Parent Directory | Revision Log
Revision 1540 - (view) (download)
1 : | edgomez | 1382 | ;/************************************************************************** |
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 : | edgomez | 1540 | ; * $Id: quantize_h263_3dne.asm,v 1.5 2004-08-29 10:02:38 edgomez Exp $ |
23 : | edgomez | 1382 | ; * |
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 : | 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 : | ;============================================================================= | ||
53 : | ; Local data | ||
54 : | ;============================================================================= | ||
55 : | |||
56 : | %ifdef FORMAT_COFF | ||
57 : | edgomez | 1519 | SECTION .rodata |
58 : | edgomez | 1382 | %else |
59 : | edgomez | 1519 | SECTION .rodata align=16 |
60 : | edgomez | 1382 | %endif |
61 : | |||
62 : | align 4 | ||
63 : | int_div: | ||
64 : | dd 0 | ||
65 : | %assign i 1 | ||
66 : | %rep 255 | ||
67 : | dd (1 << 16) / (i) + 1 | ||
68 : | %assign i i+1 | ||
69 : | %endrep | ||
70 : | |||
71 : | ALIGN 16 | ||
72 : | plus_one: | ||
73 : | times 8 dw 1 | ||
74 : | |||
75 : | ;----------------------------------------------------------------------------- | ||
76 : | ; subtract by Q/2 table | ||
77 : | ;----------------------------------------------------------------------------- | ||
78 : | |||
79 : | ALIGN 16 | ||
80 : | mmx_sub: | ||
81 : | %assign i 1 | ||
82 : | %rep 31 | ||
83 : | times 4 dw i / 2 | ||
84 : | %assign i i+1 | ||
85 : | %endrep | ||
86 : | |||
87 : | |||
88 : | ;----------------------------------------------------------------------------- | ||
89 : | ; | ||
90 : | ; divide by 2Q table | ||
91 : | ; | ||
92 : | ; use a shift of 16 to take full advantage of _pmulhw_ | ||
93 : | ; for q=1, _pmulhw_ will overflow so it is treated seperately | ||
94 : | ; (3dnow2 provides _pmulhuw_ which wont cause overflow) | ||
95 : | ; | ||
96 : | ;----------------------------------------------------------------------------- | ||
97 : | |||
98 : | ALIGN 16 | ||
99 : | mmx_div: | ||
100 : | %assign i 1 | ||
101 : | %rep 31 | ||
102 : | times 4 dw (1 << 16) / (i * 2) + 1 | ||
103 : | %assign i i+1 | ||
104 : | %endrep | ||
105 : | |||
106 : | ;----------------------------------------------------------------------------- | ||
107 : | ; add by (odd(Q) ? Q : Q - 1) table | ||
108 : | ;----------------------------------------------------------------------------- | ||
109 : | |||
110 : | ALIGN 16 | ||
111 : | mmx_add: | ||
112 : | %assign i 1 | ||
113 : | %rep 31 | ||
114 : | %if i % 2 != 0 | ||
115 : | times 4 dw i | ||
116 : | %else | ||
117 : | times 4 dw i - 1 | ||
118 : | %endif | ||
119 : | %assign i i+1 | ||
120 : | %endrep | ||
121 : | |||
122 : | ;----------------------------------------------------------------------------- | ||
123 : | ; multiple by 2Q table | ||
124 : | ;----------------------------------------------------------------------------- | ||
125 : | |||
126 : | ALIGN 16 | ||
127 : | mmx_mul: | ||
128 : | %assign i 1 | ||
129 : | %rep 31 | ||
130 : | times 4 dw i * 2 | ||
131 : | %assign i i+1 | ||
132 : | %endrep | ||
133 : | |||
134 : | ;----------------------------------------------------------------------------- | ||
