Parent Directory | Revision Log
Revision 1.16 - (view) (download)
1 : | edgomez | 1.2 | ;/************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - 3dne Quantization/Dequantization - | ||
5 : | ; * | ||
6 : | ; * Copyright (C) 2002-2003 Peter Ross <pross@xvid.org> | ||
7 : | Isibaar | 1.11 | ; * 2002-2008 Michael Militzer <michael@xvid.org> |
8 : | edgomez | 1.2 | ; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
9 : | ; * | ||
10 : | ; * This program is free software ; you can redistribute it and/or modify | ||
11 : | ; * it under the terms of the GNU General Public License as published by | ||
12 : | ; * the Free Software Foundation ; either version 2 of the License, or | ||
13 : | ; * (at your option) any later version. | ||
14 : | ; * | ||
15 : | ; * This program is distributed in the hope that it will be useful, | ||
16 : | ; * but WITHOUT ANY WARRANTY ; without even the implied warranty of | ||
17 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
18 : | ; * GNU General Public License for more details. | ||
19 : | ; * | ||
20 : | ; * You should have received a copy of the GNU General Public License | ||
21 : | ; * along with this program ; if not, write to the Free Software | ||
22 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
23 : | ; * | ||
24 : | Isibaar | 1.16 | ; * $Id: quantize_mpeg_mmx.asm,v 1.15 2008/12/04 18:30:36 Isibaar Exp $ |
25 : | edgomez | 1.2 | ; * |
26 : | ; *************************************************************************/ | ||
27 : | |||
28 : | %define SATURATE | ||
29 : | |||
30 : | Isibaar | 1.11 | %include "nasm.inc" |
31 : | edgomez | 1.2 | |
32 : | ;============================================================================= | ||
33 : | ; Local data (Read Only) | ||
34 : | ;============================================================================= | ||
35 : | |||
36 : | Isibaar | 1.11 | DATA |
37 : | edgomez | 1.2 | |
38 : | mmx_one: | ||
39 : | times 4 dw 1 | ||
40 : | |||
41 : | ;----------------------------------------------------------------------------- | ||
42 : | ; divide by 2Q table | ||
43 : | ;----------------------------------------------------------------------------- | ||
44 : | |||
45 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
46 : | edgomez | 1.2 | mmx_div: |
47 : | times 4 dw 65535 ; the div by 2 formula will overflow for the case | ||
48 : | ; quant=1 but we don't care much because quant=1 | ||
49 : | ; is handled by a different piece of code that | ||
50 : | ; doesn't use this table. | ||
51 : | %assign quant 2 | ||
52 : | %rep 30 | ||
53 : | times 4 dw (1<<17) / (quant*2) + 1 | ||
54 : | %assign quant quant+1 | ||
55 : | %endrep | ||
56 : | |||
57 : | %define VM18P 3 | ||
58 : | %define VM18Q 4 | ||
59 : | |||
60 : | |||
61 : | ;----------------------------------------------------------------------------- | ||
62 : | ; quantd table | ||
63 : | ;----------------------------------------------------------------------------- | ||
64 : | |||
65 : | quantd: | ||
66 : | %assign quant 1 | ||
67 : | %rep 31 | ||
68 : | times 4 dw ((VM18P*quant) + (VM18Q/2)) / VM18Q | ||
69 : | %assign quant quant+1 | ||
70 : | %endrep | ||
71 : | |||
72 : | ;----------------------------------------------------------------------------- | ||
73 : | ; multiple by 2Q table | ||
74 : | ;----------------------------------------------------------------------------- | ||
75 : | |||
76 : | mmx_mul_quant: | ||
77 : | %assign quant 1 | ||
78 : | %rep 31 | ||
79 : | times 4 dw quant | ||
80 : | %assign quant quant+1 | ||
81 : | %endrep | ||
82 : | |||
83 : | ;----------------------------------------------------------------------------- | ||
84 : | ; saturation limits | ||
85 : | ;----------------------------------------------------------------------------- | ||
86 : | |||
87 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
88 : | edgomez | 1.2 | |
89 : | mmx_32767_minus_2047: | ||
90 : | times 4 dw (32767-2047) | ||
91 : | mmx_32768_minus_2048: | ||
92 : | times 4 dw (32768-2048) | ||
93 : | mmx_2047: | ||
94 : | times 4 dw 2047 | ||
95 : | mmx_minus_2048: | ||
96 : | times 4 dw (-2048) | ||
97 : | zero: | ||
98 : | times 4 dw 0 | ||
99 : | |||
100 : | ;============================================================================= | ||
101 : | syskin | 1.6 | ; rounding |
102 : | ;============================================================================= | ||
103 : | |||
104 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
105 : | |||
106 : | syskin | 1.6 | mmx_rounding: |
107 : | Isibaar | 1.11 | dd (1<<13) |
108 : | dd (1<<13) | ||
109 : | syskin | 1.6 | |
110 : | ;============================================================================= | ||
111 : | edgomez | 1.2 | ; Code |
112 : | ;============================================================================= | ||
113 : | |||
114 : | Isibaar | 1.14 | TEXT |
115 : | edgomez | 1.2 | |
116 : | cglobal quant_mpeg_intra_mmx | ||
117 : | cglobal quant_mpeg_inter_mmx | ||
118 : | cglobal dequant_mpeg_intra_mmx | ||
119 : | cglobal dequant_mpeg_inter_mmx | ||
120 : | |||
121 : | syskin | 1.6 | |
122 : | %macro QUANT_MMX 1 | ||
123 : | Isibaar | 1.11 | movq mm0, [_EAX + 16*(%1)] ; data |
124 : | movq mm2, [TMP0 + 16*(%1) + 128] ; intra_matrix_rec | ||
125 : | movq mm4, [_EAX + 16*(%1) + 8] ; data | ||
126 : | movq mm6, [TMP0 + 16*(%1) + 128 + 8] ; intra_matrix_rec | ||
127 : | syskin | 1.6 | |
128 : | movq mm1, mm0 | ||
129 : | movq mm5, mm4 | ||
130 : | |||
131 : | pmullw mm0, mm2 ; low results | ||
132 : | pmulhw mm1, mm2 ; high results | ||
133 : | pmullw mm4, mm6 ; low results | ||
134 : | pmulhw mm5, mm6 ; high results | ||
135 : | |||
136 : | movq mm2, mm0 | ||
137 : | movq mm6, mm4 | ||
138 : | |||
139 : | punpckhwd mm0, mm1 | ||
140 : | punpcklwd mm2, mm1 | ||
141 : | punpckhwd mm4, mm5 | ||
142 : | punpcklwd mm6, mm5 | ||
143 : | |||
144 : | paddd mm2, mm7 | ||
145 : | paddd mm0, mm7 | ||
146 : | paddd mm6, mm7 | ||
147 : | paddd mm4, mm7 | ||
148 : | |||
149 : | psrad mm2, 14 | ||
150 : | psrad mm0, 14 | ||
151 : | psrad mm6, 14 | ||
152 : | psrad mm4, 14 | ||
153 : | |||
154 : | packssdw mm2, mm0 | ||
155 : | packssdw mm6, mm4 | ||
156 : | |||
157 : | Isibaar | 1.11 | movq [TMP1 + 16*(%1)], mm2 |
158 : | movq [TMP1 + 16*(%1)+8], mm6 | ||
159 : | syskin | 1.6 | %endmacro |
160 : | |||
161 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
162 : | ; | ||
163 : | ; uint32_t quant_mpeg_intra_mmx(int16_t * coeff, | ||
164 : | ; const int16_t const * data, | ||
165 : | ; const uint32_t quant, | ||
166 : | ; const uint32_t dcscalar, | ||
167 : | ; const uint16_t *mpeg_matrices); | ||
168 : | ; | ||
169 : | ;----------------------------------------------------------------------------- | ||
170 : | |||
171 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
172 : | edgomez | 1.2 | quant_mpeg_intra_mmx: |
173 : | |||
174 : | Isibaar | 1.11 | mov _EAX, prm2 ; data |
175 : | mov TMP0, prm5 ; mpeg_quant_matrices | ||
176 : | mov TMP1, prm1 ; coeff | ||
177 : | |||
178 : | syskin | 1.6 | movq mm7, [mmx_rounding] |
179 : | edgomez | 1.2 | |
180 : | Isibaar | 1.12 | QUANT_MMX(0) |
181 : | QUANT_MMX(1) | ||
182 : | QUANT_MMX(2) | ||
183 : | QUANT_MMX(3) | ||
184 : | QUANT_MMX(4) | ||
185 : | QUANT_MMX(5) | ||
186 : | QUANT_MMX(6) | ||
187 : | QUANT_MMX(7) | ||
188 : | syskin | 1.6 | |
189 : | ; calculate DC | ||
190 : | Isibaar | 1.15 | movsx _EAX, word [_EAX] ; data[0] |
191 : | mov TMP0, prm4 ; dcscalar | ||
192 : | Isibaar | 1.12 | mov _EDX, _EAX |
193 : | Isibaar | 1.15 | shr TMP0, 1 ; TMP0 = dcscalar/2 |
194 : | sar _EDX, 31 ; TMP1 = sign extend of _EAX (ready for division too) | ||
195 : | |||
196 : | xor TMP0, _EDX ; adjust TMP0 according to the sign of data[0] | ||
197 : | Isibaar | 1.12 | sub TMP0, _EDX |
198 : | add _EAX, TMP0 | ||
199 : | Isibaar | 1.11 | |
200 : | Isibaar | 1.12 | mov TMP0, prm4 ; dcscalar |
201 : | idiv TMP0 ; _EAX = _EDX:_EAX / dcscalar | ||
202 : | Isibaar | 1.11 | |
203 : | Isibaar | 1.12 | mov _EDX, prm1 ; coeff again |
204 : | mov word [_EDX], ax ; coeff[0] = ax | ||
205 : | syskin | 1.8 | |
206 : | Isibaar | 1.11 | xor _EAX, _EAX ; return(0); |
207 : | edgomez | 1.2 | ret |
208 : | Isibaar | 1.10 | ENDFUNC |
209 : | edgomez | 1.2 | |
210 : | |||
211 : | ;----------------------------------------------------------------------------- | ||
212 : | ; | ||
213 : | ; uint32_t quant_mpeg_inter_mmx(int16_t * coeff, | ||
214 : | ; const int16_t const * data, | ||
215 : | ; const uint32_t quant, | ||
216 : | ; const uint16_t *mpeg_matrices); | ||
217 : | ; | ||
218 : | ;----------------------------------------------------------------------------- | ||
219 : | |||
220 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
221 : | edgomez | 1.2 | quant_mpeg_inter_mmx: |
222 : | |||
223 : | Isibaar | 1.11 | mov TMP1, prm1 ; coeff |
224 : | mov _EAX, prm3 ; quant | ||
225 : | mov TMP0, prm4 ; mpeg_quant_matrices | ||
226 : | |||
227 : | push _ESI | ||
228 : | %ifdef ARCH_IS_X86_64 | ||
229 : | mov _ESI, prm2 ; data | ||
230 : | %else | ||
231 : | mov _ESI, [_ESP + 4 + 8] ; data | ||
232 : | %endif | ||
233 : | edgomez | 1.2 | |
234 : | Isibaar | 1.11 | push _EBX |
235 : | xor _EBX, _EBX | ||
236 : | edgomez | 1.2 | |
237 : | pxor mm5, mm5 ; sum | ||
238 : | |||
239 : | cmp al, 1 | ||
240 : | jz near .q1loop | ||
241 : | |||
242 : | cmp al, 2 | ||
243 : | jz near .q2loop | ||
244 : | |||
245 : | Isibaar | 1.11 | %ifdef ARCH_IS_X86_64 |
246 : | lea r9, [mmx_div] | ||
247 : | movq mm7, [r9 + _EAX * 8 - 8] | ||
248 : | %else | ||
249 : | movq mm7, [mmx_div + _EAX * 8 - 8] ; divider | ||
250 : | %endif | ||
251 : | edgomez | 1.2 | |
252 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
253 : | Isibaar | 1.10 | .loop: |
254 : | Isibaar | 1.11 | movq mm0, [_ESI + 8*_EBX] ; mm0 = [1st] |
255 : | movq mm3, [_ESI + 8*_EBX + 8] ; | ||
256 : | edgomez | 1.2 | pxor mm1, mm1 ; mm1 = 0 |
257 : | pxor mm4, mm4 ; | ||
258 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
259 : | pcmpgtw mm4, mm3 ; | ||
260 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
261 : | pxor mm3, mm4 ; | ||
262 : | psubw mm0, mm1 ; displace | ||
263 : | psubw mm3, mm4 ; | ||
264 : | psllw mm0, 4 | ||
265 : | psllw mm3, 4 | ||
266 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX] |
267 : | edgomez | 1.2 | psrlw mm2, 1 |
268 : | paddw mm0, mm2 | ||
269 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8] |
270 : | edgomez | 1.2 | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
271 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX + 8] |
272 : | edgomez | 1.2 | psrlw mm2, 1 |
273 : | paddw mm3, mm2 | ||
274 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8 + 8] |
275 : | edgomez | 1.2 | pmulhw mm3, mm2 |
276 : | pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 | ||
277 : | pmulhw mm3, mm7 ; | ||
278 : | psrlw mm0, 1 ; additional shift by 1 => 16 + 1 = 17 | ||
279 : | psrlw mm3, 1 | ||
280 : | paddw mm5, mm0 ; sum += mm0 | ||
281 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
282 : | paddw mm5, mm3 ; | ||
283 : | pxor mm3, mm4 ; | ||
284 : | psubw mm0, mm1 ; undisplace | ||
285 : | psubw mm3, mm4 | ||
286 : | Isibaar | 1.11 | movq [TMP1 + 8*_EBX], mm0 |
287 : | movq [TMP1 + 8*_EBX + 8], mm3 | ||
288 : | edgomez | 1.2 | |
289 : | Isibaar | 1.11 | add _EBX, 2 |
290 : | cmp _EBX, 16 | ||
291 : | edgomez | 1.2 | jnz near .loop |
292 : | |||
293 : | Isibaar | 1.10 | .done: |
294 : | edgomez | 1.2 | pmaddwd mm5, [mmx_one] |
295 : | movq mm0, mm5 | ||
296 : | psrlq mm5, 32 | ||
297 : | paddd mm0, mm5 | ||
298 : | movd eax, mm0 ; return sum | ||
299 : | |||
300 : | Isibaar | 1.11 | pop _EBX |
301 : | pop _ESI | ||
302 : | edgomez | 1.2 | |
303 : | ret | ||
304 : | |||
305 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
306 : | Isibaar | 1.10 | .q1loop: |
307 : | Isibaar | 1.11 | movq mm0, [_ESI + 8*_EBX] ; mm0 = [1st] |
308 : | movq mm3, [_ESI + 8*_EBX+ 8] | ||
309 : | edgomez | 1.2 | pxor mm1, mm1 ; mm1 = 0 |
310 : | pxor mm4, mm4 ; | ||
311 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
312 : | pcmpgtw mm4, mm3 ; | ||
313 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
314 : | pxor mm3, mm4 ; | ||
315 : | psubw mm0, mm1 ; displace | ||
316 : | psubw mm3, mm4 ; | ||
317 : | psllw mm0, 4 | ||
318 : | psllw mm3, 4 | ||
319 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX] |
320 : | edgomez | 1.2 | psrlw mm2, 1 |
321 : | paddw mm0, mm2 | ||
322 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8] |
323 : | edgomez | 1.2 | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
324 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX + 8] |
325 : | edgomez | 1.2 | psrlw mm2, 1 |
326 : | paddw mm3, mm2 | ||
327 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8 + 8] |
328 : | edgomez | 1.2 | pmulhw mm3, mm2 |
329 : | psrlw mm0, 1 ; mm0 >>= 1 (/2) | ||
330 : | psrlw mm3, 1 ; | ||
331 : | paddw mm5, mm0 ; sum += mm0 | ||
332 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
333 : | paddw mm5, mm3 ; | ||
334 : | pxor mm3, mm4 ; | ||
335 : | psubw mm0, mm1 ; undisplace | ||
336 : | psubw mm3, mm4 | ||
337 : | Isibaar | 1.11 | movq [TMP1 + 8*_EBX], mm0 |
338 : | movq [TMP1 + 8*_EBX + 8], mm3 | ||
339 : | edgomez | 1.2 | |
340 : | Isibaar | 1.11 | add _EBX, 2 |
341 : | cmp _EBX, 16 | ||
342 : | edgomez | 1.2 | jnz near .q1loop |
343 : | |||
344 : | jmp .done | ||
345 : | |||
346 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
347 : | Isibaar | 1.10 | .q2loop: |
348 : | Isibaar | 1.11 | movq mm0, [_ESI + 8*_EBX] ; mm0 = [1st] |
349 : | movq mm3, [_ESI + 8*_EBX+ 8] | ||
350 : | edgomez | 1.2 | pxor mm1, mm1 ; mm1 = 0 |
351 : | pxor mm4, mm4 ; | ||
352 : | pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) | ||
353 : | pcmpgtw mm4, mm3 ; | ||
354 : | pxor mm0, mm1 ; mm0 = |mm0| | ||
355 : | pxor mm3, mm4 ; | ||
356 : | psubw mm0, mm1 ; displace | ||
357 : | psubw mm3, mm4 ; | ||
358 : | psllw mm0, 4 | ||
359 : | psllw mm3, 4 | ||
360 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX] |
361 : | edgomez | 1.2 | psrlw mm2, 1 |
362 : | paddw mm0, mm2 | ||
363 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8] |
364 : | edgomez | 1.2 | pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
365 : | Isibaar | 1.11 | movq mm2, [TMP0 + 512 + 8*_EBX + 8] |
366 : | edgomez | 1.2 | psrlw mm2, 1 |
367 : | paddw mm3, mm2 | ||
368 : | Isibaar | 1.11 | movq mm2, [TMP0 + 768 + _EBX*8 + 8] |
369 : | edgomez | 1.2 | pmulhw mm3, mm2 |
370 : | psrlw mm0, 2 ; mm0 >>= 1 (/2) | ||
371 : | psrlw mm3, 2 ; | ||
372 : | paddw mm5, mm0 ; sum += mm0 | ||
373 : | pxor mm0, mm1 ; mm0 *= sign(mm0) | ||
374 : | paddw mm5, mm3 ; | ||
375 : | pxor mm3, mm4 ; | ||
376 : | psubw mm0, mm1 ; undisplace | ||
377 : | psubw mm3, mm4 | ||
378 : | Isibaar | 1.11 | movq [TMP1 + 8*_EBX], mm0 |
379 : | movq [TMP1 + 8*_EBX + 8], mm3 | ||
380 : | edgomez | 1.2 | |
381 : | Isibaar | 1.11 | add _EBX, 2 |
382 : | cmp _EBX, 16 | ||
383 : | edgomez | 1.2 | jnz near .q2loop |
384 : | |||
385 : | jmp .done | ||
386 : | Isibaar | 1.10 | ENDFUNC |
387 : | edgomez | 1.2 | |
388 : | |||
389 : | ;----------------------------------------------------------------------------- | ||
390 : | ; | ||
391 : | ; uint32_t dequant_mpeg_intra_mmx(int16_t *data, | ||
392 : | ; const int16_t const *coeff, | ||
393 : | ; const uint32_t quant, | ||
394 : | ; const uint32_t dcscalar, | ||
395 : | ; const uint16_t *mpeg_matrices); | ||
396 : | ; | ||
397 : | ;----------------------------------------------------------------------------- | ||
398 : | |||
399 : | ; Note: in order to saturate 'easily', we pre-shift the quantifier | ||
400 : | ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to | ||
401 : | ; build a saturating mask. It is non-zero only when an overflow occured. | ||
402 : | ; We thus avoid packing/unpacking toward double-word. | ||
403 : | ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., | ||
404 : | ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not | ||
405 : | ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a | ||
406 : | ; and quant in [1..31]. | ||
407 : | ; | ||
408 : | ; The original loop is: | ||
409 : | ; | ||
410 : | %if 0 | ||
411 : | Isibaar | 1.11 | movq mm0, [TMP0+8*_EAX + 8*16] ; mm0 = coeff[i] |
412 : | edgomez | 1.2 | pxor mm1, mm1 |
413 : | pcmpgtw mm1, mm0 | ||
414 : | pxor mm0, mm1 ; change sign if negative | ||
415 : | psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] | ||
416 : | |||
417 : | movq mm2, mm7 ; mm2 = quant | ||
418 : | Isibaar | 1.11 | pmullw mm2, [_EBX + 8*_EAX + 8*16 ] ; matrix[i]*quant. |
419 : | edgomez | 1.2 | |
420 : | movq mm6, mm2 | ||
421 : | pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) | ||
422 : | pmullw mm0, mm6 ; low of coeff*(matrix*quant) | ||
423 : | |||
424 : | pxor mm5, mm5 | ||
425 : | pcmpgtw mm2, mm5 ; otherflow? | ||
426 : | psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise | ||
427 : | psrlw mm0, 5 | ||
428 : | paddw mm0, mm1 ; start restoring sign | ||
429 : | por mm0, mm2 ; saturate to 2047 if needed | ||
430 : | pxor mm0, mm1 ; finish negating back | ||
431 : | |||
432 : | Isibaar | 1.11 | movq [TMP1 + 8*_EAX + 8*16], mm0 ; data[i] |
433 : | add _EAX, 1 | ||
434 : | edgomez | 1.2 | %endif |
435 : | |||
436 : | ;******************************************************************** | ||
437 : | |||
438 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
439 : | edgomez | 1.2 | dequant_mpeg_intra_mmx: |
440 : | |||
441 : | Isibaar | 1.11 | mov TMP1, prm1 ; data |
442 : | mov TMP0, prm2 ; coeff | ||
443 : | mov _EAX, prm5 ; mpeg_quant_matrices | ||
444 : | |||
445 : | push _EBX | ||
446 : | mov _EBX, _EAX | ||
447 : | %ifdef ARCH_IS_X86_64 | ||
448 : | mov _EAX, prm3 | ||
449 : | Isibaar | 1.13 | lea prm1, [mmx_mul_quant] |
450 : | movq mm7, [prm1 + _EAX*8 - 8] | ||
451 : | Isibaar | 1.11 | %else |
452 : | mov _EAX, [_ESP + 4 + 12] ; quant | ||
453 : | movq mm7, [mmx_mul_quant + _EAX*8 - 8] | ||
454 : | %endif | ||
455 : | edgomez | 1.2 | |
456 : | Isibaar | 1.11 | mov _EAX, -16 ; to keep ALIGNed, we regularly process coeff[0] |
457 : | edgomez | 1.2 | psllw mm7, 2 ; << 2. See comment. |
458 : | pxor mm6, mm6 ; this is a NOP | ||
459 : | |||
460 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
461 : | Isibaar | 1.10 | .loop: |
462 : | Isibaar | 1.11 | movq mm0, [TMP0+8*_EAX + 8*16] ; mm0 = c = coeff[i] |
463 : | movq mm3, [TMP0+8*_EAX + 8*16 +8]; mm3 = c' = coeff[i+1] | ||
464 : | edgomez | 1.2 | pxor mm1, mm1 |
465 : | pxor mm4, mm4 | ||
466 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) | ||
467 : | movq mm2, mm7 ; mm2 = quant | ||
468 : | |||
469 : | pcmpgtw mm4, mm3 ; mm4 = sgn(c') | ||
470 : | Isibaar | 1.11 | pmullw mm2, [_EBX + 8*_EAX + 8*16 ] ; matrix[i]*quant |
471 : | edgomez | 1.2 | |
472 : | pxor mm0, mm1 ; negate if negative | ||
473 : | pxor mm3, mm4 ; negate if negative | ||
474 : | |||
475 : | psubw mm0, mm1 | ||
476 : | psubw mm3, mm4 | ||
477 : | |||
478 : | ; we're short on register, here. Poor pairing... | ||
479 : | |||
480 : | movq mm5, mm2 | ||
481 : | pmullw mm2, mm0 ; low of coeff*(matrix*quant) | ||
482 : | |||
483 : | pmulhw mm0, mm5 ; high of coeff*(matrix*quant) | ||
484 : | movq mm5, mm7 ; mm2 = quant | ||
485 : | |||
486 : | Isibaar | 1.11 | pmullw mm5, [_EBX + 8*_EAX + 8*16 +8] ; matrix[i+1]*quant |
487 : | edgomez | 1.2 | |
488 : | movq mm6, mm5 | ||
489 : | Isibaar | 1.11 | add _EAX,2 ; z-flag will be tested later |
490 : | edgomez | 1.2 | |
491 : | pmullw mm6, mm3 ; low of coeff*(matrix*quant) | ||
492 : | pmulhw mm3, mm5 ; high of coeff*(matrix*quant) | ||
493 : | |||
494 : | pcmpgtw mm0, [zero] | ||
495 : | paddusw mm2, mm0 | ||
496 : | psrlw mm2, 5 | ||
497 : | |||
498 : | pcmpgtw mm3, [zero] | ||
499 : | paddusw mm6, mm3 | ||
500 : | psrlw mm6, 5 | ||
501 : | |||
502 : | pxor mm2, mm1 ; start negating back | ||
503 : | pxor mm6, mm4 ; start negating back | ||
504 : | |||
505 : | psubusw mm1, mm0 | ||
506 : | psubusw mm4, mm3 | ||
507 : | |||
508 : | psubw mm2, mm1 ; finish negating back | ||
509 : | psubw mm6, mm4 ; finish negating back | ||
510 : | |||
511 : | Isibaar | 1.11 | movq [TMP1 + 8*_EAX + 8*16 -2*8 ], mm2 ; data[i] |
512 : | movq [TMP1 + 8*_EAX + 8*16 -2*8 +8], mm6 ; data[i+1] | ||
513 : | edgomez | 1.2 | |
514 : | jnz near .loop | ||
515 : | |||
516 : | Isibaar | 1.11 | pop _EBX |
517 : | |||
518 : | edgomez | 1.2 | ; deal with DC |
519 : | Isibaar | 1.11 | movd mm0, [TMP0] |
520 : | %ifdef ARCH_IS_X86_64 | ||
521 : | movq mm6, prm4 | ||
522 : | pmullw mm0, mm6 | ||
523 : | %else | ||
524 : | pmullw mm0, prm4 ; dcscalar | ||
525 : | %endif | ||
526 : | edgomez | 1.2 | movq mm2, [mmx_32767_minus_2047] |
527 : | paddsw mm0, mm2 | ||
528 : | psubsw mm0, mm2 | ||
529 : | movq mm2, [mmx_32768_minus_2048] | ||
530 : | psubsw mm0, mm2 | ||
531 : | paddsw mm0, mm2 | ||
532 : | movd eax, mm0 | ||
533 : | Isibaar | 1.11 | mov [TMP1], ax |
534 : | edgomez | 1.2 | |
535 : | Isibaar | 1.11 | xor _EAX, _EAX |
536 : | edgomez | 1.2 | |
537 : | ret | ||
538 : | Isibaar | 1.10 | ENDFUNC |
539 : | edgomez | 1.2 | |
540 : | ;----------------------------------------------------------------------------- | ||
541 : | ; | ||
542 : | ; uint32_t dequant_mpeg_inter_mmx(int16_t * data, | ||
543 : | ; const int16_t * const coeff, | ||
544 : | ; const uint32_t quant, | ||
545 : | ; const uint16_t *mpeg_matrices); | ||
546 : | ; | ||
547 : | ;----------------------------------------------------------------------------- | ||
548 : | |||
549 : | ; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier | ||
550 : | ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. | ||
551 : | ; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. | ||
552 : | ; It's mixed with the extraction of the absolute value. | ||
553 : | |||
554 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
555 : | edgomez | 1.2 | dequant_mpeg_inter_mmx: |
556 : | |||
557 : | |||
558 : | Isibaar | 1.11 | mov TMP1, prm1 ; data |
559 : | mov TMP0, prm2 ; coeff | ||
560 : | mov _EAX, prm3 ; quant | ||
561 : | |||
562 : | push _EBX | ||
563 : | %ifdef ARCH_IS_X86_64 | ||
564 : | mov _EBX, prm4 | ||
565 : | lea r9, [mmx_mul_quant] | ||
566 : | movq mm7, [r9 + _EAX*8 - 8] | ||
567 : | %else | ||
568 : | mov _EBX, [_ESP + 4 + 16] ; mpeg_quant_matrices | ||
569 : | movq mm7, [mmx_mul_quant + _EAX*8 - 8] | ||
570 : | %endif | ||
571 : | edgomez | 1.2 | |
572 : | Isibaar | 1.11 | mov _EAX, -16 |
573 : | edgomez | 1.2 | paddw mm7, mm7 ; << 1 |
574 : | pxor mm6, mm6 ; mismatch sum | ||
575 : | |||
576 : | Isibaar | 1.11 | ALIGN SECTION_ALIGN |
577 : | Isibaar | 1.10 | .loop: |
578 : | Isibaar | 1.11 | movq mm0, [TMP0+8*_EAX + 8*16 ] ; mm0 = coeff[i] |
579 : | movq mm2, [TMP0+8*_EAX + 8*16 +8] ; mm2 = coeff[i+1] | ||
580 : | add _EAX, 2 | ||
581 : | edgomez | 1.2 | |
582 : | pxor mm1, mm1 | ||
583 : | pxor mm3, mm3 | ||
584 : | pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) | ||
585 : | pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) | ||
586 : | paddsw mm0, mm1 ; c += sgn(c) | ||
587 : | paddsw mm2, mm3 ; c += sgn(c') | ||
588 : | paddw mm0, mm0 ; c *= 2 | ||
589 : | paddw mm2, mm2 ; c'*= 2 | ||
590 : | |||
591 : | pxor mm4, mm4 | ||
592 : | pxor mm5, mm5 | ||
593 : | psubw mm4, mm0 ; -c | ||
594 : | psubw mm5, mm2 ; -c' | ||
595 : | psraw mm4, 16 ; mm4 = sgn(-c) | ||
596 : | psraw mm5, 16 ; mm5 = sgn(-c') | ||
597 : | psubsw mm0, mm4 ; c -= sgn(-c) | ||
598 : | psubsw mm2, mm5 ; c' -= sgn(-c') | ||
599 : | pxor mm0, mm1 ; finish changing sign if needed | ||
600 : | pxor mm2, mm3 ; finish changing sign if needed | ||
601 : | |||
602 : | ; we're short on register, here. Poor pairing... | ||
603 : | |||
604 : | movq mm4, mm7 ; (matrix*quant) | ||
605 : | Isibaar | 1.11 | pmullw mm4, [_EBX + 512 + 8*_EAX + 8*16 -2*8] |
606 : | edgomez | 1.2 | movq mm5, mm4 |
607 : | pmulhw mm5, mm0 ; high of c*(matrix*quant) | ||
608 : | pmullw mm0, mm4 ; low of c*(matrix*quant) | ||
609 : | |||
610 : | movq mm4, mm7 ; (matrix*quant) | ||
611 : | Isibaar | 1.11 | pmullw mm4, [_EBX + 512 + 8*_EAX + 8*16 -2*8 + 8] |
612 : | edgomez | 1.2 | |
613 : | pcmpgtw mm5, [zero] | ||
614 : | paddusw mm0, mm5 | ||
615 : | psrlw mm0, 5 | ||
616 : | pxor mm0, mm1 ; start restoring sign | ||
617 : | psubusw mm1, mm5 | ||
618 : | |||
619 : | movq mm5, mm4 | ||
620 : | pmulhw mm5, mm2 ; high of c*(matrix*quant) | ||
621 : | pmullw mm2, mm4 ; low of c*(matrix*quant) | ||
622 : | psubw mm0, mm1 ; finish restoring sign | ||
623 : | |||
624 : | pcmpgtw mm5, [zero] | ||
625 : | paddusw mm2, mm5 | ||
626 : | psrlw mm2, 5 | ||
627 : | pxor mm2, mm3 ; start restoring sign | ||
628 : | psubusw mm3, mm5 | ||
629 : | psubw mm2, mm3 ; finish restoring sign | ||
630 : | |||
631 : | pxor mm6, mm0 ; mismatch control | ||
632 : | Isibaar | 1.11 | movq [TMP1 + 8*_EAX + 8*16 -2*8 ], mm0 ; data[i] |
633 : | edgomez | 1.2 | pxor mm6, mm2 ; mismatch control |
634 : | Isibaar | 1.11 | movq [TMP1 + 8*_EAX + 8*16 -2*8 +8], mm2 ; data[i+1] |
635 : | edgomez | 1.2 | |
636 : | jnz near .loop | ||
637 : | |||
638 : | ; mismatch control | ||
639 : | |||
640 : | movq mm0, mm6 | ||
641 : | psrlq mm0, 48 | ||
642 : | movq mm1, mm6 | ||
643 : | movq mm2, mm6 | ||
644 : | psrlq mm1, 32 | ||
645 : | pxor mm6, mm0 | ||
646 : | psrlq mm2, 16 | ||
647 : | pxor mm6, mm1 | ||
648 : | pxor mm6, mm2 | ||
649 : | movd eax, mm6 | ||
650 : | Isibaar | 1.11 | and _EAX, 1 |
651 : | xor _EAX, 1 | ||
652 : | xor word [TMP1 + 2*63], ax | ||
653 : | edgomez | 1.2 | |
654 : | Isibaar | 1.11 | xor _EAX, _EAX |
655 : | edgomez | 1.2 | |
656 : | Isibaar | 1.11 | pop _EBX |
657 : | edgomez | 1.2 | |
658 : | ret | ||
659 : | Isibaar | 1.10 | ENDFUNC |
660 : | edgomez | 1.5 | |
661 : | Isibaar | 1.16 | NON_EXEC_STACK |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |