Parent Directory | Revision Log
Revision 1713 - (view) (download)
1 : | edgomez | 1382 | ;/**************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - 3dne Quantization/Dequantization - | ||
5 : | ; * | ||
6 : | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> | ||
7 : | ; * 2002 Jaan Kalda | ||
8 : | ; * | ||
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 : | syskin | 1713 | ; * $Id: quantize_mpeg_xmm.asm,v 1.6 2006-07-10 08:09:59 syskin Exp $ |
24 : | edgomez | 1382 | ; * |
25 : | ; ***************************************************************************/ | ||
26 : | |||
27 : | ; _3dne functions are compatible with iSSE, but are optimized specifically | ||
28 : | ; for K7 pipelines | ||
29 : | |||
30 : | %define SATURATE | ||
31 : | |||
32 : | BITS 32 | ||
33 : | |||
34 : | %macro cglobal 1 | ||
35 : | %ifdef PREFIX | ||
36 : | edgomez | 1535 | %ifdef MARK_FUNCS |
37 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
38 : | %define %1 _%1:function %1.endfunc-%1 | ||
39 : | edgomez | 1535 | %else |
40 : | global _%1 | ||
41 : | %define %1 _%1 | ||
42 : | %endif | ||
43 : | edgomez | 1382 | %else |
44 : | edgomez | 1535 | %ifdef MARK_FUNCS |
45 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
46 : | edgomez | 1535 | %else |
47 : | global %1 | ||
48 : | %endif | ||
49 : | edgomez | 1382 | %endif |
50 : | %endmacro | ||
51 : | |||
52 : | %macro cextern 1 | ||
53 : | %ifdef PREFIX | ||
54 : | extern _%1 | ||
55 : | %define %1 _%1 | ||
56 : | %else | ||
57 : | extern %1 | ||
58 : | %endif | ||
59 : | %endmacro | ||
60 : | |||
61 : | ;============================================================================= | ||
62 : | ; Local data | ||
63 : | ;============================================================================= | ||
64 : | |||
65 : | %ifdef FORMAT_COFF | ||
66 : | edgomez | 1519 | SECTION .rodata |
67 : | edgomez | 1382 | %else |
68 : | edgomez | 1519 | SECTION .rodata align=16 |
69 : | edgomez | 1382 | %endif |
70 : | |||
71 : | ALIGN 8 | ||
72 : | mmzero: | ||
73 : | dd 0,0 | ||
74 : | mmx_one: | ||
75 : | times 4 dw 1 | ||
76 : | |||
77 : | ;----------------------------------------------------------------------------- | ||
78 : | ; divide by 2Q table | ||
79 : | ;----------------------------------------------------------------------------- | ||
80 : | |||
81 : | ALIGN 16 | ||
82 : | mmx_divs: ;i>2 | ||
83 : | %assign i 1 | ||
84 : | %rep 31 | ||
85 : | times 4 dw ((1 << 15) / i + 1) | ||
86 : | %assign i i+1 | ||
87 : | %endrep | ||
88 : | |||
89 : | ALIGN 16 | ||
90 : | mmx_div: ;quant>2 | ||
91 : | times 4 dw 65535 ; the div by 2 formula will overflow for the case | ||
92 : | ; quant=1 but we don't care much because quant=1 | ||
93 : | ; is handled by a different piece of code that | ||
94 : | ; doesn't use this table. | ||
95 : | %assign quant 2 | ||
96 : | %rep 31 | ||
97 : | times 4 dw ((1 << 16) / quant + 1) | ||
98 : | %assign quant quant+1 | ||
99 : | %endrep | ||
100 : | |||
101 : | %macro FIXX 1 | ||
102 : | dw (1 << 16) / (%1) + 1 | ||
