd: | ||
137 : | edgomez | 1174 | %assign i 1 |
138 : | %rep 31 | ||
139 : | times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) | ||
140 : | %assign i i+1 | ||
141 : | %endrep | ||
142 : | |||
143 : | ;=========================================================================== | ||
144 : | ; | ||
145 : | ; multiple by 2Q table | ||
146 : | ; | ||
147 : | ;=========================================================================== | ||
148 : | |||
149 : | edgomez | 1176 | align 16 |
150 : | mmx_mul_quant: | ||
151 : | edgomez | 1174 | %assign i 1 |
152 : | %rep 31 | ||
153 : | times 4 dw i | ||
154 : | %assign i i+1 | ||
155 : | %endrep | ||
156 : | |||
157 : | ;=========================================================================== | ||
158 : | ; | ||
159 : | ; saturation limits | ||
160 : | ; | ||
161 : | ;=========================================================================== | ||
162 : | |||
163 : | align 16 | ||
164 : | edgomez | 1176 | mmx_32767_minus_2047: |
165 : | times 4 dw (32767-2047) | ||
166 : | mmx_32768_minus_2048: | ||
167 : | times 4 dw (32768-2048) | ||
168 : | mmx_2047: | ||
169 : | times 4 dw 2047 | ||
170 : | mmx_minus_2048: | ||
171 : | times 4 dw (-2048) | ||
172 : | zero: | ||
173 : | times 4 dw 0 | ||
174 : | edgomez | 1174 | |
175 : | edgomez | 1176 | int_div: |
176 : | edgomez | 1174 | dd 0 |
177 : | %assign i 1 | ||
178 : | %rep 255 | ||
179 : | dd (1 << 17) / ( i) + 1 | ||
180 : | %assign i i+1 | ||
181 : | %endrep | ||
182 : | |||
183 : | edgomez | 1176 | ;*************************************************************************** |
184 : | ; Code | ||
185 : | ;*************************************************************************** | ||
186 : | |||
187 : | edgomez | 1174 | section .text |
188 : | |||
189 : | ;=========================================================================== | ||
190 : | ; | ||
191 : | edgomez | 1176 | ; uint32_t quant_mpeg_intra_xmm(int16_t * coeff, |
192 : | ; const int16_t const * data, | ||
193 : | ; const uint32_t quant, | ||
194 : | ; const uint32_t dcscalar); | ||
195 : | edgomez | 1174 | ; |
196 : | ;=========================================================================== | ||
197 : | |||
198 : | align ALIGN | ||
199 : | cglobal quant_mpeg_intra_xmm | ||
200 : | quant_mpeg_intra_xmm: | ||
201 : | edgomez | 1176 | mov eax, [esp + 8] ; data |
202 : | mov ecx, [esp + 12] ; quant | ||
203 : | mov edx, [esp + 4] ; coeff | ||
204 : | edgomez | 1174 | push esi |
205 : | push edi | ||
206 : | push ebx | ||
207 : | nop | ||
208 : | edgomez | 1176 | mov edi, mmzero |
209 : | mov esi, -14 | ||
210 : | pxor mm0, mm0 | ||
211 : | pxor mm3, mm3 | ||
212 : | cmp ecx, byte 1 | ||
213 : | je near .q1loop | ||
214 : | cmp ecx, byte 19 | ||
215 : | jg near .lloop | ||
216 : | edgomez | 1174 | nop6 |
217 : | |||
218 : | align ALIGN | ||
219 : | .loop | ||
220 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
221 : | psubw mm0, mm1 ;-mm1 | ||
222 : | movq mm4, [eax + 8*esi + 120] ; | ||
223 : | psubw mm3, mm4 ;-mm4 | ||
224 : | pmaxsw mm0, mm1 ;|src| | ||
225 : | pmaxsw mm3,mm4 | ||
226 : | nop2 | ||
227 : | psraw mm1, 15 ;sign src | ||
228 : | psraw mm4, 15 | ||
229 : | psllw mm0, 4 ;level << 4 ; | ||
230 : | psllw mm3, 4 | ||
231 : | paddw mm0, [intra_matrix1 + 8*esi+112] | ||
232 : | paddw mm3, [intra_matrix1 + 8*esi+120] | ||
233 : | movq mm5, [intra_matrix_fixl + 8*esi+112] | ||
234 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
235 : | pmulhuw mm5, mm0 | ||
236 : | pmulhuw mm7, mm3 | ||
237 : | edgomez | 1176 | mov esp, esp |
238 : | edgomez | 1174 | movq mm2, [intra_matrix + 8*esi+112] |
239 : | movq mm6, [intra_matrix + 8*esi+120] | ||
240 : | pmullw mm2, mm5 | ||
241 : | pmullw mm6, mm7 | ||
242 : | psubw mm0, mm2 | ||
243 : | psubw mm3, mm6 | ||
244 : | nop4 | ||
245 : | movq mm2, [quantd + ecx * 8 - 8] | ||
246 : | movq mm6, [mmx_divs + ecx * 8 - 8] | ||
247 : | paddw mm5, mm2 | ||
248 : | paddw mm7, mm2 | ||
249 : | edgomez | 1176 | mov esp, esp |
250 : | edgomez | 1174 | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] |
251 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
252 : | paddw mm5, mm0 | ||
253 : | paddw mm7, mm3 | ||
254 : | movq mm0, [edi] | ||
255 : | movq mm3, [edi] | ||
256 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
257 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
258 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
259 : | pxor mm7, mm4 ; | ||
260 : | psubw mm5, mm1 ; undisplace | ||
261 : | psubw mm7, mm4 ; | ||
262 : | movq [edx + 8*esi+112], mm5 | ||
263 : | movq [edx + 8*esi +120], mm7 | ||
264 : | edgomez | 1176 | add esi, byte 2 |
265 : | jng near .loop | ||
266 : | edgomez | 1174 | |
267 : | .done | ||
268 : | ; calculate data[0] // (int32_t)dcscalar) | ||
269 : | edgomez | 1176 | mov esi, [esp + 12 + 16] ; dcscalar |
270 : | edgomez | 1174 | movsx ecx, word [eax] |
271 : | edgomez | 1176 | mov edi, ecx |
272 : | mov edx, [esp + 12 + 16] | ||
273 : | shr edx, 1 ; ebx = dcscalar /2 | ||
274 : | sar edi, 31 ; cdq is vectorpath | ||
275 : | xor edx, edi ; ebx = eax V -eax -1 | ||
276 : | sub ecx, edi | ||
277 : | add ecx, edx | ||
278 : | mov edx, [dword esp + 12 + 4] | ||
279 : | mov esi, [int_div+4*esi] | ||
280 : | edgomez | 1174 | imul ecx, esi |
281 : | edgomez | 1176 | sar ecx, 17 |
282 : | lea ebx, [byte ecx + 1] | ||
283 : | edgomez | 1174 | cmovs ecx, ebx |
284 : | ; idiv cx ; ecx = edi:ecx / dcscalar | ||
285 : | |||
286 : | edgomez | 1176 | mov ebx, [esp] |
287 : | mov edi, [esp+4] | ||
288 : | mov esi, [esp+8] | ||
289 : | add esp, byte 12 | ||
290 : | mov [edx], cx ; coeff[0] = ax | ||
291 : | edgomez | 1174 | |
292 : | edgomez | 1176 | xor eax, eax |
293 : | edgomez | 1174 | ret |
294 : | |||
295 : | align ALIGN | ||
296 : | .q1loop | ||
297 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
298 : | psubw mm0, mm1 ;-mm1 | ||
299 : | movq mm4, [eax + 8*esi+120] ; | ||
300 : | psubw mm3, mm4 ;-mm4 | ||
301 : | pmaxsw mm0, mm1 ;|src| | ||
302 : | pmaxsw mm3, mm4 | ||
303 : | nop2 | ||
304 : | psraw mm1, 15 ;sign src | ||
305 : | psraw mm4, 15 | ||
306 : | psllw mm0, 4 ; level << 4 | ||
307 : | psllw mm3, 4 | ||
308 : | paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided | ||
309 : | paddw mm3, [intra_matrix1 + 8*esi+120] ;intra1 contains fix for division by 1 | ||
310 : | movq mm5, [intra_matrix_fixl + 8*esi+112] ;with rounding down | ||
311 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
312 : | pmulhuw mm5, mm0 | ||
313 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
314 : | edgomez | 1176 | mov esp, esp |
315 : | edgomez | 1174 | movq mm2, [intra_matrix + 8*esi+112] |
316 : | movq mm6, [intra_matrix + 8*esi+120] ; divs for q<=16 | ||
317 : | pmullw mm2, mm5 ;test value <= original | ||
318 : | pmullw mm6, mm7 | ||
319 : | psubw mm0, mm2 ;mismatch | ||
320 : | psubw mm3, mm6 | ||
321 : | nop4 | ||
322 : | movq mm2, [quantd + ecx * 8 - 8] | ||
323 : | paddw mm5, mm2 ;first approx with quantd | ||
324 : | paddw mm7, mm2 | ||
325 : | edgomez | 1176 | mov esp, esp |
326 : | edgomez | 1174 | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
327 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
328 : | paddw mm5, mm0 ;final result with quantd | ||
329 : | paddw mm7, mm3 | ||
330 : | movq mm0, [edi] | ||
331 : | movq mm3, [edi] | ||
332 : | edgomez | 1176 | mov esp, esp |
333 : | edgomez | 1174 | psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) |
334 : | psrlw mm7, 1 | ||
335 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
336 : | pxor mm7, mm4 ; | ||
337 : | psubw mm5, mm1 ; undisplace | ||
338 : | psubw mm7, mm4 ; | ||
339 : | movq [edx + 8*esi+112], mm5 | ||
340 : | movq [edx + 8*esi +120], mm7 | ||
341 : | edgomez | 1176 | add esi, byte 2 |
342 : | jng near .q1loop | ||
343 : | jmp near .done | ||
344 : | edgomez | 1174 | |
345 : | align 8 | ||
346 : | .lloop | ||
347 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
348 : | psubw mm0, mm1 ;-mm1 | ||
349 : | movq mm4, [eax + 8*esi+120] ; | ||
350 : | psubw mm3, mm4 ;-mm4 | ||
351 : | pmaxsw mm0, mm1 ;|src| | ||
352 : | pmaxsw mm3, mm4 | ||
353 : | nop2 | ||
354 : | psraw mm1, 15 ;sign src | ||
355 : | psraw mm4, 15 | ||
356 : | psllw mm0, 4 ; level << 4 | ||
357 : | psllw mm3, 4 ; | ||
358 : | paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1 | ||
359 : | paddw mm3, [intra_matrix1 + 8*esi+120] | ||
360 : | movq mm5, [intra_matrix_fixl + 8*esi+112] | ||
361 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
362 : | pmulhuw mm5, mm0 | ||
363 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
364 : | edgomez | 1176 | mov esp, esp |
365 : | edgomez | 1174 | movq mm2, [intra_matrix + 8*esi+112] |
366 : | movq mm6, [intra_matrix + 8*esi+120] | ||
367 : | pmullw mm2, mm5 ;test value <= original | ||
368 : | pmullw mm6, mm7 | ||
369 : | psubw mm0, mm2 ;mismatch | ||
370 : | psubw mm3, mm6 | ||
371 : | nop4 | ||
372 : | movq mm2, [quantd + ecx * 8 - 8] | ||
373 : | movq mm6, [mmx_div + ecx * 8 - 8] ; divs for q<=16 | ||
374 : | paddw mm5, mm2 ;first approx with quantd | ||
375 : | paddw mm7, mm2 | ||
376 : | edgomez | 1176 | mov esp, esp |
377 : | edgomez | 1174 | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
378 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
379 : | paddw mm5, mm0 ;final result with quantd | ||
380 : | paddw mm7, mm3 | ||
381 : | movq mm0, [edi] | ||
382 : | movq mm3, [edi] | ||
383 : | edgomez | 1176 | mov esp, esp |
384 : | edgomez | 1174 | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
385 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
386 : | psrlw mm5, 1 ; (level + quantd) / (2*quant) | ||
387 : | psrlw mm7, 1 | ||
388 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
389 : | pxor mm7, mm4 ; | ||
390 : | psubw mm5, mm1 ; undisplace | ||
391 : | psubw mm7, mm4 ; | ||
392 : | movq [edx + 8*esi+112], mm5 | ||
393 : | movq [edx + 8*esi +120], mm7 | ||
394 : | edgomez | 1176 | add esi,byte 2 |
395 : | jng near .lloop | ||
396 : | jmp near .done | ||
397 : | edgomez | 1174 | |
398 : | ;=========================================================================== | ||
399 : | ; | ||
400 : | edgomez | 1176 | ; uint32_t quant_mpeg_inter_xmm(int16_t * coeff, |
401 : | ; const int16_t const * data, | ||
402 : | ; const uint32_t quant); | ||
403 : | edgomez | 1174 | ; |
404 : | ;=========================================================================== | ||
405 : | |||
406 : | align ALIGN | ||
407 : | cglobal quant_mpeg_inter_xmm | ||
408 : | quant_mpeg_inter_xmm: | ||
409 : | edgomez | 1176 | mov eax, [esp + 8] ; data |
410 : | mov ecx, [esp + 12] ; quant | ||
411 : | mov edx, [esp + 4] ; coeff | ||
412 : | edgomez | 1174 | push esi |
413 : | push edi | ||
414 : | push ebx | ||
415 : | nop | ||
416 : | edgomez | 1176 | mov edi, mmzero |
417 : | mov esi, -14 | ||
418 : | mov ebx, esp | ||
419 : | sub esp, byte 24 | ||
420 : | lea ebx, [esp+8] | ||
421 : | and ebx, byte -8 ;align 8 | ||
422 : | pxor mm0, mm0 | ||
423 : | pxor mm3, mm3 | ||
424 : | movq [byte ebx],mm0 | ||
425 : | db 0Fh, 7Fh, 44h, 23h, 8 ;movq [ebx+8],mm0 | ||
426 : | cmp ecx, byte 1 | ||
427 : | je near .q1loop | ||
428 : | cmp ecx, byte 19 | ||
429 : | jg near .lloop | ||
430 : | edgomez | 1174 | nop |
431 : | |||
432 : | align ALIGN | ||
433 : | .loop | ||
434 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
435 : | edgomez | 1176 | psubw mm0, mm1 ;-mm1 |
436 : | edgomez | 1174 | movq mm4, [eax + 8*esi + 120] ; |
437 : | edgomez | 1176 | psubw mm3, mm4 ;-mm4 |
438 : | pmaxsw mm0, mm1 ;|src| | ||
439 : | pmaxsw mm3, mm4 | ||
440 : | edgomez | 1174 | nop2 |
441 : | edgomez | 1176 | psraw mm1, 15 ;sign src |
442 : | psraw mm4, 15 | ||
443 : | psllw mm0, 4 ; level << 4 | ||
444 : | psllw mm3, 4 ; | ||
445 : | paddw mm0, [inter_matrix1 + 8*esi+112] | ||
446 : | paddw mm3, [inter_matrix1 + 8*esi+120] | ||
447 : | movq mm5, [inter_matrix_fixl + 8*esi+112] | ||
448 : | movq mm7, [inter_matrix_fixl + 8*esi+120] | ||
449 : | pmulhuw mm5, mm0 | ||
450 : | pmulhuw mm7, mm3 | ||
451 : | mov esp, esp | ||
452 : | movq mm2, [inter_matrix + 8*esi+112] | ||
453 : | movq mm6, [inter_matrix + 8*esi+120] | ||
454 : | pmullw mm2, mm5 | ||
455 : | pmullw mm6, mm7 | ||
456 : | psubw mm0, mm2 | ||
457 : | psubw mm3, mm6 | ||
458 : | movq mm2, [byte ebx] | ||
459 : | movq mm6, [mmx_divs + ecx * 8 - 8] | ||
460 : | pmulhuw mm0, [inter_matrix_fix + 8*esi+112] | ||
461 : | pmulhuw mm3, [inter_matrix_fix + 8*esi+120] | ||
462 : | paddw mm2, [ebx+8] ;sum | ||
463 : | paddw mm5, mm0 | ||
464 : | paddw mm7, mm3 | ||
465 : | movq mm0, [edi] | ||
466 : | movq mm3, [edi] | ||
467 : | edgomez | 1174 | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
468 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
469 : | edgomez | 1176 | add esi, byte 2 |
470 : | paddw mm2, mm5 ;sum += x1 | ||
471 : | movq [ebx], mm7 ;store x2 | ||
472 : | edgomez | 1174 | pxor mm5, mm1 ; mm0 *= sign(mm0) |
473 : | pxor mm7, mm4 ; | ||
474 : | psubw mm5, mm1 ; undisplace | ||
475 : | psubw mm7, mm4 ; | ||
476 : | edgomez | 1176 | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum |
477 : | edgomez | 1174 | movq [edx + 8*esi+112-16], mm5 |
478 : | movq [edx + 8*esi +120-16], mm7 | ||
479 : | jng near .loop | ||
480 : | |||
481 : | .done | ||
482 : | ; calculate data[0] // (int32_t)dcscalar) | ||
483 : | edgomez | 1176 | paddw mm2, [ebx] |
484 : | mov ebx, [esp+24] | ||
485 : | mov edi, [esp+4+24] | ||
486 : | mov esi, [esp+8+24] | ||
