Parent Directory | Revision Log
Revision 1.1.2.4 - (view) (download)
1 : | edgomez | 1.1.2.3 | ;/**************************************************************************** |
2 : | edgomez | 1.1.2.1 | ; * |
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | edgomez | 1.1.2.2 | ; * - 3dne Quantization/Dequantization - |
5 : | edgomez | 1.1.2.1 | ; * |
6 : | edgomez | 1.1.2.2 | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> |
7 : | ; * 2002 Jaan Kalda | ||
8 : | edgomez | 1.1.2.1 | ; * |
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 : | ; * $Id$ | ||
24 : | ; * | ||
25 : | edgomez | 1.1.2.3 | ; ***************************************************************************/ |
26 : | edgomez | 1.1.2.2 | |
27 : | edgomez | 1.1.2.1 | ; _3dne functions are compatible with iSSE, but are optimized specifically |
28 : | ; for K7 pipelines | ||
29 : | |||
30 : | %define SATURATE | ||
31 : | |||
32 : | edgomez | 1.1.2.3 | BITS 32 |
33 : | edgomez | 1.1.2.1 | |
34 : | edgomez | 1.1.2.3 | %macro cglobal 1 |
35 : | edgomez | 1.1.2.1 | %ifdef PREFIX |
36 : | edgomez | 1.1.2.3 | global _%1 |
37 : | edgomez | 1.1.2.1 | %define %1 _%1 |
38 : | %else | ||
39 : | global %1 | ||
40 : | %endif | ||
41 : | %endmacro | ||
42 : | |||
43 : | edgomez | 1.1.2.3 | %macro cextern 1 |
44 : | edgomez | 1.1.2.1 | %ifdef PREFIX |
45 : | edgomez | 1.1.2.3 | extern _%1 |
46 : | edgomez | 1.1.2.1 | %define %1 _%1 |
47 : | %else | ||
48 : | extern %1 | ||
49 : | %endif | ||
50 : | %endmacro | ||
51 : | |||
52 : | edgomez | 1.1.2.3 | ;============================================================================= |
53 : | edgomez | 1.1.2.2 | ; Local data |
54 : | edgomez | 1.1.2.3 | ;============================================================================= |
55 : | edgomez | 1.1.2.2 | |
56 : | edgomez | 1.1.2.4 | %ifdef FORMAT_COFF |
57 : | SECTION .rodata data | ||
58 : | %else | ||
59 : | SECTION .rodata data align=16 | ||
60 : | %endif | ||
61 : | edgomez | 1.1.2.2 | |
62 : | edgomez | 1.1.2.3 | ALIGN 8 |
63 : | edgomez | 1.1.2.2 | mmzero: |
64 : | dd 0,0 | ||
65 : | mmx_one: | ||
66 : | times 4 dw 1 | ||
67 : | edgomez | 1.1.2.1 | |
68 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
69 : | ; divide by 2Q table | ||
70 : | ;----------------------------------------------------------------------------- | ||
71 : | edgomez | 1.1.2.1 | |
72 : | edgomez | 1.1.2.3 | ALIGN 16 |
73 : | edgomez | 1.1.2.2 | mmx_divs: ;i>2 |
74 : | edgomez | 1.1.2.1 | %assign i 1 |
75 : | %rep 31 | ||
76 : | edgomez | 1.1.2.3 | times 4 dw ((1 << 15) / i + 1) |
77 : | edgomez | 1.1.2.1 | %assign i i+1 |
78 : | %endrep | ||
79 : | |||
80 : | edgomez | 1.1.2.3 | ALIGN 16 |
81 : | mmx_div: ;quant>2 | ||
82 : | times 4 dw 65535 ; the div by 2 formula will overflow for the case | ||
83 : | ; quant=1 but we don't care much because quant=1 | ||
84 : | ; is handled by a different piece of code that | ||
85 : | ; doesn't use this table. | ||
86 : | %assign quant 2 | ||
87 : | edgomez | 1.1.2.1 | %rep 31 |
88 : | edgomez | 1.1.2.3 | times 4 dw ((1 << 16) / quant + 1) |
89 : | %assign quant quant+1 | ||
90 : | edgomez | 1.1.2.1 | %endrep |
91 : | |||
92 : | |||
93 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
94 : | ; intra matrix | ||
95 : | ; (TODO remove dependency on external tables for instance safiness) | ||
96 : | ;----------------------------------------------------------------------------- | ||
97 : | edgomez | 1.1.2.1 | |
98 : | %macro FIXX 1 | ||
99 : | dw (1 << 16) / (%1) + 1 | ||
100 : | edgomez | 1.1.2.3 | %endmacro |
101 : | edgomez | 1.1.2.1 | |
102 : | cextern intra_matrix_fixl | ||
103 : | cextern intra_matrix_fix | ||
104 : | cextern intra_matrix1 | ||
105 : | cextern intra_matrix | ||
106 : | |||
107 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
108 : | edgomez | 1.1.2.1 | ; inter matrix |
109 : | edgomez | 1.1.2.3 | ; (TODO remove dependency on external tables for instance safiness) |
110 : | ;----------------------------------------------------------------------------- | ||
111 : | edgomez | 1.1.2.1 | |
112 : | cextern inter_matrix1 | ||
113 : | cextern inter_matrix | ||
114 : | cextern inter_matrix_fix | ||
115 : | cextern inter_matrix_fixl | ||
116 : | |||
117 : | |||
118 : | edgomez | 1.1.2.2 | %define nop4 db 08Dh, 074h, 026h,0 |
119 : | %define nop3 add esp, byte 0 | ||
120 : | %define nop2 mov esp, esp | ||
121 : | %define nop7 db 08dh, 02ch, 02dh,0,0,0,0 | ||
122 : | %define nop6 add ebp, dword 0 | ||
123 : | edgomez | 1.1.2.1 | |
124 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
125 : | ; quantd table | ||
126 : | ;----------------------------------------------------------------------------- | ||
127 : | edgomez | 1.1.2.1 | |
128 : | edgomez | 1.1.2.2 | %define VM18P 3 |
129 : | %define VM18Q 4 | ||
130 : | edgomez | 1.1.2.1 | |
131 : | edgomez | 1.1.2.3 | ALIGN 16 |
132 : | edgomez | 1.1.2.2 | quantd: |
133 : | edgomez | 1.1.2.1 | %assign i 1 |
134 : | %rep 31 | ||
135 : | times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) | ||
136 : | %assign i i+1 | ||
137 : | %endrep | ||
138 : | |||
139 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
140 : | edgomez | 1.1.2.1 | ; multiple by 2Q table |
141 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
142 : | edgomez | 1.1.2.1 | |
143 : | edgomez | 1.1.2.3 | ALIGN 16 |
144 : | mmx_mul_quant: | ||
145 : | edgomez | 1.1.2.1 | %assign i 1 |
146 : | %rep 31 | ||
147 : | times 4 dw i | ||
148 : | %assign i i+1 | ||
149 : | %endrep | ||
150 : | |||
151 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
152 : | ; saturation limits | ||
153 : | ;----------------------------------------------------------------------------- | ||
154 : | edgomez | 1.1.2.1 | |
155 : | edgomez | 1.1.2.3 | ALIGN 16 |
156 : | edgomez | 1.1.2.2 | mmx_32767_minus_2047: |
157 : | edgomez | 1.1.2.3 | times 4 dw (32767-2047) |
158 : | edgomez | 1.1.2.2 | mmx_32768_minus_2048: |
159 : | times 4 dw (32768-2048) | ||
160 : | mmx_2047: | ||
161 : | times 4 dw 2047 | ||
162 : | mmx_minus_2048: | ||
163 : | times 4 dw (-2048) | ||
164 : | zero: | ||
165 : | times 4 dw 0 | ||
166 : | edgomez | 1.1.2.1 | |
167 : | edgomez | 1.1.2.2 | int_div: |
168 : | edgomez | 1.1.2.1 | dd 0 |
169 : | %assign i 1 | ||
170 : | edgomez | 1.1.2.3 | %rep 255 |
171 : | edgomez | 1.1.2.1 | dd (1 << 17) / ( i) + 1 |
172 : | %assign i i+1 | ||
173 : | %endrep | ||
174 : | |||
175 : | edgomez | 1.1.2.3 | ;============================================================================= |
176 : | edgomez | 1.1.2.2 | ; Code |
177 : | edgomez | 1.1.2.3 | ;============================================================================= |
178 : | edgomez | 1.1.2.2 | |
179 : | edgomez | 1.1.2.3 | SECTION .text |
180 : | |||
181 : | cglobal quant_mpeg_intra_xmm | ||
182 : | cglobal quant_mpeg_inter_xmm | ||
183 : | cglobal dequant_mpeg_intra_3dne | ||
184 : | cglobal dequant_mpeg_inter_3dne | ||
185 : | edgomez | 1.1.2.1 | |
186 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
187 : | edgomez | 1.1.2.1 | ; |
188 : | edgomez | 1.1.2.3 | ; uint32_t quant_mpeg_intra_xmm(int16_t * coeff, |
189 : | edgomez | 1.1.2.2 | ; const int16_t const * data, |
190 : | ; const uint32_t quant, | ||
191 : | ; const uint32_t dcscalar); | ||
192 : | edgomez | 1.1.2.1 | ; |
193 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
194 : | edgomez | 1.1.2.1 | |
195 : | edgomez | 1.1.2.3 | ALIGN 16 |
196 : | edgomez | 1.1.2.1 | quant_mpeg_intra_xmm: |
197 : | edgomez | 1.1.2.3 | mov eax, [esp + 8] ; data |
198 : | mov ecx, [esp + 12] ; quant | ||
199 : | mov edx, [esp + 4] ; coeff | ||
200 : | push esi | ||
201 : | push edi | ||
202 : | push ebx | ||
203 : | nop | ||
204 : | mov edi, mmzero | ||
205 : | mov esi, -14 | ||
206 : | pxor mm0, mm0 | ||
207 : | pxor mm3, mm3 | ||