103 : | %endmacro | ||
104 : | |||
105 : | %define nop4 db 08Dh, 074h, 026h,0 | ||
106 : | %define nop3 add esp, byte 0 | ||
107 : | %define nop2 mov esp, esp | ||
108 : | %define nop7 db 08dh, 02ch, 02dh,0,0,0,0 | ||
109 : | %define nop6 add ebp, dword 0 | ||
110 : | |||
111 : | ;----------------------------------------------------------------------------- | ||
112 : | ; quantd table | ||
113 : | ;----------------------------------------------------------------------------- | ||
114 : | |||
115 : | %define VM18P 3 | ||
116 : | %define VM18Q 4 | ||
117 : | |||
118 : | ALIGN 16 | ||
119 : | quantd: | ||
120 : | %assign i 1 | ||
121 : | %rep 31 | ||
122 : | times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) | ||
123 : | %assign i i+1 | ||
124 : | %endrep | ||
125 : | |||
126 : | ;----------------------------------------------------------------------------- | ||
127 : | ; multiple by 2Q table | ||
128 : | ;----------------------------------------------------------------------------- | ||
129 : | |||
130 : | ALIGN 16 | ||
131 : | mmx_mul_quant: | ||
132 : | %assign i 1 | ||
133 : | %rep 31 | ||
134 : | times 4 dw i | ||
135 : | %assign i i+1 | ||
136 : | %endrep | ||
137 : | |||
138 : | ;----------------------------------------------------------------------------- | ||
139 : | ; saturation limits | ||
140 : | ;----------------------------------------------------------------------------- | ||
141 : | |||
142 : | ALIGN 16 | ||
143 : | mmx_32767_minus_2047: | ||
144 : | times 4 dw (32767-2047) | ||
145 : | mmx_32768_minus_2048: | ||
146 : | times 4 dw (32768-2048) | ||
147 : | mmx_2047: | ||
148 : | times 4 dw 2047 | ||
149 : | mmx_minus_2048: | ||
150 : | times 4 dw (-2048) | ||
151 : | zero: | ||
152 : | times 4 dw 0 | ||
153 : | |||
154 : | int_div: | ||
155 : | dd 0 | ||
156 : | %assign i 1 | ||
157 : | %rep 255 | ||
158 : | dd (1 << 17) / ( i) + 1 | ||
159 : | %assign i i+1 | ||
160 : | %endrep | ||
161 : | |||
162 : | ;============================================================================= | ||
163 : | ; Code | ||
164 : | ;============================================================================= | ||
165 : | |||
166 : | SECTION .text | ||
167 : | |||
168 : | cglobal quant_mpeg_inter_xmm | ||
169 : | cglobal dequant_mpeg_intra_3dne | ||
170 : | cglobal dequant_mpeg_inter_3dne | ||
171 : | |||
172 : | ;----------------------------------------------------------------------------- | ||
173 : | ; | ||
174 : | ; uint32_t quant_mpeg_inter_xmm(int16_t * coeff, | ||
175 : | ; const int16_t const * data, | ||
176 : | ; const uint32_t quant, | ||
177 : | ; const uint16_t *mpeg_matrices); | ||
178 : | ; | ||
179 : | ;----------------------------------------------------------------------------- | ||
180 : | |||
181 : | ALIGN 16 | ||
182 : | quant_mpeg_inter_xmm: | ||
183 : | mov eax, [esp + 8] ; data | ||
184 : | mov ecx, [esp + 12] ; quant | ||
185 : | mov edx, [esp + 4] ; coeff | ||
186 : | push esi | ||
187 : | push edi | ||
188 : | push ebx | ||
189 : | nop | ||