487 : | add esp, byte 12+24 | ||
488 : | pmaddwd mm2, [mmx_one] | ||
489 : | punpckldq mm0, mm2 ;get low dw to mm0:high | ||
490 : | paddd mm0,mm2 | ||
491 : | punpckhdq mm0, mm0 ;get result to low | ||
492 : | edgomez | 1174 | movd eax, mm0 |
493 : | |||
494 : | ret | ||
495 : | |||
496 : | align ALIGN | ||
497 : | .q1loop | ||
498 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
499 : | edgomez | 1176 | psubw mm0, mm1 ;-mm1 |
500 : | edgomez | 1174 | movq mm4, [eax + 8*esi+120] ; |
501 : | edgomez | 1176 | psubw mm3, mm4 ;-mm4 |
502 : | pmaxsw mm0, mm1 ;|src| | ||
503 : | pmaxsw mm3, mm4 | ||
504 : | edgomez | 1174 | nop2 |
505 : | edgomez | 1176 | psraw mm1, 15 ;sign src |
506 : | psraw mm4, 15 | ||
507 : | psllw mm0, 4 ; level << 4 | ||
508 : | psllw mm3, 4 | ||
509 : | paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided | ||
510 : | paddw mm3, [inter_matrix1 + 8*esi+120] ; inter1 contains fix for division by 1 | ||
511 : | movq mm5, [inter_matrix_fixl + 8*esi+112] ;with rounding down | ||
512 : | movq mm7, [inter_matrix_fixl + 8*esi+120] | ||
513 : | pmulhuw mm5, mm0 | ||
514 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
515 : | mov esp, esp | ||
516 : | movq mm2, [inter_matrix + 8*esi+112] | ||
517 : | movq mm6, [inter_matrix + 8*esi+120] ; divs for q<=16 | ||
518 : | pmullw mm2, mm5 ;test value <= original | ||
519 : | pmullw mm6, mm7 | ||
520 : | psubw mm0, mm2 ;mismatch | ||
521 : | psubw mm3, mm6 | ||
522 : | movq mm2, [byte ebx] | ||
523 : | pmulhuw mm0, [inter_matrix_fix + 8*esi+112] ;correction | ||
524 : | pmulhuw mm3, [inter_matrix_fix + 8*esi+120] | ||
525 : | paddw mm2, [ebx+8] ;sum | ||
526 : | paddw mm5, mm0 ;final result | ||
527 : | paddw mm7, mm3 | ||
528 : | movq mm0, [edi] | ||
529 : | movq mm3, [edi] | ||
530 : | psrlw mm5, 1 ; (level ) /2 (quant = 1) | ||
531 : | psrlw mm7, 1 | ||
532 : | add esi, byte 2 | ||
533 : | paddw mm2, mm5 ;sum += x1 | ||
534 : | movq [ebx], mm7 ;store x2 | ||
535 : | edgomez | 1174 | pxor mm5, mm1 ; mm0 *= sign(mm0) |
536 : | pxor mm7, mm4 ; | ||
537 : | psubw mm5, mm1 ; undisplace | ||
538 : | psubw mm7, mm4 ; | ||
539 : | edgomez | 1176 | movq [ebx+8], mm2 ;store sum |
540 : | edgomez | 1174 | movq [edx + 8*esi+112-16], mm5 |
541 : | movq [edx + 8*esi +120-16], mm7 | ||
542 : | edgomez | 1176 | jng near .q1loop |
543 : | jmp near .done | ||
544 : | edgomez | 1174 | |
545 : | align 8 | ||
546 : | .lloop | ||
547 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
548 : | psubw mm0,mm1 ;-mm1 | ||
549 : | movq mm4, [eax + 8*esi+120] ; | ||
550 : | psubw mm3,mm4 ;-mm4 | ||
551 : | edgomez | 1176 | pmaxsw mm0,mm1 ;|src| |
552 : | pmaxsw mm3,mm4 | ||
553 : | edgomez | 1174 | nop2 |
554 : | edgomez | 1176 | psraw mm1,15 ;sign src |
555 : | psraw mm4,15 | ||
556 : | psllw mm0, 4 ; level << 4 | ||
557 : | psllw mm3, 4 ; | ||
558 : | paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 | ||
559 : | paddw mm3, [inter_matrix1 + 8*esi+120] | ||
560 : | movq mm5,[inter_matrix_fixl + 8*esi+112] | ||
561 : | movq mm7,[inter_matrix_fixl + 8*esi+120] | ||
562 : | pmulhuw mm5,mm0 | ||
563 : | pmulhuw mm7,mm3 ;mm7: first approx of division | ||
564 : | mov esp,esp | ||
565 : | movq mm2,[inter_matrix + 8*esi+112] | ||
566 : | movq mm6,[inter_matrix + 8*esi+120] | ||
567 : | pmullw mm2,mm5 ;test value <= original | ||
568 : | pmullw mm6,mm7 | ||
569 : | psubw mm0,mm2 ;mismatch | ||
570 : | psubw mm3,mm6 | ||
571 : | movq mm2,[byte ebx] | ||
572 : | movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 | ||
573 : | pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction | ||
574 : | pmulhuw mm3,[inter_matrix_fix + 8*esi+120] | ||
575 : | paddw mm2,[ebx+8] ;sum | ||
576 : | paddw mm5,mm0 ;final result | ||
577 : | paddw mm7,mm3 | ||
578 : | movq mm0,[edi] | ||
579 : | movq mm3,[edi] | ||
580 : | edgomez | 1174 | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
581 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
582 : | edgomez | 1176 | add esi,byte 2 |
583 : | psrlw mm5, 1 ; (level ) / (2*quant) | ||
584 : | paddw mm2,mm5 ;sum += x1 | ||
585 : | psrlw mm7, 1 | ||
586 : | movq [ebx],mm7 ;store x2 | ||
587 : | edgomez | 1174 | pxor mm5, mm1 ; mm0 *= sign(mm0) |
588 : | pxor mm7, mm4 ; | ||
589 : | psubw mm5, mm1 ; undisplace | ||
590 : | psubw mm7, mm4 ; | ||
591 : | edgomez | 1176 | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum |
592 : | edgomez | 1174 | movq [edx + 8*esi+112-16], mm5 |
593 : | movq [edx + 8*esi +120-16], mm7 | ||
594 : | jng near .lloop | ||
595 : | edgomez | 1176 | jmp near .done |
596 : | edgomez | 1174 | |
597 : | |||
598 : | ;=========================================================================== | ||
599 : | ; | ||
600 : | edgomez | 1176 | ; uint32_t dequant_mpeg_intra_3dne(int16_t *data, |
601 : | ; const int16_t const *coeff, | ||
602 : | ; const uint32_t quant, | ||
603 : | ; const uint32_t dcscalar); | ||
604 : | edgomez | 1174 | ; |
605 : | ;=========================================================================== | ||
606 : | |||
607 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
608 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
609 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
610 : | ; We thus avoid packing/unpacking toward double-word. | ||
611 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
612 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
613 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
614 : | ; and quant in [1..31]. | ||
615 : | ; | ||
616 : | ;******************************************************************** | ||
617 : | edgomez | 1176 | |
618 : | edgomez | 1174 | %macro DEQUANT4INTRAMMX 1 |
619 : | edgomez | 1176 | movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] |
620 : | movq mm4, [ecx+ 16 * %1 +8]; mm3 = c' = coeff[i+1] | ||
621 : | psubw mm0, mm1 | ||
622 : | psubw mm3, mm4 | ||
623 : | pmaxsw mm0, mm1 | ||
624 : | pmaxsw mm3, mm4 | ||
625 : | psraw mm1, 15 | ||
626 : | psraw mm4, 15 | ||
627 : | edgomez | 1174 | %if %1 |
628 : | edgomez | 1176 | movq mm2, [eax+8] ;preshifted quant |
629 : | movq mm7, [eax+8] | ||
630 : | edgomez | 1174 | %endif |
631 : | edgomez | 1176 | pmullw mm2, [intra_matrix + 16 * %1 ] ; matrix[i]*quant |
632 : | pmullw mm7, [intra_matrix + 16 * %1 +8] ; matrix[i+1]*quant | ||
633 : | movq mm5, mm0 | ||
634 : | movq mm6, mm3 | ||
635 : | pmulhw mm0, mm2 ; high of coeff*(matrix*quant) | ||
636 : | pmulhw mm3, mm7 ; high of coeff*(matrix*quant) | ||
637 : | pmullw mm2, mm5 ; low of coeff*(matrix*quant) | ||
638 : | pmullw mm7, mm6 ; low of coeff*(matrix*quant) | ||
639 : | pcmpgtw mm0, [eax] | ||
640 : | pcmpgtw mm3, [eax] | ||
641 : | paddusw mm2, mm0 | ||
642 : | paddusw mm7, mm3 | ||
643 : | psrlw mm2, 5 | ||
644 : | psrlw mm7, 5 | ||
645 : | pxor mm2, mm1 ; start negating back | ||
646 : | pxor mm7, mm4 ; start negating back | ||
647 : | psubusw mm1, mm0 | ||
648 : | psubusw mm4, mm3 | ||
649 : | movq mm0, [eax] ;zero | ||
650 : | movq mm3, [eax] ;zero | ||
651 : | psubw mm2, mm1 ; finish negating back | ||
652 : | psubw mm7, mm4 ; finish negating back | ||
653 : | movq [byte edx + 16 * %1], mm2 ; data[i] | ||
654 : | movq [edx + 16 * %1 +8], mm7 ; data[i+1] | ||
655 : | edgomez | 1174 | %endmacro |
656 : | |||
657 : | align 16 | ||
658 : | cglobal dequant_mpeg_intra_3dne | ||
659 : | dequant_mpeg_intra_3dne: | ||
660 : | edgomez | 1176 | mov eax, [esp+12] ; quant |
661 : | mov ecx, [esp+8] ; coeff | ||
662 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
663 : | psllw mm7, 2 ; << 2. See comment. | ||
664 : | mov edx, [esp+4] ; data | ||
665 : | push ebx | ||
666 : | movsx ebx, word [ecx] | ||
667 : | pxor mm0, mm0 | ||
668 : | pxor mm3, mm3 | ||
669 : | push esi | ||
670 : | lea eax, [esp-28] | ||
671 : | sub esp, byte 32 | ||
672 : | and eax, byte -8 ;points to qword aligned space on stack | ||
673 : | movq [eax], mm0 | ||
674 : | movq [eax+8], mm7 | ||
675 : | imul ebx, [esp+16+8+32] ; dcscalar | ||
676 : | movq mm2, mm7 | ||
677 : | edgomez | 1174 | |
678 : | align 4 | ||
679 : | |||
680 : | DEQUANT4INTRAMMX 0 | ||
681 : | |||
682 : | edgomez | 1176 | mov esi, -2048 |
683 : | edgomez | 1174 | nop |
684 : | edgomez | 1176 | cmp ebx, esi |
685 : | edgomez | 1174 | |
686 : | DEQUANT4INTRAMMX 1 | ||
687 : | |||
688 : | edgomez | 1176 | cmovl ebx, esi |
689 : | neg esi | ||
690 : | sub esi, byte 1 ;2047 | ||
691 : | edgomez | 1174 | |
692 : | DEQUANT4INTRAMMX 2 | ||
693 : | |||
694 : | edgomez | 1176 | cmp ebx, esi |
695 : | cmovg ebx, esi | ||
696 : | lea ebp, [byte ebp] | ||
697 : | edgomez | 1174 | |
698 : | DEQUANT4INTRAMMX 3 | ||
699 : | |||
700 : | edgomez | 1176 | mov esi, [esp+32] |
701 : | mov [byte edx], bx | ||
702 : | mov ebx, [esp+32+4] | ||
703 : | edgomez | 1174 | |
704 : | DEQUANT4INTRAMMX 4 | ||
705 : | DEQUANT4INTRAMMX 5 | ||
706 : | DEQUANT4INTRAMMX 6 | ||
707 : | DEQUANT4INTRAMMX 7 | ||
708 : | |||
709 : | edgomez | 1176 | add esp, byte 32+8 |
710 : | edgomez | 1174 | |
711 : | edgomez | 1176 | xor eax, eax |
712 : | ret | ||
713 : | edgomez | 1174 | |
714 : | ;=========================================================================== | ||
715 : | ; | ||
716 : | edgomez | 1176 | ; uint32_t dequant_mpeg_inter_3dne(int16_t * data, |
717 : | ; const int16_t * const coeff, | ||
718 : | ; const uint32_t quant); | ||
719 : | edgomez | 1174 | ; |
720 : | ;=========================================================================== | ||
721 : | |||
722 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
723 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
724 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
725 : | ; It's mixed with the extraction of the absolute value. | ||
726 : | |||
727 : | align 16 | ||
728 : | cglobal dequant_mpeg_inter_3dne | ||
729 : | dequant_mpeg_inter_3dne: | ||
730 : | edgomez | 1176 | mov edx, [esp+ 4] ; data |
731 : | mov ecx, [esp+ 8] ; coeff | ||
732 : | mov eax, [esp+12] ; quant | ||
733 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
734 : | mov eax, -14 | ||
735 : | paddw mm7, mm7 ; << 1 | ||
736 : | pxor mm6, mm6 ; mismatch sum | ||
737 : | push esi | ||
738 : | mov esi, mmzero | ||
739 : | pxor mm1, mm1 | ||
740 : | pxor mm3, mm3 | ||
741 : | edgomez | 1174 | nop |
742 : | nop4 | ||
743 : | |||
744 : | align 16 | ||
745 : | .loop | ||
746 : | edgomez | 1176 | movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] |
747 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
748 : | movq mm2, [ecx+8*eax + 7*16 +8] ; mm2 = coeff[i+1] | ||
749 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
750 : | paddsw mm0, mm1 ; c += sgn(c) | ||
751 : | paddsw mm2, mm3 ; c += sgn(c') | ||
752 : | paddw mm0, mm0 ; c *= 2 | ||
753 : | paddw mm2, mm2 ; c'*= 2 | ||
754 : | edgomez | 1174 | |
755 : | edgomez | 1176 | movq mm4, [esi] |
756 : | movq mm5, [esi] | ||
757 : | psubw mm4, mm0 ; -c | ||
758 : | psubw mm5, mm2 ; -c' | ||
759 : | edgomez | 1174 | |
760 : | edgomez | 1176 | psraw mm4, 16 ; mm4 = sgn(-c) |
761 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
762 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
763 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
764 : | pxor mm0, mm1 ; finish changing sign if needed | ||
765 : | pxor mm2, mm3 ; finish changing sign if needed | ||
766 : | edgomez | 1174 | |
767 : | ; we're short on register, here. Poor pairing... | ||
768 : | |||
769 : | edgomez | 1176 | movq mm4, mm7 ; (matrix*quant) |
770 : | edgomez | 1174 | nop |
771 : | edgomez | 1176 | pmullw mm4, [inter_matrix + 8*eax + 7*16] |
772 : | movq mm5, mm4 | ||
773 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
774 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
775 : | edgomez | 1174 | |
776 : | edgomez | 1176 | movq mm4, mm7 ; (matrix*quant) |
777 : | pmullw mm4, [inter_matrix + 8*eax + 7*16 + 8] | ||
778 : | add eax, byte 2 | ||
779 : | edgomez | 1174 | |
780 : | edgomez | 1176 | pcmpgtw mm5, [esi] |
781 : | paddusw mm0, mm5 | ||
782 : | psrlw mm0, 5 | ||
783 : | pxor mm0, mm1 ; start restoring sign | ||
784 : | psubusw mm1, mm5 | ||
785 : | edgomez | 1174 | |
786 : | edgomez | 1176 | movq mm5, mm4 |
787 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
788 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
789 : | psubw mm0, mm1 ; finish restoring sign | ||
790 : | edgomez | 1174 | |
791 : | edgomez | 1176 | pcmpgtw mm5, [esi] |
792 : | paddusw mm2, mm5 | ||
793 : | psrlw mm2, 5 | ||
794 : | pxor mm2, mm3 ; start restoring sign | ||
795 : | psubusw mm3, mm5 | ||
796 : | psubw mm2, mm3 ; finish restoring sign | ||
797 : | movq mm1, [esi] | ||
798 : | movq mm3, [byte esi] | ||
799 : | pxor mm6, mm0 ; mismatch control | ||
800 : | movq [edx + 8*eax + 7*16 -2*8 ], mm0 ; data[i] | ||
801 : | pxor mm6, mm2 ; mismatch control | ||
802 : | movq [edx + 8*eax + 7*16 -2*8 +8], mm2 ; data[i+1] | ||
803 : | edgomez | 1174 | |
804 : | edgomez | 1176 | jng .loop |
805 : | edgomez | 1174 | nop |
806 : | |||
807 : | ; mismatch control | ||
808 : | |||
809 : | edgomez | 1176 | pshufw mm0, mm6, 01010101b |
810 : | pshufw mm1, mm6, 10101010b | ||
811 : | pshufw mm2, mm6, 11111111b | ||
812 : | pxor mm6, mm0 | ||
813 : | pxor mm1, mm2 | ||
814 : | pxor mm6, mm1 | ||
815 : | movd eax, mm6 | ||
816 : | and eax, byte 1 | ||
817 : | xor eax, byte 1 | ||
818 : | mov esi, [esp] | ||
819 : | add esp, byte 4 | ||
820 : | xor word [edx + 2*63], ax | ||
821 : | edgomez | 1174 | |
822 : | edgomez | 1176 | xor eax, eax |
823 : | edgomez | 1174 | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |