48 |
%define V_ADD 128 |
%define V_ADD 128 |
49 |
|
|
50 |
; Scaling used during conversion |
; Scaling used during conversion |
51 |
%define SCALEBITS 6 |
%define SCALEBITS_OUT 6 |
52 |
|
%define SCALEBITS_IN 13 |
53 |
|
|
54 |
|
%define FIX_ROUND (1<<(SCALEBITS_IN-1)) |
55 |
|
|
56 |
;============================================================================= |
;============================================================================= |
57 |
; Read only data |
; Read only data |
66 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
67 |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
68 |
|
|
69 |
bgr_y_mul: dw 25, 129, 66, 0 |
bgr_y_mul: dw 803, 4129, 2105, 0 |
70 |
bgr_u_mul: dw 112, -74, -38, 0 |
bgr_u_mul: dw 3596, -2384, -1212, 0 |
71 |
bgr_v_mul: dw -18, -94, 112, 0 |
bgr_v_mul: dw -582, -3015, 3596, 0 |
72 |
|
|
73 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
74 |
; BGR->YV12 multiplication matrices |
; BGR->YV12 multiplication matrices |
75 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
76 |
; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored |
; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored |
77 |
|
|
78 |
rgb_y_mul: dw 66, 129, 25, 0 |
rgb_y_mul: dw 2105, 4129, 803, 0 |
79 |
rgb_u_mul: dw -38, -74, 112, 0 |
rgb_u_mul: dw -1212, -2384, 3596, 0 |
80 |
rgb_v_mul: dw 112, -94, -18, 0 |
rgb_v_mul: dw 3596, -3015, -582, 0 |
81 |
|
|
82 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
83 |
; YV12->RGB data |
; YV12->RGB data |
136 |
|
|
137 |
pxor mm4, mm4 |
pxor mm4, mm4 |
138 |
pxor mm5, mm5 |
pxor mm5, mm5 |
139 |
|
%if %1 == 3 ; BGR (24-bit) |
140 |
|
movd mm1, [x_ptr+2] |
141 |
|
movd mm3, [x_ptr+x_stride+2] |
142 |
|
psrlq mm1, 8 ; src[%1...] |
143 |
|
psrlq mm3, 8 ; src[x_stride+%1...] |
144 |
|
%else |
145 |
movd mm1, [x_ptr+%1] ; src[%1...] |
movd mm1, [x_ptr+%1] ; src[%1...] |
146 |
movd mm3, [x_ptr+x_stride+%1] ; src[x_stride+%1...] |
movd mm3, [x_ptr+x_stride+%1] ; src[x_stride+%1...] |
147 |
|
%endif |
148 |
punpcklbw mm1, mm4 ; [ |b |g |r ] |
punpcklbw mm1, mm4 ; [ |b |g |r ] |
149 |
punpcklbw mm3, mm5 ; [ |b |g |r ] |
punpcklbw mm3, mm5 ; [ |b |g |r ] |
150 |
paddw mm6, mm1 ; +[ |b4|g4|r4] |
paddw mm6, mm1 ; +[ |b4|g4|r4] |
161 |
push x_stride |
push x_stride |
162 |
|
|
163 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
164 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
165 |
|
shr x_stride, SCALEBITS_IN |
166 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
167 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
168 |
|
|
169 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
170 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
171 |
|
shr x_stride, SCALEBITS_IN |
172 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
173 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
174 |
|
|
175 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
176 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
177 |
|
shr x_stride, SCALEBITS_IN |
178 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
179 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
180 |
|
|
181 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
182 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
183 |
|
shr x_stride, SCALEBITS_IN |
184 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
185 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
186 |
|
|
196 |
paddd mm2, mm6 |
paddd mm2, mm6 |
197 |
|
|
198 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
199 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
200 |
|
shr x_stride, (SCALEBITS_IN+2) |
201 |
add x_stride, U_ADD |
add x_stride, U_ADD |
202 |
mov [u_ptr], dl |
mov [u_ptr], dl |
203 |
|
|
204 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
205 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
206 |
|
shr x_stride, (SCALEBITS_IN+2) |
207 |
add x_stride, V_ADD |
add x_stride, V_ADD |
208 |
mov [v_ptr], dl |
mov [v_ptr], dl |
209 |
|
|
244 |
|
|
245 |
pxor mm4, mm4 |
pxor mm4, mm4 |
246 |
pxor mm5, mm5 |
pxor mm5, mm5 |
247 |
|
%if %1 == 3 ; BGR (24-bit) |
248 |
|
movd mm1, [x_ptr+2] |
249 |
|
movd mm3, [x_ptr+x_stride+2] |
250 |
|
psrlq mm1, 8 ; src[%1...] |
251 |
|
psrlq mm3, 8 ; src[x_stride+%1...] |
252 |
|
%else |
253 |
movd mm1, [x_ptr+%1] ; src[%1...] |
movd mm1, [x_ptr+%1] ; src[%1...] |
254 |
movd mm3, [x_ptr+x_stride+%1] ; src[x_stride+%1...] |
movd mm3, [x_ptr+x_stride+%1] ; src[x_stride+%1...] |
255 |
|
%endif |
256 |
punpcklbw mm1, mm4 ; [ |b |g |r ] |
punpcklbw mm1, mm4 ; [ |b |g |r ] |
257 |
punpcklbw mm3, mm5 ; [ |b |g |r ] |
punpcklbw mm3, mm5 ; [ |b |g |r ] |
258 |
paddw mm6, mm1 ; +[ |b4|g4|r4] |
paddw mm6, mm1 ; +[ |b4|g4|r4] |
269 |
push x_stride |
push x_stride |
270 |
|
|
271 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
272 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
273 |
|
shr x_stride, SCALEBITS_IN |
274 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
275 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
276 |
|
|
277 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
278 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
279 |
|
shr x_stride, SCALEBITS_IN |
280 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
281 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
282 |
|
|
283 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
284 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
285 |
|
shr x_stride, SCALEBITS_IN |
286 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
287 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
288 |
|
|
289 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
290 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
291 |
|
shr x_stride, SCALEBITS_IN |
292 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
293 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
294 |
|
|
304 |
paddd mm2, mm6 |
paddd mm2, mm6 |
305 |
|
|
306 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
307 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
308 |
|
shr x_stride, (SCALEBITS_IN+2) |
309 |
add x_stride, U_ADD |
add x_stride, U_ADD |
310 |
mov [u_ptr], dl |
mov [u_ptr], dl |
311 |
|
|
312 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
313 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
314 |
|
shr x_stride, (SCALEBITS_IN+2) |
315 |
add x_stride, V_ADD |
add x_stride, V_ADD |
316 |
mov [v_ptr], dl |
mov [v_ptr], dl |
317 |
|
|
370 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
371 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
372 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
373 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
374 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS_OUT |
375 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
376 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
377 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
386 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
387 |
movq mm2, mm0 |
movq mm2, mm0 |
388 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
389 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
390 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS_OUT |
391 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
392 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
393 |
movq mm0, mm4 |
movq mm0, mm4 |
397 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
398 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
399 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
400 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS_OUT |
401 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS_OUT |
402 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
403 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
404 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
407 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
408 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
409 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
410 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS_OUT |
411 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS_OUT |
412 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
413 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
414 |
movq mm0, mm5 |
movq mm0, mm5 |
416 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
417 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
418 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
419 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
420 |
psraw mm7, SCALEBITS |
psraw mm7, SCALEBITS_OUT |
421 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
422 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
423 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
424 |
psraw mm6, SCALEBITS |
psraw mm6, SCALEBITS_OUT |
425 |
psraw mm4, SCALEBITS |
psraw mm4, SCALEBITS_OUT |
426 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
427 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
428 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |
529 |
; Code |
; Code |
530 |
;============================================================================= |
;============================================================================= |
531 |
|
|
532 |
SECTION .rotext align=SECTION_ALIGN |
TEXT |
533 |
|
|
534 |
%include "colorspace_mmx.inc" |
%include "colorspace_mmx.inc" |
535 |
|
|
543 |
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 |
544 |
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 |
545 |
|
|
546 |
|
NON_EXEC_STACK |
|
%ifidn __OUTPUT_FORMAT__,elf |
|
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
|
|
%endif |
|
|
|
|