Parent Directory | Revision Log
Revision 1.2 - (view) (download)
1 : | edgomez | 1.2 | ;/************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - 3dne Quantization/Dequantization - | ||
5 : | ; * | ||
6 : | ; * Copyright(C) 2002-2003 Jaan Kalda | ||
7 : | ; * | ||
8 : | ; * This program is free software ; you can redistribute it and/or modify | ||
9 : | ; * it under the terms of the GNU General Public License as published by | ||
10 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
11 : | ; * (at your option) any later version. | ||
12 : | ; * | ||
13 : | ; * This program is distributed in the hope that it will be useful, | ||
14 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
15 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
16 : | ; * GNU General Public License for more details. | ||
17 : | ; * | ||
18 : | ; * You should have received a copy of the GNU General Public License | ||
19 : | ; * along with this program ; if not, write to the Free Software | ||
20 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
21 : | ; * | ||
22 : | ; * $Id$ | ||
23 : | ; * | ||
24 : | ; *************************************************************************/ | ||
25 : | ; | ||
26 : | ; these 3dne functions are compatible with iSSE, but are optimized specifically for | ||
27 : | ; K7 pipelines | ||
28 : | |||
29 : | ; enable dequant saturate [-2048,2047], test purposes only. | ||
30 : | %define SATURATE | ||
31 : | |||
32 : | BITS 32 | ||
33 : | |||
34 : | %macro cglobal 1 | ||
35 : | %ifdef PREFIX | ||
36 : | global _%1 | ||
37 : | %define %1 _%1 | ||
38 : | %else | ||
39 : | global %1 | ||
40 : | %endif | ||
41 : | %endmacro | ||
42 : | |||
43 : | ;============================================================================= | ||
44 : | ; Local data | ||
45 : | ;============================================================================= | ||
46 : | |||
47 : | %ifdef FORMAT_COFF | ||
48 : | SECTION .rodata data | ||
49 : | %else | ||
50 : | SECTION .rodata data align=16 | ||
51 : | %endif | ||
52 : | |||
53 : | align 4 | ||
54 : | int_div: | ||
55 : | dd 0 | ||
56 : | %assign i 1 | ||
57 : | %rep 255 | ||
58 : | dd (1 << 16) / (i) + 1 | ||
59 : | %assign i i+1 | ||
60 : | %endrep | ||
61 : | |||
62 : | ALIGN 16 | ||
63 : | plus_one: | ||
64 : | times 8 dw 1 | ||
65 : | |||
66 : | ;----------------------------------------------------------------------------- | ||
67 : | ; subtract by Q/2 table | ||
68 : | ;----------------------------------------------------------------------------- | ||
69 : | |||
70 : | ALIGN 16 | ||
71 : | mmx_sub: | ||
72 : | %assign i 1 | ||
73 : | %rep 31 | ||
74 : | times 4 dw i / 2 | ||
75 : | %assign i i+1 | ||
76 : | %endrep | ||
77 : | |||
78 : | |||
79 : | ;----------------------------------------------------------------------------- | ||
80 : | ; | ||
81 : | ; divide by 2Q table | ||
82 : | ; | ||
83 : | ; use a shift of 16 to take full advantage of _pmulhw_ | ||
84 : | ; for q=1, _pmulhw_ will overflow so it is treated seperately | ||
85 : | ; (3dnow2 provides _pmulhuw_ which wont cause overflow) | ||
86 : | ; | ||
87 : | ;----------------------------------------------------------------------------- | ||
88 : | |||
89 : | ALIGN 16 | ||
90 : | mmx_div: | ||
91 : | %assign i 1 | ||
92 : | %rep 31 | ||
93 : | times 4 dw (1 << 16) / (i * 2) + 1 | ||
94 : | %assign i i+1 | ||
95 : | %endrep | ||
96 : | |||
97 : | ;----------------------------------------------------------------------------- | ||
98 : | ; add by (odd(Q) ? Q : Q - 1) table | ||
99 : | ;----------------------------------------------------------------------------- | ||
100 : | |||
101 : | ALIGN 16 | ||
102 : | mmx_add: | ||
103 : | %assign i 1 | ||
104 : | %rep 31 | ||
105 : | %if i % 2 != 0 | ||
106 : | times 4 dw i | ||
107 : | %else | ||
108 : | times 4 dw i - 1 | ||
109 : | %endif | ||
110 : | %assign i i+1 | ||
111 : | %endrep | ||
112 : | |||
113 : | ;----------------------------------------------------------------------------- | ||
114 : | ; multiple by 2Q table | ||
115 : | ;----------------------------------------------------------------------------- | ||
116 : | |||
117 : | ALIGN 16 | ||
118 : | mmx_mul: | ||
119 : | %assign i 1 | ||
120 : | %rep 31 | ||
121 : | times 4 dw i * 2 | ||
122 : | %assign i i+1 | ||
123 : | %endrep | ||
124 : | |||
125 : | ;----------------------------------------------------------------------------- | ||
126 : | ; saturation limits | ||
127 : | ;----------------------------------------------------------------------------- | ||
128 : | |||
129 : | ALIGN 8 | ||
130 : | mmx_32768_minus_2048: | ||
131 : | times 4 dw (32768-2048) | ||
132 : | mmx_32767_minus_2047: | ||
133 : | times 4 dw (32767-2047) | ||
134 : | |||
135 : | ALIGN 16 | ||
136 : | mmx_2047: | ||
137 : | times 4 dw 2047 | ||
138 : | |||
139 : | ALIGN 8 | ||
140 : | mmzero: | ||
141 : | dd 0, 0 | ||
142 : | int2047: | ||
143 : | dd 2047 | ||
144 : | int_2048: | ||
145 : | dd -2048 | ||
146 : | |||
147 : | ;============================================================================= | ||
148 : | ; Code | ||
149 : | ;============================================================================= | ||
150 : | |||
151 : | SECTION .text | ||
152 : | |||
153 : | |||
154 : | ;----------------------------------------------------------------------------- | ||
155 : | ; | ||
156 : | ; uint32_t quant_h263_intra_3dne(int16_t * coeff, | ||
157 : | ; const int16_t const * data, | ||
158 : | ; const uint32_t quant, | ||
159 : | ; const uint32_t dcscalar, | ||
160 : | ; const uint16_t *mpeg_matrices); | ||
161 : | ; | ||
162 : | ;----------------------------------------------------------------------------- | ||
163 : | ;This is Athlon-optimized code (ca 70 clk per call) | ||
164 : | |||
165 : | %macro quant_intra1 1 | ||
166 : | psubw mm1, mm0 ;A3 | ||
167 : | psubw mm3, mm2 ;B3 | ||
168 : | %if (%1) | ||
169 : | psubw mm5, mm4 ;C8 | ||
170 : | psubw mm7, mm6 ;D8 | ||
171 : | %endif | ||
172 : | |||
173 : | ALIGN 8 | ||
174 : | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 | ||
175 : | pmaxsw mm1, mm0 ;A4 | ||
176 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
177 : | pmaxsw mm3, mm2 ;B4 | ||
178 : | |||
179 : | |||
180 : | psraw mm0, 15 ;A5 | ||
181 : | psraw mm2, 15 ;B5 | ||
182 : | %if (%1) | ||
183 : | movq [edx + %1 * 32 + 16-32], mm5 ;C9 | ||
184 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
185 : | %endif | ||
186 : | |||
187 : | psrlw mm1, 1 ;A6 | ||
188 : | psrlw mm3, 1 ;B6 | ||
189 : | movq mm5, [ebx] ;C2 | ||
190 : | movq mm7, [ebx] ;D2 | ||
191 : | |||
192 : | pxor mm1, mm0 ;A7 | ||
193 : | pxor mm3, mm2 ;B7 | ||
194 : | |||
195 : | psubw mm5, mm4 ;C3 | ||
196 : | psubw mm7, mm6 ;D3 | ||
197 : | psubw mm1, mm0 ;A8 | ||
198 : | psubw mm3, mm2 ;B8 | ||
199 : | |||
200 : | %if (%1 == 0) | ||
201 : | push ebp | ||
202 : | movq mm0, [ecx + %1 * 32 +32] | ||
203 : | %elif (%1 < 3) | ||
204 : | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 | ||
205 : | %endif | ||
206 : | pmaxsw mm5, mm4 ;C4 | ||
207 : | %if (%1 < 3) | ||
208 : | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 | ||
209 : | %else | ||
210 : | cmp esp, esp | ||
211 : | %endif | ||
212 : | pmaxsw mm7, mm6 ;D4 | ||
213 : | |||
214 : | psraw mm4, 15 ;C5 | ||
215 : | psraw mm6, 15 ;D5 | ||
216 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
217 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
218 : | |||
219 : | |||
220 : | psrlw mm5, 1 ;C6 | ||
221 : | psrlw mm7, 1 ;D6 | ||
222 : | %if (%1 < 3) | ||
223 : | movq mm1, [ebx] ;A2 | ||
224 : | movq mm3, [ebx] ;B2 | ||
225 : | %endif | ||
226 : | %if (%1 == 3) | ||
227 : | imul eax, [int_div+4*edi] | ||
228 : | %endif | ||
229 : | pxor mm5, mm4 ;C7 | ||
230 : | pxor mm7, mm6 ;D7 | ||
231 : | %endm | ||
232 : | |||
233 : | |||
234 : | %macro quant_intra 1 | ||
235 : | ; Rules for athlon: | ||
236 : | ; 1) schedule latencies | ||
237 : | ; 2) add/mul and load/store in 2:1 proportion | ||
238 : | ; 3) avoid spliting >3byte instructions over 8byte boundaries | ||
239 : | |||
240 : | psubw mm1, mm0 ;A3 | ||
241 : | psubw mm3, mm2 ;B3 | ||
242 : | %if (%1) | ||
243 : | psubw mm5, mm4 ;C8 | ||
244 : | psubw mm7, mm6 ;D8 | ||
245 : | %endif | ||
246 : | |||
247 : | ALIGN 8 | ||
248 : | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 | ||
249 : | pmaxsw mm1, mm0 ;A4 | ||
250 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
251 : | pmaxsw mm3, mm2 ;B4 | ||
252 : | |||
253 : | |||
254 : | psraw mm0, 15 ;A5 | ||
255 : | psraw mm2, 15 ;B5 | ||
256 : | %if (%1) | ||
257 : | movq [edx + %1 * 32 + 16-32], mm5 ;C9 | ||
258 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
259 : | %endif | ||
260 : | |||
261 : | pmulhw mm1, [esi] ;A6 | ||
262 : | pmulhw mm3, [esi] ;B6 | ||
263 : | movq mm5, [ebx] ;C2 | ||
264 : | movq mm7, [ebx] ;D2 | ||
265 : | |||
266 : | nop | ||
267 : | nop | ||
268 : | pxor mm1, mm0 ;A7 | ||
269 : | pxor mm3, mm2 ;B7 | ||
270 : | |||
271 : | psubw mm5, mm4 ;C3 | ||
272 : | psubw mm7, mm6 ;D3 | ||
273 : | psubw mm1, mm0 ;A8 | ||
274 : | psubw mm3, mm2 ;B8 | ||
275 : | |||
276 : | |||
277 : | %if (%1 < 3) | ||
278 : | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 | ||
279 : | %endif | ||
280 : | pmaxsw mm5, mm4 ;C4 | ||
281 : | %if (%1 < 3) | ||
282 : | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 | ||
283 : | %else | ||
284 : | cmp esp, esp | ||
285 : | %endif | ||
286 : | pmaxsw mm7,mm6 ;D4 | ||
287 : | |||
288 : | psraw mm4, 15 ;C5 | ||
289 : | psraw mm6, 15 ;D5 | ||
290 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
291 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
292 : | |||
293 : | |||
294 : | pmulhw mm5, [esi] ;C6 | ||
295 : | pmulhw mm7, [esi] ;D6 | ||
296 : | %if (%1 < 3) | ||
297 : | movq mm1, [ebx] ;A2 | ||
298 : | movq mm3, [ebx] ;B2 | ||
299 : | %endif | ||
300 : | %if (%1 == 0) | ||
301 : | push ebp | ||
302 : | %elif (%1 < 3) | ||
303 : | nop | ||
304 : | %endif | ||
305 : | nop | ||
306 : | %if (%1 == 3) | ||
307 : | imul eax, [int_div+4*edi] | ||
308 : | %endif | ||
309 : | pxor mm5, mm4 ;C7 | ||
310 : | pxor mm7, mm6 ;D7 | ||
311 : | %endmacro | ||
312 : | |||
313 : | |||
314 : | ALIGN 16 | ||
315 : | cglobal quant_h263_intra_3dne | ||
316 : | quant_h263_intra_3dne: | ||
317 : | |||
318 : | mov eax, [esp + 12] ; quant | ||
319 : | mov ecx, [esp + 8] ; data | ||
320 : | mov edx, [esp + 4] ; coeff | ||
321 : | cmp al, 1 | ||
322 : | pxor mm1, mm1 | ||
323 : | pxor mm3, mm3 | ||
324 : | movq mm0, [ecx] ; mm0 = [1st] | ||
325 : | movq mm2, [ecx + 8] | ||
326 : | push esi | ||
327 : | lea esi, [mmx_div + eax*8 - 8] | ||
328 : | |||
329 : | push ebx | ||
330 : | mov ebx, mmzero | ||
331 : | push edi | ||
332 : | jz near .q1loop | ||
333 : | |||
334 : | quant_intra 0 | ||
335 : | mov ebp, [esp + 16 + 16] ; dcscalar | ||
336 : | ; NB -- there are 3 pushes in the function preambule and one more | ||
337 : | ; in "quant_intra 0", thus an added offset of 16 bytes | ||
338 : | movsx eax, word [byte ecx] ; DC | ||
339 : | |||
340 : | quant_intra 1 | ||
341 : | mov edi, eax | ||
342 : | sar edi, 31 ; sign(DC) | ||
343 : | shr ebp, byte 1 ; ebp = dcscalar/2 | ||
344 : | |||
345 : | quant_intra 2 | ||
346 : | sub eax, edi ; DC (+1) | ||
347 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
348 : | mov edi, [esp + 16 + 16] ; dscalar | ||
349 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar/2 | ||
350 : | mov ebp, [byte esp] | ||
351 : | |||
352 : | quant_intra 3 | ||
353 : | psubw mm5, mm4 ;C8 | ||
354 : | mov esi, [esp + 12] ; pop back the register value | ||
355 : | mov edi, [esp + 4] ; pop back the register value | ||
356 : | sar eax, 16 | ||
357 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
358 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
359 : | mov [edx], ax ; coeff[0] = ax | ||
360 : | mov ebx, [esp + 8] ; pop back the register value | ||
361 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
362 : | psubw mm7, mm6 ;D8 | ||
363 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
364 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
365 : | |||
366 : | xor eax, eax | ||
367 : | ret | ||
368 : | |||
369 : | ALIGN 16 | ||
370 : | |||
371 : | .q1loop | ||
372 : | quant_intra1 0 | ||
373 : | mov ebp, [esp + 16 + 16] ; dcscalar | ||
374 : | movsx eax, word [byte ecx] ; DC | ||
375 : | |||
376 : | quant_intra1 1 | ||
377 : | mov edi, eax | ||
378 : | sar edi, 31 ; sign(DC) | ||
379 : | shr ebp, byte 1 ; ebp = dcscalar /2 | ||
380 : | |||
381 : | quant_intra1 2 | ||
382 : | sub eax, edi ; DC (+1) | ||
383 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
384 : | mov edi, [esp + 16 + 16] ; dcscalar | ||
385 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar /2 | ||
386 : | mov ebp, [byte esp] | ||
387 : | |||
388 : | quant_intra1 3 | ||
389 : | psubw mm5, mm4 ;C8 | ||
390 : | mov esi, [dword esp + 12] ; pop back the register value | ||
391 : | mov edi, [esp + 4] ; pop back the register value | ||
392 : | sar eax, 16 | ||
393 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
394 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
395 : | mov [edx], ax ; coeff[0] = ax | ||
396 : | mov ebx, [esp + 8] ; pop back the register value | ||
397 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
398 : | psubw mm7, mm6 ;D8 | ||
399 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
400 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
401 : | |||
402 : | xor eax, eax | ||
403 : | ret | ||
404 : | |||
405 : | |||
406 : | |||
407 : | |||
408 : | ;----------------------------------------------------------------------------- | ||
409 : | ; | ||
410 : | ; uint32_t quant_h263_inter_3dne(int16_t * coeff, | ||
411 : | ; const int16_t const * data, | ||
412 : | ; const uint32_t quant, | ||
413 : | ; const uint16_t *mpeg_matrices); | ||
414 : | ; | ||
415 : | ;----------------------------------------------------------------------------- | ||
416 : | ;This is Athlon-optimized code (ca 90 clk per call) | ||
417 : | ;Optimized by Jaan, 30 Nov 2002 | ||
418 : | |||
419 : | |||
420 : | %macro quantinter 1 | ||
421 : | movq mm1, [eax] ;A2 | ||
422 : | psraw mm3, 15 ;B6 | ||
423 : | %if (%1) | ||
424 : | psubw mm2, mm6 ;C10 | ||
425 : | %endif | ||
426 : | psubw mm1, mm0 ;A3 | ||
427 : | pmulhw mm4, mm7 ;B7 | ||
428 : | movq mm6, [ecx + %1*24+16] ;C1 | ||
429 : | pmaxsw mm1, mm0 ;A4 | ||
430 : | paddw mm5, mm4 ;B8 | ||
431 : | %if (%1) | ||
432 : | movq [edx + %1*24+16-24], mm2 ;C11 | ||
433 : | %endif | ||
434 : | psubusw mm1, [ebx] ;A5 mm0 -= sub (unsigned, dont go < 0) | ||
435 : | pxor mm4, mm3 ;B9 | ||
436 : | movq mm2, [eax] ;C2 | ||
437 : | psraw mm0, 15 ;A6 | ||
438 : | psubw mm4, mm3 ;B10 | ||
439 : | psubw mm2, mm6 ;C3 | ||
440 : | pmulhw mm1, mm7 ;A7 mm0 = (mm0 / 2Q) >> 24 | ||
441 : | movq mm3, [ecx + %1*24+8] ;B1 | ||
442 : | pmaxsw mm2, mm6 ;C4 | ||
443 : | paddw mm5, mm1 ;A8 sum += mm0 | ||
444 : | %if (%1) | ||
445 : | movq [edx + %1*24+8-24], mm4 ;B11 | ||
446 : | %else | ||
447 : | movq [edx + 120], mm4 ;B11 | ||
448 : | %endif | ||
449 : | psubusw mm2, [ebx] ;C5 | ||
450 : | pxor mm1, mm0 ;A9 mm0 *= sign(mm0) | ||
451 : | movq mm4, [eax] ;B2 | ||
452 : | psraw mm6, 15 ;C6 | ||
453 : | psubw mm1, mm0 ;A10 undisplace | ||
454 : | psubw mm4, mm3 ;B3 | ||
455 : | pmulhw mm2, mm7 ;C7 | ||
456 : | movq mm0, [ecx + %1*24+24] ;A1 mm0 = [1st] | ||
457 : | pmaxsw mm4, mm3 ;B4 | ||
458 : | paddw mm5, mm2 ;C8 | ||
459 : | movq [byte edx + %1*24], mm1 ;A11 | ||
460 : | psubusw mm4, [ebx] ;B5 | ||
461 : | pxor mm2, mm6 ;C9 | ||
462 : | %endmacro | ||
463 : | |||
464 : | %macro quantinter1 1 | ||
465 : | movq mm0, [byte ecx + %1*16] ;mm0 = [1st] | ||
466 : | movq mm3, [ecx + %1*16+8] ; | ||
467 : | movq mm1, [eax] | ||
468 : | movq mm4, [eax] | ||
469 : | psubw mm1, mm0 | ||
470 : | psubw mm4, mm3 | ||
471 : | pmaxsw mm1, mm0 | ||
472 : | pmaxsw mm4, mm3 | ||
473 : | psubusw mm1, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
474 : | psubusw mm4, mm6 ; | ||
475 : | psraw mm0, 15 | ||
476 : | psraw mm3, 15 | ||
477 : | psrlw mm1, 1 ; mm0 = (mm0 / 2Q) >> 16 | ||
478 : | psrlw mm4, 1 ; | ||
479 : | paddw mm5, mm1 ; sum += mm0 | ||
480 : | pxor mm1, mm0 ; mm0 *= sign(mm0) | ||
481 : | paddw mm5, mm4 | ||
482 : | pxor mm4, mm3 ; | ||
483 : | psubw mm1, mm0 ; undisplace | ||
484 : | psubw mm4, mm3 | ||
485 : | cmp esp, esp | ||
486 : | movq [byte edx + %1*16], mm1 | ||
487 : | movq [edx + %1*16+8], mm4 | ||
488 : | %endmacro | ||
489 : | |||
490 : | ALIGN 16 | ||
491 : | cglobal quant_h263_inter_3dne | ||
492 : | quant_h263_inter_3dne: | ||
493 : | mov edx, [esp + 4] ; coeff | ||
494 : | mov ecx, [esp + 8] ; data | ||
495 : | mov eax, [esp + 12] ; quant | ||
496 : | push ebx | ||
497 : | |||
498 : | pxor mm5, mm5 ; sum | ||
499 : | nop | ||
500 : | lea ebx,[mmx_sub + eax * 8 - 8] ; sub | ||
501 : | movq mm7, [mmx_div + eax * 8 - 8] ; divider | ||
502 : | |||
503 : | cmp al, 1 | ||
504 : | lea eax, [mmzero] | ||
505 : | jz near .q1loop | ||
506 : | cmp esp, esp | ||
507 : | ALIGN 8 | ||
508 : | movq mm3, [ecx + 120] ;B1 | ||
509 : | pxor mm4, mm4 ;B2 | ||
510 : | psubw mm4, mm3 ;B3 | ||
511 : | movq mm0, [ecx] ;A1 mm0 = [1st] | ||
512 : | pmaxsw mm4, mm3 ;B4 | ||
513 : | psubusw mm4, [ebx] ;B5 | ||
514 : | |||
515 : | quantinter 0 | ||
516 : | quantinter 1 | ||
517 : | quantinter 2 | ||
518 : | quantinter 3 | ||
519 : | quantinter 4 | ||
520 : | |||
521 : | psraw mm3, 15 ;B6 | ||
522 : | psubw mm2, mm6 ;C10 | ||
523 : | pmulhw mm4, mm7 ;B7 | ||
524 : | paddw mm5, mm4 ;B8 | ||
525 : | pxor mm4, mm3 ;B9 | ||
526 : | psubw mm4, mm3 ;B10 | ||
527 : | movq [edx + 4*24+16], mm2 ;C11 | ||
528 : | pop ebx | ||
529 : | movq [edx + 4*24+8], mm4 ;B11 | ||
530 : | pmaddwd mm5, [plus_one] | ||
531 : | movq mm0, mm5 | ||
532 : | punpckhdq mm5, mm5 | ||
533 : | paddd mm0, mm5 | ||
534 : | movd eax, mm0 ; return sum | ||
535 : | |||
536 : | ret | ||
537 : | |||
538 : | ALIGN 16 | ||
539 : | .q1loop | ||
540 : | movq mm6, [byte ebx] | ||
541 : | |||
542 : | quantinter1 0 | ||
543 : | quantinter1 1 | ||
544 : | quantinter1 2 | ||
545 : | quantinter1 3 | ||
546 : | quantinter1 4 | ||
547 : | quantinter1 5 | ||
548 : | quantinter1 6 | ||
549 : | quantinter1 7 | ||
550 : | |||
551 : | pmaddwd mm5, [plus_one] | ||
552 : | movq mm0, mm5 | ||
553 : | psrlq mm5, 32 | ||
554 : | paddd mm0, mm5 | ||
555 : | movd eax, mm0 ; return sum | ||
556 : | |||
557 : | pop ebx | ||
558 : | |||
559 : | ret | ||
560 : | |||
561 : | ;----------------------------------------------------------------------------- | ||
562 : | ; | ||
563 : | ; uint32_t dequant_h263_intra_3dne(int16_t *data, | ||
564 : | ; const int16_t const *coeff, | ||
565 : | ; const uint32_t quant, | ||
566 : | ; const uint32_t dcscalar, | ||
567 : | ; const uint16_t *mpeg_matrices); | ||
568 : | ; | ||
569 : | ;----------------------------------------------------------------------------- | ||
570 : | |||
571 : | ; this is the same as dequant_inter_3dne, except that we're | ||
572 : | ; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) | ||
573 : | |||
574 : | ;This is Athlon-optimized code (ca 106 clk per call) | ||
575 : | |||
576 : | %macro dequant 1 | ||
577 : | movq mm1, [ecx+%1*24] ; c = coeff[i] ;A2 | ||
578 : | psubw mm0, mm1 ;-c ;A3 (1st dep) | ||
579 : | %if (%1) | ||
580 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
581 : | %endif | ||
582 : | pmaxsw mm0, mm1 ;|c| ;A4 (2nd) | ||
583 : | %if (%1) | ||
584 : | mov ebp, ebp | ||
585 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) later | ||
586 : | %endif | ||
587 : | movq mm6, [esi] ;0 ;A5 mm6 in use | ||
588 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
589 : | %if (%1) | ||
590 : | pxor mm5, mm4 ;C13 (6th+) 1later | ||
591 : | %endif | ||
592 : | movq mm4, [esi] ;C1 ;0 | ||
593 : | mov esp, esp | ||
594 : | pcmpeqw mm6, [ecx+%1*24] ;A6 (c ==0) ? -1 : 0 (1st) | ||
595 : | ALIGN 4 | ||
596 : | psraw mm1, 15 ; sign(c) ;A7 (2nd) | ||
597 : | %if (%1) | ||
598 : | movq [edx+%1*24+16-24], mm5 ; C14 (7th) 2later | ||
599 : | %endif | ||
600 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
601 : | pmullw mm0, [edi] ;*= 2Q ;A8 (3rd+) | ||
602 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
603 : | lea ebp, [byte ebp] | ||
604 : | movq mm5, [ecx+%1*24+16] ;C2 ; c = coeff[i] | ||
605 : | psubw mm4, mm5 ;-c ;C3 (1st dep) | ||
606 : | pandn mm6, [eax] ;A9 offset = isZero ? 0 : quant_add (2nd) | ||
607 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
608 : | pxor mm3, mm2 ;B13 (6th+) | ||
609 : | movq mm2, [byte esi] ;B1 ;0 | ||
610 : | %if (%1) | ||
611 : | movq [edx+%1*24+8-24], mm3 ;B14 (7th) | ||
612 : | %else | ||
613 : | movq [edx+120], mm3 | ||
614 : | %endif | ||
615 : | pmaxsw mm4, mm5 ;|c| ;C4 (2nd) | ||
616 : | paddw mm6, mm1 ;A10 offset +negate back (3rd) | ||
617 : | movq mm3, [ecx+%1*24 + 8] ;B2 ; c = coeff[i] | ||
618 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
619 : | paddw mm0, mm6 ;A11 mm6 free (4th+) | ||
620 : | movq mm6, [byte esi] ;0 ;C5 mm6 in use | ||
621 : | pcmpeqw mm6, [ecx+%1*24+16] ;C6 (c ==0) ? -1 : 0 (1st) | ||
622 : | pminsw mm0, [ebx] ;A12 saturates to +2047 (5th+) | ||
623 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
624 : | pxor mm1, mm0 ;A13 (6th+) | ||
625 : | pmullw mm4, [edi] ;*= 2Q ;C8 (3rd+) | ||
626 : | psraw mm5, 15 ; sign(c) ;C7 (2nd) | ||
627 : | movq mm7, [byte esi] ;0 ;B5 mm7 in use | ||
628 : | pcmpeqw mm7, [ecx+%1*24 + 8] ;B6 (c ==0) ? -1 : 0 (1st) | ||
629 : | %if (%1 < 4) | ||
630 : | movq mm0, [byte esi] ;A1 ;0 | ||
631 : | %endif | ||
632 : | pandn mm6, [byte eax] ;C9 offset = isZero ? 0 : quant_add (2nd) | ||
633 : | psraw mm3, 15 ;sign(c) ;B7 (2nd) | ||
634 : | movq [byte edx+%1*24], mm1 ;A14 (7th) | ||
635 : | paddw mm6, mm5 ;C10 offset +negate back (3rd) | ||
636 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
637 : | mov esp, esp | ||
638 : | %endmacro | ||
639 : | |||
640 : | |||
641 : | ALIGN 16 | ||
642 : | cglobal dequant_h263_intra_3dne | ||
643 : | dequant_h263_intra_3dne: | ||
644 : | mov ecx, [esp+ 8] ; coeff | ||
645 : | mov eax, [esp+12] ; quant | ||
646 : | pxor mm0, mm0 | ||
647 : | pxor mm2, mm2 | ||
648 : | push edi | ||
649 : | push ebx | ||
650 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
651 : | push ebp | ||
652 : | mov ebx, mmx_2047 | ||
653 : | movsx ebp, word [ecx] | ||
654 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
655 : | push esi | ||
656 : | mov esi, mmzero | ||
657 : | pxor mm7, mm7 | ||
658 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
659 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
660 : | |||
661 : | imul ebp, [esp+16+16] ; dcscalar | ||
662 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
663 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
664 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
665 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
666 : | mov edx, [esp+ 4+16] ; data | ||
667 : | |||
668 : | ALIGN 8 | ||
669 : | dequant 0 | ||
670 : | |||
671 : | cmp ebp, -2048 | ||
672 : | mov esp, esp | ||
673 : | |||
674 : | dequant 1 | ||
675 : | |||
676 : | cmovl ebp, [int_2048] | ||
677 : | nop | ||
678 : | |||
679 : | dequant 2 | ||
680 : | |||
681 : | cmp ebp, 2047 | ||
682 : | mov esp, esp | ||
683 : | |||
684 : | dequant 3 | ||
685 : | |||
686 : | cmovg ebp, [int2047] | ||
687 : | nop | ||
688 : | |||
689 : | dequant 4 | ||
690 : | |||
691 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
692 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
693 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
694 : | mov eax, ebp | ||
695 : | mov esi, [esp] | ||
696 : | mov ebp, [esp+4] | ||
697 : | pxor mm5, mm4 ;C13 (6th+) | ||
698 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
699 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
700 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
701 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
702 : | mov ebx, [esp+8] | ||
703 : | mov edi, [esp+12] | ||
704 : | add esp, byte 16 | ||
705 : | pxor mm3, mm2 ;B13 (6th+) | ||
706 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
707 : | mov [edx], ax | ||
708 : | |||
709 : | xor eax, eax | ||
710 : | ret | ||
711 : | |||
712 : | ;----------------------------------------------------------------------------- | ||
713 : | ; | ||
714 : | ; uint32_t dequant_h263_inter_3dne(int16_t * data, | ||
715 : | ; const int16_t * const coeff, | ||
716 : | ; const uint32_t quant, | ||
717 : | ; const uint16_t *mpeg_matrices); | ||
718 : | ; | ||
719 : | ;----------------------------------------------------------------------------- | ||
720 : | |||
721 : | ; this is the same as dequant_inter_3dne, | ||
722 : | ; except that we're saturating using 'pminsw' (saves 2 cycles/loop) | ||
723 : | ; This is Athlon-optimized code (ca 100 clk per call) | ||
724 : | |||
725 : | ALIGN 16 | ||
726 : | cglobal dequant_h263_inter_3dne | ||
727 : | dequant_h263_inter_3dne: | ||
728 : | mov ecx, [esp+ 8] ; coeff | ||
729 : | mov eax, [esp+12] ; quant | ||
730 : | pxor mm0, mm0 | ||
731 : | pxor mm2, mm2 | ||
732 : | push edi | ||
733 : | push ebx | ||
734 : | push esi | ||
735 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
736 : | mov ebx, mmx_2047 | ||
737 : | pxor mm7, mm7 | ||
738 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
739 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
740 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
741 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
742 : | mov esi, mmzero | ||
743 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
744 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
745 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
746 : | mov edx, [dword esp+ 4+12] ; data | ||
747 : | |||
748 : | ALIGN 8 | ||
749 : | |||
750 : | dequant 0 | ||
751 : | dequant 1 | ||
752 : | dequant 2 | ||
753 : | dequant 3 | ||
754 : | dequant 4 | ||
755 : | |||
756 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
757 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
758 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
759 : | mov esi, [esp] | ||
760 : | pxor mm5, mm4 ;C13 (6th+) | ||
761 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
762 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
763 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
764 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
765 : | mov ebx, [esp+4] | ||
766 : | mov edi, [esp+8] | ||
767 : | add esp, byte 12 | ||
768 : | pxor mm3, mm2 ;B13 (6th+) | ||
769 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
770 : | |||
771 : | xor eax, eax | ||
772 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |