1 |
|
;/***************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * - MPEG4 Quantization H263 implementation / MMX optimized - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
7 |
|
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
8 |
|
; * |
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 |
|
; ****************************************************************************/ |
26 |
|
|
27 |
|
; enable dequant saturate [-2048,2047], test purposes only. |
28 |
|
%define SATURATE |
29 |
|
|
30 |
|
BITS 32 |
31 |
|
|
32 |
|
%macro cglobal 1 |
33 |
|
%ifdef PREFIX |
34 |
|
global _%1 |
35 |
|
%define %1 _%1 |
36 |
|
%else |
37 |
|
global %1 |
38 |
|
%endif |
39 |
|
%endmacro |
40 |
|
|
41 |
|
;============================================================================= |
42 |
|
; Read only Local data |
43 |
|
;============================================================================= |
44 |
|
|
45 |
|
SECTION .rodata |
46 |
|
|
47 |
|
ALIGN 16 |
48 |
|
plus_one: |
49 |
|
times 8 dw 1 |
50 |
|
|
51 |
|
;----------------------------------------------------------------------------- |
52 |
|
; |
53 |
|
; subtract by Q/2 table |
54 |
|
; |
55 |
|
;----------------------------------------------------------------------------- |
56 |
|
|
57 |
|
ALIGN 16 |
58 |
|
mmx_sub: |
59 |
|
%assign quant 1 |
60 |
|
%rep 31 |
61 |
|
times 4 dw quant / 2 |
62 |
|
%assign quant quant+1 |
63 |
|
%endrep |
64 |
|
|
65 |
|
;----------------------------------------------------------------------------- |
66 |
|
; |
67 |
|
; divide by 2Q table |
68 |
|
; |
69 |
|
; use a shift of 16 to take full advantage of _pmulhw_ |
70 |
|
; for q=1, _pmulhw_ will overflow so it is treated seperately |
71 |
|
; (3dnow2 provides _pmulhuw_ which wont cause overflow) |
72 |
|
; |
73 |
|
;----------------------------------------------------------------------------- |
74 |
|
|
75 |
|
ALIGN 16 |
76 |
|
mmx_div: |
77 |
|
%assign quant 1 |
78 |
|
%rep 31 |
79 |
|
times 4 dw (1<<16) / (quant*2) + 1 |
80 |
|
%assign quant quant+1 |
81 |
|
%endrep |
82 |
|
|
83 |
|
;----------------------------------------------------------------------------- |
84 |
|
; |
85 |
|
; add by (odd(Q) ? Q : Q - 1) table |
86 |
|
; |
87 |
|
;----------------------------------------------------------------------------- |
88 |
|
|
89 |
|
ALIGN 16 |
90 |
|
mmx_add: |
91 |
|
%assign quant 1 |
92 |
|
%rep 31 |
93 |
|
%if quant % 2 != 0 |
94 |
|
times 4 dw quant |
95 |
|
%else |
96 |
|
times 4 dw quant - 1 |
97 |
|
%endif |
98 |
|
%assign quant quant+1 |
99 |
|
%endrep |
100 |
|
|
101 |
|
;----------------------------------------------------------------------------- |
102 |
|
; |
103 |
|
; multiple by 2Q table |
104 |
|
; |
105 |
|
;----------------------------------------------------------------------------- |
106 |
|
|
107 |
|
ALIGN 16 |
108 |
|
mmx_mul: |
109 |
|
%assign quant 1 |
110 |
|
%rep 31 |
111 |
|
times 4 dw quant*2 |
112 |
|
%assign quant quant+1 |
113 |
|
%endrep |
114 |
|
|
115 |
|
;----------------------------------------------------------------------------- |
116 |
|
; |
117 |
|
; saturation limits |
118 |
|
; |
119 |
|
;----------------------------------------------------------------------------- |
120 |
|
|
121 |
|
ALIGN 16 |
122 |
|
sse2_2047: |
123 |
|
times 8 dw 2047 |
124 |
|
|
125 |
|
ALIGN 16 |
126 |
|
mmx_2047: |
127 |
|
times 4 dw 2047 |
128 |
|
|
129 |
|
ALIGN 8 |
130 |
|
mmx_32768_minus_2048: |
131 |
|
times 4 dw (32768-2048) |
132 |
|
|
133 |
|
mmx_32767_minus_2047: |
134 |
|
times 4 dw (32767-2047) |
135 |
|
|
136 |
|
|
137 |
|
;============================================================================= |
138 |
|
; Code |
139 |
|
;============================================================================= |
140 |
|
|
141 |
|
SECTION .text |
142 |
|
|
143 |
|
cglobal quant_h263_intra_mmx |
144 |
|
cglobal quant_h263_intra_sse2 |
145 |
|
cglobal quant_h263_inter_mmx |
146 |
|
cglobal quant_h263_inter_sse2 |
147 |
|
cglobal dequant_h263_intra_mmx |
148 |
|
cglobal dequant_h263_intra_xmm |
149 |
|
cglobal dequant_h263_intra_sse2 |
150 |
|
cglobal dequant_h263_inter_mmx |
151 |
|
cglobal dequant_h263_inter_xmm |
152 |
|
cglobal dequant_h263_inter_sse2 |
153 |
|
|
154 |
|
;----------------------------------------------------------------------------- |
155 |
|
; |
156 |
|
; uint32_t quant_h263_intra_mmx(int16_t * coeff, |
157 |
|
; const int16_t const * data, |
158 |
|
; const uint32_t quant, |
159 |
|
; const uint32_t dcscalar); |
160 |
|
; |
161 |
|
;----------------------------------------------------------------------------- |
162 |
|
|
163 |
|
ALIGN 16 |
164 |
|
quant_h263_intra_mmx: |
165 |
|
|
166 |
|
push ecx |
167 |
|
push esi |
168 |
|
push edi |
169 |
|
|
170 |
|
mov edi, [esp + 12 + 4] ; coeff |
171 |
|
mov esi, [esp + 12 + 8] ; data |
172 |
|
mov eax, [esp + 12 + 12] ; quant |
173 |
|
|
174 |
|
xor ecx, ecx |
175 |
|
cmp al, 1 |
176 |
|
jz .q1loop |
177 |
|
|
178 |
|
movq mm7, [mmx_div + eax * 8 - 8] |
179 |
|
|
180 |
|
ALIGN 16 |
181 |
|
.loop |
182 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
183 |
|
movq mm3, [esi + 8*ecx + 8] |
184 |
|
pxor mm1, mm1 ; mm1 = 0 |
185 |
|
pxor mm4, mm4 ; |
186 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
187 |
|
pcmpgtw mm4, mm3 ; |
188 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
189 |
|
pxor mm3, mm4 ; |
190 |
|
psubw mm0, mm1 ; displace |
191 |
|
psubw mm3, mm4 ; |
192 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
193 |
|
pmulhw mm3, mm7 ; |
194 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
195 |
|
pxor mm3, mm4 ; |
196 |
|
psubw mm0, mm1 ; undisplace |
197 |
|
psubw mm3, mm4 ; |
198 |
|
movq [edi + 8*ecx], mm0 |
199 |
|
movq [edi + 8*ecx + 8], mm3 |
200 |
|
|
201 |
|
add ecx, 2 |
202 |
|
cmp ecx, 16 |
203 |
|
jnz .loop |
204 |
|
|
205 |
|
.done |
206 |
|
|
207 |
|
; caclulate data[0] // (int32_t)dcscalar) |
208 |
|
mov ecx, [esp + 12 + 16] ; dcscalar |
209 |
|
mov edx, ecx |
210 |
|
movsx eax, word [esi] ; data[0] |
211 |
|
shr edx, 1 ; edx = dcscalar /2 |
212 |
|
cmp eax, 0 |
213 |
|
jg .gtzero |
214 |
|
|
215 |
|
sub eax, edx |
216 |
|
jmp short .mul |
217 |
|
|
218 |
|
.gtzero |
219 |
|
add eax, edx |
220 |
|
.mul |
221 |
|
cdq ; expand eax -> edx:eax |
222 |
|
idiv ecx ; eax = edx:eax / dcscalar |
223 |
|
mov [edi], ax ; coeff[0] = ax |
224 |
|
|
225 |
|
xor eax, eax ; return(0); |
226 |
|
pop edi |
227 |
|
pop esi |
228 |
|
pop ecx |
229 |
|
|
230 |
|
ret |
231 |
|
|
232 |
|
ALIGN 16 |
233 |
|
.q1loop |
234 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
235 |
|
movq mm3, [esi + 8*ecx + 8] |
236 |
|
pxor mm1, mm1 ; mm1 = 0 |
237 |
|
pxor mm4, mm4 ; |
238 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
239 |
|
pcmpgtw mm4, mm3 ; |
240 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
241 |
|
pxor mm3, mm4 ; |
242 |
|
psubw mm0, mm1 ; displace |
243 |
|
psubw mm3, mm4 ; |
244 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
245 |
|
psrlw mm3, 1 ; |
246 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
247 |
|
pxor mm3, mm4 |
248 |
|
psubw mm0, mm1 ; undisplace |
249 |
|
psubw mm3, mm4 ; |
250 |
|
movq [edi + 8*ecx], mm0 |
251 |
|
movq [edi + 8*ecx + 8], mm3 |
252 |
|
|
253 |
|
add ecx, 2 |
254 |
|
cmp ecx, 16 |
255 |
|
jnz .q1loop |
256 |
|
jmp short .done |
257 |
|
|
258 |
|
|
259 |
|
|
260 |
|
;----------------------------------------------------------------------------- |
261 |
|
; |
262 |
|
; uint32_t quant_h263_intra_sse2(int16_t * coeff, |
263 |
|
; const int16_t const * data, |
264 |
|
; const uint32_t quant, |
265 |
|
; const uint32_t dcscalar); |
266 |
|
; |
267 |
|
;----------------------------------------------------------------------------- |
268 |
|
|
269 |
|
ALIGN 16 |
270 |
|
quant_h263_intra_sse2: |
271 |
|
|
272 |
|
push esi |
273 |
|
push edi |
274 |
|
|
275 |
|
mov edi, [esp + 8 + 4] ; coeff |
276 |
|
mov esi, [esp + 8 + 8] ; data |
277 |
|
mov eax, [esp + 8 + 12] ; quant |
278 |
|
|
279 |
|
xor ecx, ecx |
280 |
|
cmp al, 1 |
281 |
|
jz near .qas2_q1loop |
282 |
|
|
283 |
|
.qas2_not1 |
284 |
|
movq mm7, [mmx_div + eax*8 - 8] |
285 |
|
movq2dq xmm7, mm7 |
286 |
|
movlhps xmm7, xmm7 |
287 |
|
|
288 |
|
ALIGN 16 |
289 |
|
.qas2_loop |
290 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
291 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
292 |
|
pxor xmm1, xmm1 |
293 |
|
pxor xmm4, xmm4 |
294 |
|
pcmpgtw xmm1, xmm0 |
295 |
|
pcmpgtw xmm4, xmm3 |
296 |
|
pxor xmm0, xmm1 |
297 |
|
pxor xmm3, xmm4 |
298 |
|
psubw xmm0, xmm1 |
299 |
|
psubw xmm3, xmm4 |
300 |
|
pmulhw xmm0, xmm7 |
301 |
|
pmulhw xmm3, xmm7 |
302 |
|
pxor xmm0, xmm1 |
303 |
|
pxor xmm3, xmm4 |
304 |
|
psubw xmm0, xmm1 |
305 |
|
psubw xmm3, xmm4 |
306 |
|
movdqa [edi + ecx*8], xmm0 |
307 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
308 |
|
|
309 |
|
add ecx, 4 |
310 |
|
cmp ecx, 16 |
311 |
|
jnz .qas2_loop |
312 |
|
|
313 |
|
.qas2_done |
314 |
|
mov ecx, [esp + 8 + 16] ; dcscalar |
315 |
|
mov edx, ecx |
316 |
|
movsx eax, word [esi] |
317 |
|
shr edx, 1 |
318 |
|
cmp eax, 0 |
319 |
|
jg .qas2_gtzero |
320 |
|
|
321 |
|
sub eax, edx |
322 |
|
jmp short .qas2_mul |
323 |
|
|
324 |
|
.qas2_gtzero |
325 |
|
add eax, edx |
326 |
|
|
327 |
|
.qas2_mul |
328 |
|
cdq |
329 |
|
idiv ecx |
330 |
|
|
331 |
|
mov [edi], ax |
332 |
|
|
333 |
|
xor eax, eax ; return(0); |
334 |
|
pop edi |
335 |
|
pop esi |
336 |
|
|
337 |
|
ret |
338 |
|
|
339 |
|
ALIGN 16 |
340 |
|
.qas2_q1loop |
341 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
342 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
343 |
|
pxor xmm1, xmm1 |
344 |
|
pxor xmm4, xmm4 |
345 |
|
pcmpgtw xmm1, xmm0 |
346 |
|
pcmpgtw xmm4, xmm3 |
347 |
|
pxor xmm0, xmm1 |
348 |
|
pxor xmm3, xmm4 |
349 |
|
psubw xmm0, xmm1 |
350 |
|
psubw xmm3, xmm4 |
351 |
|
psrlw xmm0, 1 |
352 |
|
psrlw xmm3, 1 |
353 |
|
pxor xmm0, xmm1 |
354 |
|
pxor xmm3, xmm4 |
355 |
|
psubw xmm0, xmm1 |
356 |
|
psubw xmm3, xmm4 |
357 |
|
movdqa [edi + ecx*8], xmm0 |
358 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
359 |
|
|
360 |
|
add ecx, 4 |
361 |
|
cmp ecx, 16 |
362 |
|
jnz .qas2_q1loop |
363 |
|
jmp near .qas2_done |
364 |
|
|
365 |
|
|
366 |
|
|
367 |
|
;----------------------------------------------------------------------------- |
368 |
|
; |
369 |
|
; uint32_t quant_h263_inter_mmx(int16_t * coeff, |
370 |
|
; const int16_t const * data, |
371 |
|
; const uint32_t quant); |
372 |
|
; |
373 |
|
;----------------------------------------------------------------------------- |
374 |
|
|
375 |
|
ALIGN 16 |
376 |
|
quant_h263_inter_mmx: |
377 |
|
|
378 |
|
push ecx |
379 |
|
push esi |
380 |
|
push edi |
381 |
|
|
382 |
|
mov edi, [esp + 12 + 4] ; coeff |
383 |
|
mov esi, [esp + 12 + 8] ; data |
384 |
|
mov eax, [esp + 12 + 12] ; quant |
385 |
|
|
386 |
|
xor ecx, ecx |
387 |
|
|
388 |
|
pxor mm5, mm5 ; sum |
389 |
|
movq mm6, [mmx_sub + eax * 8 - 8] ; sub |
390 |
|
|
391 |
|
cmp al, 1 |
392 |
|
jz .q1loop |
393 |
|
|
394 |
|
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
395 |
|
|
396 |
|
ALIGN 8 |
397 |
|
.loop |
398 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
399 |
|
movq mm3, [esi + 8*ecx + 8] |
400 |
|
pxor mm1, mm1 ; mm1 = 0 |
401 |
|
pxor mm4, mm4 ; |
402 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
403 |
|
pcmpgtw mm4, mm3 ; |
404 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
405 |
|
pxor mm3, mm4 ; |
406 |
|
psubw mm0, mm1 ; displace |
407 |
|
psubw mm3, mm4 ; |
408 |
|
psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
409 |
|
psubusw mm3, mm6 ; |
410 |
|
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
411 |
|
pmulhw mm3, mm7 ; |
412 |
|
paddw mm5, mm0 ; sum += mm0 |
413 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
414 |
|
paddw mm5, mm3 ; |
415 |
|
pxor mm3, mm4 ; |
416 |
|
psubw mm0, mm1 ; undisplace |
417 |
|
psubw mm3, mm4 |
418 |
|
movq [edi + 8*ecx], mm0 |
419 |
|
movq [edi + 8*ecx + 8], mm3 |
420 |
|
|
421 |
|
add ecx, 2 |
422 |
|
cmp ecx, 16 |
423 |
|
jnz .loop |
424 |
|
|
425 |
|
.done |
426 |
|
pmaddwd mm5, [plus_one] |
427 |
|
movq mm0, mm5 |
428 |
|
psrlq mm5, 32 |
429 |
|
paddd mm0, mm5 |
430 |
|
|
431 |
|
movd eax, mm0 ; return sum |
432 |
|
pop edi |
433 |
|
pop esi |
434 |
|
pop ecx |
435 |
|
|
436 |
|
ret |
437 |
|
|
438 |
|
ALIGN 8 |
439 |
|
.q1loop |
440 |
|
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
441 |
|
movq mm3, [esi + 8*ecx+ 8] ; |
442 |
|
pxor mm1, mm1 ; mm1 = 0 |
443 |
|
pxor mm4, mm4 ; |
444 |
|
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
445 |
|
pcmpgtw mm4, mm3 ; |
446 |
|
pxor mm0, mm1 ; mm0 = |mm0| |
447 |
|
pxor mm3, mm4 ; |
448 |
|
psubw mm0, mm1 ; displace |
449 |
|
psubw mm3, mm4 ; |
450 |
|
psubusw mm0, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
451 |
|
psubusw mm3, mm6 ; |
452 |
|
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
453 |
|
psrlw mm3, 1 ; |
454 |
|
paddw mm5, mm0 ; sum += mm0 |
455 |
|
pxor mm0, mm1 ; mm0 *= sign(mm0) |
456 |
|
paddw mm5, mm3 ; |
457 |
|
pxor mm3, mm4 ; |
458 |
|
psubw mm0, mm1 ; undisplace |
459 |
|
psubw mm3, mm4 |
460 |
|
movq [edi + 8*ecx], mm0 |
461 |
|
movq [edi + 8*ecx + 8], mm3 |
462 |
|
|
463 |
|
add ecx, 2 |
464 |
|
cmp ecx, 16 |
465 |
|
jnz .q1loop |
466 |
|
|
467 |
|
jmp .done |
468 |
|
|
469 |
|
|
470 |
|
|
471 |
|
;----------------------------------------------------------------------------- |
472 |
|
; |
473 |
|
; uint32_t quant_h263_inter_sse2(int16_t * coeff, |
474 |
|
; const int16_t const * data, |
475 |
|
; const uint32_t quant); |
476 |
|
; |
477 |
|
;----------------------------------------------------------------------------- |
478 |
|
|
479 |
|
ALIGN 16 |
480 |
|
quant_h263_inter_sse2: |
481 |
|
|
482 |
|
push esi |
483 |
|
push edi |
484 |
|
|
485 |
|
mov edi, [esp + 8 + 4] ; coeff |
486 |
|
mov esi, [esp + 8 + 8] ; data |
487 |
|
mov eax, [esp + 8 + 12] ; quant |
488 |
|
|
489 |
|
xor ecx, ecx |
490 |
|
|
491 |
|
pxor xmm5, xmm5 ; sum |
492 |
|
|
493 |
|
movq mm0, [mmx_sub + eax*8 - 8] ; sub |
494 |
|
movq2dq xmm6, mm0 ; load into low 8 bytes |
495 |
|
movlhps xmm6, xmm6 ; duplicate into high 8 bytes |
496 |
|
|
497 |
|
cmp al, 1 |
498 |
|
jz near .qes2_q1loop |
499 |
|
|
500 |
|
.qes2_not1 |
501 |
|
movq mm0, [mmx_div + eax*8 - 8] ; divider |
502 |
|
movq2dq xmm7, mm0 |
503 |
|
movlhps xmm7, xmm7 |
504 |
|
|
505 |
|
ALIGN 16 |
506 |
|
.qes2_loop |
507 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
508 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
509 |
|
pxor xmm1, xmm1 |
510 |
|
pxor xmm4, xmm4 |
511 |
|
pcmpgtw xmm1, xmm0 |
512 |
|
pcmpgtw xmm4, xmm3 |
513 |
|
pxor xmm0, xmm1 |
514 |
|
pxor xmm3, xmm4 |
515 |
|
psubw xmm0, xmm1 |
516 |
|
psubw xmm3, xmm4 |
517 |
|
psubusw xmm0, xmm6 |
518 |
|
psubusw xmm3, xmm6 |
519 |
|
pmulhw xmm0, xmm7 |
520 |
|
pmulhw xmm3, xmm7 |
521 |
|
paddw xmm5, xmm0 |
522 |
|
pxor xmm0, xmm1 |
523 |
|
paddw xmm5, xmm3 |
524 |
|
pxor xmm3, xmm4 |
525 |
|
psubw xmm0, xmm1 |
526 |
|
psubw xmm3, xmm4 |
527 |
|
movdqa [edi + ecx*8], xmm0 |
528 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
529 |
|
|
530 |
|
add ecx, 4 |
531 |
|
cmp ecx, 16 |
532 |
|
jnz .qes2_loop |
533 |
|
|
534 |
|
.qes2_done |
535 |
|
movdqu xmm6, [plus_one] |
536 |
|
pmaddwd xmm5, xmm6 |
537 |
|
movhlps xmm6, xmm5 |
538 |
|
paddd xmm5, xmm6 |
539 |
|
movdq2q mm0, xmm5 |
540 |
|
|
541 |
|
movq mm5, mm0 |
542 |
|
psrlq mm5, 32 |
543 |
|
paddd mm0, mm5 |
544 |
|
|
545 |
|
movd eax, mm0 ; return sum |
546 |
|
|
547 |
|
pop edi |
548 |
|
pop esi |
549 |
|
|
550 |
|
ret |
551 |
|
|
552 |
|
ALIGN 16 |
553 |
|
.qes2_q1loop |
554 |
|
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
555 |
|
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
556 |
|
pxor xmm1, xmm1 |
557 |
|
pxor xmm4, xmm4 |
558 |
|
pcmpgtw xmm1, xmm0 |
559 |
|
pcmpgtw xmm4, xmm3 |
560 |
|
pxor xmm0, xmm1 |
561 |
|
pxor xmm3, xmm4 |
562 |
|
psubw xmm0, xmm1 |
563 |
|
psubw xmm3, xmm4 |
564 |
|
psubusw xmm0, xmm6 |
565 |
|
psubusw xmm3, xmm6 |
566 |
|
psrlw xmm0, 1 |
567 |
|
psrlw xmm3, 1 |
568 |
|
paddw xmm5, xmm0 |
569 |
|
pxor xmm0, xmm1 |
570 |
|
paddw xmm5, xmm3 |
571 |
|
pxor xmm3, xmm4 |
572 |
|
psubw xmm0, xmm1 |
573 |
|
psubw xmm3, xmm4 |
574 |
|
movdqa [edi + ecx*8], xmm0 |
575 |
|
movdqa [edi + ecx*8 + 16], xmm3 |
576 |
|
|
577 |
|
add ecx, 4 |
578 |
|
cmp ecx, 16 |
579 |
|
jnz .qes2_q1loop |
580 |
|
jmp .qes2_done |
581 |
|
|
582 |
|
|
583 |
|
;----------------------------------------------------------------------------- |
584 |
|
; |
585 |
|
; uint32_t dequant_h263_intra_mmx(int16_t *data, |
586 |
|
; const int16_t const *coeff, |
587 |
|
; const uint32_t quant, |
588 |
|
; const uint32_t dcscalar); |
589 |
|
; |
590 |
|
;----------------------------------------------------------------------------- |
591 |
|
|
592 |
|
; note: we only saturate to +2047 *before* restoring the sign. |
593 |
|
; Hence, final clamp really is [-2048,2047] |
594 |
|
|
595 |
|
ALIGN 16 |
596 |
|
dequant_h263_intra_mmx: |
597 |
|
|
598 |
|
mov edx, [esp+ 4] ; data |
599 |
|
mov ecx, [esp+ 8] ; coeff |
600 |
|
mov eax, [esp+12] ; quant |
601 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
602 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
603 |
|
mov eax, -16 |
604 |
|
|
605 |
|
ALIGN 16 |
606 |
|
.loop |
607 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
608 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
609 |
|
pxor mm1, mm1 |
610 |
|
pxor mm4, mm4 |
611 |
|
pcmpgtw mm1, mm0 ; sign(c) |
612 |
|
pcmpgtw mm4, mm3 ; sign(c') |
613 |
|
pxor mm2, mm2 |
614 |
|
pxor mm5, mm5 |
615 |
|
pcmpeqw mm2, mm0 ; c is zero |
616 |
|
pcmpeqw mm5, mm3 ; c' is zero |
617 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
618 |
|
pandn mm5, mm6 |
619 |
|
pxor mm0, mm1 ; negate if negative |
620 |
|
pxor mm3, mm4 ; negate if negative |
621 |
|
psubw mm0, mm1 |
622 |
|
psubw mm3, mm4 |
623 |
|
pmullw mm0, mm7 ; *= 2Q |
624 |
|
pmullw mm3, mm7 ; *= 2Q |
625 |
|
paddw mm0, mm2 ; + offset |
626 |
|
paddw mm3, mm5 ; + offset |
627 |
|
paddw mm0, mm1 ; negate back |
628 |
|
paddw mm3, mm4 ; negate back |
629 |
|
|
630 |
|
; saturates to +2047 |
631 |
|
movq mm2, [mmx_32767_minus_2047] |
632 |
|
add eax, 2 |
633 |
|
paddsw mm0, mm2 |
634 |
|
paddsw mm3, mm2 |
635 |
|
psubsw mm0, mm2 |
636 |
|
psubsw mm3, mm2 |
637 |
|
|
638 |
|
pxor mm0, mm1 |
639 |
|
pxor mm3, mm4 |
640 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
641 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
642 |
|
jnz near .loop |
643 |
|
|
644 |
|
; deal with DC |
645 |
|
movd mm0, [ecx] |
646 |
|
pmullw mm0, [esp+16] ; dcscalar |
647 |
|
movq mm2, [mmx_32767_minus_2047] |
648 |
|
paddsw mm0, mm2 |
649 |
|
psubsw mm0, mm2 |
650 |
|
movq mm3, [mmx_32768_minus_2048] |
651 |
|
psubsw mm0, mm3 |
652 |
|
paddsw mm0, mm3 |
653 |
|
movd eax, mm0 |
654 |
|
mov [edx], ax |
655 |
|
|
656 |
|
xor eax, eax ; return(0); |
657 |
|
ret |
658 |
|
|
659 |
|
;----------------------------------------------------------------------------- |
660 |
|
; |
661 |
|
; uint32_t dequant_h263_intra_xmm(int16_t *data, |
662 |
|
; const int16_t const *coeff, |
663 |
|
; const uint32_t quant, |
664 |
|
; const uint32_t dcscalar); |
665 |
|
; |
666 |
|
;----------------------------------------------------------------------------- |
667 |
|
|
668 |
|
; this is the same as dequant_inter_mmx, except that we're |
669 |
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
670 |
|
|
671 |
|
ALIGN 16 |
672 |
|
dequant_h263_intra_xmm: |
673 |
|
|
674 |
|
mov edx, [esp+ 4] ; data |
675 |
|
mov ecx, [esp+ 8] ; coeff |
676 |
|
mov eax, [esp+12] ; quant |
677 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
678 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
679 |
|
mov eax, -16 |
680 |
|
|
681 |
|
ALIGN 16 |
682 |
|
.loop |
683 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
684 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
685 |
|
pxor mm1, mm1 |
686 |
|
pxor mm4, mm4 |
687 |
|
pcmpgtw mm1, mm0 ; sign(c) |
688 |
|
pcmpgtw mm4, mm3 ; sign(c') |
689 |
|
pxor mm2, mm2 |
690 |
|
pxor mm5, mm5 |
691 |
|
pcmpeqw mm2, mm0 ; c is zero |
692 |
|
pcmpeqw mm5, mm3 ; c' is zero |
693 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
694 |
|
pandn mm5, mm6 |
695 |
|
pxor mm0, mm1 ; negate if negative |
696 |
|
pxor mm3, mm4 ; negate if negative |
697 |
|
psubw mm0, mm1 |
698 |
|
psubw mm3, mm4 |
699 |
|
pmullw mm0, mm7 ; *= 2Q |
700 |
|
pmullw mm3, mm7 ; *= 2Q |
701 |
|
paddw mm0, mm2 ; + offset |
702 |
|
paddw mm3, mm5 ; + offset |
703 |
|
paddw mm0, mm1 ; negate back |
704 |
|
paddw mm3, mm4 ; negate back |
705 |
|
|
706 |
|
; saturates to +2047 |
707 |
|
movq mm2, [mmx_2047] |
708 |
|
pminsw mm0, mm2 |
709 |
|
add eax, 2 |
710 |
|
pminsw mm3, mm2 |
711 |
|
|
712 |
|
pxor mm0, mm1 |
713 |
|
pxor mm3, mm4 |
714 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
715 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
716 |
|
jnz near .loop |
717 |
|
|
718 |
|
; deal with DC |
719 |
|
movd mm0, [ecx] |
720 |
|
pmullw mm0, [esp+16] ; dcscalar |
721 |
|
movq mm2, [mmx_32767_minus_2047] |
722 |
|
paddsw mm0, mm2 |
723 |
|
psubsw mm0, mm2 |
724 |
|
movq mm2, [mmx_32768_minus_2048] |
725 |
|
psubsw mm0, mm2 |
726 |
|
paddsw mm0, mm2 |
727 |
|
movd eax, mm0 |
728 |
|
mov [edx], ax |
729 |
|
|
730 |
|
xor eax, eax |
731 |
|
ret |
732 |
|
|
733 |
|
|
734 |
|
;----------------------------------------------------------------------------- |
735 |
|
; |
736 |
|
; uint32_t dequant_h263_intra_sse2(int16_t *data, |
737 |
|
; const int16_t const *coeff, |
738 |
|
; const uint32_t quant, |
739 |
|
; const uint32_t dcscalar); |
740 |
|
; |
741 |
|
;----------------------------------------------------------------------------- |
742 |
|
|
743 |
|
ALIGN 16 |
744 |
|
dequant_h263_intra_sse2: |
745 |
|
mov edx, [esp+ 4] ; data |
746 |
|
mov ecx, [esp+ 8] ; coeff |
747 |
|
mov eax, [esp+12] ; quant |
748 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
749 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
750 |
|
movq2dq xmm6, mm6 |
751 |
|
movq2dq xmm7, mm7 |
752 |
|
movlhps xmm6, xmm6 |
753 |
|
movlhps xmm7, xmm7 |
754 |
|
mov eax, -16 |
755 |
|
|
756 |
|
ALIGN 16 |
757 |
|
.loop |
758 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
759 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax+ 16] |
760 |
|
pxor xmm1, xmm1 |
761 |
|
pxor xmm4, xmm4 |
762 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
763 |
|
pcmpgtw xmm4, xmm3 |
764 |
|
pxor xmm2, xmm2 |
765 |
|
pxor xmm5, xmm5 |
766 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
767 |
|
pcmpeqw xmm5, xmm3 |
768 |
|
pandn xmm2, xmm6 ; offset = isZero ? 0 : quant_add |
769 |
|
pandn xmm5, xmm6 |
770 |
|
pxor xmm0, xmm1 ; negate if negative |
771 |
|
pxor xmm3, xmm4 |
772 |
|
psubw xmm0, xmm1 |
773 |
|
psubw xmm3, xmm4 |
774 |
|
pmullw xmm0, xmm7 ; *= 2Q |
775 |
|
pmullw xmm3, xmm7 |
776 |
|
paddw xmm0, xmm2 ; + offset |
777 |
|
paddw xmm3, xmm5 |
778 |
|
paddw xmm0, xmm1 ; negate back |
779 |
|
paddw xmm3, xmm4 |
780 |
|
|
781 |
|
; saturates to +2047 |
782 |
|
movdqa xmm2, [sse2_2047] |
783 |
|
pminsw xmm0, xmm2 |
784 |
|
add eax, 4 |
785 |
|
pminsw xmm3, xmm2 |
786 |
|
|
787 |
|
pxor xmm0, xmm1 |
788 |
|
pxor xmm3, xmm4 |
789 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
790 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
791 |
|
jnz near .loop |
792 |
|
|
793 |
|
; deal with DC |
794 |
|
movd mm0, [ecx] |
795 |
|
pmullw mm0, [esp+16] ; dcscalar |
796 |
|
movq mm2, [mmx_32767_minus_2047] |
797 |
|
paddsw mm0, mm2 |
798 |
|
psubsw mm0, mm2 |
799 |
|
movq mm2, [mmx_32768_minus_2048] |
800 |
|
psubsw mm0, mm2 |
801 |
|
paddsw mm0, mm2 |
802 |
|
movd eax, mm0 |
803 |
|
mov [edx], ax |
804 |
|
|
805 |
|
xor eax, eax |
806 |
|
ret |
807 |
|
|
808 |
|
;----------------------------------------------------------------------------- |
809 |
|
; |
810 |
|
; uint32t dequant_h263_inter_mmx(int16_t * data, |
811 |
|
; const int16_t * const coeff, |
812 |
|
; const uint32_t quant); |
813 |
|
; |
814 |
|
;----------------------------------------------------------------------------- |
815 |
|
|
816 |
|
ALIGN 16 |
817 |
|
dequant_h263_inter_mmx: |
818 |
|
|
819 |
|
mov edx, [esp+ 4] ; data |
820 |
|
mov ecx, [esp+ 8] ; coeff |
821 |
|
mov eax, [esp+12] ; quant |
822 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
823 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
824 |
|
mov eax, -16 |
825 |
|
|
826 |
|
ALIGN 16 |
827 |
|
.loop |
828 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
829 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
830 |
|
pxor mm1, mm1 |
831 |
|
pxor mm4, mm4 |
832 |
|
pcmpgtw mm1, mm0 ; sign(c) |
833 |
|
pcmpgtw mm4, mm3 ; sign(c') |
834 |
|
pxor mm2, mm2 |
835 |
|
pxor mm5, mm5 |
836 |
|
pcmpeqw mm2, mm0 ; c is zero |
837 |
|
pcmpeqw mm5, mm3 ; c' is zero |
838 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
839 |
|
pandn mm5, mm6 |
840 |
|
pxor mm0, mm1 ; negate if negative |
841 |
|
pxor mm3, mm4 ; negate if negative |
842 |
|
psubw mm0, mm1 |
843 |
|
psubw mm3, mm4 |
844 |
|
pmullw mm0, mm7 ; *= 2Q |
845 |
|
pmullw mm3, mm7 ; *= 2Q |
846 |
|
paddw mm0, mm2 ; + offset |
847 |
|
paddw mm3, mm5 ; + offset |
848 |
|
paddw mm0, mm1 ; negate back |
849 |
|
paddw mm3, mm4 ; negate back |
850 |
|
; saturates to +2047 |
851 |
|
movq mm2, [mmx_32767_minus_2047] |
852 |
|
add eax, 2 |
853 |
|
paddsw mm0, mm2 |
854 |
|
paddsw mm3, mm2 |
855 |
|
psubsw mm0, mm2 |
856 |
|
psubsw mm3, mm2 |
857 |
|
|
858 |
|
pxor mm0, mm1 |
859 |
|
pxor mm3, mm4 |
860 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
861 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
862 |
|
jnz near .loop |
863 |
|
|
864 |
|
xor eax, eax |
865 |
|
ret |
866 |
|
|
867 |
|
;----------------------------------------------------------------------------- |
868 |
|
; |
869 |
|
; uint32_t dequant_h263_inter_xmm(int16_t * data, |
870 |
|
; const int16_t * const coeff, |
871 |
|
; const uint32_t quant); |
872 |
|
; |
873 |
|
;----------------------------------------------------------------------------- |
874 |
|
|
875 |
|
; this is the same as dequant_inter_mmx, |
876 |
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
877 |
|
|
878 |
|
ALIGN 16 |
879 |
|
dequant_h263_inter_xmm: |
880 |
|
|
881 |
|
mov edx, [esp+ 4] ; data |
882 |
|
mov ecx, [esp+ 8] ; coeff |
883 |
|
mov eax, [esp+12] ; quant |
884 |
|
movq mm6, [mmx_add + eax*8 - 8] ; quant or quant-1 |
885 |
|
movq mm7, [mmx_mul + eax*8 - 8] ; 2*quant |
886 |
|
mov eax, -16 |
887 |
|
|
888 |
|
ALIGN 16 |
889 |
|
.loop |
890 |
|
movq mm0, [ecx+8*eax+8*16] ; c = coeff[i] |
891 |
|
movq mm3, [ecx+8*eax+8*16 + 8] ; c' = coeff[i+1] |
892 |
|
pxor mm1, mm1 |
893 |
|
pxor mm4, mm4 |
894 |
|
pcmpgtw mm1, mm0 ; sign(c) |
895 |
|
pcmpgtw mm4, mm3 ; sign(c') |
896 |
|
pxor mm2, mm2 |
897 |
|
pxor mm5, mm5 |
898 |
|
pcmpeqw mm2, mm0 ; c is zero |
899 |
|
pcmpeqw mm5, mm3 ; c' is zero |
900 |
|
pandn mm2, mm6 ; offset = isZero ? 0 : quant_add |
901 |
|
pandn mm5, mm6 |
902 |
|
pxor mm0, mm1 ; negate if negative |
903 |
|
pxor mm3, mm4 ; negate if negative |
904 |
|
psubw mm0, mm1 |
905 |
|
psubw mm3, mm4 |
906 |
|
pmullw mm0, mm7 ; *= 2Q |
907 |
|
pmullw mm3, mm7 ; *= 2Q |
908 |
|
paddw mm0, mm2 ; + offset |
909 |
|
paddw mm3, mm5 ; + offset |
910 |
|
paddw mm0, mm1 ; start restoring sign |
911 |
|
paddw mm3, mm4 ; start restoring sign |
912 |
|
; saturates to +2047 |
913 |
|
movq mm2, [mmx_2047] |
914 |
|
pminsw mm0, mm2 |
915 |
|
add eax, 2 |
916 |
|
pminsw mm3, mm2 |
917 |
|
|
918 |
|
pxor mm0, mm1 ; finish restoring sign |
919 |
|
pxor mm3, mm4 ; finish restoring sign |
920 |
|
movq [edx + 8*eax + 8*16 - 2*8], mm0 |
921 |
|
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
922 |
|
jnz near .loop |
923 |
|
|
924 |
|
xor eax, eax |
925 |
|
ret |
926 |
|
|
927 |
|
;----------------------------------------------------------------------------- |
928 |
|
; |
929 |
|
; uint32_t dequant_h263_inter_sse2(int16_t * data, |
930 |
|
; const int16_t * const coeff, |
931 |
|
; const uint32_t quant); |
932 |
|
; |
933 |
|
;----------------------------------------------------------------------------- |
934 |
|
|
935 |
|
ALIGN 16 |
936 |
|
dequant_h263_inter_sse2: |
937 |
|
mov edx, [esp + 4] ; data |
938 |
|
mov ecx, [esp + 8] ; coeff |
939 |
|
mov eax, [esp + 12] ; quant |
940 |
|
movq mm6, [mmx_add + eax * 8 - 8] |
941 |
|
movq mm7, [mmx_mul + eax * 8 - 8] |
942 |
|
movq2dq xmm6, mm6 |
943 |
|
movq2dq xmm7, mm7 |
944 |
|
movlhps xmm6, xmm6 |
945 |
|
movlhps xmm7, xmm7 |
946 |
|
mov eax, -16 |
947 |
|
|
948 |
|
ALIGN 16 |
949 |
|
.loop |
950 |
|
movdqa xmm0, [ecx + 8*16 + 8*eax] ; c = coeff[i] |
951 |
|
movdqa xmm3, [ecx + 8*16 + 8*eax + 16] |
952 |
|
|
953 |
|
pxor xmm1, xmm1 |
954 |
|
pxor xmm4, xmm4 |
955 |
|
pcmpgtw xmm1, xmm0 ; sign(c) |
956 |
|
pcmpgtw xmm4, xmm3 |
957 |
|
pxor xmm2, xmm2 |
958 |
|
pxor xmm5, xmm5 |
959 |
|
pcmpeqw xmm2, xmm0 ; c is zero |
960 |
|
pcmpeqw xmm5, xmm3 |
961 |
|
pandn xmm2, xmm6 |
962 |
|
pandn xmm5, xmm6 |
963 |
|
pxor xmm0, xmm1 ; negate if negative |
964 |
|
pxor xmm3, xmm4 |
965 |
|
psubw xmm0, xmm1 |
966 |
|
psubw xmm3, xmm4 |
967 |
|
pmullw xmm0, xmm7 ; *= 2Q |
968 |
|
pmullw xmm3, xmm7 |
969 |
|
paddw xmm0, xmm2 ; + offset |
970 |
|
paddw xmm3, xmm5 |
971 |
|
|
972 |
|
paddw xmm0, xmm1 ; start restoring sign |
973 |
|
paddw xmm3, xmm4 |
974 |
|
|
975 |
|
; saturates to +2047 |
976 |
|
movdqa xmm2, [sse2_2047] |
977 |
|
pminsw xmm0, xmm2 |
978 |
|
add eax, 4 |
979 |
|
pminsw xmm3, xmm2 |
980 |
|
|
981 |
|
pxor xmm0, xmm1 ; finish restoring sign |
982 |
|
pxor xmm3, xmm4 |
983 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0 |
984 |
|
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
985 |
|
jnz near .loop |
986 |
|
|
987 |
|
xor eax, eax |
988 |
|
ret |