190 : | mov edi, [esp + 12 + 16] | ||
191 : | mov esi, -14 | ||
192 : | mov ebx, esp | ||
193 : | sub esp, byte 24 | ||
194 : | lea ebx, [esp+8] | ||
195 : | and ebx, byte -8 ;ALIGN 8 | ||
196 : | pxor mm0, mm0 | ||
197 : | pxor mm3, mm3 | ||
198 : | movq [byte ebx],mm0 | ||
199 : | db 0Fh, 7Fh, 44h, 23h, 8 ;movq [ebx+8],mm0 | ||
200 : | cmp ecx, byte 1 | ||
201 : | je near .q1loop | ||
202 : | cmp ecx, byte 19 | ||
203 : | jg near .lloop | ||
204 : | nop | ||
205 : | |||
206 : | ALIGN 16 | ||
207 : | .loop | ||
208 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
209 : | psubw mm0, mm1 ;-mm1 | ||
210 : | movq mm4, [eax + 8*esi + 120] ; | ||
211 : | psubw mm3, mm4 ;-mm4 | ||
212 : | pmaxsw mm0, mm1 ;|src| | ||
213 : | pmaxsw mm3, mm4 | ||
214 : | nop2 | ||
215 : | psraw mm1, 15 ;sign src | ||
216 : | psraw mm4, 15 | ||
217 : | psllw mm0, 4 ; level << 4 | ||
218 : | psllw mm3, 4 ; | ||
219 : | paddw mm0, [edi + 640 + 8*esi+112] | ||
220 : | paddw mm3, [edi + 640 + 8*esi+120] | ||
221 : | movq mm5, [edi + 896 + 8*esi+112] | ||
222 : | movq mm7, [edi + 896 + 8*esi+120] | ||
223 : | pmulhuw mm5, mm0 | ||
224 : | pmulhuw mm7, mm3 | ||
225 : | mov esp, esp | ||
226 : | movq mm2, [edi + 512 + 8*esi+112] | ||
227 : | movq mm6, [edi + 512 + 8*esi+120] | ||
228 : | pmullw mm2, mm5 | ||
229 : | pmullw mm6, mm7 | ||
230 : | psubw mm0, mm2 | ||
231 : | psubw mm3, mm6 | ||
232 : | movq mm2, [byte ebx] | ||
233 : | movq mm6, [mmx_divs + ecx * 8 - 8] | ||
234 : | pmulhuw mm0, [edi + 768 + 8*esi+112] | ||
235 : | pmulhuw mm3, [edi + 768 + 8*esi+120] | ||
236 : | paddw mm2, [ebx+8] ;sum | ||
237 : | paddw mm5, mm0 | ||
238 : | paddw mm7, mm3 | ||
239 : | pxor mm0, mm0 | ||
240 : | pxor mm3, mm3 | ||
241 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
242 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
243 : | add esi, byte 2 | ||
244 : | paddw mm2, mm5 ;sum += x1 | ||
245 : | movq [ebx], mm7 ;store x2 | ||
246 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
247 : | pxor mm7, mm4 ; | ||
248 : | psubw mm5, mm1 ; undisplace | ||
249 : | psubw mm7, mm4 ; | ||
250 : | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
251 : | movq [edx + 8*esi+112-16], mm5 | ||
252 : | movq [edx + 8*esi +120-16], mm7 | ||
253 : | jng near .loop | ||
254 : | |||
255 : | .done | ||
256 : | ; calculate data[0] // (int32_t)dcscalar) | ||
257 : | paddw mm2, [ebx] | ||
258 : | mov ebx, [esp+24] | ||
259 : | mov edi, [esp+4+24] | ||
260 : | mov esi, [esp+8+24] | ||
261 : | add esp, byte 12+24 | ||
262 : | pmaddwd mm2, [mmx_one] | ||
263 : | punpckldq mm0, mm2 ;get low dw to mm0:high | ||
264 : | paddd mm0,mm2 | ||
265 : | punpckhdq mm0, mm0 ;get result to low | ||
266 : | movd eax, mm0 | ||
267 : | |||
268 : | ret | ||
269 : | |||
270 : | ALIGN 16 | ||
271 : | .q1loop | ||
272 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
273 : | psubw mm0, mm1 ;-mm1 | ||
274 : | movq mm4, [eax + 8*esi+120] | ||
275 : | psubw mm3, mm4 ;-mm4 | ||
276 : | pmaxsw mm0, mm1 ;|src| | ||
277 : | pmaxsw mm3, mm4 | ||
278 : | nop2 | ||
279 : | psraw mm1, 15 ; sign src | ||
280 : | psraw mm4, 15 | ||
281 : | psllw mm0, 4 ; level << 4 | ||
282 : | psllw mm3, 4 | ||
283 : | paddw mm0, [edi + 640 + 8*esi+112] ;mm0 is to be divided | ||
284 : | paddw mm3, [edi + 640 + 8*esi+120] ; inter1 contains fix for division by 1 | ||
285 : | movq mm5, [edi + 896 + 8*esi+112] ;with rounding down | ||
286 : | movq mm7, [edi + 896 + 8*esi+120] | ||
287 : | pmulhuw mm5, mm0 | ||
288 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
289 : | mov esp, esp | ||
290 : | movq mm2, [edi + 512 + 8*esi+112] | ||
291 : | movq mm6, [edi + 512 + 8*esi+120] ; divs for q<=16 | ||
292 : | pmullw mm2, mm5 ;test value <= original | ||
293 : | pmullw mm6, mm7 | ||
294 : | psubw mm0, mm2 ;mismatch | ||
295 : | psubw mm3, mm6 | ||
296 : | movq mm2, [byte ebx] | ||
297 : | pmulhuw mm0, [edi + 768 + 8*esi+112] ;correction | ||
298 : | pmulhuw mm3, [edi + 768 + 8*esi+120] | ||
299 : | paddw mm2, [ebx+8] ;sum | ||
300 : | paddw mm5, mm0 ;final result | ||
301 : | paddw mm7, mm3 | ||
302 : | pxor mm0, mm0 | ||
303 : | pxor mm3, mm3 | ||
304 : | psrlw mm5, 1 ; (level ) /2 (quant = 1) | ||
305 : | psrlw mm7, 1 | ||
306 : | add esi, byte 2 | ||
307 : | paddw mm2, mm5 ;sum += x1 | ||
308 : | movq [ebx], mm7 ;store x2 | ||
309 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
310 : | pxor mm7, mm4 ; | ||
311 : | psubw mm5, mm1 ; undisplace | ||
312 : | psubw mm7, mm4 ; | ||
313 : | movq [ebx+8], mm2 ;store sum | ||
314 : | movq [edx + 8*esi+112-16], mm5 | ||
315 : | movq [edx + 8*esi +120-16], mm7 | ||
316 : | jng near .q1loop | ||
317 : | jmp near .done | ||
318 : | |||
319 : | ALIGN 8 | ||
320 : | .lloop | ||
321 : | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] | ||
322 : | psubw mm0,mm1 ;-mm1 | ||
323 : | movq mm4, [eax + 8*esi+120] | ||
324 : | psubw mm3,mm4 ;-mm4 | ||
325 : | pmaxsw mm0,mm1 ;|src| | ||
326 : | pmaxsw mm3,mm4 | ||
327 : | nop2 | ||
328 : | psraw mm1,15 ;sign src | ||
329 : | psraw mm4,15 | ||
330 : | psllw mm0, 4 ; level << 4 | ||
331 : | psllw mm3, 4 ; | ||
332 : | paddw mm0, [edi + 640 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 | ||
333 : | paddw mm3, [edi + 640 + 8*esi+120] | ||
334 : | movq mm5,[edi + 896 + 8*esi+112] | ||
335 : | movq mm7,[edi + 896 + 8*esi+120] | ||
336 : | pmulhuw mm5,mm0 | ||
337 : | pmulhuw mm7,mm3 ;mm7: first approx of division | ||
338 : | mov esp,esp | ||
339 : | movq mm2,[edi + 512 + 8*esi+112] | ||
340 : | movq mm6,[edi + 512 + 8*esi+120] | ||
341 : | pmullw mm2,mm5 ;test value <= original | ||
342 : | pmullw mm6,mm7 | ||
343 : | psubw mm0,mm2 ;mismatch | ||
344 : | psubw mm3,mm6 | ||
345 : | movq mm2,[byte ebx] | ||
346 : | movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 | ||
347 : | pmulhuw mm0,[edi + 768 + 8*esi+112] ;correction | ||
348 : | pmulhuw mm3,[edi + 768 + 8*esi+120] | ||
349 : | paddw mm2,[ebx+8] ;sum | ||
350 : | paddw mm5,mm0 ;final result | ||
351 : | paddw mm7,mm3 | ||
352 : | pxor mm0,mm0 | ||
353 : | pxor mm3,mm3 | ||
354 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
355 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
356 : | add esi,byte 2 | ||
357 : | psrlw mm5, 1 ; (level ) / (2*quant) | ||
358 : | paddw mm2,mm5 ;sum += x1 | ||
359 : | psrlw mm7, 1 | ||
360 : | movq [ebx],mm7 ;store x2 | ||
361 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
362 : | pxor mm7, mm4 ; | ||
363 : | psubw mm5, mm1 ; undisplace | ||
364 : | psubw mm7, mm4 ; | ||
365 : | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
366 : | movq [edx + 8*esi+112-16], mm5 | ||
367 : | movq [edx + 8*esi +120-16], mm7 | ||
368 : | jng near .lloop | ||
369 : | jmp near .done | ||
370 : | edgomez | 1540 | .endfunc |
371 : | edgomez | 1382 | |
372 : | |||
373 : | ;----------------------------------------------------------------------------- | ||
374 : | ; | ||
375 : | ; uint32_t dequant_mpeg_intra_3dne(int16_t *data, | ||
376 : | ; const int16_t const *coeff, | ||
377 : | ; const uint32_t quant, | ||
378 : | ; const uint32_t dcscalar, | ||
379 : | ; const uint16_t *mpeg_matrices); | ||
380 : | ; | ||
381 : | ;----------------------------------------------------------------------------- | ||
382 : | |||
383 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
384 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
385 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
386 : | ; We thus avoid packing/unpacking toward double-word. | ||
387 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
388 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
389 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
390 : | ; and quant in [1..31]. | ||
391 : | ; | ||
392 : | |||
393 : | %macro DEQUANT4INTRAMMX 1 | ||
394 : | movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] | ||
395 : | movq mm4, [ecx+ 16 * %1 +8] ; mm3 = c' = coeff[i+1] | ||
396 : | psubw mm0, mm1 | ||
397 : | psubw mm3, mm4 | ||
398 : | pmaxsw mm0, mm1 | ||
399 : | pmaxsw mm3, mm4 | ||
400 : | psraw mm1, 15 | ||
401 : | psraw mm4, 15 | ||
402 : | %if %1 | ||
403 : | movq mm2, [eax+8] ;preshifted quant | ||
404 : | movq mm7, [eax+8] | ||
405 : | %endif | ||
406 : | pmullw mm2, [edi + 16 * %1 ] ; matrix[i]*quant | ||
407 : | pmullw mm7, [edi + 16 * %1 +8] ; matrix[i+1]*quant | ||
408 : | movq mm5, mm0 | ||
409 : | movq mm6, mm3 | ||
410 : | pmulhw mm0, mm2 ; high of coeff*(matrix*quant) | ||
411 : | pmulhw mm3, mm7 ; high of coeff*(matrix*quant) | ||
412 : | pmullw mm2, mm5 ; low of coeff*(matrix*quant) | ||
413 : | pmullw mm7, mm6 ; low of coeff*(matrix*quant) | ||
414 : | pcmpgtw mm0, [eax] | ||
415 : | pcmpgtw mm3, [eax] | ||
416 : | paddusw mm2, mm0 | ||
417 : | paddusw mm7, mm3 | ||
418 : | psrlw mm2, 5 | ||
419 : | psrlw mm7, 5 | ||
420 : | pxor mm2, mm1 ; start negating back | ||
421 : | pxor mm7, mm4 ; start negating back | ||
422 : | psubusw mm1, mm0 | ||
423 : | psubusw mm4, mm3 | ||
424 : | movq mm0, [eax] ;zero | ||
425 : | movq mm3, [eax] ;zero | ||
426 : | psubw mm2, mm1 ; finish negating back | ||
427 : | psubw mm7, mm4 ; finish negating back | ||
428 : | movq [byte edx + 16 * %1], mm2 ; data[i] | ||
429 : | movq [edx + 16 * %1 +8], mm7 ; data[i+1] | ||
430 : | %endmacro | ||
431 : | |||
432 : | ALIGN 16 | ||
433 : | dequant_mpeg_intra_3dne: | ||
434 : | mov eax, [esp+12] ; quant | ||
435 : | mov ecx, [esp+8] ; coeff | ||
436 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
437 : | psllw mm7, 2 ; << 2. See comment. | ||
438 : | mov edx, [esp+4] ; data | ||
439 : | push ebx | ||
440 : | movsx ebx, word [ecx] | ||
441 : | pxor mm0, mm0 | ||
442 : | pxor mm3, mm3 | ||
443 : | push esi | ||
444 : | lea eax, [esp-28] | ||
445 : | sub esp, byte 32 | ||
446 : | and eax, byte -8 ;points to qword ALIGNed space on stack | ||
447 : | movq [eax], mm0 | ||
448 : | movq [eax+8], mm7 | ||
449 : | imul ebx, [esp+16+8+32] ; dcscalar | ||
450 : | movq mm2, mm7 | ||
451 : | push edi | ||
452 : | mov edi, [esp + 32 + 12 + 20] ; mpeg_quant_matrices | ||
453 : | ALIGN 4 | ||
454 : | |||
455 : | DEQUANT4INTRAMMX 0 | ||
456 : | |||
457 : | mov esi, -2048 | ||
458 : | nop | ||
459 : | cmp ebx, esi | ||
460 : | |||
461 : | DEQUANT4INTRAMMX 1 | ||
462 : | |||
463 : | cmovl ebx, esi | ||
464 : | neg esi | ||
465 : | sub esi, byte 1 ;2047 | ||
466 : | |||
467 : | DEQUANT4INTRAMMX 2 | ||
468 : | |||
469 : | cmp ebx, esi | ||
470 : | cmovg ebx, esi | ||
471 : | lea ebp, [byte ebp] | ||
472 : | |||
473 : | DEQUANT4INTRAMMX 3 | ||
474 : | |||
475 : | mov esi, [esp+36] | ||
476 : | mov [byte edx], bx | ||
477 : | mov ebx, [esp+36+4] | ||
478 : | |||
479 : | DEQUANT4INTRAMMX 4 | ||
480 : | DEQUANT4INTRAMMX 5 | ||
481 : | DEQUANT4INTRAMMX 6 | ||
482 : | DEQUANT4INTRAMMX 7 | ||
483 : | |||
484 : | pop edi | ||
485 : | |||
486 : | add esp, byte 32+8 | ||
487 : | |||
488 : | xor eax, eax | ||
489 : | ret | ||
490 : | edgomez | 1540 | .endfunc |
491 : | edgomez | 1382 | |
492 : | ;----------------------------------------------------------------------------- | ||
493 : | ; | ||
494 : | ; uint32_t dequant_mpeg_inter_3dne(int16_t * data, | ||
495 : | ; const int16_t * const coeff, | ||
496 : | ; const uint32_t quant, | ||
497 : | ; const uint16_t *mpeg_matrices); | ||
498 : | ; | ||
499 : | ;----------------------------------------------------------------------------- | ||
500 : | |||
501 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
502 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
503 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
504 : | ; It's mixed with the extraction of the absolute value. | ||
505 : | |||
506 : | ALIGN 16 | ||
507 : | dequant_mpeg_inter_3dne: | ||
508 : | mov edx, [esp+ 4] ; data | ||
509 : | mov ecx, [esp+ 8] ; coeff | ||
510 : | mov eax, [esp+12] ; quant | ||
511 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
512 : | mov eax, -14 | ||
513 : | paddw mm7, mm7 ; << 1 | ||
514 : | pxor mm6, mm6 ; mismatch sum | ||
515 : | push esi | ||
516 : | push edi | ||
517 : | mov esi, mmzero | ||
518 : | pxor mm1, mm1 | ||
519 : | pxor mm3, mm3 | ||
520 : | mov edi, [esp + 8 + 16] ; mpeg_quant_matrices | ||
521 : | nop | ||
522 : | nop4 | ||
523 : | |||
524 : | ALIGN 16 | ||
525 : | .loop | ||
526 : | movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] | ||
527 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
528 : | movq mm2, [ecx+8*eax + 7*16 +8] ; mm2 = coeff[i+1] | ||
529 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
530 : | paddsw mm0, mm1 ; c += sgn(c) | ||
531 : | paddsw mm2, mm3 ; c += sgn(c') | ||
532 : | paddw mm0, mm0 ; c *= 2 | ||
533 : | paddw mm2, mm2 ; c'*= 2 | ||
534 : | |||
535 : | movq mm4, [esi] | ||
536 : | movq mm5, [esi] | ||
537 : | psubw mm4, mm0 ; -c | ||
538 : | psubw mm5, mm2 ; -c' | ||
539 : | |||
540 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
541 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
542 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
543 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
544 : | pxor mm0, mm1 ; finish changing sign if needed | ||
545 : | pxor mm2, mm3 ; finish changing sign if needed | ||
546 : | |||
547 : | ; we're short on register, here. Poor pairing... | ||
548 : | |||
549 : | movq mm4, mm7 ; (matrix*quant) | ||
550 : | nop | ||
551 : | pmullw mm4, [edi + 512 + 8*eax + 7*16] | ||
552 : | movq mm5, mm4 | ||
553 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
554 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
555 : | |||
556 : | movq mm4, mm7 ; (matrix*quant) | ||
557 : | pmullw mm4, [edi + 512 + 8*eax + 7*16 + 8] | ||
558 : | add eax, byte 2 | ||
559 : | |||
560 : | pcmpgtw mm5, [esi] | ||
561 : | paddusw mm0, mm5 | ||
562 : | psrlw mm0, 5 | ||
563 : | pxor mm0, mm1 ; start restoring sign | ||
564 : | psubusw mm1, mm5 | ||
565 : | |||
566 : | movq mm5, mm4 | ||
567 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
568 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
569 : | psubw mm0, mm1 ; finish restoring sign | ||
570 : | |||
571 : | pcmpgtw mm5, [esi] | ||
572 : | paddusw mm2, mm5 | ||
573 : | psrlw mm2, 5 | ||
574 : | pxor mm2, mm3 ; start restoring sign | ||
575 : | psubusw mm3, mm5 | ||
576 : | psubw mm2, mm3 ; finish restoring sign | ||
577 : | movq mm1, [esi] | ||
578 : | movq mm3, [byte esi] | ||
579 : | pxor mm6, mm0 ; mismatch control | ||
580 : | movq [edx + 8*eax + 7*16 -2*8 ], mm0 ; data[i] | ||
581 : | pxor mm6, mm2 ; mismatch control | ||
582 : | movq [edx + 8*eax + 7*16 -2*8 +8], mm2 ; data[i+1] | ||
583 : | |||
584 : | jng .loop | ||
585 : | nop | ||
586 : | |||
587 : | ; mismatch control | ||
588 : | |||
589 : | pshufw mm0, mm6, 01010101b | ||
590 : | pshufw mm1, mm6, 10101010b | ||
591 : | pshufw mm2, mm6, 11111111b | ||
592 : | pxor mm6, mm0 | ||
593 : | pxor mm1, mm2 | ||
594 : | pxor mm6, mm1 | ||
595 : | movd eax, mm6 | ||
596 : | pop edi | ||
597 : | and eax, byte 1 | ||
598 : | xor eax, byte 1 | ||
599 : | mov esi, [esp] | ||
600 : | add esp, byte 4 | ||
601 : | xor word [edx + 2*63], ax | ||
602 : | |||
603 : | xor eax, eax | ||
604 : | ret | ||
605 : | edgomez | 1540 | .endfunc |
606 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |