1 |
|
;/************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * - mmx quantization/dequantization - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001-2003 XviD Team <xvid-devel@xvid.org> |
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 |
|
; * quant4 bugs have been fixed: (a) overflow bug for matrix elements |
27 |
|
; * equal to 1 or 2 is fixed by substituting pmulhw with pmulhuw (iSSE) |
28 |
|
; * and using multiplier 0ffffh instead of 10001h (for matrix element = 1; |
29 |
|
; * in that case, 1 is added before multiplying, that additional 1 comes |
30 |
|
; * from intra_matrix1; (b) rounding error for large coefficients and matrix |
31 |
|
; * elements is fixed by two-step approach: first approximation (rounded |
32 |
|
; * down) is found as usual; the result is multiplied by the matrix element |
33 |
|
; * and mismatch is used to calculate the correction. |
34 |
|
; *************************************************************************/ |
35 |
|
; _3dne functions are compatible with iSSE, but are optimized specifically |
36 |
|
; for K7 pipelines |
37 |
|
; |
38 |
|
;--------------------------------------------------------------------------- |
39 |
|
; 09.12.2002 Athlon optimizations contributed by Jaan Kalda |
40 |
|
;--------------------------------------------------------------------------- |
41 |
|
|
42 |
|
|
43 |
|
; data/text alignment |
44 |
|
%define ALIGN 8 |
45 |
|
%define SATURATE |
46 |
|
|
47 |
|
bits 32 |
48 |
|
|
49 |
|
%ifdef FORMAT_COFF |
50 |
|
SECTION .data data |
51 |
|
%else |
52 |
|
SECTION .data data align=8 |
53 |
|
%endif |
54 |
|
|
55 |
|
%macro cglobal 1 |
56 |
|
%ifdef PREFIX |
57 |
|
global _%1 |
58 |
|
%define %1 _%1 |
59 |
|
%else |
60 |
|
global %1 |
61 |
|
%endif |
62 |
|
%endmacro |
63 |
|
|
64 |
|
%macro cextern 1 |
65 |
|
%ifdef PREFIX |
66 |
|
extern _%1 |
67 |
|
%define %1 _%1 |
68 |
|
%else |
69 |
|
extern %1 |
70 |
|
%endif |
71 |
|
%endmacro |
72 |
|
align 8 |
73 |
|
mmzero dd 0,0 |
74 |
|
|
75 |
|
mmx_one times 4 dw 1 |
76 |
|
|
77 |
|
;=========================================================================== |
78 |
|
; |
79 |
|
; divide by 2Q table |
80 |
|
; |
81 |
|
;=========================================================================== |
82 |
|
|
83 |
|
align ALIGN |
84 |
|
mmx_divs ;i>2 |
85 |
|
%assign i 1 |
86 |
|
%rep 31 |
87 |
|
times 4 dw ((1 << 15) / i + 1) |
88 |
|
%assign i i+1 |
89 |
|
%endrep |
90 |
|
|
91 |
|
align ALIGN |
92 |
|
mmx_div ;i>2 |
93 |
|
%assign i 1 |
94 |
|
%rep 31 |
95 |
|
times 4 dw ((1 << 16) / i + 1) |
96 |
|
%assign i i+1 |
97 |
|
%endrep |
98 |
|
|
99 |
|
|
100 |
|
;=========================================================================== |
101 |
|
; |
102 |
|
; intra matrix |
103 |
|
; |
104 |
|
;=========================================================================== |
105 |
|
|
106 |
|
%macro FIXX 1 |
107 |
|
dw (1 << 16) / (%1) + 1 |
108 |
|
%endmacro |
109 |
|
|
110 |
|
cextern intra_matrix_fixl |
111 |
|
cextern intra_matrix_fix |
112 |
|
cextern intra_matrix1 |
113 |
|
cextern intra_matrix |
114 |
|
|
115 |
|
;=========================================================================== |
116 |
|
; |
117 |
|
; inter matrix |
118 |
|
; |
119 |
|
;=========================================================================== |
120 |
|
|
121 |
|
cextern inter_matrix1 |
122 |
|
cextern inter_matrix |
123 |
|
cextern inter_matrix_fix |
124 |
|
cextern inter_matrix_fixl |
125 |
|
|
126 |
|
|
127 |
|
%define VM18P 3 |
128 |
|
%define VM18Q 4 |
129 |
|
%define nop4 db 08Dh,074h,026h,0 |
130 |
|
%define nop3 add esp,byte 0 |
131 |
|
%define nop2 mov esp,esp |
132 |
|
%define nop7 db 08dh,02ch,02dh,0,0,0,0 |
133 |
|
%define nop6 add ebp,dword 0 |
134 |
|
|
135 |
|
;=========================================================================== |
136 |
|
; |
137 |
|
; quantd table |
138 |
|
; |
139 |
|
;=========================================================================== |
140 |
|
|
141 |
|
|
142 |
|
quantd |
143 |
|
%assign i 1 |
144 |
|
%rep 31 |
145 |
|
times 4 dw (((VM18P*i) + (VM18Q/2)) / VM18Q) |
146 |
|
%assign i i+1 |
147 |
|
%endrep |
148 |
|
|
149 |
|
;=========================================================================== |
150 |
|
; |
151 |
|
; multiple by 2Q table |
152 |
|
; |
153 |
|
;=========================================================================== |
154 |
|
|
155 |
|
|
156 |
|
mmx_mul_quant |
157 |
|
%assign i 1 |
158 |
|
%rep 31 |
159 |
|
times 4 dw i |
160 |
|
%assign i i+1 |
161 |
|
%endrep |
162 |
|
|
163 |
|
;=========================================================================== |
164 |
|
; |
165 |
|
; saturation limits |
166 |
|
; |
167 |
|
;=========================================================================== |
168 |
|
|
169 |
|
align 16 |
170 |
|
|
171 |
|
mmx_32767_minus_2047 times 4 dw (32767-2047) |
172 |
|
mmx_32768_minus_2048 times 4 dw (32768-2048) |
173 |
|
mmx_2047 times 4 dw 2047 |
174 |
|
mmx_minus_2048 times 4 dw (-2048) |
175 |
|
zero times 4 dw 0 |
176 |
|
|
177 |
|
int_div |
178 |
|
dd 0 |
179 |
|
%assign i 1 |
180 |
|
%rep 255 |
181 |
|
dd (1 << 17) / ( i) + 1 |
182 |
|
%assign i i+1 |
183 |
|
%endrep |
184 |
|
|
185 |
|
section .text |
186 |
|
|
187 |
|
;=========================================================================== |
188 |
|
; |
189 |
|
; void quant4_intra_xmm(int16_t * coeff, |
190 |
|
; const int16_t const * data, |
191 |
|
; const uint32_t quant, |
192 |
|
; const uint32_t dcscalar); |
193 |
|
; |
194 |
|
;=========================================================================== |
195 |
|
|
196 |
|
align ALIGN |
197 |
|
cglobal quant_mpeg_intra_xmm |
198 |
|
quant_mpeg_intra_xmm: |
199 |
|
mov eax, [esp + 8] ; data |
200 |
|
mov ecx, [esp + 12] ; quant |
201 |
|
mov edx, [esp + 4] ; coeff |
202 |
|
push esi |
203 |
|
push edi |
204 |
|
push ebx |
205 |
|
nop |
206 |
|
mov edi,mmzero |
207 |
|
mov esi,-14 |
208 |
|
pxor mm0,mm0 |
209 |
|
pxor mm3,mm3 |
210 |
|
cmp ecx,byte 1 |
211 |
|
je near .q1loop |
212 |
|
cmp ecx,byte 19 |
213 |
|
jg near .lloop |
214 |
|
nop6 |
215 |
|
|
216 |
|
|
217 |
|
align ALIGN |
218 |
|
.loop |
219 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
220 |
|
psubw mm0, mm1 ;-mm1 |
221 |
|
movq mm4, [eax + 8*esi + 120] ; |
222 |
|
psubw mm3, mm4 ;-mm4 |
223 |
|
pmaxsw mm0, mm1 ;|src| |
224 |
|
pmaxsw mm3,mm4 |
225 |
|
nop2 |
226 |
|
psraw mm1, 15 ;sign src |
227 |
|
psraw mm4, 15 |
228 |
|
psllw mm0, 4 ;level << 4 ; |
229 |
|
psllw mm3, 4 |
230 |
|
paddw mm0, [intra_matrix1 + 8*esi+112] |
231 |
|
paddw mm3, [intra_matrix1 + 8*esi+120] |
232 |
|
movq mm5, [intra_matrix_fixl + 8*esi+112] |
233 |
|
movq mm7, [intra_matrix_fixl + 8*esi+120] |
234 |
|
pmulhuw mm5, mm0 |
235 |
|
pmulhuw mm7, mm3 |
236 |
|
mov esp, esp |
237 |
|
movq mm2, [intra_matrix + 8*esi+112] |
238 |
|
movq mm6, [intra_matrix + 8*esi+120] |
239 |
|
pmullw mm2, mm5 |
240 |
|
pmullw mm6, mm7 |
241 |
|
psubw mm0, mm2 |
242 |
|
psubw mm3, mm6 |
243 |
|
nop4 |
244 |
|
movq mm2, [quantd + ecx * 8 - 8] |
245 |
|
movq mm6, [mmx_divs + ecx * 8 - 8] |
246 |
|
paddw mm5, mm2 |
247 |
|
paddw mm7, mm2 |
248 |
|
mov esp, esp |
249 |
|
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] |
250 |
|
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
251 |
|
paddw mm5, mm0 |
252 |
|
paddw mm7, mm3 |
253 |
|
movq mm0, [edi] |
254 |
|
movq mm3, [edi] |
255 |
|
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
256 |
|
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
257 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
258 |
|
pxor mm7, mm4 ; |
259 |
|
psubw mm5, mm1 ; undisplace |
260 |
|
psubw mm7, mm4 ; |
261 |
|
movq [edx + 8*esi+112], mm5 |
262 |
|
movq [edx + 8*esi +120], mm7 |
263 |
|
add esi, byte 2 |
264 |
|
jng near .loop |
265 |
|
|
266 |
|
.done |
267 |
|
; calculate data[0] // (int32_t)dcscalar) |
268 |
|
mov esi, [esp + 12 + 16] ; dcscalar |
269 |
|
movsx ecx, word [eax] |
270 |
|
mov edi, ecx |
271 |
|
mov edx, [esp + 12 + 16] |
272 |
|
shr edx, 1 ; ebx = dcscalar /2 |
273 |
|
sar edi, 31 ; cdq is vectorpath |
274 |
|
xor edx, edi ; ebx = eax V -eax -1 |
275 |
|
sub ecx, edi |
276 |
|
add ecx, edx |
277 |
|
mov edx, [dword esp + 12 + 4] |
278 |
|
mov esi, [int_div+4*esi] |
279 |
|
imul ecx, esi |
280 |
|
sar ecx, 17 |
281 |
|
lea ebx, [byte ecx + 1] |
282 |
|
cmovs ecx, ebx |
283 |
|
; idiv cx ; ecx = edi:ecx / dcscalar |
284 |
|
|
285 |
|
mov ebx, [esp] |
286 |
|
mov edi, [esp+4] |
287 |
|
mov esi, [esp+8] |
288 |
|
add esp, byte 12 |
289 |
|
mov [edx], cx ; coeff[0] = ax |
290 |
|
|
291 |
|
ret |
292 |
|
|
293 |
|
align ALIGN |
294 |
|
.q1loop |
295 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
296 |
|
psubw mm0, mm1 ;-mm1 |
297 |
|
movq mm4, [eax + 8*esi+120] ; |
298 |
|
psubw mm3, mm4 ;-mm4 |
299 |
|
pmaxsw mm0, mm1 ;|src| |
300 |
|
pmaxsw mm3, mm4 |
301 |
|
nop2 |
302 |
|
psraw mm1, 15 ;sign src |
303 |
|
psraw mm4, 15 |
304 |
|
psllw mm0, 4 ; level << 4 |
305 |
|
psllw mm3, 4 |
306 |
|
paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided |
307 |
|
paddw mm3, [intra_matrix1 + 8*esi+120] ;intra1 contains fix for division by 1 |
308 |
|
movq mm5, [intra_matrix_fixl + 8*esi+112] ;with rounding down |
309 |
|
movq mm7, [intra_matrix_fixl + 8*esi+120] |
310 |
|
pmulhuw mm5, mm0 |
311 |
|
pmulhuw mm7, mm3 ;mm7: first approx of division |
312 |
|
mov esp, esp |
313 |
|
movq mm2, [intra_matrix + 8*esi+112] |
314 |
|
movq mm6, [intra_matrix + 8*esi+120] ; divs for q<=16 |
315 |
|
pmullw mm2, mm5 ;test value <= original |
316 |
|
pmullw mm6, mm7 |
317 |
|
psubw mm0, mm2 ;mismatch |
318 |
|
psubw mm3, mm6 |
319 |
|
nop4 |
320 |
|
movq mm2, [quantd + ecx * 8 - 8] |
321 |
|
paddw mm5, mm2 ;first approx with quantd |
322 |
|
paddw mm7, mm2 |
323 |
|
mov esp, esp |
324 |
|
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
325 |
|
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
326 |
|
paddw mm5, mm0 ;final result with quantd |
327 |
|
paddw mm7, mm3 |
328 |
|
movq mm0, [edi] |
329 |
|
movq mm3, [edi] |
330 |
|
mov esp, esp |
331 |
|
psrlw mm5, 1 ; (level + quantd) /2 (quant = 1) |
332 |
|
psrlw mm7, 1 |
333 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
334 |
|
pxor mm7, mm4 ; |
335 |
|
psubw mm5, mm1 ; undisplace |
336 |
|
psubw mm7, mm4 ; |
337 |
|
movq [edx + 8*esi+112], mm5 |
338 |
|
movq [edx + 8*esi +120], mm7 |
339 |
|
add esi, byte 2 |
340 |
|
jng near .q1loop |
341 |
|
jmp near .done |
342 |
|
|
343 |
|
align 8 |
344 |
|
.lloop |
345 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
346 |
|
psubw mm0, mm1 ;-mm1 |
347 |
|
movq mm4, [eax + 8*esi+120] ; |
348 |
|
psubw mm3, mm4 ;-mm4 |
349 |
|
pmaxsw mm0, mm1 ;|src| |
350 |
|
pmaxsw mm3, mm4 |
351 |
|
nop2 |
352 |
|
psraw mm1, 15 ;sign src |
353 |
|
psraw mm4, 15 |
354 |
|
psllw mm0, 4 ; level << 4 |
355 |
|
psllw mm3, 4 ; |
356 |
|
paddw mm0, [intra_matrix1 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1 |
357 |
|
paddw mm3, [intra_matrix1 + 8*esi+120] |
358 |
|
movq mm5, [intra_matrix_fixl + 8*esi+112] |
359 |
|
movq mm7, [intra_matrix_fixl + 8*esi+120] |
360 |
|
pmulhuw mm5, mm0 |
361 |
|
pmulhuw mm7, mm3 ;mm7: first approx of division |
362 |
|
mov esp, esp |
363 |
|
movq mm2, [intra_matrix + 8*esi+112] |
364 |
|
movq mm6, [intra_matrix + 8*esi+120] |
365 |
|
pmullw mm2, mm5 ;test value <= original |
366 |
|
pmullw mm6, mm7 |
367 |
|
psubw mm0, mm2 ;mismatch |
368 |
|
psubw mm3, mm6 |
369 |
|
nop4 |
370 |
|
movq mm2, [quantd + ecx * 8 - 8] |
371 |
|
movq mm6, [mmx_div + ecx * 8 - 8] ; divs for q<=16 |
372 |
|
paddw mm5, mm2 ;first approx with quantd |
373 |
|
paddw mm7, mm2 |
374 |
|
mov esp, esp |
375 |
|
pmulhuw mm0, [intra_matrix_fix + 8*esi+112] ;correction |
376 |
|
pmulhuw mm3, [intra_matrix_fix + 8*esi+120] |
377 |
|
paddw mm5, mm0 ;final result with quantd |
378 |
|
paddw mm7, mm3 |
379 |
|
movq mm0, [edi] |
380 |
|
movq mm3, [edi] |
381 |
|
mov esp, esp |
382 |
|
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
383 |
|
pmulhuw mm7, mm6 ; (level + quantd) / quant (0<quant<32) |
384 |
|
psrlw mm5, 1 ; (level + quantd) / (2*quant) |
385 |
|
psrlw mm7, 1 |
386 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
387 |
|
pxor mm7, mm4 ; |
388 |
|
psubw mm5, mm1 ; undisplace |
389 |
|
psubw mm7, mm4 ; |
390 |
|
movq [edx + 8*esi+112], mm5 |
391 |
|
movq [edx + 8*esi +120], mm7 |
392 |
|
add esi,byte 2 |
393 |
|
jng near .lloop |
394 |
|
jmp near .done |
395 |
|
|
396 |
|
;=========================================================================== |
397 |
|
; |
398 |
|
; uint32_t quant4_inter_xmm(int16_t * coeff, |
399 |
|
; const int16_t const * data, |
400 |
|
; const uint32_t quant); |
401 |
|
; |
402 |
|
;=========================================================================== |
403 |
|
|
404 |
|
align ALIGN |
405 |
|
cglobal quant_mpeg_inter_xmm |
406 |
|
quant_mpeg_inter_xmm: |
407 |
|
mov eax, [esp + 8] ; data |
408 |
|
mov ecx, [esp + 12] ; quant |
409 |
|
mov edx, [esp + 4] ; coeff |
410 |
|
push esi |
411 |
|
push edi |
412 |
|
push ebx |
413 |
|
nop |
414 |
|
mov edi,mmzero |
415 |
|
mov esi,-14 |
416 |
|
mov ebx,esp |
417 |
|
sub esp,byte 24 |
418 |
|
lea ebx,[esp+8] |
419 |
|
and ebx,byte -8 ;align 8 |
420 |
|
pxor mm0,mm0 |
421 |
|
pxor mm3,mm3 |
422 |
|
movq [byte ebx],mm0 |
423 |
|
db 0Fh, 7Fh, 44h, 23h, 8 ;movq [ebx+8],mm0 |
424 |
|
cmp ecx,byte 1 |
425 |
|
je near .q1loop |
426 |
|
cmp ecx,byte 19 |
427 |
|
jg near .lloop |
428 |
|
nop |
429 |
|
|
430 |
|
align ALIGN |
431 |
|
.loop |
432 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
433 |
|
psubw mm0,mm1 ;-mm1 |
434 |
|
movq mm4, [eax + 8*esi + 120] ; |
435 |
|
psubw mm3,mm4 ;-mm4 |
436 |
|
pmaxsw mm0,mm1 ;|src| |
437 |
|
pmaxsw mm3,mm4 |
438 |
|
nop2 |
439 |
|
psraw mm1,15 ;sign src |
440 |
|
psraw mm4,15 |
441 |
|
psllw mm0, 4 ; level << 4 |
442 |
|
psllw mm3, 4 ; |
443 |
|
paddw mm0, [inter_matrix1 + 8*esi+112] |
444 |
|
paddw mm3, [inter_matrix1 + 8*esi+120] |
445 |
|
movq mm5,[inter_matrix_fixl + 8*esi+112] |
446 |
|
movq mm7,[inter_matrix_fixl + 8*esi+120] |
447 |
|
pmulhuw mm5,mm0 |
448 |
|
pmulhuw mm7,mm3 |
449 |
|
mov esp,esp |
450 |
|
movq mm2,[inter_matrix + 8*esi+112] |
451 |
|
movq mm6,[inter_matrix + 8*esi+120] |
452 |
|
pmullw mm2,mm5 |
453 |
|
pmullw mm6,mm7 |
454 |
|
psubw mm0,mm2 |
455 |
|
psubw mm3,mm6 |
456 |
|
movq mm2,[byte ebx] |
457 |
|
movq mm6,[mmx_divs + ecx * 8 - 8] |
458 |
|
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] |
459 |
|
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
460 |
|
paddw mm2,[ebx+8] ;sum |
461 |
|
paddw mm5,mm0 |
462 |
|
paddw mm7,mm3 |
463 |
|
movq mm0,[edi] |
464 |
|
movq mm3,[edi] |
465 |
|
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
466 |
|
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
467 |
|
add esi,byte 2 |
468 |
|
paddw mm2,mm5 ;sum += x1 |
469 |
|
movq [ebx],mm7 ;store x2 |
470 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
471 |
|
pxor mm7, mm4 ; |
472 |
|
psubw mm5, mm1 ; undisplace |
473 |
|
psubw mm7, mm4 ; |
474 |
|
db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum |
475 |
|
movq [edx + 8*esi+112-16], mm5 |
476 |
|
movq [edx + 8*esi +120-16], mm7 |
477 |
|
jng near .loop |
478 |
|
|
479 |
|
.done |
480 |
|
; calculate data[0] // (int32_t)dcscalar) |
481 |
|
paddw mm2,[ebx] |
482 |
|
mov ebx,[esp+24] |
483 |
|
mov edi,[esp+4+24] |
484 |
|
mov esi,[esp+8+24] |
485 |
|
add esp,byte 12+24 |
486 |
|
pmaddwd mm2, [mmx_one] |
487 |
|
punpckldq mm0,mm2 ;get low dw to mm0:high |
488 |
|
paddd mm0,mm2 |
489 |
|
punpckhdq mm0,mm0 ;get result to low |
490 |
|
movd eax, mm0 |
491 |
|
|
492 |
|
ret |
493 |
|
|
494 |
|
align ALIGN |
495 |
|
.q1loop |
496 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
497 |
|
psubw mm0,mm1 ;-mm1 |
498 |
|
movq mm4, [eax + 8*esi+120] ; |
499 |
|
psubw mm3,mm4 ;-mm4 |
500 |
|
pmaxsw mm0,mm1 ;|src| |
501 |
|
pmaxsw mm3,mm4 |
502 |
|
nop2 |
503 |
|
psraw mm1,15 ;sign src |
504 |
|
psraw mm4,15 |
505 |
|
psllw mm0, 4 ; level << 4 |
506 |
|
psllw mm3, 4 |
507 |
|
paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided |
508 |
|
paddw mm3, [inter_matrix1 + 8*esi+120] ; inter1 contains fix for division by 1 |
509 |
|
movq mm5,[inter_matrix_fixl + 8*esi+112] ;with rounding down |
510 |
|
movq mm7,[inter_matrix_fixl + 8*esi+120] |
511 |
|
pmulhuw mm5,mm0 |
512 |
|
pmulhuw mm7,mm3 ;mm7: first approx of division |
513 |
|
mov esp,esp |
514 |
|
movq mm2,[inter_matrix + 8*esi+112] |
515 |
|
movq mm6,[inter_matrix + 8*esi+120] ; divs for q<=16 |
516 |
|
pmullw mm2,mm5 ;test value <= original |
517 |
|
pmullw mm6,mm7 |
518 |
|
psubw mm0,mm2 ;mismatch |
519 |
|
psubw mm3,mm6 |
520 |
|
movq mm2,[byte ebx] |
521 |
|
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction |
522 |
|
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
523 |
|
paddw mm2,[ebx+8] ;sum |
524 |
|
paddw mm5,mm0 ;final result |
525 |
|
paddw mm7,mm3 |
526 |
|
movq mm0,[edi] |
527 |
|
movq mm3,[edi] |
528 |
|
psrlw mm5, 1 ; (level ) /2 (quant = 1) |
529 |
|
psrlw mm7, 1 |
530 |
|
add esi,byte 2 |
531 |
|
paddw mm2,mm5 ;sum += x1 |
532 |
|
movq [ebx],mm7 ;store x2 |
533 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
534 |
|
pxor mm7, mm4 ; |
535 |
|
psubw mm5, mm1 ; undisplace |
536 |
|
psubw mm7, mm4 ; |
537 |
|
movq [ebx+8],mm2 ;store sum |
538 |
|
movq [edx + 8*esi+112-16], mm5 |
539 |
|
movq [edx + 8*esi +120-16], mm7 |
540 |
|
jng near .q1loop |
541 |
|
jmp near .done |
542 |
|
|
543 |
|
align 8 |
544 |
|
.lloop |
545 |
|
movq mm1, [eax + 8*esi+112] ; mm0 = [1st] |
546 |
|
psubw mm0,mm1 ;-mm1 |
547 |
|
movq mm4, [eax + 8*esi+120] ; |
548 |
|
psubw mm3,mm4 ;-mm4 |
549 |
|
pmaxsw mm0,mm1 ;|src| |
550 |
|
pmaxsw mm3,mm4 |
551 |
|
nop2 |
552 |
|
psraw mm1,15 ;sign src |
553 |
|
psraw mm4,15 |
554 |
|
psllw mm0, 4 ; level << 4 |
555 |
|
psllw mm3, 4 ; |
556 |
|
paddw mm0, [inter_matrix1 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1 |
557 |
|
paddw mm3, [inter_matrix1 + 8*esi+120] |
558 |
|
movq mm5,[inter_matrix_fixl + 8*esi+112] |
559 |
|
movq mm7,[inter_matrix_fixl + 8*esi+120] |
560 |
|
pmulhuw mm5,mm0 |
561 |
|
pmulhuw mm7,mm3 ;mm7: first approx of division |
562 |
|
mov esp,esp |
563 |
|
movq mm2,[inter_matrix + 8*esi+112] |
564 |
|
movq mm6,[inter_matrix + 8*esi+120] |
565 |
|
pmullw mm2,mm5 ;test value <= original |
566 |
|
pmullw mm6,mm7 |
567 |
|
psubw mm0,mm2 ;mismatch |
568 |
|
psubw mm3,mm6 |
569 |
|
movq mm2,[byte ebx] |
570 |
|
movq mm6,[mmx_div + ecx * 8 - 8] ; divs for q<=16 |
571 |
|
pmulhuw mm0,[inter_matrix_fix + 8*esi+112] ;correction |
572 |
|
pmulhuw mm3,[inter_matrix_fix + 8*esi+120] |
573 |
|
paddw mm2,[ebx+8] ;sum |
574 |
|
paddw mm5,mm0 ;final result |
575 |
|
paddw mm7,mm3 |
576 |
|
movq mm0,[edi] |
577 |
|
movq mm3,[edi] |
578 |
|
pmulhuw mm5, mm6 ; mm0 = (mm0 / 2Q) >> 16 |
579 |
|
pmulhuw mm7, mm6 ; (level ) / quant (0<quant<32) |
580 |
|
add esi,byte 2 |
581 |
|
psrlw mm5, 1 ; (level ) / (2*quant) |
582 |
|
paddw mm2,mm5 ;sum += x1 |
583 |
|
psrlw mm7, 1 |
584 |
|
movq [ebx],mm7 ;store x2 |
585 |
|
pxor mm5, mm1 ; mm0 *= sign(mm0) |
586 |
|
pxor mm7, mm4 ; |
587 |
|
psubw mm5, mm1 ; undisplace |
588 |
|
psubw mm7, mm4 ; |
589 |
|
db 0Fh, 7Fh, 54h, 23h, 08 ;movq [ebx+8],mm2 ;store sum |
590 |
|
movq [edx + 8*esi+112-16], mm5 |
591 |
|
movq [edx + 8*esi +120-16], mm7 |
592 |
|
jng near .lloop |
593 |
|
jmp near .done |
594 |
|
|
595 |
|
|
596 |
|
;=========================================================================== |
597 |
|
; |
598 |
|
; void dequant4_intra_mmx(int16_t *data, |
599 |
|
; const int16_t const *coeff, |
600 |
|
; const uint32_t quant, |
601 |
|
; const uint32_t dcscalar); |
602 |
|
; |
603 |
|
;=========================================================================== |
604 |
|
|
605 |
|
; Note: in order to saturate 'easily', we pre-shift the quantifier |
606 |
|
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
607 |
|
; build a saturating mask. It is non-zero only when an overflow occured. |
608 |
|
; We thus avoid packing/unpacking toward double-word. |
609 |
|
; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., |
610 |
|
; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not |
611 |
|
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
612 |
|
; and quant in [1..31]. |
613 |
|
; |
614 |
|
;******************************************************************** |
615 |
|
%macro DEQUANT4INTRAMMX 1 |
616 |
|
movq mm1, [byte ecx+ 16 * %1] ; mm0 = c = coeff[i] |
617 |
|
movq mm4, [ecx+ 16 * %1 +8]; mm3 = c' = coeff[i+1] |
618 |
|
psubw mm0,mm1 |
619 |
|
psubw mm3,mm4 |
620 |
|
pmaxsw mm0,mm1 |
621 |
|
pmaxsw mm3,mm4 |
622 |
|
psraw mm1,15 |
623 |
|
psraw mm4,15 |
624 |
|
%if %1 |
625 |
|
movq mm2,[eax+8] ;preshifted quant |
626 |
|
movq mm7,[eax+8] |
627 |
|
%endif |
628 |
|
pmullw mm2, [intra_matrix + 16 * %1 ] ; matrix[i]*quant |
629 |
|
pmullw mm7, [intra_matrix + 16 * %1 +8] ; matrix[i+1]*quant |
630 |
|
movq mm5,mm0 |
631 |
|
movq mm6,mm3 |
632 |
|
pmulhw mm0, mm2 ; high of coeff*(matrix*quant) |
633 |
|
pmulhw mm3, mm7 ; high of coeff*(matrix*quant) |
634 |
|
pmullw mm2, mm5 ; low of coeff*(matrix*quant) |
635 |
|
pmullw mm7, mm6 ; low of coeff*(matrix*quant) |
636 |
|
pcmpgtw mm0, [eax] |
637 |
|
pcmpgtw mm3, [eax] |
638 |
|
paddusw mm2, mm0 |
639 |
|
paddusw mm7, mm3 |
640 |
|
psrlw mm2, 5 |
641 |
|
psrlw mm7, 5 |
642 |
|
pxor mm2, mm1 ; start negating back |
643 |
|
pxor mm7, mm4 ; start negating back |
644 |
|
psubusw mm1, mm0 |
645 |
|
psubusw mm4, mm3 |
646 |
|
movq mm0,[eax] ;zero |
647 |
|
movq mm3,[eax] ;zero |
648 |
|
psubw mm2, mm1 ; finish negating back |
649 |
|
psubw mm7, mm4 ; finish negating back |
650 |
|
movq [byte edx + 16 * %1], mm2 ; data[i] |
651 |
|
movq [edx + 16 * %1 +8], mm7 ; data[i+1] |
652 |
|
%endmacro |
653 |
|
|
654 |
|
align 16 |
655 |
|
cglobal dequant_mpeg_intra_3dne |
656 |
|
dequant_mpeg_intra_3dne: |
657 |
|
mov eax, [esp+12] ; quant |
658 |
|
mov ecx, [esp+8] ; coeff |
659 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
660 |
|
psllw mm7, 2 ; << 2. See comment. |
661 |
|
mov edx, [esp+4] ; data |
662 |
|
push ebx |
663 |
|
movsx ebx,word [ecx] |
664 |
|
pxor mm0, mm0 |
665 |
|
pxor mm3, mm3 |
666 |
|
push esi |
667 |
|
lea eax,[esp-28] |
668 |
|
sub esp,byte 32 |
669 |
|
and eax,byte -8 ;points to qword aligned space on stack |
670 |
|
movq [eax],mm0 |
671 |
|
movq [eax+8],mm7 |
672 |
|
imul ebx,[esp+16+8+32] ; dcscalar |
673 |
|
movq mm2,mm7 |
674 |
|
|
675 |
|
|
676 |
|
align 4 |
677 |
|
|
678 |
|
DEQUANT4INTRAMMX 0 |
679 |
|
|
680 |
|
mov esi,-2048 |
681 |
|
nop |
682 |
|
cmp ebx,esi |
683 |
|
|
684 |
|
DEQUANT4INTRAMMX 1 |
685 |
|
|
686 |
|
cmovl ebx, esi |
687 |
|
neg esi |
688 |
|
sub esi, byte 1 ;2047 |
689 |
|
|
690 |
|
DEQUANT4INTRAMMX 2 |
691 |
|
|
692 |
|
cmp ebx, esi |
693 |
|
cmovg ebx, esi |
694 |
|
lea ebp, [byte ebp] |
695 |
|
|
696 |
|
DEQUANT4INTRAMMX 3 |
697 |
|
|
698 |
|
mov esi, [esp+32] |
699 |
|
mov [byte edx], bx |
700 |
|
mov ebx, [esp+32+4] |
701 |
|
|
702 |
|
DEQUANT4INTRAMMX 4 |
703 |
|
DEQUANT4INTRAMMX 5 |
704 |
|
DEQUANT4INTRAMMX 6 |
705 |
|
DEQUANT4INTRAMMX 7 |
706 |
|
|
707 |
|
add esp, byte 32+8 |
708 |
|
|
709 |
|
ret |
710 |
|
|
711 |
|
;=========================================================================== |
712 |
|
; |
713 |
|
; void dequant4_inter_3dne(int16_t * data, |
714 |
|
; const int16_t * const coeff, |
715 |
|
; const uint32_t quant); |
716 |
|
; |
717 |
|
;=========================================================================== |
718 |
|
|
719 |
|
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
720 |
|
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
721 |
|
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
722 |
|
; It's mixed with the extraction of the absolute value. |
723 |
|
|
724 |
|
align 16 |
725 |
|
cglobal dequant_mpeg_inter_3dne |
726 |
|
dequant_mpeg_inter_3dne: |
727 |
|
mov edx, [esp+ 4] ; data |
728 |
|
mov ecx, [esp+ 8] ; coeff |
729 |
|
mov eax, [esp+12] ; quant |
730 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
731 |
|
mov eax, -14 |
732 |
|
paddw mm7, mm7 ; << 1 |
733 |
|
pxor mm6, mm6 ; mismatch sum |
734 |
|
push esi |
735 |
|
mov esi,mmzero |
736 |
|
pxor mm1,mm1 |
737 |
|
pxor mm3,mm3 |
738 |
|
nop |
739 |
|
nop4 |
740 |
|
|
741 |
|
align 16 |
742 |
|
.loop |
743 |
|
movq mm0, [ecx+8*eax + 7*16 ] ; mm0 = coeff[i] |
744 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
745 |
|
movq mm2, [ecx+8*eax + 7*16 +8] ; mm2 = coeff[i+1] |
746 |
|
pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) |
747 |
|
paddsw mm0, mm1 ; c += sgn(c) |
748 |
|
paddsw mm2, mm3 ; c += sgn(c') |
749 |
|
paddw mm0, mm0 ; c *= 2 |
750 |
|
paddw mm2, mm2 ; c'*= 2 |
751 |
|
|
752 |
|
movq mm4, [esi] |
753 |
|
movq mm5, [esi] |
754 |
|
psubw mm4, mm0 ; -c |
755 |
|
psubw mm5, mm2 ; -c' |
756 |
|
|
757 |
|
psraw mm4, 16 ; mm4 = sgn(-c) |
758 |
|
psraw mm5, 16 ; mm5 = sgn(-c') |
759 |
|
psubsw mm0, mm4 ; c -= sgn(-c) |
760 |
|
psubsw mm2, mm5 ; c' -= sgn(-c') |
761 |
|
pxor mm0, mm1 ; finish changing sign if needed |
762 |
|
pxor mm2, mm3 ; finish changing sign if needed |
763 |
|
|
764 |
|
; we're short on register, here. Poor pairing... |
765 |
|
|
766 |
|
movq mm4, mm7 ; (matrix*quant) |
767 |
|
nop |
768 |
|
pmullw mm4, [inter_matrix + 8*eax + 7*16] |
769 |
|
movq mm5, mm4 |
770 |
|
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
771 |
|
pmullw mm0, mm4 ; low of c*(matrix*quant) |
772 |
|
|
773 |
|
movq mm4, mm7 ; (matrix*quant) |
774 |
|
pmullw mm4, [inter_matrix + 8*eax + 7*16 + 8] |
775 |
|
add eax,byte 2 |
776 |
|
|
777 |
|
pcmpgtw mm5, [esi] |
778 |
|
paddusw mm0, mm5 |
779 |
|
psrlw mm0, 5 |
780 |
|
pxor mm0, mm1 ; start restoring sign |
781 |
|
psubusw mm1, mm5 |
782 |
|
|
783 |
|
movq mm5, mm4 |
784 |
|
pmulhw mm5, mm2 ; high of c*(matrix*quant) |
785 |
|
pmullw mm2, mm4 ; low of c*(matrix*quant) |
786 |
|
psubw mm0, mm1 ; finish restoring sign |
787 |
|
|
788 |
|
pcmpgtw mm5, [esi] |
789 |
|
paddusw mm2, mm5 |
790 |
|
psrlw mm2, 5 |
791 |
|
pxor mm2, mm3 ; start restoring sign |
792 |
|
psubusw mm3, mm5 |
793 |
|
psubw mm2, mm3 ; finish restoring sign |
794 |
|
movq mm1, [esi] |
795 |
|
movq mm3, [byte esi] |
796 |
|
pxor mm6, mm0 ; mismatch control |
797 |
|
movq [edx + 8*eax + 7*16 -2*8 ], mm0 ; data[i] |
798 |
|
pxor mm6, mm2 ; mismatch control |
799 |
|
movq [edx + 8*eax + 7*16 -2*8 +8], mm2 ; data[i+1] |
800 |
|
|
801 |
|
jng .loop |
802 |
|
nop |
803 |
|
|
804 |
|
; mismatch control |
805 |
|
|
806 |
|
pshufw mm0,mm6,01010101b |
807 |
|
pshufw mm1,mm6,10101010b |
808 |
|
pshufw mm2,mm6,11111111b |
809 |
|
pxor mm6, mm0 |
810 |
|
pxor mm1, mm2 |
811 |
|
pxor mm6, mm1 |
812 |
|
movd eax, mm6 |
813 |
|
and eax,byte 1 |
814 |
|
xor eax,byte 1 |
815 |
|
mov esi,[esp] |
816 |
|
add esp,byte 4 |
817 |
|
xor word [edx + 2*63], ax |
818 |
|
|
819 |
|
ret |