1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * colorspace |
; * - 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 |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
22 |
; * |
; * |
23 |
; * 10.10.2001 initial version; (c)2002 peter ross <pross@xvid.org> |
; ****************************************************************************/ |
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
24 |
|
|
25 |
bits 32 |
BITS 32 |
26 |
|
|
27 |
%macro cglobal 1 |
%macro cglobal 1 |
28 |
%ifdef PREFIX |
%ifdef PREFIX |
33 |
%endif |
%endif |
34 |
%endmacro |
%endmacro |
35 |
|
|
36 |
|
;============================================================================= |
37 |
|
; Some constants |
38 |
|
;============================================================================= |
39 |
|
|
40 |
section .data |
;----------------------------------------------------------------------------- |
|
align 16 |
|
|
|
|
|
;=========================================================================== |
|
41 |
; RGB->YV12 yuv constants |
; RGB->YV12 yuv constants |
42 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
43 |
|
|
44 |
%define Y_R 0.257 |
%define Y_R 0.257 |
45 |
%define Y_G 0.504 |
%define Y_G 0.504 |
46 |
%define Y_B 0.098 |
%define Y_B 0.098 |
56 |
%define V_B 0.071 |
%define V_B 0.071 |
57 |
%define V_ADD 128 |
%define V_ADD 128 |
58 |
|
|
59 |
;=========================================================================== |
; Scaling used during conversion |
60 |
; 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 |
|
61 |
|
|
62 |
|
;============================================================================= |
63 |
|
; Read only data |
64 |
|
;============================================================================= |
65 |
|
|
66 |
;=========================================================================== |
%ifdef FORMAT_COFF |
67 |
; YV12->RGB data |
SECTION .rodata |
68 |
;=========================================================================== |
%else |
69 |
%define SCALEBITS 6 |
SECTION .rodata align=16 |
70 |
Y_SUB dw 16, 16, 16, 16 |
%endif |
71 |
U_SUB dw 128, 128, 128, 128 |
ALIGN 16 |
|
V_SUB dw 128, 128, 128, 128 |
|
72 |
|
|
73 |
Y_MUL dw 74, 74, 74, 74 |
;----------------------------------------------------------------------------- |
74 |
|
; RGB->YV12 multiplication matrices |
75 |
|
;----------------------------------------------------------------------------- |
76 |
|
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
77 |
|
|
78 |
UG_MUL dw 25, 25, 25, 25 |
y_mul: dw 25, 129, 66, 0 |
79 |
VG_MUL dw 52, 52, 52, 52 |
u_mul: dw 112, -74, -38, 0 |
80 |
|
v_mul: dw -18, -94, 112, 0 |
81 |
|
|
|
UB_MUL dw 129, 129, 129, 129 |
|
|
VR_MUL dw 102, 102, 102, 102 |
|
82 |
|
|
83 |
|
;----------------------------------------------------------------------------- |
84 |
|
; YV12->RGB data |
85 |
|
;----------------------------------------------------------------------------- |
86 |
|
|
87 |
|
Y_SUB: dw 16, 16, 16, 16 |
88 |
|
U_SUB: dw 128, 128, 128, 128 |
89 |
|
V_SUB: dw 128, 128, 128, 128 |
90 |
|
|
91 |
section .text |
Y_MUL: dw 74, 74, 74, 74 |
92 |
|
|
93 |
%include "colorspace_mmx.inc" |
UG_MUL: dw 25, 25, 25, 25 |
94 |
|
VG_MUL: dw 52, 52, 52, 52 |
95 |
|
|
96 |
|
UB_MUL: dw 129, 129, 129, 129 |
97 |
|
VR_MUL: dw 102, 102, 102, 102 |
98 |
|
|
99 |
|
BRIGHT: db 128, 128, 128, 128, 128, 128, 128, 128 |
100 |
|
|
101 |
|
;============================================================================= |
102 |
|
; Helper macros used with the colorspace_mmx.inc file |
103 |
|
;============================================================================= |
104 |
|
|
105 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
106 |
; RGB_TO_YV12( BYTES ) |
; BGR_TO_YV12( BYTES ) |
107 |
; |
; |
108 |
; BYTES 3=rgb24, 4=rgb32 |
; BYTES 3=bgr(24bit), 4=bgra(32-bit) |
109 |
; |
; |
110 |
; bytes=3/4, pixels = 2, vpixels=2 |
; bytes=3/4, pixels = 2, vpixels=2 |
111 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
112 |
%macro RGB_TO_YV12_INIT 2 |
|
113 |
|
%macro BGR_TO_YV12_INIT 2 |
114 |
movq mm7, [y_mul] |
movq mm7, [y_mul] |
115 |
%endmacro |
%endmacro |
116 |
|
|
117 |
|
|
118 |
%macro RGB_TO_YV12 2 |
%macro BGR_TO_YV12 2 |
119 |
; y_out |
; y_out |
120 |
pxor mm4, mm4 |
pxor mm4, mm4 |
121 |
pxor mm5, mm5 |
pxor mm5, mm5 |
174 |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
175 |
|
|
176 |
; u_ptr, v_ptr |
; u_ptr, v_ptr |
|
|
|
177 |
movq mm0, mm6 ; = [ |b4|g4|r4] |
movq mm0, mm6 ; = [ |b4|g4|r4] |
178 |
pmaddwd mm6, [v_mul] ; *= V_MUL |
pmaddwd mm6, [v_mul] ; *= V_MUL |
179 |
pmaddwd mm0, [u_mul] ; *= U_MUL |
pmaddwd mm0, [u_mul] ; *= U_MUL |
196 |
|
|
197 |
pop edx |
pop edx |
198 |
%endmacro |
%endmacro |
|
;------------------------------------------------------------------------------ |
|
|
|
|
|
|
|
|
|
|
|
|
|
199 |
|
|
200 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
201 |
; YV12_TO_RGB( BYTES ) |
; YV12_TO_BGR( BYTES ) |
202 |
; |
; |
203 |
; BYTES 3=rgb24, 4=rgb32 |
; BYTES 3=bgr(24-bit), 4=bgra(32-bit) |
204 |
; |
; |
205 |
; bytes=3/4, pixels = 8, vpixels=2 |
; bytes=3/4, pixels = 8, vpixels=2 |
206 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
207 |
%macro YV12_TO_RGB_INIT 2 |
|
208 |
|
%macro YV12_TO_BGR_INIT 2 |
209 |
pxor mm7, mm7 ; clear mm7 |
pxor mm7, mm7 ; clear mm7 |
210 |
%endmacro |
%endmacro |
211 |
|
|
212 |
|
%macro YV12_TO_BGR 2 |
|
%macro YV12_TO_RGB 2 |
|
213 |
%define TEMP_Y1 esp |
%define TEMP_Y1 esp |
214 |
%define TEMP_Y2 esp + 8 |
%define TEMP_Y2 esp + 8 |
215 |
%define TEMP_G1 esp + 16 |
%define TEMP_G1 esp + 16 |
216 |
%define TEMP_G2 esp + 24 |
%define TEMP_G2 esp + 24 |
217 |
%define TEMP_B1 esp + 32 |
%define TEMP_B1 esp + 32 |
218 |
%define TEMP_B2 esp + 40 |
%define TEMP_B2 esp + 40 |
219 |
|
|
220 |
movd mm2, [ebx] ; u_ptr[0] |
movd mm2, [ebx] ; u_ptr[0] |
221 |
movd mm3, [ecx] ; v_ptr[0] |
movd mm3, [ecx] ; v_ptr[0] |
|
|
|
222 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
223 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
|
|
|
224 |
psubsw mm2, [U_SUB] ; U - 128 |
psubsw mm2, [U_SUB] ; U - 128 |
225 |
psubsw mm3, [V_SUB] ; V - 128 |
psubsw mm3, [V_SUB] ; V - 128 |
|
|
|
226 |
movq mm4, mm2 |
movq mm4, mm2 |
227 |
movq mm5, mm3 |
movq mm5, mm3 |
|
|
|
228 |
pmullw mm2, [UG_MUL] |
pmullw mm2, [UG_MUL] |
229 |
pmullw mm3, [VG_MUL] |
pmullw mm3, [VG_MUL] |
|
|
|
230 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
231 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
232 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
|
|
|
233 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
|
|
|
234 |
movq mm0, mm3 |
movq mm0, mm3 |
235 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
236 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
|
|
|
237 |
paddsw mm2, mm3 |
paddsw mm2, mm3 |
238 |
paddsw mm6, mm0 |
paddsw mm6, mm0 |
|
|
|
239 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
|
|
|
240 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
241 |
movq mm1, mm0 |
movq mm1, mm0 |
242 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
243 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
244 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
245 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
246 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
247 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
248 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
249 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
|
|
|
250 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
251 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
252 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
253 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
254 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
255 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
|
|
|
256 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
257 |
movq mm1, mm0 |
movq mm1, mm0 |
|
|
|
258 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
259 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
260 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
261 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
262 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
263 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
264 |
movq mm3, mm1 |
movq mm3, mm1 |
265 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
|
|
|
266 |
movq mm2, mm0 |
movq mm2, mm0 |
267 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
268 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
269 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
270 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
271 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
|
|
|
272 |
movq mm0, mm4 |
movq mm0, mm4 |
273 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
274 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
|
|
|
275 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
276 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
277 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
|
|
|
278 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
279 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
280 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
281 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
282 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
|
|
|
283 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
284 |
movq mm2, [TEMP_Y1] |
movq mm2, [TEMP_Y1] |
|
|
|
285 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
286 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
287 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
288 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
289 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
290 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
291 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
292 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
|
|
|
293 |
movq mm0, mm5 |
movq mm0, mm5 |
294 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
295 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
|
|
|
296 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
297 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
|
|
|
298 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
299 |
psraw mm7, SCALEBITS |
psraw mm7, SCALEBITS |
|
|
|
300 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
|
|
|
301 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
302 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
|
|
|
303 |
psraw mm6, SCALEBITS |
psraw mm6, SCALEBITS |
304 |
psraw mm4, SCALEBITS |
psraw mm4, SCALEBITS |
|
|
|
305 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
|
|
|
306 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
307 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |
|
|
|
308 |
movq mm6, mm7 |
movq mm6, mm7 |
|
|
|
309 |
movq mm2, mm0 |
movq mm2, mm0 |
310 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
311 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
312 |
pxor mm7, mm7 |
pxor mm7, mm7 |
|
|
|
313 |
movq mm3, mm1 |
movq mm3, mm1 |
314 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
315 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
316 |
movq mm4, mm2 |
movq mm4, mm2 |
317 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
318 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
319 |
movq mm5, mm0 |
movq mm5, mm0 |
320 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
321 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
322 |
|
%if %1 == 3 ; BGR (24-bit) |
|
%if %1 == 3 ; RGB24 |
|
323 |
movd [edi], mm2 |
movd [edi], mm2 |
324 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
325 |
movd [edi + 3], mm2 |
movd [edi + 3], mm2 |
326 |
movd [edi + 6], mm4 |
movd [edi + 6], mm4 |
|
|
|
327 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
328 |
movd [edi + 9], mm4 |
movd [edi + 9], mm4 |
329 |
movd [edi + 12], mm0 |
movd [edi + 12], mm0 |
|
|
|
330 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
331 |
movd [edi + 15], mm0 |
movd [edi + 15], mm0 |
332 |
movd [edi + 18], mm5 |
movq mm2, mm5 |
333 |
|
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
334 |
psrlq mm5, 32 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
335 |
|
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
336 |
movd [edi + 21], mm5 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
337 |
|
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
338 |
|
psllq mm5, 40 ; r7g7b700000 -> mm5 |
339 |
|
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
340 |
|
movq [edi + 16], mm5 |
341 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
342 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
343 |
movq mm2, mm0 |
movq mm2, mm0 |
344 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
345 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
346 |
movq mm3, mm1 |
movq mm3, mm1 |
347 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
348 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
349 |
movq mm4, mm2 |
movq mm4, mm2 |
350 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
351 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
352 |
movq mm5, mm0 |
movq mm5, mm0 |
353 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
354 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
355 |
movd [edi+edx], mm2 |
movd [edi+edx], mm2 |
356 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
357 |
movd [edi+edx + 3], mm2 |
movd [edi+edx + 3], mm2 |
358 |
movd [edi+edx + 6], mm4 |
movd [edi+edx + 6], mm4 |
|
|
|
359 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
360 |
movd [edi+edx + 9], mm4 |
movd [edi+edx + 9], mm4 |
361 |
movd [edi+edx + 12], mm0 |
movd [edi+edx + 12], mm0 |
|
|
|
362 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
363 |
movd [edi+edx + 15], mm0 |
movd [edi+edx + 15], mm0 |
364 |
movd [edi+edx + 18], mm5 |
movq mm2, mm5 |
365 |
|
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
366 |
psrlq mm5, 32 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
367 |
|
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
368 |
movd [edi+edx + 21], mm5 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
369 |
|
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
370 |
|
psllq mm5, 40 ; r7g7b700000 -> mm5 |
371 |
|
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
372 |
|
movq [edi + edx + 16], mm5 |
373 |
|
|
374 |
%else ; RGB32 |
%else ; BGRA (32-bit) |
375 |
movq [edi], mm2 |
movq [edi], mm2 |
376 |
movq [edi + 8], mm4 |
movq [edi + 8], mm4 |
377 |
movq [edi + 16], mm0 |
movq [edi + 16], mm0 |
378 |
movq [edi + 24], mm5 |
movq [edi + 24], mm5 |
|
|
|
379 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
380 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
381 |
movq mm2, mm0 |
movq mm2, mm0 |
382 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
383 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
384 |
movq mm3, mm1 |
movq mm3, mm1 |
385 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
386 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
387 |
movq mm4, mm2 |
movq mm4, mm2 |
388 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
389 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
390 |
movq mm5, mm0 |
movq mm5, mm0 |
391 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
392 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
393 |
movq [edi + edx], mm2 |
movq [edi + edx], mm2 |
394 |
movq [edi + edx + 8], mm4 |
movq [edi + edx + 8], mm4 |
395 |
movq [edi + edx + 16], mm0 |
movq [edi + edx + 16], mm0 |
403 |
%undef TEMP_B1 |
%undef TEMP_B1 |
404 |
%undef TEMP_B2 |
%undef TEMP_B2 |
405 |
%endmacro |
%endmacro |
|
;------------------------------------------------------------------------------ |
|
406 |
|
|
407 |
|
;============================================================================= |
408 |
|
; Code |
409 |
|
;============================================================================= |
410 |
|
|
411 |
|
SECTION .text |
412 |
|
|
413 |
|
%include "colorspace_mmx.inc" |
414 |
|
|
415 |
; input |
; input |
416 |
|
MAKE_COLORSPACE bgr_to_yv12_mmx,0, 3,2,2, BGR_TO_YV12, 3, -1 |
417 |
MAKE_COLORSPACE rgb24_to_yv12_mmx,0, 3,2,2, RGB_TO_YV12, 3, -1 |
MAKE_COLORSPACE bgra_to_yv12_mmx,0, 4,2,2, BGR_TO_YV12, 4, -1 |
|
MAKE_COLORSPACE rgb32_to_yv12_mmx,0, 4,2,2, RGB_TO_YV12, 4, -1 |
|
418 |
|
|
419 |
; output |
; output |
420 |
|
MAKE_COLORSPACE yv12_to_bgr_mmx,48, 3,8,2, YV12_TO_BGR, 3, -1 |
421 |
MAKE_COLORSPACE yv12_to_rgb24_mmx,48, 3,8,2, YV12_TO_RGB, 3, -1 |
MAKE_COLORSPACE yv12_to_bgra_mmx,48, 4,8,2, YV12_TO_BGR, 4, -1 |
|
MAKE_COLORSPACE yv12_to_rgb32_mmx,48, 4,8,2, YV12_TO_RGB, 4, -1 |
|
422 |
|
|