208 : | cmp ecx, byte 1 | ||
209 : | je near .q1loop | ||
210 : | cmp ecx, byte 19 | ||
211 : | jg near .lloop | ||
212 : | nop6 | ||
213 : | |||
214 : | ALIGN 16 | ||
215 : | edgomez | 1.1.2.1 | .loop |
216 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
217 : | psubw mm0, mm1 ;-mm1 | ||
218 : | movq mm4, [eax + 8*esi + 120] ; | ||
219 : | psubw mm3, mm4 ;-mm4 | ||
220 : | pmaxsw mm0, mm1 ;|src| | ||
221 : | pmaxsw mm3,mm4 | ||
222 : | nop2 | ||
223 : | psraw mm1, 15 ;sign src | ||
224 : | psraw mm4, 15 | ||
225 : | psllw mm0, 4 ;level << 4 ; | ||
226 : | psllw mm3, 4 | ||
227 : | paddw mm0, [intra_matrix1 + 8*esi+112] | ||
228 : | paddw mm3, [intra_matrix1 + 8*esi+120] | ||
229 : | movq mm5, [intra_matrix_fixl + 8*esi+112] | ||
230 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
231 : | pmulhuw mm5, mm0 | ||
232 : | pmulhuw mm7, mm3 | ||
233 : | mov esp, esp | ||
234 : | movq mm2, [intra_matrix + 8*esi+112] | ||
235 : | movq mm6, [intra_matrix + 8*esi+120] | ||
236 : | pmullw mm2, mm5 | ||
237 : | pmullw mm6, mm7 | ||
238 : | psubw mm0, mm2 | ||
239 : | psubw mm3, mm6 | ||
240 : | nop4 | ||
241 : | movq mm2, [quantd + ecx * 8 - 8] | ||
242 : | movq mm6, [mmx_divs + ecx * 8 - 8] | ||
243 : | paddw mm5, mm2 | ||
244 : | paddw mm7, mm2 | ||
245 : | mov esp, esp | ||
246 : | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] | ||
247 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
248 : | paddw mm5, mm0 | ||
249 : | paddw mm7, mm3 | ||
250 : | movq mm0, [edi] | ||
251 : | movq mm3, [edi] | ||
252 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
253 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
254 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
255 : | pxor mm7, mm4 ; | ||
256 : | psubw mm5, mm1 ; undisplace | ||
257 : | psubw mm7, mm4 ; | ||
258 : | movq [edx + 8*esi+112], mm5 | ||
259 : | movq [edx + 8*esi +120], mm7 | ||
260 : | add esi, byte 2 | ||
261 : | jng near .loop | ||
262 : | edgomez | 1.1.2.1 | |
263 : | .done | ||
264 : | ; calculate data[0] // (int32_t)dcscalar) | ||
265 : | edgomez | 1.1.2.3 | mov esi, [esp + 12 + 16] ; dcscalar |
266 : | movsx ecx, word [eax] | ||
267 : | mov edi, ecx | ||
268 : | mov edx, [esp + 12 + 16] | ||
269 : | shr edx, 1 ; ebx = dcscalar /2 | ||
270 : | sar edi, 31 ; cdq is vectorpath | ||
271 : | xor edx, edi ; ebx = eax V -eax -1 | ||
272 : | sub ecx, edi | ||
273 : | add ecx, edx | ||
274 : | mov edx, [dword esp + 12 + 4] | ||
275 : | mov esi, [int_div+4*esi] | ||
276 : | imul ecx, esi | ||
277 : | sar ecx, 17 | ||
278 : | lea ebx, [byte ecx + 1] | ||
279 : | cmovs ecx, ebx | ||
280 : | ; idiv cx ; ecx = edi:ecx / dcscalar | ||
281 : | |||
282 : | mov ebx, [esp] | ||
283 : | mov edi, [esp+4] | ||
284 : | mov esi, [esp+8] | ||
285 : | add esp, byte 12 | ||
286 : | mov [edx], cx ; coeff[0] = ax | ||
287 : | edgomez | 1.1.2.1 | |
288 : | edgomez | 1.1.2.3 | xor eax, eax |
289 : | ret | ||
290 : | edgomez | 1.1.2.1 | |
291 : | edgomez | 1.1.2.3 | ALIGN 16 |
292 : | edgomez | 1.1.2.1 | .q1loop |
293 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
294 : | psubw mm0, mm1 ;-mm1 | ||
295 : | movq mm4, [eax + 8*esi+120] ; | ||
296 : | psubw mm3, mm4 ;-mm4 | ||
297 : | pmaxsw mm0, mm1 ;|src| | ||
298 : | pmaxsw mm3, mm4 | ||
299 : | nop2 | ||
300 : | psraw mm1, 15 ;sign src | ||
301 : | psraw mm4, 15 | ||
302 : | psllw mm0, 4 ; level << 4 | ||
303 : | psllw mm3, 4 | ||
304 : | paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided | ||
305 : | paddw mm3, [intra_matrix1 + 8*esi+120] ;intra1 contains fix for division by 1 | ||
306 : | movq mm5, [intra_matrix_fixl + 8*esi+112] ;with rounding down | ||
307 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
308 : | pmulhuw mm5, mm0 | ||
309 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
310 : | mov esp, esp | ||
311 : | movq mm2, [intra_matrix + 8*esi+112] | ||
312 : | movq mm6, [intra_matrix + 8*esi+120] ; divs for q<=16 | ||
