1 |
|
;/****************************************************************************** |
2 |
|
; * * |
3 |
|
; * This file is part of XviD, a free MPEG-4 video encoder/decoder * |
4 |
|
; * * |
5 |
|
; * XviD is an implementation of a part of one or more MPEG-4 Video tools * |
6 |
|
; * as specified in ISO/IEC 14496-2 standard. Those intending to use this * |
7 |
|
; * software module in hardware or software products are advised that its * |
8 |
|
; * use may infringe existing patents or copyrights, and any such use * |
9 |
|
; * would be at such party's own risk. The original developer of this * |
10 |
|
; * software module and his/her company, and subsequent editors and their * |
11 |
|
; * companies, will have no liability for use of this software or * |
12 |
|
; * modifications or derivatives thereof. * |
13 |
|
; * * |
14 |
|
; * XviD is free software; you can redistribute it and/or modify it * |
15 |
|
; * under the terms of the GNU General Public License as published by * |
16 |
|
; * the Free Software Foundation; either version 2 of the License, or * |
17 |
|
; * (at your option) any later version. * |
18 |
|
; * * |
19 |
|
; * XviD is distributed in the hope that it will be useful, but * |
20 |
|
; * WITHOUT ANY WARRANTY; without even the implied warranty of * |
21 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the * |
22 |
|
; * GNU General Public License for more details. * |
23 |
|
; * * |
24 |
|
; * You should have received a copy of the GNU General Public License * |
25 |
|
; * along with this program; if not, write to the Free Software * |
26 |
|
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * |
27 |
|
; * * |
28 |
|
; ******************************************************************************/ |
29 |
|
; |
30 |
|
;/****************************************************************************** |
31 |
|
; * * |
32 |
|
; * quantize4.asm, MMX optimized MPEG quantization/dequantization * |
33 |
|
; * * |
34 |
|
; * Copyright (C) 2002 - Peter Ross <pross@cs.rmit.edu.au> * |
35 |
|
; * Copyright (C) 2002 - Michael Militzer <isibaar@xvid.org> * |
36 |
|
; * * |
37 |
|
; * For more information visit the XviD homepage: http://www.xvid.org * |
38 |
|
; * * |
39 |
|
; ******************************************************************************/ |
40 |
|
; |
41 |
|
;/****************************************************************************** |
42 |
|
; * * |
43 |
|
; * Revision history: * |
44 |
|
; * * |
45 |
|
; * 14.06.2002 mmx dequant4_* funcs revamped -Skal- * |
46 |
|
; * 22.01.2002 initial version * |
47 |
|
; * * |
48 |
|
; ******************************************************************************/ |
49 |
|
|
50 |
|
; data/text alignment |
51 |
|
%define ALIGN 8 |
52 |
|
|
53 |
|
%define SATURATE |
54 |
|
|
55 |
|
bits 32 |
56 |
|
|
57 |
|
%ifdef FORMAT_COFF |
58 |
|
SECTION .data data |
59 |
|
%else |
60 |
|
SECTION .data data align=8 |
61 |
|
%endif |
62 |
|
|
63 |
|
|
64 |
|
%macro cglobal 1 |
65 |
|
%ifdef PREFIX |
66 |
|
global _%1 |
67 |
|
%define %1 _%1 |
68 |
|
%else |
69 |
|
global %1 |
70 |
|
%endif |
71 |
|
%endmacro |
72 |
|
|
73 |
|
%macro cextern 1 |
74 |
|
%ifdef PREFIX |
75 |
|
extern _%1 |
76 |
|
%define %1 _%1 |
77 |
|
%else |
78 |
|
extern %1 |
79 |
|
%endif |
80 |
|
%endmacro |
81 |
|
|
82 |
|
mmx_one times 4 dw 1 |
83 |
|
|
84 |
|
;=========================================================================== |
85 |
|
; |
86 |
|