135 : | ; saturation limits | ||
136 : | ;----------------------------------------------------------------------------- | ||
137 : | |||
138 : | ALIGN 8 | ||
139 : | mmx_32768_minus_2048: | ||
140 : | times 4 dw (32768-2048) | ||
141 : | mmx_32767_minus_2047: | ||
142 : | times 4 dw (32767-2047) | ||
143 : | |||
144 : | ALIGN 16 | ||
145 : | mmx_2047: | ||
146 : | times 4 dw 2047 | ||
147 : | |||
148 : | ALIGN 8 | ||
149 : | mmzero: | ||
150 : | dd 0, 0 | ||
151 : | int2047: | ||
152 : | dd 2047 | ||
153 : | int_2048: | ||
154 : | dd -2048 | ||
155 : | |||
156 : | ;============================================================================= | ||
157 : | ; Code | ||
158 : | ;============================================================================= | ||
159 : | |||
160 : | SECTION .text | ||
161 : | |||
162 : | |||
163 : | ;----------------------------------------------------------------------------- | ||
164 : | ; | ||
165 : | ; uint32_t quant_h263_intra_3dne(int16_t * coeff, | ||
166 : | ; const int16_t const * data, | ||
167 : | ; const uint32_t quant, | ||
168 : | ; const uint32_t dcscalar, | ||
169 : | ; const uint16_t *mpeg_matrices); | ||
170 : | ; | ||
171 : | ;----------------------------------------------------------------------------- | ||
172 : | ;This is Athlon-optimized code (ca 70 clk per call) | ||
173 : | |||
174 : | %macro quant_intra1 1 | ||
175 : | psubw mm1, mm0 ;A3 | ||
176 : | psubw mm3, mm2 ;B3 | ||
177 : | %if (%1) | ||
178 : | psubw mm5, mm4 ;C8 | ||
179 : | psubw mm7, mm6 ;D8 | ||
180 : | %endif | ||
181 : | |||
182 : | ALIGN 8 | ||
183 : | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 | ||
184 : | pmaxsw mm1, mm0 ;A4 | ||
185 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
186 : | pmaxsw mm3, mm2 ;B4 | ||
187 : | |||
188 : | |||
189 : | psraw mm0, 15 ;A5 | ||
190 : | psraw mm2, 15 ;B5 | ||
191 : | %if (%1) | ||
192 : | movq [edx + %1 * 32 + 16-32], mm5 ;C9 | ||
193 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
194 : | %endif | ||
195 : | |||
196 : | psrlw mm1, 1 ;A6 | ||
197 : | psrlw mm3, 1 ;B6 | ||
198 : | movq mm5, [ebx] ;C2 | ||
199 : | movq mm7, [ebx] ;D2 | ||
200 : | |||
201 : | pxor mm1, mm0 ;A7 | ||
202 : | pxor mm3, mm2 ;B7 | ||
203 : | |||
204 : | psubw mm5, mm4 ;C3 | ||
205 : | psubw mm7, mm6 ;D3 | ||
206 : | psubw mm1, mm0 ;A8 | ||
207 : | psubw mm3, mm2 ;B8 | ||
208 : | |||
209 : | %if (%1 == 0) | ||
210 : | push ebp | ||
211 : | movq mm0, [ecx + %1 * 32 +32] | ||
212 : | %elif (%1 < 3) | ||
213 : | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 | ||
214 : | %endif | ||
215 : | pmaxsw mm5, mm4 ;C4 | ||
216 : | %if (%1 < 3) | ||
217 : | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 | ||
218 : | %else | ||
219 : | cmp esp, esp | ||
220 : | %endif | ||
221 : | pmaxsw mm7, mm6 ;D4 | ||
222 : | |||
223 : | psraw mm4, 15 ;C5 | ||
224 : | psraw mm6, 15 ;D5 | ||
225 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
226 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
227 : | |||
228 : | |||
229 : | psrlw mm5, 1 ;C6 | ||
230 : | psrlw mm7, 1 ;D6 | ||
231 : | %if (%1 < 3) | ||
232 : | movq mm1, [ebx] ;A2 | ||
233 : | movq mm3, [ebx] ;B2 | ||
234 : | %endif | ||
235 : | %if (%1 == 3) | ||
236 : | imul eax, [int_div+4*edi] | ||
237 : | %endif | ||
238 : | pxor mm5, mm4 ;C7 | ||
239 : | pxor mm7, mm6 ;D7 | ||
240 : | %endm | ||
241 : | |||
242 : | |||
243 : | %macro quant_intra 1 | ||
244 : | ; Rules for athlon: | ||
245 : | ; 1) schedule latencies | ||
246 : | ; 2) add/mul and load/store in 2:1 proportion | ||
247 : | ; 3) avoid spliting >3byte instructions over 8byte boundaries | ||
248 : | |||
249 : | psubw mm1, mm0 ;A3 | ||
250 : | psubw mm3, mm2 ;B3 | ||
251 : | %if (%1) | ||
252 : | psubw mm5, mm4 ;C8 | ||
253 : | psubw mm7, mm6 ;D8 | ||
254 : | %endif | ||
255 : | |||
256 : | ALIGN 8 | ||
257 : | db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 | ||
258 : | pmaxsw mm1, mm0 ;A4 | ||
259 : | db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 | ||
260 : | pmaxsw mm3, mm2 ;B4 | ||
261 : | |||
262 : | |||
263 : | psraw mm0, 15 ;A5 | ||
264 : | psraw mm2, 15 ;B5 | ||
265 : | %if (%1) | ||
266 : | movq [edx + %1 * 32 + 16-32], mm5 ;C9 | ||
267 : | movq [edx + %1 * 32 + 24-32], mm7 ;D9 | ||
268 : | %endif | ||
269 : | |||
270 : | pmulhw mm1, [esi] ;A6 | ||
271 : | pmulhw mm3, [esi] ;B6 | ||
272 : | movq mm5, [ebx] ;C2 | ||
273 : | movq mm7, [ebx] ;D2 | ||
274 : | |||
275 : | nop | ||
276 : | nop | ||
277 : | pxor mm1, mm0 ;A7 | ||
278 : | pxor mm3, mm2 ;B7 | ||
279 : | |||
280 : | psubw mm5, mm4 ;C3 | ||
281 : | psubw mm7, mm6 ;D3 | ||
282 : | psubw mm1, mm0 ;A8 | ||
283 : | psubw mm3, mm2 ;B8 | ||
284 : | |||
285 : | |||
286 : | %if (%1 < 3) | ||
287 : | db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 | ||
288 : | %endif | ||
289 : | pmaxsw mm5, mm4 ;C4 | ||
290 : | %if (%1 < 3) | ||
291 : | db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 | ||
292 : | %else | ||
293 : | cmp esp, esp | ||
294 : | %endif | ||
295 : | pmaxsw mm7,mm6 ;D4 | ||
296 : | |||
297 : | psraw mm4, 15 ;C5 | ||
298 : | psraw mm6, 15 ;D5 | ||
299 : | movq [byte edx + %1 * 32], mm1 ;A9 | ||
300 : | movq [edx + %1 * 32+8], mm3 ;B9 | ||
301 : | |||
302 : | |||
303 : | pmulhw mm5, [esi] ;C6 | ||
304 : | pmulhw mm7, [esi] ;D6 | ||
305 : | %if (%1 < 3) | ||
306 : | movq mm1, [ebx] ;A2 | ||
307 : | movq mm3, [ebx] ;B2 | ||
308 : | %endif | ||
309 : | %if (%1 == 0) | ||
310 : | push ebp | ||
311 : | %elif (%1 < 3) | ||
312 : | nop | ||
313 : | %endif | ||
314 : | nop | ||
315 : | %if (%1 == 3) | ||
316 : | imul eax, [int_div+4*edi] | ||
317 : | %endif | ||
318 : | pxor mm5, mm4 ;C7 | ||
319 : | pxor mm7, mm6 ;D7 | ||
320 : | %endmacro | ||
321 : | |||
322 : | |||
323 : | ALIGN 16 | ||
324 : | cglobal quant_h263_intra_3dne | ||
325 : | quant_h263_intra_3dne: | ||
326 : | |||
327 : | mov eax, [esp + 12] ; quant | ||
328 : | mov ecx, [esp + 8] ; data | ||
329 : | mov edx, [esp + 4] ; coeff | ||
330 : | cmp al, 1 | ||
331 : | pxor mm1, mm1 | ||
332 : | pxor mm3, mm3 | ||
333 : | movq mm0, [ecx] ; mm0 = [1st] | ||
334 : | movq mm2, [ecx + 8] | ||
335 : | push esi | ||
336 : | lea esi, [mmx_div + eax*8 - 8] | ||
337 : | |||
338 : | push ebx | ||
339 : | mov ebx, mmzero | ||
340 : | push edi | ||
341 : | jz near .q1loop | ||
342 : | |||
343 : | quant_intra 0 | ||
344 : | mov ebp, [esp + 16 + 16] ; dcscalar | ||
345 : | ; NB -- there are 3 pushes in the function preambule and one more | ||
346 : | ; in "quant_intra 0", thus an added offset of 16 bytes | ||
347 : | movsx eax, word [byte ecx] ; DC | ||
348 : | |||
349 : | quant_intra 1 | ||
350 : | mov edi, eax | ||
351 : | sar edi, 31 ; sign(DC) | ||
352 : | shr ebp, byte 1 ; ebp = dcscalar/2 | ||
353 : | |||
354 : | quant_intra 2 | ||
355 : | sub eax, edi ; DC (+1) | ||
356 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
357 : | mov edi, [esp + 16 + 16] ; dscalar | ||
358 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar/2 | ||
359 : | mov ebp, [byte esp] | ||
360 : | |||
361 : | quant_intra 3 | ||
362 : | psubw mm5, mm4 ;C8 | ||
363 : | mov esi, [esp + 12] ; pop back the register value | ||
364 : | mov edi, [esp + 4] ; pop back the register value | ||
365 : | sar eax, 16 | ||
366 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
367 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
368 : | mov [edx], ax ; coeff[0] = ax | ||
369 : | mov ebx, [esp + 8] ; pop back the register value | ||
370 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
371 : | psubw mm7, mm6 ;D8 | ||
372 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
373 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
374 : | |||
375 : | xor eax, eax | ||
376 : | ret | ||
377 : | |||
378 : | ALIGN 16 | ||
379 : | |||
380 : | .q1loop | ||
381 : | quant_intra1 0 | ||
382 : | mov ebp, [esp + 16 + 16] ; dcscalar | ||
383 : | movsx eax, word [byte ecx] ; DC | ||
384 : | |||
385 : | quant_intra1 1 | ||
386 : | mov edi, eax | ||
387 : | sar edi, 31 ; sign(DC) | ||
388 : | shr ebp, byte 1 ; ebp = dcscalar /2 | ||
389 : | |||
390 : | quant_intra1 2 | ||
391 : | sub eax, edi ; DC (+1) | ||
392 : | xor ebp, edi ; sign(DC) dcscalar /2 (-1) | ||
393 : | mov edi, [esp + 16 + 16] ; dcscalar | ||
394 : | lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar /2 | ||
395 : | mov ebp, [byte esp] | ||
396 : | |||
397 : | quant_intra1 3 | ||
398 : | psubw mm5, mm4 ;C8 | ||
399 : | mov esi, [dword esp + 12] ; pop back the register value | ||
400 : | mov edi, [esp + 4] ; pop back the register value | ||
401 : | sar eax, 16 | ||
402 : | lea ebx, [byte eax + 1] ; workaround for eax < 0 | ||
403 : | cmovs eax, ebx ; conditionnaly move the corrected value | ||
404 : | mov [edx], ax ; coeff[0] = ax | ||
405 : | mov ebx, [esp + 8] ; pop back the register value | ||
406 : | add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 | ||
407 : | psubw mm7, mm6 ;D8 | ||
408 : | movq [edx + 3 * 32 + 16], mm5 ;C9 | ||
409 : | movq [edx + 3 * 32 + 24], mm7 ;D9 | ||
410 : | |||
411 : | xor eax, eax | ||
412 : | ret | ||
413 : | edgomez | 1540 | .endfunc |
414 : | edgomez | 1382 | |
415 : | |||
416 : | |||
417 : | |||
418 : | ;----------------------------------------------------------------------------- | ||
419 : | ; | ||
420 : | ; uint32_t quant_h263_inter_3dne(int16_t * coeff, | ||
421 : | ; const int16_t const * data, | ||
422 : | ; const uint32_t quant, | ||
423 : | ; const uint16_t *mpeg_matrices); | ||
424 : | ; | ||
425 : | ;----------------------------------------------------------------------------- | ||
426 : | ;This is Athlon-optimized code (ca 90 clk per call) | ||
427 : | ;Optimized by Jaan, 30 Nov 2002 | ||
428 : | |||
429 : | |||
430 : | %macro quantinter 1 | ||
431 : | movq mm1, [eax] ;A2 | ||
432 : | psraw mm3, 15 ;B6 | ||
433 : | %if (%1) | ||
434 : | psubw mm2, mm6 ;C10 | ||
435 : | %endif | ||
436 : | psubw mm1, mm0 ;A3 | ||
437 : | pmulhw mm4, mm7 ;B7 | ||
438 : | movq mm6, [ecx + %1*24+16] ;C1 | ||
439 : | pmaxsw mm1, mm0 ;A4 | ||
440 : | paddw mm5, mm4 ;B8 | ||
441 : | %if (%1) | ||
442 : | movq [edx + %1*24+16-24], mm2 ;C11 | ||
443 : | %endif | ||
444 : | psubusw mm1, [ebx] ;A5 mm0 -= sub (unsigned, dont go < 0) | ||
445 : | pxor mm4, mm3 ;B9 | ||
446 : | movq mm2, [eax] ;C2 | ||
447 : | psraw mm0, 15 ;A6 | ||
448 : | psubw mm4, mm3 ;B10 | ||
449 : | psubw mm2, mm6 ;C3 | ||
450 : | pmulhw mm1, mm7 ;A7 mm0 = (mm0 / 2Q) >> 24 | ||
451 : | movq mm3, [ecx + %1*24+8] ;B1 | ||
452 : | pmaxsw mm2, mm6 ;C4 | ||
453 : | paddw mm5, mm1 ;A8 sum += mm0 | ||
454 : | %if (%1) | ||
455 : | movq [edx + %1*24+8-24], mm4 ;B11 | ||
456 : | %else | ||
457 : | movq [edx + 120], mm4 ;B11 | ||
458 : | %endif | ||
459 : | psubusw mm2, [ebx] ;C5 | ||
460 : | pxor mm1, mm0 ;A9 mm0 *= sign(mm0) | ||
461 : | movq mm4, [eax] ;B2 | ||
462 : | psraw mm6, 15 ;C6 | ||
463 : | psubw mm1, mm0 ;A10 undisplace | ||
464 : | psubw mm4, mm3 ;B3 | ||
465 : | pmulhw mm2, mm7 ;C7 | ||
466 : | movq mm0, [ecx + %1*24+24] ;A1 mm0 = [1st] | ||
467 : | pmaxsw mm4, mm3 ;B4 | ||
468 : | paddw mm5, mm2 ;C8 | ||
469 : | movq [byte edx + %1*24], mm1 ;A11 | ||
470 : | psubusw mm4, [ebx] ;B5 | ||
471 : | pxor mm2, mm6 ;C9 | ||
472 : | %endmacro | ||
473 : | |||
474 : | %macro quantinter1 1 | ||
475 : | movq mm0, [byte ecx + %1*16] ;mm0 = [1st] | ||
476 : | movq mm3, [ecx + %1*16+8] ; | ||
477 : | movq mm1, [eax] | ||
478 : | movq mm4, [eax] | ||
479 : | psubw mm1, mm0 | ||
480 : | psubw mm4, mm3 | ||
481 : | pmaxsw mm1, mm0 | ||
482 : | pmaxsw mm4, mm3 | ||
483 : | psubusw mm1, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
484 : | psubusw mm4, mm6 ; | ||
485 : | psraw mm0, 15 | ||
486 : | psraw mm3, 15 | ||
487 : | psrlw mm1, 1 ; mm0 = (mm0 / 2Q) >> 16 | ||
488 : | psrlw mm4, 1 ; | ||
489 : | paddw mm5, mm1 ; sum += mm0 | ||
490 : | pxor mm1, mm0 ; mm0 *= sign(mm0) | ||
491 : | paddw mm5, mm4 | ||
492 : | pxor mm4, mm3 ; | ||
493 : | psubw mm1, mm0 ; undisplace | ||
494 : | psubw mm4, mm3 | ||
495 : | cmp esp, esp | ||
496 : | movq [byte edx + %1*16], mm1 | ||
497 : | movq [edx + %1*16+8], mm4 | ||
498 : | %endmacro | ||
499 : | |||
500 : | ALIGN 16 | ||
501 : | cglobal quant_h263_inter_3dne | ||
502 : | quant_h263_inter_3dne: | ||
503 : | mov edx, [esp + 4] ; coeff | ||
504 : | mov ecx, [esp + 8] ; data | ||
505 : | mov eax, [esp + 12] ; quant | ||
506 : | push ebx | ||
507 : | |||
508 : | pxor mm5, mm5 ; sum | ||
509 : | nop | ||
510 : | lea ebx,[mmx_sub + eax * 8 - 8] ; sub | ||
511 : | movq mm7, [mmx_div + eax * 8 - 8] ; divider | ||
512 : | |||
513 : | cmp al, 1 | ||
514 : | lea eax, [mmzero] | ||
515 : | jz near .q1loop | ||
516 : | cmp esp, esp | ||
517 : | ALIGN 8 | ||
518 : | movq mm3, [ecx + 120] ;B1 | ||
519 : | pxor mm4, mm4 ;B2 | ||
520 : | psubw mm4, mm3 ;B3 | ||
521 : | movq mm0, [ecx] ;A1 mm0 = [1st] | ||
522 : | pmaxsw mm4, mm3 ;B4 | ||
523 : | psubusw mm4, [ebx] ;B5 | ||
524 : | |||
525 : | quantinter 0 | ||
526 : | quantinter 1 | ||
527 : | quantinter 2 | ||
528 : | quantinter 3 | ||
529 : | quantinter 4 | ||
530 : | |||
531 : | psraw mm3, 15 ;B6 | ||
532 : | psubw mm2, mm6 ;C10 | ||
533 : | pmulhw mm4, mm7 ;B7 | ||
534 : | paddw mm5, mm4 ;B8 | ||
535 : | pxor mm4, mm3 ;B9 | ||
536 : | psubw mm4, mm3 ;B10 | ||
537 : | movq [edx + 4*24+16], mm2 ;C11 | ||
538 : | pop ebx | ||
539 : | movq [edx + 4*24+8], mm4 ;B11 | ||
540 : | pmaddwd mm5, [plus_one] | ||
541 : | movq mm0, mm5 | ||
542 : | punpckhdq mm5, mm5 | ||
543 : | paddd mm0, mm5 | ||
544 : | movd eax, mm0 ; return sum | ||
545 : | |||
546 : | ret | ||
547 : | |||
548 : | ALIGN 16 | ||
549 : | .q1loop | ||
550 : | movq mm6, [byte ebx] | ||
551 : | |||
552 : | quantinter1 0 | ||
553 : | quantinter1 1 | ||
554 : | quantinter1 2 | ||
555 : | quantinter1 3 | ||
556 : | quantinter1 4 | ||
557 : | quantinter1 5 | ||
558 : | quantinter1 6 | ||
559 : | quantinter1 7 | ||
560 : | |||
561 : | pmaddwd mm5, [plus_one] | ||
562 : | movq mm0, mm5 | ||
563 : | psrlq mm5, 32 | ||
564 : | paddd mm0, mm5 | ||
565 : | movd eax, mm0 ; return sum | ||
566 : | |||
567 : | pop ebx | ||
568 : | |||
569 : | ret | ||
570 : | edgomez | 1540 | .endfunc |
571 : | edgomez | 1382 | |
572 : | ;----------------------------------------------------------------------------- | ||
573 : | ; | ||
574 : | ; uint32_t dequant_h263_intra_3dne(int16_t *data, | ||
575 : | ; const int16_t const *coeff, | ||
576 : | ; const uint32_t quant, | ||
577 : | ; const uint32_t dcscalar, | ||
578 : | ; const uint16_t *mpeg_matrices); | ||
579 : | ; | ||
580 : | ;----------------------------------------------------------------------------- | ||
581 : | |||
582 : | ; this is the same as dequant_inter_3dne, except that we're | ||
583 : | ; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) | ||
584 : | |||
585 : | ;This is Athlon-optimized code (ca 106 clk per call) | ||
586 : | |||
587 : | %macro dequant 1 | ||
588 : | movq mm1, [ecx+%1*24] ; c = coeff[i] ;A2 | ||
589 : | psubw mm0, mm1 ;-c ;A3 (1st dep) | ||
590 : | %if (%1) | ||
591 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
592 : | %endif | ||
593 : | pmaxsw mm0, mm1 ;|c| ;A4 (2nd) | ||
594 : | %if (%1) | ||
595 : | mov ebp, ebp | ||
596 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) later | ||
597 : | %endif | ||
598 : | movq mm6, [esi] ;0 ;A5 mm6 in use | ||
599 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
600 : | %if (%1) | ||
601 : | pxor mm5, mm4 ;C13 (6th+) 1later | ||
602 : | %endif | ||
603 : | movq mm4, [esi] ;C1 ;0 | ||
604 : | mov esp, esp | ||
605 : | pcmpeqw mm6, [ecx+%1*24] ;A6 (c ==0) ? -1 : 0 (1st) | ||
606 : | ALIGN 4 | ||
607 : | psraw mm1, 15 ; sign(c) ;A7 (2nd) | ||
608 : | %if (%1) | ||
609 : | movq [edx+%1*24+16-24], mm5 ; C14 (7th) 2later | ||
610 : | %endif | ||
611 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
612 : | pmullw mm0, [edi] ;*= 2Q ;A8 (3rd+) | ||
613 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
614 : | lea ebp, [byte ebp] | ||
615 : | movq mm5, [ecx+%1*24+16] ;C2 ; c = coeff[i] | ||
616 : | psubw mm4, mm5 ;-c ;C3 (1st dep) | ||
617 : | pandn mm6, [eax] ;A9 offset = isZero ? 0 : quant_add (2nd) | ||
618 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
619 : | pxor mm3, mm2 ;B13 (6th+) | ||
620 : | movq mm2, [byte esi] ;B1 ;0 | ||
621 : | %if (%1) | ||
622 : | movq [edx+%1*24+8-24], mm3 ;B14 (7th) | ||
623 : | %else | ||
624 : | movq [edx+120], mm3 | ||
625 : | %endif | ||
626 : | pmaxsw mm4, mm5 ;|c| ;C4 (2nd) | ||
627 : | paddw mm6, mm1 ;A10 offset +negate back (3rd) | ||
628 : | movq mm3, [ecx+%1*24 + 8] ;B2 ; c = coeff[i] | ||
629 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
630 : | paddw mm0, mm6 ;A11 mm6 free (4th+) | ||
631 : | movq mm6, [byte esi] ;0 ;C5 mm6 in use | ||
632 : | pcmpeqw mm6, [ecx+%1*24+16] ;C6 (c ==0) ? -1 : 0 (1st) | ||
633 : | pminsw mm0, [ebx] ;A12 saturates to +2047 (5th+) | ||
634 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
635 : | pxor mm1, mm0 ;A13 (6th+) | ||
636 : | pmullw mm4, [edi] ;*= 2Q ;C8 (3rd+) | ||
637 : | psraw mm5, 15 ; sign(c) ;C7 (2nd) | ||
638 : | movq mm7, [byte esi] ;0 ;B5 mm7 in use | ||
639 : | pcmpeqw mm7, [ecx+%1*24 + 8] ;B6 (c ==0) ? -1 : 0 (1st) | ||
640 : | %if (%1 < 4) | ||
641 : | movq mm0, [byte esi] ;A1 ;0 | ||
642 : | %endif | ||
643 : | pandn mm6, [byte eax] ;C9 offset = isZero ? 0 : quant_add (2nd) | ||
644 : | psraw mm3, 15 ;sign(c) ;B7 (2nd) | ||
645 : | movq [byte edx+%1*24], mm1 ;A14 (7th) | ||
646 : | paddw mm6, mm5 ;C10 offset +negate back (3rd) | ||
647 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
648 : | mov esp, esp | ||
649 : | %endmacro | ||
650 : | |||
651 : | |||
652 : | ALIGN 16 | ||
653 : | cglobal dequant_h263_intra_3dne | ||
654 : | dequant_h263_intra_3dne: | ||
655 : | mov ecx, [esp+ 8] ; coeff | ||
656 : | mov eax, [esp+12] ; quant | ||
657 : | pxor mm0, mm0 | ||
658 : | pxor mm2, mm2 | ||
659 : | push edi | ||
660 : | push ebx | ||
661 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
662 : | push ebp | ||
663 : | mov ebx, mmx_2047 | ||
664 : | movsx ebp, word [ecx] | ||
665 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
666 : | push esi | ||
667 : | mov esi, mmzero | ||
668 : | pxor mm7, mm7 | ||
669 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
670 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
671 : | |||
672 : | imul ebp, [esp+16+16] ; dcscalar | ||
673 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
674 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
675 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
676 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
677 : | mov edx, [esp+ 4+16] ; data | ||
678 : | |||
679 : | ALIGN 8 | ||
680 : | dequant 0 | ||
681 : | |||
682 : | cmp ebp, -2048 | ||
683 : | mov esp, esp | ||
684 : | |||
685 : | dequant 1 | ||
686 : | |||
687 : | cmovl ebp, [int_2048] | ||
688 : | nop | ||
689 : | |||
690 : | dequant 2 | ||
691 : | |||
692 : | cmp ebp, 2047 | ||
693 : | mov esp, esp | ||
694 : | |||
695 : | dequant 3 | ||
696 : | |||
697 : | cmovg ebp, [int2047] | ||
698 : | nop | ||
699 : | |||
700 : | dequant 4 | ||
701 : | |||
702 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
703 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
704 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
705 : | mov eax, ebp | ||
706 : | mov esi, [esp] | ||
707 : | mov ebp, [esp+4] | ||
708 : | pxor mm5, mm4 ;C13 (6th+) | ||
709 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
710 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
711 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
712 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
713 : | mov ebx, [esp+8] | ||
714 : | mov edi, [esp+12] | ||
715 : | add esp, byte 16 | ||
716 : | pxor mm3, mm2 ;B13 (6th+) | ||
717 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
718 : | mov [edx], ax | ||
719 : | |||
720 : | xor eax, eax | ||
721 : | ret | ||
722 : | edgomez | 1540 | .endfunc |
723 : | edgomez | 1382 | |
724 : | ;----------------------------------------------------------------------------- | ||
725 : | ; | ||
726 : | ; uint32_t dequant_h263_inter_3dne(int16_t * data, | ||
727 : | ; const int16_t * const coeff, | ||
728 : | ; const uint32_t quant, | ||
729 : | ; const uint16_t *mpeg_matrices); | ||
730 : | ; | ||
731 : | ;----------------------------------------------------------------------------- | ||
732 : | |||
733 : | ; this is the same as dequant_inter_3dne, | ||
734 : | ; except that we're saturating using 'pminsw' (saves 2 cycles/loop) | ||
735 : | ; This is Athlon-optimized code (ca 100 clk per call) | ||
736 : | |||
737 : | ALIGN 16 | ||
738 : | cglobal dequant_h263_inter_3dne | ||
739 : | dequant_h263_inter_3dne: | ||
740 : | mov ecx, [esp+ 8] ; coeff | ||
741 : | mov eax, [esp+12] ; quant | ||
742 : | pxor mm0, mm0 | ||
743 : | pxor mm2, mm2 | ||
744 : | push edi | ||
745 : | push ebx | ||
746 : | push esi | ||
747 : | lea edi, [mmx_mul + eax*8 - 8] ; 2*quant | ||
748 : | mov ebx, mmx_2047 | ||
749 : | pxor mm7, mm7 | ||
750 : | movq mm3, [ecx+120] ;B2 ; c = coeff[i] | ||
751 : | pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
752 : | lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 | ||
753 : | psubw mm2, mm3 ;-c ;B3 (1st dep) | ||
754 : | mov esi, mmzero | ||
755 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
756 : | pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) | ||
757 : | psraw mm3, 15 ; sign(c) ;B7 (2nd) | ||
758 : | mov edx, [dword esp+ 4+12] ; data | ||
759 : | |||
760 : | ALIGN 8 | ||
761 : | |||
762 : | dequant 0 | ||
763 : | dequant 1 | ||
764 : | dequant 2 | ||
765 : | dequant 3 | ||
766 : | dequant 4 | ||
767 : | |||
768 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
769 : | pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) | ||
770 : | pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
771 : | mov esi, [esp] | ||
772 : | pxor mm5, mm4 ;C13 (6th+) | ||
773 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
774 : | movq [edx+4*24+16], mm5 ;C14 (7th) | ||
775 : | paddw mm2, mm7 ;B11 mm7 free (4th+) | ||
776 : | pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) | ||
777 : | mov ebx, [esp+4] | ||
778 : | mov edi, [esp+8] | ||
779 : | add esp, byte 12 | ||
780 : | pxor mm3, mm2 ;B13 (6th+) | ||
781 : | movq [edx+4*24+8], mm3 ;B14 (7th) | ||
782 : | |||
783 : | xor eax, eax | ||
784 : | ret | ||
785 : | edgomez | 1540 | .endfunc |
786 : |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |