22 |
; * |
; * |
23 |
; ****************************************************************************/ |
; ****************************************************************************/ |
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 |
|
%define SCALEBITS 6 |
61 |
|
|
62 |
|
;============================================================================= |
63 |
|
; Read only data |
64 |
|
;============================================================================= |
65 |
|
|
66 |
|
SECTION .rodata |
67 |
|
ALIGN 16 |
68 |
|
|
69 |
|
;----------------------------------------------------------------------------- |
70 |
; RGB->YV12 multiplication matrices |
; RGB->YV12 multiplication matrices |
71 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
72 |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
|
y_mul dw 25, 129, 66, 0 |
|
|
u_mul dw 112, -74, -38, 0 |
|
|
v_mul dw -18, -94, 112, 0 |
|
73 |
|
|
74 |
|
y_mul: dw 25, 129, 66, 0 |
75 |
|
u_mul: dw 112, -74, -38, 0 |
76 |
|
v_mul: dw -18, -94, 112, 0 |
77 |
|
|
|
;=========================================================================== |
|
|
; YV12->RGB data |
|
|
;=========================================================================== |
|
|
%define SCALEBITS 6 |
|
|
Y_SUB dw 16, 16, 16, 16 |
|
|
U_SUB dw 128, 128, 128, 128 |
|
|
V_SUB dw 128, 128, 128, 128 |
|
78 |
|
|
79 |
Y_MUL dw 74, 74, 74, 74 |
;----------------------------------------------------------------------------- |
80 |
|
; YV12->RGB data |
81 |
|
;----------------------------------------------------------------------------- |
82 |
|
|
83 |
UG_MUL dw 25, 25, 25, 25 |
Y_SUB: dw 16, 16, 16, 16 |
84 |
VG_MUL dw 52, 52, 52, 52 |
U_SUB: dw 128, 128, 128, 128 |
85 |
|
V_SUB: dw 128, 128, 128, 128 |
86 |
|
|
87 |
UB_MUL dw 129, 129, 129, 129 |
Y_MUL: dw 74, 74, 74, 74 |
|
VR_MUL dw 102, 102, 102, 102 |
|
88 |
|
|
89 |
BRIGHT db 128, 128, 128, 128, 128, 128, 128, 128 |
UG_MUL: dw 25, 25, 25, 25 |
90 |
|
VG_MUL: dw 52, 52, 52, 52 |
91 |
|
|
92 |
section .text |
UB_MUL: dw 129, 129, 129, 129 |
93 |
|
VR_MUL: dw 102, 102, 102, 102 |
94 |
|
|
95 |
%include "colorspace_mmx.inc" |
BRIGHT: db 128, 128, 128, 128, 128, 128, 128, 128 |
96 |
|
|
97 |
|
;============================================================================= |
98 |
|
; Helper macros used with the colorspace_mmx.inc file |
99 |
|
;============================================================================= |
100 |
|
|
101 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
102 |
; BGR_TO_YV12( BYTES ) |
; BGR_TO_YV12( BYTES ) |
105 |
; |
; |
106 |
; bytes=3/4, pixels = 2, vpixels=2 |
; bytes=3/4, pixels = 2, vpixels=2 |
107 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
108 |
|
|
109 |
%macro BGR_TO_YV12_INIT 2 |
%macro BGR_TO_YV12_INIT 2 |
110 |
movq mm7, [y_mul] |
movq mm7, [y_mul] |
111 |
%endmacro |
%endmacro |
170 |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] |
171 |
|
|
172 |
; u_ptr, v_ptr |
; u_ptr, v_ptr |
|
|
|
173 |
movq mm0, mm6 ; = [ |b4|g4|r4] |
movq mm0, mm6 ; = [ |b4|g4|r4] |
174 |
pmaddwd mm6, [v_mul] ; *= V_MUL |
pmaddwd mm6, [v_mul] ; *= V_MUL |
175 |
pmaddwd mm0, [u_mul] ; *= U_MUL |
pmaddwd mm0, [u_mul] ; *= U_MUL |
192 |
|
|
193 |
pop edx |
pop edx |
194 |
%endmacro |
%endmacro |
|
;------------------------------------------------------------------------------ |
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
|
|
196 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
197 |
; YV12_TO_BGR( BYTES ) |
; YV12_TO_BGR( BYTES ) |
200 |
; |
; |
201 |
; bytes=3/4, pixels = 8, vpixels=2 |
; bytes=3/4, pixels = 8, vpixels=2 |
202 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
203 |
|
|
204 |
%macro YV12_TO_BGR_INIT 2 |
%macro YV12_TO_BGR_INIT 2 |
205 |
pxor mm7, mm7 ; clear mm7 |
pxor mm7, mm7 ; clear mm7 |
206 |
%endmacro |
%endmacro |
215 |
|
|
216 |
movd mm2, [ebx] ; u_ptr[0] |
movd mm2, [ebx] ; u_ptr[0] |
217 |
movd mm3, [ecx] ; v_ptr[0] |
movd mm3, [ecx] ; v_ptr[0] |
|
|
|
218 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 |
219 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 |
|
|
|
220 |
psubsw mm2, [U_SUB] ; U - 128 |
psubsw mm2, [U_SUB] ; U - 128 |
221 |
psubsw mm3, [V_SUB] ; V - 128 |
psubsw mm3, [V_SUB] ; V - 128 |
|
|
|
222 |
movq mm4, mm2 |
movq mm4, mm2 |
223 |
movq mm5, mm3 |
movq mm5, mm3 |
|
|
|
224 |
pmullw mm2, [UG_MUL] |
pmullw mm2, [UG_MUL] |
225 |
pmullw mm3, [VG_MUL] |
pmullw mm3, [VG_MUL] |
|
|
|
226 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
movq mm6, mm2 ; u3u2u1u0 -> mm6 |
227 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 |
228 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 |
|
|
|
229 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
pmullw mm4, [UB_MUL] ; B_ADD -> mm4 |
|
|
|
230 |
movq mm0, mm3 |
movq mm0, mm3 |
231 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 |
232 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 |
|
|
|
233 |
paddsw mm2, mm3 |
paddsw mm2, mm3 |
234 |
paddsw mm6, mm0 |
paddsw mm6, mm0 |
|
|
|
235 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
pmullw mm5, [VR_MUL] ; R_ADD -> mm5 |
|
|
|
236 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
237 |
movq mm1, mm0 |
movq mm1, mm0 |
238 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
239 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
240 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
241 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
242 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
243 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
244 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 |
245 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
|
|
|
246 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
247 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
248 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
249 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
250 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
251 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
|
|
|
252 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 |
|
|
|
253 |
movq mm1, mm0 |
movq mm1, mm0 |
|
|
|
254 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 |
255 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 |
|
|
|
256 |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
psubsw mm0, [Y_SUB] ; Y - Y_SUB |
257 |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
psubsw mm1, [Y_SUB] ; Y - Y_SUB |
|
|
|
258 |
pmullw mm1, [Y_MUL] |
pmullw mm1, [Y_MUL] |
259 |
pmullw mm0, [Y_MUL] |
pmullw mm0, [Y_MUL] |
|
|
|
260 |
movq mm3, mm1 |
movq mm3, mm1 |
261 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
|
|
|
262 |
movq mm2, mm0 |
movq mm2, mm0 |
263 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
|
|
|
264 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
265 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS |
|
|
|
266 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
|
|
|
267 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
|
|
|
268 |
movq mm0, mm4 |
movq mm0, mm4 |
269 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 |
270 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 |
|
|
|
271 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
movq mm1, mm3 ; y7y6y5y4 -> mm1 |
272 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
273 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
|
|
|
274 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
275 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
276 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
277 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
278 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
|
|
|
279 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
280 |
movq mm2, [TEMP_Y1] |
movq mm2, [TEMP_Y1] |
|
|
|
281 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
movq mm6, mm3 ; TEMP_Y2 -> mm6 |
282 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
|
|
|
283 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
284 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
|
|
|
285 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS |
286 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS |
|
|
|
287 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
|
|
|
288 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
|
|
|
289 |
movq mm0, mm5 |
movq mm0, mm5 |
290 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 |
291 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
|
|
|
292 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
293 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
|
|
|
294 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS |
295 |
psraw mm7, SCALEBITS |
psraw mm7, SCALEBITS |
|
|
|
296 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
|
|
|
297 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
298 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
|
|
|
299 |
psraw mm6, SCALEBITS |
psraw mm6, SCALEBITS |
300 |
psraw mm4, SCALEBITS |
psraw mm4, SCALEBITS |
|
|
|
301 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
|
|
|
302 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
303 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |
|
|
|
304 |
movq mm6, mm7 |
movq mm6, mm7 |
|
|
|
305 |
movq mm2, mm0 |
movq mm2, mm0 |
306 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 |
307 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
308 |
pxor mm7, mm7 |
pxor mm7, mm7 |
|
|
|
309 |
movq mm3, mm1 |
movq mm3, mm1 |
310 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
311 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
312 |
movq mm4, mm2 |
movq mm4, mm2 |
313 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
314 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
315 |
movq mm5, mm0 |
movq mm5, mm0 |
316 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
317 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
318 |
%if %1 == 3 ; BGR (24-bit) |
%if %1 == 3 ; BGR (24-bit) |
319 |
movd [edi], mm2 |
movd [edi], mm2 |
320 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
321 |
movd [edi + 3], mm2 |
movd [edi + 3], mm2 |
322 |
movd [edi + 6], mm4 |
movd [edi + 6], mm4 |
|
|
|
323 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
324 |
movd [edi + 9], mm4 |
movd [edi + 9], mm4 |
325 |
movd [edi + 12], mm0 |
movd [edi + 12], mm0 |
|
|
|
326 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
327 |
movd [edi + 15], mm0 |
movd [edi + 15], mm0 |
328 |
movq mm2, mm5 |
movq mm2, mm5 |
|
|
|
329 |
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
330 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
|
|
|
331 |
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
332 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
|
|
|
333 |
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
334 |
psllq mm5, 40 ; r7g7b700000 -> mm5 |
psllq mm5, 40 ; r7g7b700000 -> mm5 |
|
|
|
335 |
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
336 |
movq [edi + 16], mm5 |
movq [edi + 16], mm5 |
|
|
|
337 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
338 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
339 |
movq mm2, mm0 |
movq mm2, mm0 |
340 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
341 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
342 |
movq mm3, mm1 |
movq mm3, mm1 |
343 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
344 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
345 |
movq mm4, mm2 |
movq mm4, mm2 |
346 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
347 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
348 |
movq mm5, mm0 |
movq mm5, mm0 |
349 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
350 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
351 |
movd [edi+edx], mm2 |
movd [edi+edx], mm2 |
352 |
psrlq mm2, 32 |
psrlq mm2, 32 |
|
|
|
353 |
movd [edi+edx + 3], mm2 |
movd [edi+edx + 3], mm2 |
354 |
movd [edi+edx + 6], mm4 |
movd [edi+edx + 6], mm4 |
|
|
|
355 |
psrlq mm4, 32 |
psrlq mm4, 32 |
|
|
|
356 |
movd [edi+edx + 9], mm4 |
movd [edi+edx + 9], mm4 |
357 |
movd [edi+edx + 12], mm0 |
movd [edi+edx + 12], mm0 |
|
|
|
358 |
psrlq mm0, 32 |
psrlq mm0, 32 |
|
|
|
359 |
movd [edi+edx + 15], mm0 |
movd [edi+edx + 15], mm0 |
360 |
movq mm2, mm5 |
movq mm2, mm5 |
|
|
|
361 |
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
psrlq mm0, 8 ; 000000r5g5 -> mm0 |
362 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
psllq mm2, 32 ; 0r6g6b60000 -> mm2 |
|
|
|
363 |
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
psrlq mm5, 32 ; 00000r7g7b7 -> mm5 |
364 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
psrlq mm2, 16 ; 000r6g6b600 -> mm2 |
|
|
|
365 |
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 |
366 |
psllq mm5, 40 ; r7g7b700000 -> mm5 |
psllq mm5, 40 ; r7g7b700000 -> mm5 |
|
|
|
367 |
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 |
368 |
movq [edi + edx + 16], mm5 |
movq [edi + edx + 16], mm5 |
369 |
|
|
372 |
movq [edi + 8], mm4 |
movq [edi + 8], mm4 |
373 |
movq [edi + 16], mm0 |
movq [edi + 16], mm0 |
374 |
movq [edi + 24], mm5 |
movq [edi + 24], mm5 |
|
|
|
375 |
movq mm0, [TEMP_B2] |
movq mm0, [TEMP_B2] |
376 |
movq mm1, [TEMP_G2] |
movq mm1, [TEMP_G2] |
|
|
|
377 |
movq mm2, mm0 |
movq mm2, mm0 |
378 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 |
379 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 |
|
|
|
380 |
movq mm3, mm1 |
movq mm3, mm1 |
381 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 |
382 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 |
|
|
|
383 |
movq mm4, mm2 |
movq mm4, mm2 |
384 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 |
385 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 |
|
|
|
386 |
movq mm5, mm0 |
movq mm5, mm0 |
387 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 |
388 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 |
|
|
|
389 |
movq [edi + edx], mm2 |
movq [edi + edx], mm2 |
390 |
movq [edi + edx + 8], mm4 |
movq [edi + edx + 8], mm4 |
391 |
movq [edi + edx + 16], mm0 |
movq [edi + edx + 16], mm0 |
399 |
%undef TEMP_B1 |
%undef TEMP_B1 |
400 |
%undef TEMP_B2 |
%undef TEMP_B2 |
401 |
%endmacro |
%endmacro |
|
;------------------------------------------------------------------------------ |
|
402 |
|
|
403 |
|
;============================================================================= |
404 |
|
; Code |
405 |
|
;============================================================================= |
406 |
|
|
407 |
|
SECTION .text |
408 |
|
|
409 |
|
%include "colorspace_mmx.inc" |
410 |
|
|
411 |
; input |
; input |
|
|
|
412 |
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 |
413 |
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 |
414 |
|
|
415 |
; output |
; output |
|
|
|
416 |
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 |
417 |
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 |
418 |
|
|