1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * colorspace rgb |
; * - RGB colorspace conversions - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2002-2003 Michael Militzer <isibaar@xvid.org> |
7 |
|
; * 2002-2003 Peter Ross <pross@xvid.org> |
8 |
; * |
; * |
9 |
; * This program is free software; you can redistribute it and/or modify |
; * 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 |
; * it under the terms of the GNU General Public License as published by |
18 |
; * |
; * |
19 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
20 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
21 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; *************************************************************************/ |
; ****************************************************************************/ |
24 |
|
|
25 |
;/************************************************************************** |
BITS 32 |
|
; * |
|
|
; * History: |
|
|
; * |
|
|
; * 10.10.2001 initial version; (c)2002 peter ross <pross@xvid.org> |
|
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
|
|
|
bits 32 |
|
26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
29 |
|
%ifdef MARK_FUNCS |
30 |
|
global _%1:function %1.endfunc-%1 |
31 |
|
%define %1 _%1:function %1.endfunc-%1 |
32 |
|
%define ENDFUNC .endfunc |
33 |
|
%else |
34 |
global _%1 |
global _%1 |
35 |
%define %1 _%1 |
%define %1 _%1 |
36 |
|
%define ENDFUNC |
37 |
|
%endif |
38 |
|
%else |
39 |
|
%ifdef MARK_FUNCS |
40 |
|
global %1:function %1.endfunc-%1 |
41 |
|
%define ENDFUNC .endfunc |
42 |
%else |
%else |
43 |
global %1 |
global %1 |
44 |
|
%define ENDFUNC |
45 |
|
%endif |
46 |
%endif |
%endif |
47 |
%endmacro |
%endmacro |
48 |
|
|
49 |
|
;============================================================================= |
50 |
|
; Some constants |
51 |
|
;============================================================================= |
52 |
|
|
53 |
section .data |
;----------------------------------------------------------------------------- |
|
align 16 |
|
|
|
|
|
;=========================================================================== |
|
54 |
; RGB->YV12 yuv constants |
; RGB->YV12 yuv constants |
55 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
56 |
|
|
57 |
%define Y_R 0.257 |
%define Y_R 0.257 |
58 |
%define Y_G 0.504 |
%define Y_G 0.504 |
59 |
%define Y_B 0.098 |
%define Y_B 0.098 |
69 |
%define V_B 0.071 |
%define V_B 0.071 |
70 |
%define V_ADD 128 |
%define V_ADD 128 |
71 |
|
|
72 |
;=========================================================================== |
; Scaling used during conversion |
73 |
; RGB->YV12 multiplication matrices |
%define SCALEBITS 6 |
|
;=========================================================================== |
|
|
; FIX(Y_B) FIX(Y_G) FIX(Y_R) |
|
|
y_mul dw 25, 129, 66, 0 |
|
|
u_mul dw 112, -74, -38, 0 |
|
|
v_mul dw -18, -94, 112, 0 |
|
74 |
|
|
75 |
|
;============================================================================= |
76 |
|
; Read only data |
77 |
|
;============================================================================= |
78 |
|
|
79 |
;=========================================================================== |
%ifdef FORMAT_COFF |
80 |
; YV12->RGB data |
SECTION .rodata |
81 |
;=========================================================================== |
%else |
82 |
%define SCALEBITS 6 |
SECTION .rodata align=16 |
83 |
Y_SUB dw 16, 16, 16, 16 |
%endif |
84 |
U_SUB dw 128, 128, 128, 128 |
ALIGN 16 |
85 |
V_SUB dw 128, 128, 128, 128 |
|
86 |
|
;----------------------------------------------------------------------------- |
87 |
|
; RGB->YV12 multiplication matrices |
88 |
|
;----------------------------------------------------------------------------- |
89 |
|
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
90 |
|
|
91 |
Y_MUL dw 74, 74, 74, 74 |
bgr_y_mul: dw 25, 129, 66, 0 |
92 |
|
bgr_u_mul: dw 112, -74, -38, 0 |
93 |
|
bgr_v_mul: dw -18, -94, 112, 0 |
94 |
|
|
95 |
|
;----------------------------------------------------------------------------- |
96 |
|
; BGR->YV12 multiplication matrices |
97 |
|
;----------------------------------------------------------------------------- |
98 |
|
; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored |
99 |
|
|
100 |
|
rgb_y_mul: dw 66, 129, 25, 0 |
101 |
|
rgb_u_mul: dw -38, -74, 112, 0 |
102 |
|
rgb_v_mul: dw 112, -94, -18, 0 |
103 |
|
|
104 |
UG_MUL dw 25, 25, 25, 25 |
;----------------------------------------------------------------------------- |
105 |
VG_MUL dw 52, 52, 52, 52 |
; YV12->RGB data |
106 |
|
;----------------------------------------------------------------------------- |
107 |
|
|
108 |
UB_MUL dw 129, 129, 129, 129 |
Y_SUB: dw 16, 16, 16, 16 |
109 |
VR_MUL dw 102, 102, 102, 102 |
U_SUB: dw 128, 128, 128, 128 |
110 |
|
V_SUB: dw 128, 128, 128, 128 |
111 |
|
|
112 |
|
Y_MUL: dw 74, 74, 74, 74 |
113 |
|
|
114 |
|
UG_MUL: dw 25, 25, 25, 25 |
115 |
|
VG_MUL: dw 52, 52, 52, 52 |
116 |
|
|
117 |
section .text |
UB_MUL: dw 129, 129, 129, 129 |
118 |
|
VR_MUL: dw 102, 102, 102, 102 |
119 |
|
|
120 |
%include "colorspace_mmx.inc" |
BRIGHT: db 128, 128, 128, 128, 128, 128, 128, 128 |
121 |
|
|
122 |
|
;============================================================================= |
123 |
|
; Helper macros used with the colorspace_mmx.inc file |
124 |
|
;============================================================================= |
125 |
|
|
126 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
127 |
; BGR_TO_YV12( BYTES ) |
; BGR_TO_YV12( BYTES ) |
130 |
; |
; |
131 |
; bytes=3/4, pixels = 2, vpixels=2 |
; bytes=3/4, pixels = 2, vpixels=2 |
132 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
133 |
|
|
134 |
%macro BGR_TO_YV12_INIT 2 |
%macro BGR_TO_YV12_INIT 2 |
135 |
movq mm7, [y_mul] |
movq mm7, [bgr_y_mul] |
136 |
%endmacro |
%endmacro |
137 |
|
|
138 |
|
|
195 |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
196 |
|
|
197 |
; u_ptr, v_ptr |
; u_ptr, v_ptr |
|
|
|
198 |
movq mm0, mm6 ; = [ |b4|g4|r4] |
movq mm0, mm6 ; = [ |b4|g4|r4] |
199 |
pmaddwd mm6, [v_mul] ; *= V_MUL |
pmaddwd mm6, [bgr_v_mul] ; *= V_MUL |
200 |
pmaddwd mm0, [u_mul] ; *= U_MUL |
pmaddwd mm0, [bgr_u_mul] ; *= U_MUL |
201 |
movq mm1, mm0 |
movq mm1, mm0 |
202 |
movq mm2, mm6 |
movq mm2, mm6 |
203 |
psrlq mm1, 32 |
psrlq mm1, 32 |
217 |
|
|
218 |
pop edx |
pop edx |
219 |
%endmacro |
%endmacro |
220 |
|
|
221 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
222 |
|
; RGB_TO_YV12( BYTES ) |
223 |
|
; |
224 |
|
; BYTES 3=rgb(24bit), 4=rgba(32-bit) |
225 |
|
; |
226 |
|
; bytes=3/4, pixels = 2, vpixels=2 |
227 |
|
;------------------------------------------------------------------------------ |
228 |
|
|
229 |
|
%macro RGB_TO_YV12_INIT 2 |
230 |
|
movq mm7, [rgb_y_mul] |
231 |
|
%endmacro |
232 |
|
|
233 |
|
|
234 |
|
%macro RGB_TO_YV12 2 |
235 |
|
; y_out |
236 |
|
pxor mm4, mm4 |
237 |
|
pxor mm5, mm5 |
238 |
|
movd mm0, [edi] ; x_ptr[0...] |
239 |
|
movd mm2, [edi+edx] ; x_ptr[x_stride...] |
240 |
|
punpcklbw mm0, mm4 ; [ |b |g |r ] |
241 |
|
punpcklbw mm2, mm5 ; [ |b |g |r ] |
242 |
|
movq mm6, mm0 ; = [ |b4|g4|r4] |
243 |
|
paddw mm6, mm2 ; +[ |b4|g4|r4] |
244 |
|
pmaddwd mm0, mm7 ; *= Y_MUL |
245 |
|
pmaddwd mm2, mm7 ; *= Y_MUL |
246 |
|
movq mm4, mm0 ; [r] |
247 |
|
movq mm5, mm2 ; [r] |
248 |
|
psrlq mm4, 32 ; +[g] |
249 |
|
psrlq mm5, 32 ; +[g] |
250 |
|
paddd mm0, mm4 ; +[b] |
251 |
|
paddd mm2, mm5 ; +[b] |
252 |
|
|
253 |
|
pxor mm4, mm4 |
254 |
|
pxor mm5, mm5 |
255 |
|
movd mm1, [edi+%1] ; src[%1...] |
256 |
|
movd mm3, [edi+edx+%1] ; src[x_stride+%1...] |
257 |
|
punpcklbw mm1, mm4 ; [ |b |g |r ] |
258 |
|
punpcklbw mm3, mm5 ; [ |b |g |r ] |
259 |
|
paddw mm6, mm1 ; +[ |b4|g4|r4] |
260 |
|
paddw mm6, mm3 ; +[ |b4|g4|r4] |
261 |
|
pmaddwd mm1, mm7 ; *= Y_MUL |
262 |
|
pmaddwd mm3, mm7 ; *= Y_MUL |
263 |
|
movq mm4, mm1 ; [r] |
264 |
|
movq mm5, mm3 ; [r] |
265 |
|
psrlq mm4, 32 ; +[g] |
266 |
|
psrlq mm5, 32 ; +[g] |
267 |
|
paddd mm1, mm4 ; +[b] |
268 |
|
paddd mm3, mm5 ; +[b] |
269 |
|
|
270 |
|
push edx |
271 |
|
|
272 |
|
movd edx, mm0 |
273 |
|
shr edx, 8 |
274 |
|
add edx, Y_ADD |
275 |
|
mov [esi], dl ; y_ptr[0] |
276 |
|
|
277 |
|
movd edx, mm1 |
278 |
|
shr edx, 8 |
279 |
|
add edx, Y_ADD |
280 |
|
mov [esi + 1], dl ; y_ptr[1] |
281 |
|
|
282 |
|
movd edx, mm2 |
283 |
|
shr edx, 8 |
284 |
|
add edx, Y_ADD |
285 |
|
mov [esi + eax + 0], dl ; y_ptr[y_stride + 0] |
286 |
|
|
287 |
|
movd edx, mm3 |
288 |
|
shr edx, 8 |
289 |
|
add edx, Y_ADD |
290 |
|
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
291 |
|
|
292 |
|
; u_ptr, v_ptr |
293 |
|
movq mm0, mm6 ; = [ |b4|g4|r4] |
294 |
|
pmaddwd mm6, [rgb_v_mul] ; *= V_MUL |
295 |
|
pmaddwd mm0, [rgb_u_mul] ; *= U_MUL |
296 |
|
movq mm1, mm0 |
297 |
|
movq mm2, mm6 |
298 |
|
psrlq mm1, 32 |
299 |
|
psrlq mm2, 32 |
300 |
|
paddd mm0, mm1 |
301 |
|
paddd mm2, mm6 |
302 |
|
|
303 |
|
movd edx, mm0 |
304 |
|
shr edx, 10 |
305 |
|
add edx, U_ADD |
306 |
|
mov [ebx], dl |
307 |
|
|
308 |
|
movd edx, mm2 |
309 |
|
shr edx, 10 |
310 |
|
add edx, V_ADD |
311 |
|
mov [ecx], dl |
312 |
|
|
313 |
|
pop edx |
314 |
|
%endmacro |
315 |
|
|
316 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
317 |
; YV12_TO_BGR( BYTES ) |
; YV12_TO_BGR( BYTES ) |
320 |
; |
; |
321 |
; bytes=3/4, pixels = 8, vpixels=2 |
; bytes=3/4, pixels = 8, vpixels=2 |
322 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
323 |
|
|
324 |
%macro YV12_TO_BGR_INIT 2 |
%macro YV12_TO_BGR_INIT 2 |
325 |
pxor mm7, mm7 ; clear mm7 |
pxor mm7, mm7 ; clear mm7 |
326 |
%endmacro |
%endmacro |
327 |
|
|
|
|
|
328 |
%macro YV12_TO_BGR 2 |
%macro YV12_TO_BGR 2 |
329 |
%define TEMP_Y1 esp |
%define TEMP_Y1 esp |
330 |
%define TEMP_Y2 esp + 8 |
%define TEMP_Y2 esp + 8 |
332 |
%define TEMP_G2 esp + 24 |
%define TEMP_G2 esp + 24 |
333 |
%define TEMP_B1 esp + 32 |
%define TEMP_B1 esp + 32 |
334 |
%define TEMP_B2 esp + 40 |
%define TEMP_B2 esp + 40 |
335 |
|
|
336 |
movd mm2, [ebx] ; u_ptr[0] |
movd mm2, [ebx] ; u_ptr[0] |
337 |
movd mm3, [ecx] ; v_ptr[0] |
movd mm3, [ecx] ; v_ptr[0] |
|
|
|
338 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
339 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
|
|
|
340 |
psubsw mm2, [U_SUB] ; U - 128 |
psubsw mm2, [U_SUB] ; U - 128 |
341 |
psubsw mm3, [V_SUB] ; V - 128 |
psubsw mm3, [V_SUB] ; V - 128 |
|
|
|
342 |
movq mm4, mm2 |
movq mm4, mm2 |
343 |
movq mm5, mm3 |
movq mm5, mm3 |
|
|
|
344 |
pmullw mm2, [UG_MUL] |
pmullw mm2, [UG_MUL] |
345 |
pmullw mm3, [VG_MUL] |
pmullw mm3, [VG_MUL] |
|
|
|
346 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
347 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
348 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
|
|
|
349 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
|
|
|
350 |
movq mm0, mm3 |
movq mm0, mm3 |
351 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
352 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
|
|
|
353 |
paddsw mm2, mm3 |
paddsw mm2, mm3 |
354 |
paddsw mm6, mm0 |
paddsw mm6, mm0 |
|
|
|
355 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
|
|
|
356 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
357 |
movq mm1, mm0 |
movq mm1, mm0 |
358 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
359 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
360 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
361 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
362 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
363 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
364 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
365 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
|
|
|
366 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
367 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
368 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
369 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
370 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
371 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
|
|
|
372 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
373 |
movq mm1, mm0 |
movq mm1, mm0 |
|
|
|
374 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
375 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
376 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
377 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
378 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
379 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
380 |
movq mm3, mm1 |
movq mm3, mm1 |
381 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
|
|
|
382 |
movq mm2, mm0 |
movq mm2, mm0 |
383 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
384 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
385 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
386 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
387 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
|
|
|
388 |
movq mm0, mm4 |
movq mm0, mm4 |
389 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
390 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
|
|
|
391 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
392 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
393 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
|
|
|
394 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
395 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
396 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
397 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
398 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
|
|
|
399 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
400 |
movq mm2, [TEMP_Y1] |
movq mm2, [TEMP_Y1] |
|
|
|
401 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
402 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
403 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
404 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
405 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
406 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
407 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
408 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
|
|
|
409 |
movq mm0, mm5 |
movq mm0, mm5 |
410 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
411 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
|
|
|
412 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
413 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
|
|
|
414 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
415 |
psraw mm7, SCALEBITS |
psraw mm7, SCALEBITS |
|
|
|
416 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
|
|
|
417 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
418 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
|
|
|
419 |
psraw mm6, SCALEBITS |
psraw mm6, SCALEBITS |
420 |
psraw mm4, SCALEBITS |
psraw mm4, SCALEBITS |
|
|
|
421 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
|
|
|
422 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
423 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |
|
|
|
424 |
movq mm6, mm7 |
movq mm6, mm7 |
|
|
|
425 |
movq mm2, mm0 |
movq mm2, mm0 |
426 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
427 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
428 |
pxor mm7, mm7 |
pxor mm7, mm7 |
|
|
|
429 |
movq mm3, mm1 |
movq mm3, mm1 |
430 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
431 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
432 |
movq mm4, mm2 |
movq mm4, mm2 |
433 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
434 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
435 |
movq mm5, mm0 |
movq mm5, mm0 |
436 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
437 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
438 |
%if %1 == 3 ; BGR (24-bit) |
%if %1 == 3 ; BGR (24-bit) |
439 |
movd [edi], mm2 |
movd [edi], mm2 |
440 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
441 |
movd [edi + 3], mm2 |
movd [edi + 3], mm2 |
442 |
movd [edi + 6], mm4 |
movd [edi + 6], mm4 |
|
|
|
443 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
444 |
movd [edi + 9], mm4 |
movd [edi + 9], mm4 |
445 |
movd [edi + 12], mm0 |
movd [edi + 12], mm0 |
|
|
|
446 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
447 |
movd [edi + 15], mm0 |
movd [edi + 15], mm0 |
448 |
movd [edi + 18], mm5 |
movq mm2, mm5 |
449 |
|
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
450 |
psrlq mm5, 32 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
451 |
|
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
452 |
movd [edi + 21], mm5 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
453 |
|
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
454 |
|
psllq mm5, 40 ; r7g7b700000 -> mm5 |
455 |
|
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
456 |
|
movq [edi + 16], mm5 |
457 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
458 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
459 |
movq mm2, mm0 |
movq mm2, mm0 |
460 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
461 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
462 |
movq mm3, mm1 |
movq mm3, mm1 |
463 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
464 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
465 |
movq mm4, mm2 |
movq mm4, mm2 |
466 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
467 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
468 |
movq mm5, mm0 |
movq mm5, mm0 |
469 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
470 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
471 |
movd [edi+edx], mm2 |
movd [edi+edx], mm2 |
472 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
473 |
movd [edi+edx + 3], mm2 |
movd [edi+edx + 3], mm2 |
474 |
movd [edi+edx + 6], mm4 |
movd [edi+edx + 6], mm4 |
|
|
|
475 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
476 |
movd [edi+edx + 9], mm4 |
movd [edi+edx + 9], mm4 |
477 |
movd [edi+edx + 12], mm0 |
movd [edi+edx + 12], mm0 |
|
|
|
478 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
479 |
movd [edi+edx + 15], mm0 |
movd [edi+edx + 15], mm0 |
480 |
movd [edi+edx + 18], mm5 |
movq mm2, mm5 |
481 |
|
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
482 |
psrlq mm5, 32 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
483 |
|
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
484 |
movd [edi+edx + 21], mm5 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
485 |
|
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
486 |
|
psllq mm5, 40 ; r7g7b700000 -> mm5 |
487 |
|
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
488 |
|
movq [edi + edx + 16], mm5 |
489 |
|
|
490 |
%else ; BGRA (32-bit) |
%else ; BGRA (32-bit) |
491 |
movq [edi], mm2 |
movq [edi], mm2 |
492 |
movq [edi + 8], mm4 |
movq [edi + 8], mm4 |
493 |
movq [edi + 16], mm0 |
movq [edi + 16], mm0 |
494 |
movq [edi + 24], mm5 |
movq [edi + 24], mm5 |
|
|
|
495 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
496 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
497 |
movq mm2, mm0 |
movq mm2, mm0 |
498 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
499 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
500 |
movq mm3, mm1 |
movq mm3, mm1 |
501 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
502 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
503 |
movq mm4, mm2 |
movq mm4, mm2 |
504 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
505 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
506 |
movq mm5, mm0 |
movq mm5, mm0 |
507 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
508 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
509 |
movq [edi + edx], mm2 |
movq [edi + edx], mm2 |
510 |
movq [edi + edx + 8], mm4 |
movq [edi + edx + 8], mm4 |
511 |
movq [edi + edx + 16], mm0 |
movq [edi + edx + 16], mm0 |
519 |
%undef TEMP_B1 |
%undef TEMP_B1 |
520 |
%undef TEMP_B2 |
%undef TEMP_B2 |
521 |
%endmacro |
%endmacro |
|
;------------------------------------------------------------------------------ |
|
522 |
|
|
523 |
|
;============================================================================= |
524 |
|
; Code |
525 |
|
;============================================================================= |
526 |
|
|
527 |
|
SECTION .text |
528 |
|
|
529 |
|
%include "colorspace_mmx.inc" |
530 |
|
|
531 |
; input |
; input |
|
|
|
532 |
MAKE_COLORSPACE bgr_to_yv12_mmx,0, 3,2,2, BGR_TO_YV12, 3, -1 |
MAKE_COLORSPACE bgr_to_yv12_mmx,0, 3,2,2, BGR_TO_YV12, 3, -1 |
533 |
MAKE_COLORSPACE bgra_to_yv12_mmx,0, 4,2,2, BGR_TO_YV12, 4, -1 |
MAKE_COLORSPACE bgra_to_yv12_mmx,0, 4,2,2, BGR_TO_YV12, 4, -1 |
534 |
|
MAKE_COLORSPACE rgb_to_yv12_mmx,0, 3,2,2, RGB_TO_YV12, 3, -1 |
535 |
|
MAKE_COLORSPACE rgba_to_yv12_mmx,0, 4,2,2, RGB_TO_YV12, 4, -1 |
536 |
|
|
537 |
; output |
; output |
|
|
|
538 |
MAKE_COLORSPACE yv12_to_bgr_mmx,48, 3,8,2, YV12_TO_BGR, 3, -1 |
MAKE_COLORSPACE yv12_to_bgr_mmx,48, 3,8,2, YV12_TO_BGR, 3, -1 |
539 |
MAKE_COLORSPACE yv12_to_bgra_mmx,48, 4,8,2, YV12_TO_BGR, 4, -1 |
MAKE_COLORSPACE yv12_to_bgra_mmx,48, 4,8,2, YV12_TO_BGR, 4, -1 |
540 |
|
|
541 |
|
|
542 |
|
%ifidn __OUTPUT_FORMAT__,elf |
543 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
544 |
|
%endif |
545 |
|
|