1 |
|
;/************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * colorspace rgb |
5 |
|
; * |
6 |
|
; * This program is free software; you can redistribute it and/or modify |
7 |
|
; * it under the terms of the GNU General Public License as published by |
8 |
|
; * the Free Software Foundation; either version 2 of the License, or |
9 |
|
; * (at your option) any later version. |
10 |
|
; * |
11 |
|
; * This program is distributed in the hope that it will be useful, |
12 |
|
; * but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
14 |
|
; * GNU General Public License for more details. |
15 |
|
; * |
16 |
|
; * You should have received a copy of the GNU General Public License |
17 |
|
; * along with this program; if not, write to the Free Software |
18 |
|
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
19 |
|
; * |
20 |
|
; *************************************************************************/ |
21 |
|
|
22 |
|
;/************************************************************************** |
23 |
|
; * |
24 |
|
; * History: |
25 |
|
; * |
26 |
|
; * 10.10.2001 initial version; (c)2002 peter ross <pross@xvid.org> |
27 |
|
; * |
28 |
|
; *************************************************************************/ |
29 |
|
|
30 |
|
|
31 |
|
bits 32 |
32 |
|
|
33 |
|
%macro cglobal 1 |
34 |
|
%ifdef PREFIX |
35 |
|
global _%1 |
36 |
|
%define %1 _%1 |
37 |
|
%else |
38 |
|
global %1 |
39 |
|
%endif |
40 |
|
%endmacro |
41 |
|
|
42 |
|
|
43 |
|
section .data |
44 |
|
align 16 |
45 |
|
|
46 |
|
;=========================================================================== |
47 |
|
; RGB->YV12 yuv constants |
48 |
|
;=========================================================================== |
49 |
|
%define Y_R 0.257 |
50 |
|
%define Y_G 0.504 |
51 |
|
%define Y_B 0.098 |
52 |
|
%define Y_ADD 16 |
53 |
|
|
54 |
|
%define U_R 0.148 |
55 |
|
%define U_G 0.291 |
56 |
|
%define U_B 0.439 |
57 |
|
%define U_ADD 128 |
58 |
|
|
59 |
|
%define V_R 0.439 |
60 |
|
%define V_G 0.368 |
61 |
|
%define V_B 0.071 |
62 |
|
%define V_ADD 128 |
63 |
|
|
64 |
|
;=========================================================================== |
65 |
|
; RGB->YV12 multiplication matrices |
66 |
|
;=========================================================================== |
67 |
|
; FIX(Y_B) FIX(Y_G) FIX(Y_R) |
68 |
|
y_mul dw 25, 129, 66, 0 |
69 |
|
u_mul dw 112, -74, -38, 0 |
70 |
|
v_mul dw -18, -94, 112, 0 |
71 |
|
|
72 |
|
|
73 |
|
;=========================================================================== |
74 |
|
; YV12->RGB data |
75 |
|
;=========================================================================== |
76 |
|
%define SCALEBITS 6 |
77 |
|
Y_SUB dw 16, 16, 16, 16 |
78 |
|
U_SUB dw 128, 128, 128, 128 |
79 |
|
V_SUB dw 128, 128, 128, 128 |
80 |
|
|
81 |
|
Y_MUL dw 74, 74, 74, 74 |
82 |
|
|
83 |
|
UG_MUL dw 25, 25, 25, 25 |
84 |
|
VG_MUL dw 52, 52, 52, 52 |
85 |
|
|
86 |
|
UB_MUL dw 129, 129, 129, 129 |
87 |
|
VR_MUL dw 102, 102, 102, 102 |
88 |
|
|
89 |
|
|
90 |
|
|
91 |
|
section .text |
92 |
|
|
93 |
|
%include "colorspace_mmx.inc" |
94 |
|
|
95 |
|
|
96 |
|
;------------------------------------------------------------------------------ |
97 |
|
; BGR_TO_YV12( BYTES ) |
98 |
|
; |
99 |
|
; BYTES 3=bgr(24bit), 4=bgra(32-bit) |
100 |
|
; |
101 |
|
; bytes=3/4, pixels = 2, vpixels=2 |
102 |
|
;------------------------------------------------------------------------------ |
103 |
|
%macro BGR_TO_YV12_INIT 2 |
104 |
|
movq mm7, [y_mul] |
105 |
|
%endmacro |
106 |
|
|
107 |
|
|
108 |
|
%macro BGR_TO_YV12 2 |
109 |
|
; y_out |
110 |
|
pxor mm4, mm4 |
111 |
|
pxor mm5, mm5 |
112 |
|
movd mm0, [edi] ; x_ptr[0...] |
113 |
|
movd mm2, [edi+edx] ; x_ptr[x_stride...] |
114 |
|
punpcklbw mm0, mm4 ; [ |b |g |r ] |
115 |
|
punpcklbw mm2, mm5 ; [ |b |g |r ] |
116 |
|
movq mm6, mm0 ; = [ |b4|g4|r4] |
117 |
|
paddw mm6, mm2 ; +[ |b4|g4|r4] |
118 |
|
pmaddwd mm0, mm7 ; *= Y_MUL |
119 |
|
pmaddwd mm2, mm7 ; *= Y_MUL |
120 |
|
movq mm4, mm0 ; [r] |
121 |
|
movq mm5, mm2 ; [r] |
122 |
|
psrlq mm4, 32 ; +[g] |
123 |
|
psrlq mm5, 32 ; +[g] |
124 |
|
paddd mm0, mm4 ; +[b] |
125 |
|
paddd mm2, mm5 ; +[b] |
126 |
|
|
127 |
|
pxor mm4, mm4 |
128 |
|
pxor mm5, mm5 |
129 |
|
movd mm1, [edi+%1] ; src[%1...] |
130 |
|
movd mm3, [edi+edx+%1] ; src[x_stride+%1...] |
131 |
|
punpcklbw mm1, mm4 ; [ |b |g |r ] |
132 |
|
punpcklbw mm3, mm5 ; [ |b |g |r ] |
133 |
|
paddw mm6, mm1 ; +[ |b4|g4|r4] |
134 |
|
paddw mm6, mm3 ; +[ |b4|g4|r4] |
135 |
|
pmaddwd mm1, mm7 ; *= Y_MUL |
136 |
|
pmaddwd mm3, mm7 ; *= Y_MUL |
137 |
|
movq mm4, mm1 ; [r] |
138 |
|
movq mm5, mm3 ; [r] |
139 |
|
psrlq mm4, 32 ; +[g] |
140 |
|
psrlq mm5, 32 ; +[g] |
141 |
|
paddd mm1, mm4 ; +[b] |
142 |
|
paddd mm3, mm5 ; +[b] |
143 |
|
|
144 |
|
push edx |
145 |
|
|
146 |
|
movd edx, mm0 |
147 |
|
shr edx, 8 |
148 |
|
add edx, Y_ADD |
149 |
|
mov [esi], dl ; y_ptr[0] |
150 |
|
|
151 |
|
movd edx, mm1 |
152 |
|
shr edx, 8 |
153 |
|
add edx, Y_ADD |
154 |
|
mov [esi + 1], dl ; y_ptr[1] |
155 |
|
|
156 |
|
movd edx, mm2 |
157 |
|
shr edx, 8 |
158 |
|
add edx, Y_ADD |
159 |
|
mov [esi + eax + 0], dl ; y_ptr[y_stride + 0] |
160 |
|
|
161 |
|
movd edx, mm3 |
162 |
|
shr edx, 8 |
163 |
|
add edx, Y_ADD |
164 |
|
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
165 |
|
|
166 |
|
; u_ptr, v_ptr |
167 |
|
|
168 |
|
movq mm0, mm6 ; = [ |b4|g4|r4] |
169 |
|
pmaddwd mm6, [v_mul] ; *= V_MUL |
170 |
|
pmaddwd mm0, [u_mul] ; *= U_MUL |
171 |
|
movq mm1, mm0 |
172 |
|
movq mm2, mm6 |
173 |
|
psrlq mm1, 32 |
174 |
|
psrlq mm2, 32 |
175 |
|
paddd mm0, mm1 |
176 |
|
paddd mm2, mm6 |
177 |
|
|
178 |
|
movd edx, mm0 |
179 |
|
shr edx, 10 |
180 |
|
add edx, U_ADD |
181 |
|
mov [ebx], dl |
182 |
|
|
183 |
|
movd edx, mm2 |
184 |
|
shr edx, 10 |
185 |
|
add edx, V_ADD |
186 |
|
mov [ecx], dl |
187 |
|
|
188 |
|
pop edx |
189 |
|
%endmacro |
190 |
|
;------------------------------------------------------------------------------ |
191 |
|
|
192 |
|
|
193 |
|
|
194 |
|
|
195 |
|
|
196 |
|
;------------------------------------------------------------------------------ |
197 |
|
; YV12_TO_BGR( BYTES ) |
198 |
|
; |
199 |
|
; BYTES 3=bgr(24-bit), 4=bgra(32-bit) |
200 |
|
; |
201 |
|
; bytes=3/4, pixels = 8, vpixels=2 |
202 |
|
;------------------------------------------------------------------------------ |
203 |
|
%macro YV12_TO_BGR_INIT 2 |
204 |
|
pxor mm7, mm7 ; clear mm7 |
205 |
|
%endmacro |
206 |
|
|
207 |
|
|
208 |
|
%macro YV12_TO_BGR 2 |
209 |
|
%define TEMP_Y1 esp |
210 |
|
%define TEMP_Y2 esp + 8 |
211 |
|
%define TEMP_G1 esp + 16 |
212 |
|
%define TEMP_G2 esp + 24 |
213 |
|
%define TEMP_B1 esp + 32 |
214 |
|
%define TEMP_B2 esp + 40 |
215 |
|
movd mm2, [ebx] ; u_ptr[0] |
216 |
|
movd mm3, [ecx] ; v_ptr[0] |
217 |
|
|
218 |
|
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
219 |
|
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
220 |
|
|
221 |
|
psubsw mm2, [U_SUB] ; U - 128 |
222 |
|
psubsw mm3, [V_SUB] ; V - 128 |
223 |
|
|
224 |
|
movq mm4, mm2 |
225 |
|
movq mm5, mm3 |
226 |
|
|
227 |
|
pmullw mm2, [UG_MUL] |
228 |
|
pmullw mm3, [VG_MUL] |
229 |
|
|
230 |
|
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
231 |
|
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
232 |
|
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
233 |
|
|
234 |
|
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
235 |
|
|
236 |
|
movq mm0, mm3 |
237 |
|
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
238 |
|
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
239 |
|
|
240 |
|
paddsw mm2, mm3 |
241 |
|
paddsw mm6, mm0 |
242 |
|
|
243 |
|
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
244 |
|
|
245 |
|
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
246 |
|
|
247 |
|
movq mm1, mm0 |
248 |
|
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
249 |
|
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
250 |
|
|
251 |
|
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
252 |
|
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
253 |
|
|
254 |
|
pmullw mm1, [Y_MUL] |
255 |
|
pmullw mm0, [Y_MUL] |
256 |
|
|
257 |
|
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
258 |
|
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
259 |
|
|
260 |
|
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
261 |
|
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
262 |
|
|
263 |
|
psraw mm1, SCALEBITS |
264 |
|
psraw mm0, SCALEBITS |
265 |
|
|
266 |
|
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
267 |
|
|
268 |
|
movq [TEMP_G1], mm0 |
269 |
|
|
270 |
|
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
271 |
|
|
272 |
|
movq mm1, mm0 |
273 |
|
|
274 |
|
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
275 |
|
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
276 |
|
|
277 |
|
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
278 |
|
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
279 |
|
|
280 |
|
pmullw mm1, [Y_MUL] |
281 |
|
pmullw mm0, [Y_MUL] |
282 |
|
|
283 |
|
movq mm3, mm1 |
284 |
|
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
285 |
|
|
286 |
|
movq mm2, mm0 |
287 |
|
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
288 |
|
|
289 |
|
psraw mm1, SCALEBITS |
290 |
|
psraw mm0, SCALEBITS |
291 |
|
|
292 |
|
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
293 |
|
|
294 |
|
movq [TEMP_G2], mm0 |
295 |
|
|
296 |
|
movq mm0, mm4 |
297 |
|
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
298 |
|
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
299 |
|
|
300 |
|
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
301 |
|
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
302 |
|
|
303 |
|
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
304 |
|
|
305 |
|
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
306 |
|
|
307 |
|
psraw mm3, SCALEBITS |
308 |
|
psraw mm2, SCALEBITS |
309 |
|
|
310 |
|
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
311 |
|
|
312 |
|
movq [TEMP_B2], mm2 |
313 |
|
|
314 |
|
movq mm3, [TEMP_Y2] |
315 |
|
movq mm2, [TEMP_Y1] |
316 |
|
|
317 |
|
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
318 |
|
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
319 |
|
|
320 |
|
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
321 |
|
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
322 |
|
|
323 |
|
psraw mm3, SCALEBITS |
324 |
|
psraw mm2, SCALEBITS |
325 |
|
|
326 |
|
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
327 |
|
|
328 |
|
movq [TEMP_B1], mm2 |
329 |
|
|
330 |
|
movq mm0, mm5 |
331 |
|
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
332 |
|
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
333 |
|
|
334 |
|
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
335 |
|
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
336 |
|
|
337 |
|
psraw mm1, SCALEBITS |
338 |
|
psraw mm7, SCALEBITS |
339 |
|
|
340 |
|
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
341 |
|
|
342 |
|
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
343 |
|
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
344 |
|
|
345 |
|
psraw mm6, SCALEBITS |
346 |
|
psraw mm4, SCALEBITS |
347 |
|
|
348 |
|
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
349 |
|
|
350 |
|
movq mm0, [TEMP_B1] |
351 |
|
movq mm1, [TEMP_G1] |
352 |
|
|
353 |
|
movq mm6, mm7 |
354 |
|
|
355 |
|
movq mm2, mm0 |
356 |
|
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
357 |
|
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
358 |
|
|
359 |
|
pxor mm7, mm7 |
360 |
|
|
361 |
|
movq mm3, mm1 |
362 |
|
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
363 |
|
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
364 |
|
|
365 |
|
movq mm4, mm2 |
366 |
|
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
367 |
|
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
368 |
|
|
369 |
|
movq mm5, mm0 |
370 |
|
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
371 |
|
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
372 |
|
|
373 |
|
%if %1 == 3 ; BGR (24-bit) |
374 |
|
movd [edi], mm2 |
375 |
|
psrlq mm2, 32 |
376 |
|
|
377 |
|
movd [edi + 3], mm2 |
378 |
|
movd [edi + 6], mm4 |
379 |
|
|
380 |
|
psrlq mm4, 32 |
381 |
|
|
382 |
|
movd [edi + 9], mm4 |
383 |
|
movd [edi + 12], mm0 |
384 |
|
|
385 |
|
psrlq mm0, 32 |
386 |
|
|
387 |
|
movd [edi + 15], mm0 |
388 |
|
movd [edi + 18], mm5 |
389 |
|
|
390 |
|
psrlq mm5, 32 |
391 |
|
|
392 |
|
movd [edi + 21], mm5 |
393 |
|
|
394 |
|
movq mm0, [TEMP_B2] |
395 |
|
movq mm1, [TEMP_G2] |
396 |
|
|
397 |
|
movq mm2, mm0 |
398 |
|
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
399 |
|
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
400 |
|
|
401 |
|
movq mm3, mm1 |
402 |
|
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
403 |
|
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
404 |
|
|
405 |
|
movq mm4, mm2 |
406 |
|
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
407 |
|
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
408 |
|
|
409 |
|
movq mm5, mm0 |
410 |
|
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
411 |
|
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
412 |
|
|
413 |
|
movd [edi+edx], mm2 |
414 |
|
psrlq mm2, 32 |
415 |
|
|
416 |
|
movd [edi+edx + 3], mm2 |
417 |
|
movd [edi+edx + 6], mm4 |
418 |
|
|
419 |
|
psrlq mm4, 32 |
420 |
|
|
421 |
|
movd [edi+edx + 9], mm4 |
422 |
|
movd [edi+edx + 12], mm0 |
423 |
|
|
424 |
|
psrlq mm0, 32 |
425 |
|
|
426 |
|
movd [edi+edx + 15], mm0 |
427 |
|
movd [edi+edx + 18], mm5 |
428 |
|
|
429 |
|
psrlq mm5, 32 |
430 |
|
|
431 |
|
movd [edi+edx + 21], mm5 |
432 |
|
|
433 |
|
%else ; BGRA (32-bit) |
434 |
|
movq [edi], mm2 |
435 |
|
movq [edi + 8], mm4 |
436 |
|
movq [edi + 16], mm0 |
437 |
|
movq [edi + 24], mm5 |
438 |
|
|
439 |
|
movq mm0, [TEMP_B2] |
440 |
|
movq mm1, [TEMP_G2] |
441 |
|
|
442 |
|
movq mm2, mm0 |
443 |
|
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
444 |
|
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
445 |
|
|
446 |
|
movq mm3, mm1 |
447 |
|
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
448 |
|
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
449 |
|
|
450 |
|
movq mm4, mm2 |
451 |
|
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
452 |
|
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
453 |
|
|
454 |
|
movq mm5, mm0 |
455 |
|
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
456 |
|
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
457 |
|
|
458 |
|
movq [edi + edx], mm2 |
459 |
|
movq [edi + edx + 8], mm4 |
460 |
|
movq [edi + edx + 16], mm0 |
461 |
|
movq [edi + edx + 24], mm5 |
462 |
|
%endif |
463 |
|
|
464 |
|
%undef TEMP_Y1 |
465 |
|
%undef TEMP_Y2 |
466 |
|
%undef TEMP_G1 |
467 |
|
%undef TEMP_G2 |
468 |
|
%undef TEMP_B1 |
469 |
|
%undef TEMP_B2 |
470 |
|
%endmacro |
471 |
|
;------------------------------------------------------------------------------ |
472 |
|
|
473 |
|
|
474 |
|
|
475 |
|
|
476 |
|
; input |
477 |
|
|
478 |
|
MAKE_COLORSPACE bgr_to_yv12_mmx,0, 3,2,2, BGR_TO_YV12, 3, -1 |
479 |
|
MAKE_COLORSPACE bgra_to_yv12_mmx,0, 4,2,2, BGR_TO_YV12, 4, -1 |
480 |
|
|
481 |
|
; output |
482 |
|
|
483 |
|
MAKE_COLORSPACE yv12_to_bgr_mmx,48, 3,8,2, YV12_TO_BGR, 3, -1 |
484 |
|
MAKE_COLORSPACE yv12_to_bgra_mmx,48, 4,8,2, YV12_TO_BGR, 4, -1 |
485 |
|
|