; divide by 2Q table |
87 |
|
; |
88 |
|
;=========================================================================== |
89 |
|
|
90 |
|
%macro MMX_DIV 1 |
91 |
|
times 4 dw (1 << 17) / (%1 * 2) + 1 |
92 |
|
%endmacro |
93 |
|
|
94 |
|
align ALIGN |
95 |
|
mmx_div |
96 |
|
MMX_DIV 1 |
97 |
|
MMX_DIV 2 |
98 |
|
MMX_DIV 3 |
99 |
|
MMX_DIV 4 |
100 |
|
MMX_DIV 5 |
101 |
|
MMX_DIV 6 |
102 |
|
MMX_DIV 7 |
103 |
|
MMX_DIV 8 |
104 |
|
MMX_DIV 9 |
105 |
|
MMX_DIV 10 |
106 |
|
MMX_DIV 11 |
107 |
|
MMX_DIV 12 |
108 |
|
MMX_DIV 13 |
109 |
|
MMX_DIV 14 |
110 |
|
MMX_DIV 15 |
111 |
|
MMX_DIV 16 |
112 |
|
MMX_DIV 17 |
113 |
|
MMX_DIV 18 |
114 |
|
MMX_DIV 19 |
115 |
|
MMX_DIV 20 |
116 |
|
MMX_DIV 21 |
117 |
|
MMX_DIV 22 |
118 |
|
MMX_DIV 23 |
119 |
|
MMX_DIV 24 |
120 |
|
MMX_DIV 25 |
121 |
|
MMX_DIV 26 |
122 |
|
MMX_DIV 27 |
123 |
|
MMX_DIV 28 |
124 |
|
MMX_DIV 29 |
125 |
|
MMX_DIV 30 |
126 |
|
MMX_DIV 31 |
127 |
|
|
128 |
|
|
129 |
|
;=========================================================================== |
130 |
|
; |
131 |
|
; intra matrix |
132 |
|
; |
133 |
|
;=========================================================================== |
134 |
|
|
135 |
|
cextern intra_matrix |
136 |
|
cextern intra_matrix_fix |
137 |
|
|
138 |
|
;=========================================================================== |
139 |
|
; |
140 |
|
; inter matrix |
141 |
|
; |
142 |
|
;=========================================================================== |
143 |
|
|
144 |
|
cextern inter_matrix |
145 |
|
cextern inter_matrix_fix |
146 |
|
|
147 |
|
|
148 |
|
%define VM18P 3 |
149 |
|
%define VM18Q 4 |
150 |
|
|
151 |
|
|
152 |
|
;=========================================================================== |
153 |
|
; |
154 |
|
; quantd table |
155 |
|
; |
156 |
|
;=========================================================================== |
157 |
|
|
158 |
|
%macro MMX_QUANTD 1 |
159 |
|
times 4 dw ((VM18P*%1) + (VM18Q/2)) / VM18Q |
160 |
|
%endmacro |
161 |
|
|
162 |
|
quantd |
163 |
|
MMX_QUANTD 1 |
164 |
|
MMX_QUANTD 2 |
165 |
|
MMX_QUANTD 3 |
166 |
|
MMX_QUANTD 4 |
167 |
|
MMX_QUANTD 5 |
168 |
|
MMX_QUANTD 6 |
169 |
|
MMX_QUANTD 7 |
170 |
|
MMX_QUANTD 8 |
171 |
|
MMX_QUANTD 9 |
172 |
|
MMX_QUANTD 10 |
173 |
|
MMX_QUANTD 11 |
174 |
|
MMX_QUANTD 12 |
175 |
|
MMX_QUANTD 13 |
176 |
|
MMX_QUANTD 14 |
177 |
|
MMX_QUANTD 15 |
178 |
|
MMX_QUANTD 16 |
179 |
|
MMX_QUANTD 17 |
180 |
|
MMX_QUANTD 18 |
181 |
|
MMX_QUANTD 19 |
182 |
|
MMX_QUANTD 20 |
183 |
|
MMX_QUANTD 21 |
184 |
|
MMX_QUANTD 22 |
185 |
|
MMX_QUANTD 23 |
186 |
|
MMX_QUANTD 24 |
187 |
|
MMX_QUANTD 25 |
188 |
|
MMX_QUANTD 26 |
189 |
|
MMX_QUANTD 27 |
190 |
|
MMX_QUANTD 28 |
191 |
|
MMX_QUANTD 29 |
192 |
|
MMX_QUANTD 30 |
193 |
|
MMX_QUANTD 31 |
194 |
|
|
195 |
|
|
196 |
|
;=========================================================================== |
197 |
|
; |
198 |
|
; multiple by 2Q table |
199 |
|
; |
200 |
|
;=========================================================================== |
201 |
|
|
202 |
|
%macro MMX_MUL_QUANT 1 |
203 |
|
times 4 dw %1 |
204 |
|
%endmacro |
205 |
|
|
206 |
|
mmx_mul_quant |
207 |
|
MMX_MUL_QUANT 1 |
208 |
|
MMX_MUL_QUANT 2 |
209 |
|
MMX_MUL_QUANT 3 |
210 |
|
MMX_MUL_QUANT 4 |
211 |
|
MMX_MUL_QUANT 5 |
212 |
|
MMX_MUL_QUANT 6 |
213 |
|
MMX_MUL_QUANT 7 |
214 |
|
MMX_MUL_QUANT 8 |
215 |
|
MMX_MUL_QUANT 9 |
216 |
|
MMX_MUL_QUANT 10 |
217 |
|
MMX_MUL_QUANT 11 |
218 |
|
MMX_MUL_QUANT 12 |
219 |
|
MMX_MUL_QUANT 13 |
220 |
|
MMX_MUL_QUANT 14 |
221 |
|
MMX_MUL_QUANT 15 |
222 |
|
MMX_MUL_QUANT 16 |
223 |
|
MMX_MUL_QUANT 17 |
224 |
|
MMX_MUL_QUANT 18 |
225 |
|
MMX_MUL_QUANT 19 |
226 |
|
MMX_MUL_QUANT 20 |
227 |
|
MMX_MUL_QUANT 21 |
228 |
|
MMX_MUL_QUANT 22 |
229 |
|
MMX_MUL_QUANT 23 |
230 |
|
MMX_MUL_QUANT 24 |
231 |
|
MMX_MUL_QUANT 25 |
232 |
|
MMX_MUL_QUANT 26 |
233 |
|
MMX_MUL_QUANT 27 |
234 |
|
MMX_MUL_QUANT 28 |
235 |
|
MMX_MUL_QUANT 29 |
236 |
|
MMX_MUL_QUANT 30 |
237 |
|
MMX_MUL_QUANT 31 |
238 |
|
|
239 |
|
;=========================================================================== |
240 |
|
; |
241 |
|
; saturation limits |
242 |
|
; |
243 |
|
;=========================================================================== |
244 |
|
|
245 |
|
align 16 |
246 |
|
|
247 |
|
mmx_32767_minus_2047 times 4 dw (32767-2047) |
248 |
|
mmx_32768_minus_2048 times 4 dw (32768-2048) |
249 |
|
mmx_2047 times 4 dw 2047 |
250 |
|
mmx_minus_2048 times 4 dw (-2048) |
251 |
|
zero times 4 dw 0 |
252 |
|
|
253 |
|
section .text |
254 |
|
|
255 |
|
;=========================================================================== |
256 |
|
; |
257 |
|
; void quant_intra4_mmx(int16_t * coeff, |
258 |
|
; const int16_t const * data, |
259 |
|
; const uint32_t quant, |
260 |
|
; const uint32_t dcscalar); |
261 |
|
; |
262 |
|
;=========================================================================== |
263 |
|
|
264 |
|
align ALIGN |
265 |
|
cglobal quant_mpeg_intra_mmx |
266 |
|
quant_mpeg_intra_mmx: |
267 |
|
|
268 |
|
push ecx |
269 |
|
push esi |
270 |
|
push edi |
271 |
|
|
272 |
|
mov edi, [esp + 12 + 4] ; coeff |
273 |
|
mov esi, [esp + 12 + 8] ; data |
274 |
|
mov eax, [esp + 12 + 12] ; quant |
275 |
|
|
276 |
|
movq mm5, [quantd + eax * 8 - 8] ; quantd -> mm5 |
277 |
|
|
278 |
|
xor ecx, ecx |
279 |
|
cmp al, 1 |
280 |
|
jz near .q1loop |
281 |
|
|
282 |
|
cmp al, 2 |
283 |
|
jz near .q2loop |
284 |
|
|
285 |
|
movq mm7, [mmx_div + eax * 8 - 8] ; multipliers[quant] -> mm7 |
286 |
|
|
287 |
|
align ALIGN |
288 |
|
.loop |
289 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
290 |
|
movq mm3, [esi + 8*ecx + 8] ; |
291 |
|
|
292 |
|
pxor mm1, mm1 ; mm1 = 0 |
293 |
|
pxor mm4, mm4 |
294 |
|
|
295 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
296 |
|
pcmpgtw mm4, mm3 |
297 |
|
|
298 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
299 |
|
pxor mm3, mm4 ; |
300 |
|
psubw mm0, mm1 ; displace |
301 |
|
psubw mm3, mm4 ; |
302 |
|
|
303 |
|
psllw mm0, 4 ; level << 4 |
304 |
|
psllw mm3, 4 ; |
305 |
|
|
306 |
|
movq mm2, [intra_matrix + 8*ecx] |
307 |
|
psrlw mm2, 1 ; intra_matrix[i]>>1 |
308 |
|
paddw mm0, mm2 |
309 |
|
|
310 |
|
movq mm2, [intra_matrix_fix + ecx*8] |
311 |
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
312 |
|
|
313 |
|
movq mm2, [intra_matrix + 8*ecx + 8] |
314 |
|
psrlw mm2, 1 |
315 |
|
paddw mm3, mm2 |
316 |
|
|
317 |
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
318 |
|
pmulhw mm3, mm2 |
319 |
|
|
320 |
|
paddw mm0, mm5 ; + quantd |
321 |
|
paddw mm3, mm5 |
322 |
|
|
323 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
324 |
|
pmulhw mm3, mm7 ; |
325 |
|
psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 |
326 |
|
psrlw mm3, 1 |
327 |
|
|
328 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
329 |
|
pxor mm3, mm4 ; |
330 |
|
psubw mm0, mm1 ; undisplace |
331 |
|
psubw mm3, mm4 ; |
332 |
|
|
333 |
|
movq [edi + 8*ecx], mm0 |
334 |
|
movq [edi + 8*ecx + 8], mm3 |
335 |
|
|
336 |
|
add ecx,2 |
337 |
|
cmp ecx,16 |
338 |
|
jnz near .loop |
339 |
|
|
340 |
|
.done |
341 |
|
; caclulate data[0] // (int32_t)dcscalar) |
342 |
|
|
343 |
|
mov ecx, [esp + 12 + 16] ; dcscalar |
344 |
|
mov edx, ecx |
345 |
|
movsx eax, word [esi] ; data[0] |
346 |
|
shr edx, 1 ; edx = dcscalar /2 |
347 |
|
cmp eax, 0 |
348 |
|
jg .gtzero |
349 |
|
|
350 |
|
sub eax, edx |
351 |
|
jmp short .mul |
352 |
|
.gtzero |
353 |
|
add eax, edx |
354 |
|
.mul |
355 |
|
cdq ; expand eax -> edx:eax |
356 |
|
idiv ecx ; eax = edx:eax / dcscalar |
357 |
|
|
358 |
|
mov [edi], ax ; coeff[0] = ax |
359 |
|
|
360 |
|
pop edi |
361 |
|
pop esi |
362 |
|
pop ecx |
363 |
|
|
364 |
|
ret |
365 |
|
|
366 |
|
align ALIGN |
367 |
|
.q1loop |
368 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
369 |
|
movq mm3, [esi + 8*ecx + 8] ; |
370 |
|
|
371 |
|
pxor mm1, mm1 ; mm1 = 0 |
372 |
|
pxor mm4, mm4 ; |
373 |
|
|
374 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
375 |
|
pcmpgtw mm4, mm3 ; |
376 |
|
|
377 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
378 |
|
pxor mm3, mm4 ; |
379 |
|
psubw mm0, mm1 ; displace |
380 |
|
psubw mm3, mm4 ; |
381 |
|
|
382 |
|
psllw mm0, 4 |
383 |
|
psllw mm3, 4 |
384 |
|
|
385 |
|
movq mm2, [intra_matrix + 8*ecx] |
386 |
|
psrlw mm2, 1 |
387 |
|
paddw mm0, mm2 |
388 |
|
|
389 |
|
movq mm2, [intra_matrix_fix + ecx*8] |
390 |
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
391 |
|
|
392 |
|
movq mm2, [intra_matrix + 8*ecx + 8] |
393 |
|
psrlw mm2, 1 |
394 |
|
paddw mm3, mm2 |
395 |
|
|
396 |
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
397 |
|
pmulhw mm3, mm2 |
398 |
|
|
399 |
|
paddw mm0, mm5 |
400 |
|
paddw mm3, mm5 |
401 |
|
|
402 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
403 |
|
psrlw mm3, 1 ; |
404 |
|
|
405 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
406 |
|
pxor mm3, mm4 ; |
407 |
|
psubw mm0, mm1 ; undisplace |
408 |
|
psubw mm3, mm4 ; |
409 |
|
|
410 |
|
movq [edi + 8*ecx], mm0 |
411 |
|
movq [edi + 8*ecx + 8], mm3 |
412 |
|
|
413 |
|
add ecx,2 |
414 |
|
cmp ecx,16 |
415 |
|
jnz near .q1loop |
416 |
|
jmp near .done |
417 |
|
|
418 |
|
|
419 |
|
align ALIGN |
420 |
|
.q2loop |
421 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
422 |
|
movq mm3, [esi + 8*ecx + 8] ; |
423 |
|
|
424 |
|
pxor mm1, mm1 ; mm1 = 0 |
425 |
|
pxor mm4, mm4 ; |
426 |
|
|
427 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
428 |
|
pcmpgtw mm4, mm3 ; |
429 |
|
|
430 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
431 |
|
pxor mm3, mm4 ; |
432 |
|
psubw mm0, mm1 ; displace |
433 |
|
psubw mm3, mm4 ; |
434 |
|
|
435 |
|
psllw mm0, 4 |
436 |
|
psllw mm3, 4 |
437 |
|
|
438 |
|
movq mm2, [intra_matrix + 8*ecx] |
439 |
|
psrlw mm2, 1 |
440 |
|
paddw mm0, mm2 |
441 |
|
|
442 |
|
movq mm2, [intra_matrix_fix + ecx*8] |
443 |
|
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
444 |
|
|
445 |
|
movq mm2, [intra_matrix + 8*ecx + 8] |
446 |
|
psrlw mm2, 1 |
447 |
|
paddw mm3, mm2 |
448 |
|
|
449 |
|
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
450 |
|
pmulhw mm3, mm2 |
451 |
|
|
452 |
|
paddw mm0, mm5 |
453 |
|
paddw mm3, mm5 |
454 |
|
|
455 |
|
psrlw mm0, 2 ; mm0 >>= 1 (/4) |
456 |
|
psrlw mm3, 2 ; |
457 |
|
|
458 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
459 |
|
pxor mm3, mm4 ; |
460 |
|
psubw mm0, mm1 ; undisplace |
461 |
|
psubw mm3, mm4 ; |
462 |
|
|
463 |
|
movq [edi + 8*ecx], mm0 |
464 |
|
movq [edi + 8*ecx + 8], mm3 |
465 |
|
|
466 |
|
add ecx,2 |
467 |
|
cmp ecx,16 |
468 |
|
jnz near .q2loop |
469 |
|
jmp near .done |
470 |
|
|
471 |
|
|
472 |
|
;=========================================================================== |
473 |
|
; |
474 |
|
; uint32_t quant4_inter_mmx(int16_t * coeff, |
475 |
|
; const int16_t const * data, |
476 |
|
; const uint32_t quant); |
477 |
|
; |
478 |
|
;=========================================================================== |
479 |
|
|
480 |
|
align ALIGN |
481 |
|
cglobal quant_mpeg_inter_mmx |
482 |
|
quant_mpeg_inter_mmx: |
483 |
|
|
484 |
|
push ecx |
485 |
|
push esi |
486 |
|
push edi |
487 |
|
|
488 |
|
mov edi, [esp + 12 + 4] ; coeff |
489 |
|
mov esi, [esp + 12 + 8] ; data |
490 |
|
mov eax, [esp + 12 + 12] ; quant |
491 |
|
|
492 |
|
xor ecx, ecx |
493 |
|
|
494 |
|
pxor mm5, mm5 ; sum |
495 |
|
|
496 |
|
cmp al, 1 |
497 |
|
jz near .q1loop |
498 |
|
|
499 |
|
cmp al, 2 |
500 |
|
jz near .q2loop |
501 |
|
|
502 |
|
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
503 |
|
|
504 |
|
align ALIGN |
505 |
|
.loop |
506 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
507 |
|
movq mm3, [esi + 8*ecx + 8] ; |
508 |
|
pxor mm1, mm1 ; mm1 = 0 |
509 |
|
pxor mm4, mm4 ; |
510 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
511 |
|
pcmpgtw mm4, mm3 ; |
512 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
513 |
|
pxor mm3, mm4 ; |
514 |
|
psubw mm0, mm1 ; displace |
515 |
|
psubw mm3, mm4 ; |
516 |
|
|
517 |
|
psllw mm0, 4 |
518 |
|
psllw mm3, 4 |
519 |
|
|
520 |
|
movq mm2, [inter_matrix + 8*ecx] |
521 |
|
psrlw mm2, 1 |
522 |
|
paddw mm0, mm2 |
523 |
|
|
524 |
|
movq mm2, [inter_matrix_fix + ecx*8] |
525 |
|
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
526 |
|
|
527 |
|
movq mm2, [inter_matrix + 8*ecx + 8] |
528 |
|
psrlw mm2, 1 |
529 |
|
paddw mm3, mm2 |
530 |
|
|
531 |
|
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
532 |
|
pmulhw mm3, mm2 |
533 |
|
|
534 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
535 |
|
pmulhw mm3, mm7 ; |
536 |
|
psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 |
537 |
|
psrlw mm3, 1 |
538 |
|
|
539 |
|
paddw mm5, mm0 ; sum += mm0 |
540 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
541 |
|
paddw mm5, mm3 ; |
542 |
|
pxor mm3, mm4 ; |
543 |
|
psubw mm0, mm1 ; undisplace |
544 |
|
psubw mm3, mm4 |
545 |
|
movq [edi + 8*ecx], mm0 |
546 |
|
movq [edi + 8*ecx + 8], mm3 |
547 |
|
|
548 |
|
add ecx, 2 |
549 |
|
cmp ecx, 16 |
550 |
|
jnz near .loop |
551 |
|
|
552 |
|
.done |
553 |
|
pmaddwd mm5, [mmx_one] |
554 |
|
movq mm0, mm5 |
555 |
|
psrlq mm5, 32 |
556 |
|
paddd mm0, mm5 |
557 |
|
movd eax, mm0 ; return sum |
558 |
|
|
559 |
|
pop edi |
560 |
|
pop esi |
561 |
|
pop ecx |
562 |
|
|
563 |
|
ret |
564 |
|
|
565 |
|
align ALIGN |
566 |
|
.q1loop |
567 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
568 |
|
movq mm3, [esi + 8*ecx+ 8] |
569 |
|
; |
570 |
|
pxor mm1, mm1 ; mm1 = 0 |
571 |
|
pxor mm4, mm4 ; |
572 |
|
|
573 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
574 |
|
pcmpgtw mm4, mm3 ; |
575 |
|
|
576 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
577 |
|
pxor mm3, mm4 ; |
578 |
|
psubw mm0, mm1 ; displace |
579 |
|
psubw mm3, mm4 ; |
580 |
|
|
581 |
|
psllw mm0, 4 |
582 |
|
psllw mm3, 4 |
583 |
|
|
584 |
|
movq mm2, [inter_matrix + 8*ecx] |
585 |
|
psrlw mm2, 1 |
586 |
|
paddw mm0, mm2 |
587 |
|
|
588 |
|
movq mm2, [inter_matrix_fix + ecx*8] |
589 |
|
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
590 |
|
|
591 |
|
movq mm2, [inter_matrix + 8*ecx + 8] |
592 |
|
psrlw mm2, 1 |
593 |
|
paddw mm3, mm2 |
594 |
|
|
595 |
|
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
596 |
|
pmulhw mm3, mm2 |
597 |
|
|
598 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
599 |
|
psrlw mm3, 1 ; |
600 |
|
|
601 |
|
paddw mm5, mm0 ; sum += mm0 |
602 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
603 |
|
paddw mm5, mm3 ; |
604 |
|
pxor mm3, mm4 ; |
605 |
|
psubw mm0, mm1 ; undisplace |
606 |
|
psubw mm3, mm4 |
607 |
|
|
608 |
|
movq [edi + 8*ecx], mm0 |
609 |
|
movq [edi + 8*ecx + 8], mm3 |
610 |
|
|
611 |
|
add ecx,2 |
612 |
|
cmp ecx,16 |
613 |
|
jnz near .q1loop |
614 |
|
|
615 |
|
jmp .done |
616 |
|
|
617 |
|
|
618 |
|
align ALIGN |
619 |
|
.q2loop |
620 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
621 |
|
movq mm3, [esi + 8*ecx+ 8] |
622 |
|
; |
623 |
|
pxor mm1, mm1 ; mm1 = 0 |
624 |
|
pxor mm4, mm4 ; |
625 |
|
|
626 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
627 |
|
pcmpgtw mm4, mm3 ; |
628 |
|
|
629 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
630 |
|
pxor mm3, mm4 ; |
631 |
|
psubw mm0, mm1 ; displace |
632 |
|
psubw mm3, mm4 ; |
633 |
|
|
634 |
|
psllw mm0, 4 |
635 |
|
psllw mm3, 4 |
636 |
|
|
637 |
|
movq mm2, [inter_matrix + 8*ecx] |
638 |
|
psrlw mm2, 1 |
639 |
|
paddw mm0, mm2 |
640 |
|
|
641 |
|
movq mm2, [inter_matrix_fix + ecx*8] |
642 |
|
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
643 |
|
|
644 |
|
movq mm2, [inter_matrix + 8*ecx + 8] |
645 |
|
psrlw mm2, 1 |
646 |
|
paddw mm3, mm2 |
647 |
|
|
648 |
|
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
649 |
|
pmulhw mm3, mm2 |
650 |
|
|
651 |
|
psrlw mm0, 2 ; mm0 >>= 1 (/2) |
652 |
|
psrlw mm3, 2 ; |
653 |
|
|
654 |
|
paddw mm5, mm0 ; sum += mm0 |
655 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
656 |
|
paddw mm5, mm3 ; |
657 |
|
pxor mm3, mm4 ; |
658 |
|
psubw mm0, mm1 ; undisplace |
659 |
|
psubw mm3, mm4 |
660 |
|
|
661 |
|
movq [edi + 8*ecx], mm0 |
662 |
|
movq [edi + 8*ecx + 8], mm3 |
663 |
|
|
664 |
|
add ecx,2 |
665 |
|
cmp ecx,16 |
666 |
|
jnz near .q2loop |
667 |
|
|
668 |
|
jmp .done |
669 |
|
|
670 |
|
|
671 |
|
;=========================================================================== |
672 |
|
; |
673 |
|
; void dequant4_intra_mmx(int16_t *data, |
674 |
|
; const int16_t const *coeff, |
675 |
|
; const uint32_t quant, |
676 |
|
; const uint32_t dcscalar); |
677 |
|
; |
678 |
|
;=========================================================================== |
679 |
|
|
680 |
|
; Note: in order to saturate 'easily', we pre-shift the quantifier |
681 |
|
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
682 |
|
; build a saturating mask. It is non-zero only when an overflow occured. |
683 |
|
; We thus avoid packing/unpacking toward double-word. |
684 |
|
; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., |
685 |
|
; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not |
686 |
|
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
687 |
|
; and quant in [1..31]. |
688 |
|
; |
689 |
|
; The original loop is: |
690 |
|
; |
691 |
|
%if 0 |
692 |
|
movq mm0, [ecx+8*eax + 8*16] ; mm0 = coeff[i] |
693 |
|
pxor mm1, mm1 |
694 |
|
pcmpgtw mm1, mm0 |
695 |
|
pxor mm0, mm1 ; change sign if negative |
696 |
|
psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] |
697 |
|
|
698 |
|
movq mm2, mm7 ; mm2 = quant |
699 |
|
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant. |
700 |
|
|
701 |
|
movq mm6, mm2 |
702 |
|
pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) |
703 |
|
pmullw mm0, mm6 ; low of coeff*(matrix*quant) |
704 |
|
|
705 |
|
pxor mm5, mm5 |
706 |
|
pcmpgtw mm2, mm5 ; otherflow? |
707 |
|
psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise |
708 |
|
psrlw mm0, 5 |
709 |
|
paddw mm0, mm1 ; start restoring sign |
710 |
|
por mm0, mm2 ; saturate to 2047 if needed |
711 |
|
pxor mm0, mm1 ; finish negating back |
712 |
|
|
713 |
|
movq [edx + 8*eax + 8*16], mm0 ; data[i] |
714 |
|
add eax, 1 |
715 |
|
%endif |
716 |
|
|
717 |
|
;******************************************************************** |
718 |
|
|
719 |
|
align 16 |
720 |
|
cglobal dequant_mpeg_intra_mmx |
721 |
|
dequant_mpeg_intra_mmx: |
722 |
|
|
723 |
|
mov edx, [esp+4] ; data |
724 |
|
mov ecx, [esp+8] ; coeff |
725 |
|
mov eax, [esp+12] ; quant |
726 |
|
|
727 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
728 |
|
mov eax, -16 ; to keep aligned, we regularly process coeff[0] |
729 |
|
psllw mm7, 2 ; << 2. See comment. |
730 |
|
pxor mm6, mm6 ; this is a NOP |
731 |
|
|
732 |
|
align 16 |
733 |
|
.loop |
734 |
|
movq mm0, [ecx+8*eax + 8*16] ; mm0 = c = coeff[i] |
735 |
|
movq mm3, [ecx+8*eax + 8*16 +8]; mm3 = c' = coeff[i+1] |
736 |
|
pxor mm1, mm1 |
737 |
|
pxor mm4, mm4 |
738 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) |
739 |
|
movq mm2, mm7 ; mm2 = quant |
740 |
|
|
741 |
|
pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
742 |
|
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant |
743 |
|
|
744 |
|
pxor mm0, mm1 ; negate if negative |
745 |
|
pxor mm3, mm4 ; negate if negative |
746 |
|
|
747 |
|
psubw mm0, mm1 |
748 |
|
psubw mm3, mm4 |
749 |
|
|
750 |
|
; we're short on register, here. Poor pairing... |
751 |
|
|
752 |
|
movq mm5, mm2 |
753 |
|
pmullw mm2, mm0 ; low of coeff*(matrix*quant) |
754 |
|
|
755 |
|
pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
756 |
|
movq mm5, mm7 ; mm2 = quant |
757 |
|
|
758 |
|
pmullw mm5, [intra_matrix + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
759 |
|
|
760 |
|
movq mm6, mm5 |
761 |
|
add eax,2 ; z-flag will be tested later |
762 |
|
|
763 |
|
pmullw mm6, mm3 ; low of coeff*(matrix*quant) |
764 |
|
pmulhw mm3, mm5 ; high of coeff*(matrix*quant) |
765 |
|
|
766 |
|
pcmpgtw mm0, [zero] |
767 |
|
paddusw mm2, mm0 |
768 |
|
psrlw mm2, 5 |
769 |
|
|
770 |
|
pcmpgtw mm3, [zero] |
771 |
|
paddusw mm6, mm3 |
772 |
|
psrlw mm6, 5 |
773 |
|
|
774 |
|
pxor mm2, mm1 ; start negating back |
775 |
|
pxor mm6, mm4 ; start negating back |
776 |
|
|
777 |
|
psubusw mm1, mm0 |
778 |
|
psubusw mm4, mm3 |
779 |
|
|
780 |
|
psubw mm2, mm1 ; finish negating back |
781 |
|
psubw mm6, mm4 ; finish negating back |
782 |
|
|
783 |
|
movq [edx + 8*eax + 8*16 -2*8 ], mm2 ; data[i] |
784 |
|
movq [edx + 8*eax + 8*16 -2*8 +8], mm6 ; data[i+1] |
785 |
|
|
786 |
|
jnz near .loop |
787 |
|
|
788 |
|
; deal with DC |
789 |
|
|
790 |
|
movd mm0, [ecx] |
791 |
|
pmullw mm0, [esp+16] ; dcscalar |
792 |
|
movq mm2, [mmx_32767_minus_2047] |
793 |
|
paddsw mm0, mm2 |
794 |
|
psubsw mm0, mm2 |
795 |
|
movq mm2, [mmx_32768_minus_2048] |
796 |
|
psubsw mm0, mm2 |
797 |
|
paddsw mm0, mm2 |
798 |
|
movd eax, mm0 |
799 |
|
mov [edx], ax |
800 |
|
|
801 |
|
ret |
802 |
|
|
803 |
|
;=========================================================================== |
804 |
|
; |
805 |
|
; void dequant4_inter_mmx(int16_t * data, |
806 |
|
; const int16_t * const coeff, |
807 |
|
; const uint32_t quant); |
808 |
|
; |
809 |
|
;=========================================================================== |
810 |
|
|
811 |
|
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
812 |
|
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
813 |
|
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
814 |
|
; It's mixed with the extraction of the absolute value. |
815 |
|
|
816 |
|
align 16 |
817 |
|
cglobal dequant_mpeg_inter_mmx |
818 |
|
dequant_mpeg_inter_mmx: |
819 |
|
|
820 |
|
mov edx, [esp+ 4] ; data |
821 |
|
mov ecx, [esp+ 8] ; coeff |
822 |
|
mov eax, [esp+12] ; quant |
823 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
824 |
|
mov eax, -16 |
825 |
|
paddw mm7, mm7 ; << 1 |
826 |
|
pxor mm6, mm6 ; mismatch sum |
827 |
|
|
828 |
|
align 16 |
829 |
|
.loop |
830 |
|
movq mm0, [ecx+8*eax + 8*16 ] ; mm0 = coeff[i] |
831 |
|
movq mm2, [ecx+8*eax + 8*16 +8] ; mm2 = coeff[i+1] |
832 |
|
add eax,2 |
833 |
|
|
834 |
|
pxor mm1, mm1 |
835 |
|
pxor mm3, mm3 |
836 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
837 |
|
pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) |
838 |
|
paddsw mm0, mm1 ; c += sgn(c) |
839 |
|
paddsw mm2, mm3 ; c += sgn(c') |
840 |
|
paddw mm0, mm0 ; c *= 2 |
841 |
|
paddw mm2, mm2 ; c'*= 2 |
842 |
|
|
843 |
|
pxor mm4, mm4 |
844 |
|
pxor mm5, mm5 |
845 |
|
psubw mm4, mm0 ; -c |
846 |
|
psubw mm5, mm2 ; -c' |
847 |
|
psraw mm4, 16 ; mm4 = sgn(-c) |
848 |
|
psraw mm5, 16 ; mm5 = sgn(-c') |
849 |
|
psubsw mm0, mm4 ; c -= sgn(-c) |
850 |
|
psubsw mm2, mm5 ; c' -= sgn(-c') |
851 |
|
pxor mm0, mm1 ; finish changing sign if needed |
852 |
|
pxor mm2, mm3 ; finish changing sign if needed |
853 |
|
|
854 |
|
; we're short on register, here. Poor pairing... |
855 |
|
|
856 |
|
movq mm4, mm7 ; (matrix*quant) |
857 |
|
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8] |
858 |
|
movq mm5, mm4 |
859 |
|
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
860 |
|
pmullw mm0, mm4 ; low of c*(matrix*quant) |
861 |
|
|
862 |
|
movq mm4, mm7 ; (matrix*quant) |
863 |
|
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8 + 8] |
864 |
|
|
865 |
|
pcmpgtw mm5, [zero] |
866 |
|
paddusw mm0, mm5 |
867 |
|
psrlw mm0, 5 |
868 |
|
pxor mm0, mm1 ; start restoring sign |
869 |
|
psubusw mm1, mm5 |
870 |
|
|
871 |
|
movq mm5, mm4 |
872 |
|
pmulhw mm5, mm2 ; high of c*(matrix*quant) |
873 |
|
pmullw mm2, mm4 ; low of c*(matrix*quant) |
874 |
|
psubw mm0, mm1 ; finish restoring sign |
875 |
|
|
876 |
|
pcmpgtw mm5, [zero] |
877 |
|
paddusw mm2, mm5 |
878 |
|
psrlw mm2, 5 |
879 |
|
pxor mm2, mm3 ; start restoring sign |
880 |
|
psubusw mm3, mm5 |
881 |
|
psubw mm2, mm3 ; finish restoring sign |
882 |
|
|
883 |
|
pxor mm6, mm0 ; mismatch control |
884 |
|
movq [edx + 8*eax + 8*16 -2*8 ], mm0 ; data[i] |
885 |
|
pxor mm6, mm2 ; mismatch control |
886 |
|
movq [edx + 8*eax + 8*16 -2*8 +8], mm2 ; data[i+1] |
887 |
|
|
888 |
|
jnz near .loop |
889 |
|
|
890 |
|
; mismatch control |
891 |
|
|
892 |
|
movq mm0, mm6 |
893 |
|
psrlq mm0, 48 |
894 |
|
movq mm1, mm6 |
895 |
|
movq mm2, mm6 |
896 |
|
psrlq mm1, 32 |
897 |
|
pxor mm6, mm0 |
898 |
|
psrlq mm2, 16 |
899 |
|
pxor mm6, mm1 |
900 |
|
pxor mm6, mm2 |
901 |
|
movd eax, mm6 |
902 |
|
and eax, 1 |
903 |
|
xor eax, 1 |
904 |
|
xor word [edx + 2*63], ax |
905 |
|
|
906 |
|
ret |
907 |
|
|