48 |
%define V_ADD 128 |
%define V_ADD 128 |
49 |
|
|
50 |
; Scaling used during conversion |
; Scaling used during conversion |
51 |
%define SCALEBITS_OUT 6 |
%define SCALEBITS 6 |
|
%define SCALEBITS_IN 13 |
|
|
|
|
|
%define FIX_ROUND (1<<(SCALEBITS_IN-1)) |
|
52 |
|
|
53 |
;============================================================================= |
;============================================================================= |
54 |
; Read only data |
; Read only data |
63 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
64 |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored |
65 |
|
|
66 |
bgr_y_mul: dw 803, 4129, 2105, 0 |
bgr_y_mul: dw 25, 129, 66, 0 |
67 |
bgr_u_mul: dw 3596, -2384, -1212, 0 |
bgr_u_mul: dw 112, -74, -38, 0 |
68 |
bgr_v_mul: dw -582, -3015, 3596, 0 |
bgr_v_mul: dw -18, -94, 112, 0 |
69 |
|
|
70 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
71 |
; BGR->YV12 multiplication matrices |
; BGR->YV12 multiplication matrices |
72 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
73 |
; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored |
; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored |
74 |
|
|
75 |
rgb_y_mul: dw 2105, 4129, 803, 0 |
rgb_y_mul: dw 66, 129, 25, 0 |
76 |
rgb_u_mul: dw -1212, -2384, 3596, 0 |
rgb_u_mul: dw -38, -74, 112, 0 |
77 |
rgb_v_mul: dw 3596, -3015, -582, 0 |
rgb_v_mul: dw 112, -94, -18, 0 |
78 |
|
|
79 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
80 |
; YV12->RGB data |
; YV12->RGB data |
151 |
push x_stride |
push x_stride |
152 |
|
|
153 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
154 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
155 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
156 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
157 |
|
|
158 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
159 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
160 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
161 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
162 |
|
|
163 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
164 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
165 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
166 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
167 |
|
|
168 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
169 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
170 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
171 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
172 |
|
|
182 |
paddd mm2, mm6 |
paddd mm2, mm6 |
183 |
|
|
184 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
185 |
add x_stride, 4*FIX_ROUND |
shr x_stride, 10 |
|
shr x_stride, (SCALEBITS_IN+2) |
|
186 |
add x_stride, U_ADD |
add x_stride, U_ADD |
187 |
mov [u_ptr], dl |
mov [u_ptr], dl |
188 |
|
|
189 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
190 |
add x_stride, 4*FIX_ROUND |
shr x_stride, 10 |
|
shr x_stride, (SCALEBITS_IN+2) |
|
191 |
add x_stride, V_ADD |
add x_stride, V_ADD |
192 |
mov [v_ptr], dl |
mov [v_ptr], dl |
193 |
|
|
246 |
push x_stride |
push x_stride |
247 |
|
|
248 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
249 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
250 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
251 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
252 |
|
|
253 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
254 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
255 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
256 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
257 |
|
|
258 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
259 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
260 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
261 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
262 |
|
|
263 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
264 |
add x_stride, FIX_ROUND |
shr x_stride, 8 |
|
shr x_stride, SCALEBITS_IN |
|
265 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
266 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
267 |
|
|
277 |
paddd mm2, mm6 |
paddd mm2, mm6 |
278 |
|
|
279 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
280 |
add x_stride, 4*FIX_ROUND |
shr x_stride, 10 |
|
shr x_stride, (SCALEBITS_IN+2) |
|
281 |
add x_stride, U_ADD |
add x_stride, U_ADD |
282 |
mov [u_ptr], dl |
mov [u_ptr], dl |
283 |
|
|
284 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
285 |
add x_stride, 4*FIX_ROUND |
shr x_stride, 10 |
|
shr x_stride, (SCALEBITS_IN+2) |
|
286 |
add x_stride, V_ADD |
add x_stride, V_ADD |
287 |
mov [v_ptr], dl |
mov [v_ptr], dl |
288 |
|
|
341 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
342 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
343 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
344 |
psraw mm1, SCALEBITS_OUT |
psraw mm1, SCALEBITS |
345 |
psraw mm0, SCALEBITS_OUT |
psraw mm0, SCALEBITS |
346 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
347 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
348 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
357 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
358 |
movq mm2, mm0 |
movq mm2, mm0 |
359 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
360 |
psraw mm1, SCALEBITS_OUT |
psraw mm1, SCALEBITS |
361 |
psraw mm0, SCALEBITS_OUT |
psraw mm0, SCALEBITS |
362 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
363 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
364 |
movq mm0, mm4 |
movq mm0, mm4 |
368 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
369 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
370 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
371 |
psraw mm3, SCALEBITS_OUT |
psraw mm3, SCALEBITS |
372 |
psraw mm2, SCALEBITS_OUT |
psraw mm2, SCALEBITS |
373 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
374 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
375 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
378 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
379 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
380 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
381 |
psraw mm3, SCALEBITS_OUT |
psraw mm3, SCALEBITS |
382 |
psraw mm2, SCALEBITS_OUT |
psraw mm2, SCALEBITS |
383 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
384 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
385 |
movq mm0, mm5 |
movq mm0, mm5 |
387 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
388 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
389 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
390 |
psraw mm1, SCALEBITS_OUT |
psraw mm1, SCALEBITS |
391 |
psraw mm7, SCALEBITS_OUT |
psraw mm7, SCALEBITS |
392 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
393 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
394 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
395 |
psraw mm6, SCALEBITS_OUT |
psraw mm6, SCALEBITS |
396 |
psraw mm4, SCALEBITS_OUT |
psraw mm4, SCALEBITS |
397 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
398 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
399 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |
514 |
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 |
515 |
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 |
516 |
|
|
517 |
NON_EXEC_STACK |
|
518 |
|
%ifidn __OUTPUT_FORMAT__,elf |
519 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
520 |
|
%endif |
521 |
|
|