313 : | pmullw mm2, mm5 ;test value <= original | ||
314 : | pmullw mm6, mm7 | ||
315 : | psubw mm0, mm2 ;mismatch | ||
316 : | psubw mm3, mm6 | ||
317 : | nop4 | ||
318 : | movq mm2, [quantd + ecx * 8 - 8] | ||
319 : | paddw mm5, mm2 ;first approx with quantd | ||
320 : | paddw mm7, mm2 | ||
321 : | mov esp, esp | ||
322 : | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction | ||
323 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
324 : | paddw mm5, mm0 ;final result with quantd | ||
325 : | paddw mm7, mm3 | ||
326 : | movq mm0, [edi] | ||
327 : | movq mm3, [edi] | ||
328 : | mov esp, esp | ||
329 : | psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) | ||
330 : | psrlw mm7, 1 | ||
331 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
332 : | pxor mm7, mm4 ; | ||
333 : | psubw mm5, mm1 ; undisplace | ||
334 : | psubw mm7, mm4 ; | ||
335 : | movq [edx + 8*esi+112], mm5 | ||
336 : | movq [edx + 8*esi +120], mm7 | ||
337 : | add esi, byte 2 | ||
338 : | jng near .q1loop | ||
339 : | jmp near .done | ||
340 : | edgomez | 1.1.2.1 | |
341 : | edgomez | 1.1.2.3 | ALIGN 8 |
342 : | edgomez | 1.1.2.1 | .lloop |
343 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
344 : | psubw mm0, mm1 ;-mm1 | ||
345 : | movq mm4, [eax + 8*esi+120] | ||
346 : | psubw mm3, mm4 ;-mm4 | ||
347 : | pmaxsw mm0, mm1 ;|src| | ||
348 : | pmaxsw mm3, mm4 | ||
349 : | nop2 | ||
350 : | psraw mm1, 15 ;sign src | ||
351 : | psraw mm4, 15 | ||
352 : | psllw mm0, 4 ; level << 4 | ||
353 : | psllw mm3, 4 ; | ||
354 : | paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1 | ||
355 : | paddw mm3, [intra_matrix1 + 8*esi+120] | ||
356 : | movq mm5, [intra_matrix_fixl + 8*esi+112] | ||
357 : | movq mm7, [intra_matrix_fixl + 8*esi+120] | ||
358 : | pmulhuw mm5, mm0 | ||
359 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
360 : | mov esp, esp | ||
361 : | movq mm2, [intra_matrix + 8*esi+112] | ||
362 : | movq mm6, [intra_matrix + 8*esi+120] | ||
363 : | pmullw mm2, mm5 ;test value <= original | ||
364 : | pmullw mm6, mm7 | ||
365 : | psubw mm0, mm2 ;mismatch | ||
366 : | psubw mm3, mm6 | ||
367 : | nop4 | ||
368 : | movq mm2, [quantd + ecx * 8 - 8] | ||
369 : | movq mm6, [mmx_div + ecx * 8 - 8] ; divs for q<=16 | ||
370 : | paddw mm5, mm2 ;first approx with quantd | ||
371 : | paddw mm7, mm2 | ||
372 : | mov esp, esp | ||
373 : | pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction | ||
374 : | pmulhuw mm3, [intra_matrix_fix + 8*esi+120] | ||
375 : | paddw mm5, mm0 ;final result with quantd | ||
376 : | paddw mm7, mm3 | ||
377 : | movq mm0, [edi] | ||
378 : | movq mm3, [edi] | ||
379 : | mov esp, esp | ||
380 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
381 : | pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) | ||
382 : | psrlw mm5, 1 ; (level + quantd) / (2*quant) | ||
383 : | psrlw mm7, 1 | ||
384 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
385 : | pxor mm7, mm4 ; | ||
386 : | psubw mm5, mm1 ; undisplace | ||
387 : | psubw mm7, mm4 ; | ||
388 : | movq [edx + 8*esi+112], mm5 | ||
389 : | movq [edx + 8*esi +120], mm7 | ||
390 : | add esi,byte 2 | ||
391 : | jng near .lloop | ||
392 : | jmp near .done | ||
393 : | edgomez | 1.1.2.1 | |
394 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
395 : | edgomez | 1.1.2.1 | ; |
396 : | edgomez | 1.1.2.2 | ; uint32_t quant_mpeg_inter_xmm(int16_t * coeff, |
397 : | ; const int16_t const * data, | ||
398 : | ; const uint32_t quant); | ||
399 : | edgomez | 1.1.2.1 | ; |
400 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
401 : | edgomez | 1.1.2.1 | |
402 : | edgomez | 1.1.2.3 | ALIGN 16 |
403 : | edgomez | 1.1.2.1 | quant_mpeg_inter_xmm: |
404 : | edgomez | 1.1.2.3 | mov eax, [esp + 8] ; data |
405 : | mov ecx, [esp + 12] ; quant | ||
406 : | mov edx, [esp + 4] ; coeff | ||
407 : | push esi | ||
408 : | push edi | ||
409 : | push ebx | ||
410 : | nop | ||
411 : | mov edi, mmzero | ||
412 : | mov esi, -14 | ||
413 : | mov ebx, esp | ||
414 : | sub esp, byte 24 | ||
415 : | lea ebx, [esp+8] | ||
416 : | and ebx, byte -8 ;ALIGN 8 | ||
417 : | pxor mm0, mm0 | ||
418 : | pxor mm3, mm3 | ||
419 : | movq [byte ebx],mm0 | ||
420 : | db 0Fh, 7Fh, 44h, 23h, 8 ;movq [ebx+8],mm0 | ||
421 : | cmp ecx, byte 1 | ||
422 : | je near .q1loop | ||
423 : | cmp ecx, byte 19 | ||
424 : | jg near .lloop | ||
425 : | nop | ||
426 : | |||
427 : | ALIGN 16 | ||
428 : | edgomez | 1.1.2.1 | .loop |
429 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
430 : | psubw mm0, mm1 ;-mm1 | ||
431 : | movq mm4, [eax + 8*esi + 120] ; | ||
432 : | psubw mm3, mm4 ;-mm4 | ||
433 : | pmaxsw mm0, mm1 ;|src| | ||
434 : | pmaxsw mm3, mm4 | ||
435 : | nop2 | ||
436 : | psraw mm1, 15 ;sign src | ||
437 : | psraw mm4, 15 | ||
438 : | psllw mm0, 4 ; level << 4 | ||
439 : | psllw mm3, 4 ; | ||
440 : | paddw mm0, [inter_matrix1 + 8*esi+112] | ||
441 : | paddw mm3, [inter_matrix1 + 8*esi+120] | ||
442 : | movq mm5, [inter_matrix_fixl + 8*esi+112] | ||
443 : | movq mm7, [inter_matrix_fixl + 8*esi+120] | ||
444 : | pmulhuw mm5, mm0 | ||
445 : | pmulhuw mm7, mm3 | ||
446 : | mov esp, esp | ||
447 : | movq mm2, [inter_matrix + 8*esi+112] | ||
448 : | movq mm6, [inter_matrix + 8*esi+120] | ||
449 : | pmullw mm2, mm5 | ||
450 : | pmullw mm6, mm7 | ||
451 : | psubw mm0, mm2 | ||
452 : | psubw mm3, mm6 | ||
453 : | movq mm2, [byte ebx] | ||
454 : | movq mm6, [mmx_divs + ecx * 8 - 8] | ||
455 : | pmulhuw mm0, [inter_matrix_fix + 8*esi+112] | ||
456 : | pmulhuw mm3, [inter_matrix_fix + 8*esi+120] | ||
457 : | paddw mm2, [ebx+8] ;sum | ||
458 : | paddw mm5, mm0 | ||
459 : | paddw mm7, mm3 | ||
460 : | movq mm0, [edi] | ||
461 : | movq mm3, [edi] | ||
462 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
463 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
464 : | add esi, byte 2 | ||
465 : | paddw mm2, mm5 ;sum += x1 | ||
466 : | movq [ebx], mm7 ;store x2 | ||
467 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
468 : | pxor mm7, mm4 ; | ||
469 : | psubw mm5, mm1 ; undisplace | ||
470 : | psubw mm7, mm4 ; | ||
471 : | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
472 : | movq [edx + 8*esi+112-16], mm5 | ||
473 : | movq [edx + 8*esi +120-16], mm7 | ||
474 : | jng near .loop | ||
475 : | edgomez | 1.1.2.1 | |
476 : | .done | ||
477 : | ; calculate data[0] // (int32_t)dcscalar) | ||
478 : | edgomez | 1.1.2.3 | paddw mm2, [ebx] |
479 : | mov ebx, [esp+24] | ||
480 : | mov edi, [esp+4+24] | ||
481 : | mov esi, [esp+8+24] | ||
482 : | add esp, byte 12+24 | ||
483 : | pmaddwd mm2, [mmx_one] | ||
484 : | punpckldq mm0, mm2 ;get low dw to mm0:high | ||
485 : | paddd mm0,mm2 | ||
486 : | punpckhdq mm0, mm0 ;get result to low | ||
487 : | movd eax, mm0 | ||
488 : | edgomez | 1.1.2.1 | |
489 : | edgomez | 1.1.2.3 | ret |
490 : | edgomez | 1.1.2.1 | |
491 : | edgomez | 1.1.2.3 | ALIGN 16 |
492 : | edgomez | 1.1.2.1 | .q1loop |
493 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
494 : | psubw mm0, mm1 ;-mm1 | ||
495 : | movq mm4, [eax + 8*esi+120] | ||
496 : | psubw mm3, mm4 ;-mm4 | ||
497 : | pmaxsw mm0, mm1 ;|src| | ||
498 : | pmaxsw mm3, mm4 | ||
499 : | nop2 | ||
500 : | psraw mm1, 15 ; sign src | ||
501 : | psraw mm4, 15 | ||
502 : | psllw mm0, 4 ; level << 4 | ||
503 : | psllw mm3, 4 | ||
504 : | paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided | ||
505 : | paddw mm3, [inter_matrix1 + 8*esi+120] ; inter1 contains fix for division by 1 | ||
506 : | movq mm5, [inter_matrix_fixl + 8*esi+112] ;with rounding down | ||
507 : | movq mm7, [inter_matrix_fixl + 8*esi+120] | ||
508 : | pmulhuw mm5, mm0 | ||
509 : | pmulhuw mm7, mm3 ;mm7: first approx of division | ||
510 : | mov esp, esp | ||
511 : | movq mm2, [inter_matrix + 8*esi+112] | ||
512 : | movq mm6, [inter_matrix + 8*esi+120] ; divs for q<=16 | ||
513 : | pmullw mm2, mm5 ;test value <= original | ||
514 : | pmullw mm6, mm7 | ||
515 : | psubw mm0, mm2 ;mismatch | ||
516 : | psubw mm3, mm6 | ||
517 : | movq mm2, [byte ebx] | ||
518 : | pmulhuw mm0, [inter_matrix_fix + 8*esi+112] ;correction | ||
519 : | pmulhuw mm3, [inter_matrix_fix + 8*esi+120] | ||
520 : | paddw mm2, [ebx+8] ;sum | ||
521 : | paddw mm5, mm0 ;final result | ||
522 : | paddw mm7, mm3 | ||
523 : | movq mm0, [edi] | ||
524 : | movq mm3, [edi] | ||
525 : | psrlw mm5, 1 ; (level ) /2 (quant = 1) | ||
526 : | psrlw mm7, 1 | ||
527 : | add esi, byte 2 | ||
528 : | paddw mm2, mm5 ;sum += x1 | ||
529 : | movq [ebx], mm7 ;store x2 | ||
530 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
531 : | pxor mm7, mm4 ; | ||
532 : | psubw mm5, mm1 ; undisplace | ||
533 : | psubw mm7, mm4 ; | ||
534 : | movq [ebx+8], mm2 ;store sum | ||
535 : | movq [edx + 8*esi+112-16], mm5 | ||
536 : | movq [edx + 8*esi +120-16], mm7 | ||
537 : | jng near .q1loop | ||
538 : | jmp near .done | ||
539 : | edgomez | 1.1.2.1 | |
540 : | edgomez | 1.1.2.3 | ALIGN 8 |
541 : | edgomez | 1.1.2.1 | .lloop |
542 : | edgomez | 1.1.2.3 | movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
543 : | psubw mm0,mm1 ;-mm1 | ||
544 : | movq mm4, [eax + 8*esi+120] | ||
545 : | psubw mm3,mm4 ;-mm4 | ||
546 : | pmaxsw mm0,mm1 ;|src| | ||
547 : | pmaxsw mm3,mm4 | ||
548 : | nop2 | ||
549 : | psraw mm1,15 ;sign src | ||
550 : | psraw mm4,15 | ||
551 : | psllw mm0, 4 ; level << 4 | ||
552 : | psllw mm3, 4 ; | ||
553 : | paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 | ||
554 : | paddw mm3, [inter_matrix1 + 8*esi+120] | ||
555 : | movq mm5,[inter_matrix_fixl + 8*esi+112] | ||
556 : | movq mm7,[inter_matrix_fixl + 8*esi+120] | ||
557 : | pmulhuw mm5,mm0 | ||
558 : | pmulhuw mm7,mm3 ;mm7: first approx of division | ||
559 : | mov esp,esp | ||
560 : | movq mm2,[inter_matrix + 8*esi+112] | ||
561 : | movq mm6,[inter_matrix + 8*esi+120] | ||
562 : | pmullw mm2,mm5 ;test value <= original | ||
563 : | pmullw mm6,mm7 | ||
564 : | psubw mm0,mm2 ;mismatch | ||
565 : | psubw mm3,mm6 | ||
566 : | movq mm2,[byte ebx] | ||
567 : | movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 | ||
568 : | pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction | ||
569 : | pmulhuw mm3,[inter_matrix_fix + 8*esi+120] | ||
570 : | paddw mm2,[ebx+8] ;sum | ||
571 : | paddw mm5,mm0 ;final result | ||
572 : | paddw mm7,mm3 | ||
573 : | movq mm0,[edi] | ||
574 : | movq mm3,[edi] | ||
575 : | pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 | ||
576 : | pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) | ||
577 : | add esi,byte 2 | ||
578 : | psrlw mm5, 1 ; (level ) / (2*quant) | ||
579 : | paddw mm2,mm5 ;sum += x1 | ||
580 : | psrlw mm7, 1 | ||
581 : | movq [ebx],mm7 ;store x2 | ||
582 : | pxor mm5, mm1 ; mm0 *= sign(mm0) | ||
583 : | pxor mm7, mm4 ; | ||
584 : | psubw mm5, mm1 ; undisplace | ||
585 : | psubw mm7, mm4 ; | ||
586 : | db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum | ||
587 : | movq [edx + 8*esi+112-16], mm5 | ||
588 : | movq [edx + 8*esi +120-16], mm7 | ||
589 : | jng near .lloop | ||
590 : | jmp near .done | ||
591 : | edgomez | 1.1.2.1 | |
592 : | |||
593 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
594 : | edgomez | 1.1.2.1 | ; |
595 : | edgomez | 1.1.2.2 | ; uint32_t dequant_mpeg_intra_3dne(int16_t *data, |
596 : | ; const int16_t const *coeff, | ||
597 : | ; const uint32_t quant, | ||
598 : | ; const uint32_t dcscalar); | ||
599 : | edgomez | 1.1.2.1 | ; |
600 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
601 : | edgomez | 1.1.2.1 | |
602 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
603 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
604 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
605 : | ; We thus avoid packing/unpacking toward double-word. | ||
606 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
607 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
608 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
609 : | edgomez | 1.1.2.3 | ; and quant in [1..31]. |
610 : | edgomez | 1.1.2.1 | ; |
611 : | edgomez | 1.1.2.2 | |
612 : | edgomez | 1.1.2.1 | %macro DEQUANT4INTRAMMX 1 |
613 : | edgomez | 1.1.2.3 | movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] |
614 : | movq mm4, [ecx+ 16 * %1 +8] ; mm3 = c' = coeff[i+1] | ||
615 : | psubw mm0, mm1 | ||
616 : | psubw mm3, mm4 | ||
617 : | pmaxsw mm0, mm1 | ||
618 : | pmaxsw mm3, mm4 | ||
619 : | psraw mm1, 15 | ||
620 : | psraw mm4, 15 | ||
621 : | %if %1 | ||
622 : | movq mm2, [eax+8] ;preshifted quant | ||
623 : | movq mm7, [eax+8] | ||
624 : | edgomez | 1.1.2.1 | %endif |
625 : | edgomez | 1.1.2.3 | pmullw mm2, [intra_matrix + 16 * %1 ] ; matrix[i]*quant |
626 : | pmullw mm7, [intra_matrix + 16 * %1 +8] ; matrix[i+1]*quant | ||
627 : | movq mm5, mm0 | ||
628 : | movq mm6, mm3 | ||
629 : | pmulhw mm0, mm2 ; high of coeff*(matrix*quant) | ||
630 : | pmulhw mm3, mm7 ; high of coeff*(matrix*quant) | ||
631 : | pmullw mm2, mm5 ; low of coeff*(matrix*quant) | ||
632 : | pmullw mm7, mm6 ; low of coeff*(matrix*quant) | ||
633 : | pcmpgtw mm0, [eax] | ||
634 : | pcmpgtw mm3, [eax] | ||
635 : | paddusw mm2, mm0 | ||
636 : | paddusw mm7, mm3 | ||
637 : | psrlw mm2, 5 | ||
638 : | psrlw mm7, 5 | ||
639 : | pxor mm2, mm1 ; start negating back | ||
640 : | pxor mm7, mm4 ; start negating back | ||
641 : | psubusw mm1, mm0 | ||
642 : | psubusw mm4, mm3 | ||
643 : | movq mm0, [eax] ;zero | ||
644 : | movq mm3, [eax] ;zero | ||
645 : | psubw mm2, mm1 ; finish negating back | ||
646 : | psubw mm7, mm4 ; finish negating back | ||
647 : | movq [byte edx + 16 * %1], mm2 ; data[i] | ||
648 : | movq [edx + 16 * %1 +8], mm7 ; data[i+1] | ||
649 : | edgomez | 1.1.2.1 | %endmacro |
650 : | |||
651 : | edgomez | 1.1.2.3 | ALIGN 16 |
652 : | edgomez | 1.1.2.1 | dequant_mpeg_intra_3dne: |
653 : | edgomez | 1.1.2.3 | mov eax, [esp+12] ; quant |
654 : | mov ecx, [esp+8] ; coeff | ||
655 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
656 : | psllw mm7, 2 ; << 2. See comment. | ||
657 : | mov edx, [esp+4] ; data | ||
658 : | push ebx | ||
659 : | movsx ebx, word [ecx] | ||
660 : | pxor mm0, mm0 | ||
661 : | pxor mm3, mm3 | ||
662 : | push esi | ||
663 : | lea eax, [esp-28] | ||
664 : | sub esp, byte 32 | ||
665 : | and eax, byte -8 ;points to qword ALIGNed space on stack | ||
666 : | movq [eax], mm0 | ||
667 : | movq [eax+8], mm7 | ||
668 : | imul ebx, [esp+16+8+32] ; dcscalar | ||
669 : | movq mm2, mm7 | ||
670 : | |||
671 : | ALIGN 4 | ||
672 : | |||
673 : | DEQUANT4INTRAMMX 0 | ||
674 : | |||
675 : | mov esi, -2048 | ||
676 : | nop | ||
677 : | cmp ebx, esi | ||
678 : | |||
679 : | DEQUANT4INTRAMMX 1 | ||
680 : | |||
681 : | cmovl ebx, esi | ||
682 : | neg esi | ||
683 : | sub esi, byte 1 ;2047 | ||
684 : | |||
685 : | DEQUANT4INTRAMMX 2 | ||
686 : | |||
687 : | cmp ebx, esi | ||
688 : | cmovg ebx, esi | ||
689 : | lea ebp, [byte ebp] | ||
690 : | |||
691 : | DEQUANT4INTRAMMX 3 | ||
692 : | |||
693 : | mov esi, [esp+32] | ||
694 : | mov [byte edx], bx | ||
695 : | mov ebx, [esp+32+4] | ||
696 : | |||
697 : | DEQUANT4INTRAMMX 4 | ||
698 : | DEQUANT4INTRAMMX 5 | ||
699 : | DEQUANT4INTRAMMX 6 | ||
700 : | DEQUANT4INTRAMMX 7 | ||
701 : | edgomez | 1.1.2.1 | |
702 : | edgomez | 1.1.2.3 | add esp, byte 32+8 |
703 : | edgomez | 1.1.2.1 | |
704 : | edgomez | 1.1.2.3 | xor eax, eax |
705 : | ret | ||
706 : | edgomez | 1.1.2.1 | |
707 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
708 : | edgomez | 1.1.2.1 | ; |
709 : | edgomez | 1.1.2.2 | ; uint32_t dequant_mpeg_inter_3dne(int16_t * data, |
710 : | ; const int16_t * const coeff, | ||
711 : | ; const uint32_t quant); | ||
712 : | edgomez | 1.1.2.1 | ; |
713 : | edgomez | 1.1.2.3 | ;----------------------------------------------------------------------------- |
714 : | edgomez | 1.1.2.1 | |
715 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
716 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
717 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
718 : | ; It's mixed with the extraction of the absolute value. | ||
719 : | |||
720 : | edgomez | 1.1.2.3 | ALIGN 16 |
721 : | edgomez | 1.1.2.1 | dequant_mpeg_inter_3dne: |
722 : | edgomez | 1.1.2.3 | mov edx, [esp+ 4] ; data |
723 : | mov ecx, [esp+ 8] ; coeff | ||
724 : | mov eax, [esp+12] ; quant | ||
725 : | movq mm7, [mmx_mul_quant + eax*8 - 8] | ||
726 : | mov eax, -14 | ||
727 : | paddw mm7, mm7 ; << 1 | ||
728 : | pxor mm6, mm6 ; mismatch sum | ||
729 : | push esi | ||
730 : | mov esi, mmzero | ||
731 : | pxor mm1, mm1 | ||
732 : | pxor mm3, mm3 | ||
733 : | nop | ||
734 : | nop4 | ||
735 : | edgomez | 1.1.2.1 | |
736 : | edgomez | 1.1.2.3 | ALIGN 16 |
737 : | edgomez | 1.1.2.1 | .loop |
738 : | edgomez | 1.1.2.3 | movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] |
739 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
740 : | movq mm2, [ecx+8*eax + 7*16 +8] ; mm2 = coeff[i+1] | ||
741 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
742 : | paddsw mm0, mm1 ; c += sgn(c) | ||
743 : | paddsw mm2, mm3 ; c += sgn(c') | ||
744 : | paddw mm0, mm0 ; c *= 2 | ||
745 : | paddw mm2, mm2 ; c'*= 2 | ||
746 : | |||
747 : | movq mm4, [esi] | ||
748 : | movq mm5, [esi] | ||
749 : | psubw mm4, mm0 ; -c | ||
750 : | psubw mm5, mm2 ; -c' | ||
751 : | |||
752 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
753 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
754 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
755 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
756 : | pxor mm0, mm1 ; finish changing sign if needed | ||
757 : | pxor mm2, mm3 ; finish changing sign if needed | ||
758 : | |||
759 : | ; we're short on register, here. Poor pairing... | ||
760 : | |||
761 : | movq mm4, mm7 ; (matrix*quant) | ||
762 : | nop | ||
763 : | pmullw mm4, [inter_matrix + 8*eax + 7*16] | ||
764 : | movq mm5, mm4 | ||
765 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
766 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
767 : | |||
768 : | movq mm4, mm7 ; (matrix*quant) | ||
769 : | pmullw mm4, [inter_matrix + 8*eax + 7*16 + 8] | ||
770 : | add eax, byte 2 | ||
771 : | |||
772 : | pcmpgtw mm5, [esi] | ||
773 : | paddusw mm0, mm5 | ||
774 : | psrlw mm0, 5 | ||
775 : | pxor mm0, mm1 ; start restoring sign | ||
776 : | psubusw mm1, mm5 | ||
777 : | |||
778 : | movq mm5, mm4 | ||
779 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
780 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
781 : | psubw mm0, mm1 ; finish restoring sign | ||
782 : | |||
783 : | pcmpgtw mm5, [esi] | ||
784 : | paddusw mm2, mm5 | ||
785 : | psrlw mm2, 5 | ||
786 : | pxor mm2, mm3 ; start restoring sign | ||
787 : | psubusw mm3, mm5 | ||
788 : | psubw mm2, mm3 ; finish restoring sign | ||
789 : | movq mm1, [esi] | ||
790 : | movq mm3, [byte esi] | ||
791 : | pxor mm6, mm0 ; mismatch control | ||
792 : | movq [edx + 8*eax + 7*16 -2*8 ], mm0 ; data[i] | ||
793 : | pxor mm6, mm2 ; mismatch control | ||
794 : | movq [edx + 8*eax + 7*16 -2*8 +8], mm2 ; data[i+1] | ||
795 : | |||
796 : | jng .loop | ||
797 : | nop | ||
798 : | |||
799 : | ; mismatch control | ||
800 : | |||
801 : | pshufw mm0, mm6, 01010101b | ||
802 : | pshufw mm1, mm6, 10101010b | ||
803 : | pshufw mm2, mm6, 11111111b | ||
804 : | pxor mm6, mm0 | ||
805 : | pxor mm1, mm2 | ||
806 : | pxor mm6, mm1 | ||
807 : | movd eax, mm6 | ||
808 : | and eax, byte 1 | ||
809 : | xor eax, byte 1 | ||
810 : | mov esi, [esp] | ||
811 : | add esp, byte 4 | ||
812 : | xor word [edx + 2*63], ax | ||
813 : | edgomez | 1.1.2.1 | |
814 : | edgomez | 1.1.2.3 | xor eax, eax |
815 : | ret |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |