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 |
154 |
push x_stride |
push x_stride |
155 |
|
|
156 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
157 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
158 |
|
shr x_stride, SCALEBITS_IN |
159 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
160 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
161 |
|
|
162 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
163 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
164 |
|
shr x_stride, SCALEBITS_IN |
165 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
166 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
167 |
|
|
168 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
169 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
170 |
|
shr x_stride, SCALEBITS_IN |
171 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
172 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
173 |
|
|
174 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
175 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
176 |
|
shr x_stride, SCALEBITS_IN |
177 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
178 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
179 |
|
|
189 |
paddd mm2, mm6 |
paddd mm2, mm6 |
190 |
|
|
191 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
192 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
193 |
|
shr x_stride, (SCALEBITS_IN+2) |
194 |
add x_stride, U_ADD |
add x_stride, U_ADD |
195 |
mov [u_ptr], dl |
mov [u_ptr], dl |
196 |
|
|
197 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
198 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
199 |
|
shr x_stride, (SCALEBITS_IN+2) |
200 |
add x_stride, V_ADD |
add x_stride, V_ADD |
201 |
mov [v_ptr], dl |
mov [v_ptr], dl |
202 |
|
|
255 |
push x_stride |
push x_stride |
256 |
|
|
257 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
258 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
259 |
|
shr x_stride, SCALEBITS_IN |
260 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
261 |
mov [y_ptr], dl ; y_ptr[0] |
mov [y_ptr], dl ; y_ptr[0] |
262 |
|
|
263 |
movd x_stride_d, mm1 |
movd x_stride_d, mm1 |
264 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
265 |
|
shr x_stride, SCALEBITS_IN |
266 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
267 |
mov [y_ptr + 1], dl ; y_ptr[1] |
mov [y_ptr + 1], dl ; y_ptr[1] |
268 |
|
|
269 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
270 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
271 |
|
shr x_stride, SCALEBITS_IN |
272 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
273 |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
mov [y_ptr + y_stride + 0], dl ; y_ptr[y_stride + 0] |
274 |
|
|
275 |
movd x_stride_d, mm3 |
movd x_stride_d, mm3 |
276 |
shr x_stride, 8 |
add x_stride, FIX_ROUND |
277 |
|
shr x_stride, SCALEBITS_IN |
278 |
add x_stride, Y_ADD |
add x_stride, Y_ADD |
279 |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
mov [y_ptr + y_stride + 1], dl ; y_ptr[y_stride + 1] |
280 |
|
|
290 |
paddd mm2, mm6 |
paddd mm2, mm6 |
291 |
|
|
292 |
movd x_stride_d, mm0 |
movd x_stride_d, mm0 |
293 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
294 |
|
shr x_stride, (SCALEBITS_IN+2) |
295 |
add x_stride, U_ADD |
add x_stride, U_ADD |
296 |
mov [u_ptr], dl |
mov [u_ptr], dl |
297 |
|
|
298 |
movd x_stride_d, mm2 |
movd x_stride_d, mm2 |
299 |
shr x_stride, 10 |
add x_stride, 4*FIX_ROUND |
300 |
|
shr x_stride, (SCALEBITS_IN+2) |
301 |
add x_stride, V_ADD |
add x_stride, V_ADD |
302 |
mov [v_ptr], dl |
mov [v_ptr], dl |
303 |
|
|
356 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 |
357 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
358 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
359 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
360 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS_OUT |
361 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 |
362 |
movq [TEMP_G1], mm0 |
movq [TEMP_G1], mm0 |
363 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
movq mm0, [y_ptr+y_stride] ; y7y6y5y4y3y2y1y0 -> mm0 |
372 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
psubsw mm1, mm2 ; g7g6g5g4 -> mm1 |
373 |
movq mm2, mm0 |
movq mm2, mm0 |
374 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
psubsw mm0, mm6 ; g3g2g1g0 -> mm0 |
375 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
376 |
psraw mm0, SCALEBITS |
psraw mm0, SCALEBITS_OUT |
377 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 |
378 |
movq [TEMP_G2], mm0 |
movq [TEMP_G2], mm0 |
379 |
movq mm0, mm4 |
movq mm0, mm4 |
383 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
384 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
movq mm7, mm2 ; y3y2y1y0 -> mm7 |
385 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
386 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS_OUT |
387 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS_OUT |
388 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
389 |
movq [TEMP_B2], mm2 |
movq [TEMP_B2], mm2 |
390 |
movq mm3, [TEMP_Y2] |
movq mm3, [TEMP_Y2] |
393 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
paddsw mm3, mm4 ; b7b6b5b4 -> mm3 |
394 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
movq mm4, mm2 ; TEMP_Y1 -> mm4 |
395 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
paddsw mm2, mm0 ; b3b2b1b0 -> mm2 |
396 |
psraw mm3, SCALEBITS |
psraw mm3, SCALEBITS_OUT |
397 |
psraw mm2, SCALEBITS |
psraw mm2, SCALEBITS_OUT |
398 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 |
399 |
movq [TEMP_B1], mm2 |
movq [TEMP_B1], mm2 |
400 |
movq mm0, mm5 |
movq mm0, mm5 |
402 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 |
403 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
paddsw mm1, mm5 ; r7r6r5r4 -> mm1 |
404 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
paddsw mm7, mm0 ; r3r2r1r0 -> mm7 |
405 |
psraw mm1, SCALEBITS |
psraw mm1, SCALEBITS_OUT |
406 |
psraw mm7, SCALEBITS |
psraw mm7, SCALEBITS_OUT |
407 |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) |
408 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
paddsw mm6, mm5 ; r7r6r5r4 -> mm6 |
409 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
paddsw mm4, mm0 ; r3r2r1r0 -> mm4 |
410 |
psraw mm6, SCALEBITS |
psraw mm6, SCALEBITS_OUT |
411 |
psraw mm4, SCALEBITS |
psraw mm4, SCALEBITS_OUT |
412 |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) |
413 |
movq mm0, [TEMP_B1] |
movq mm0, [TEMP_B1] |
414 |
movq mm1, [TEMP_G1] |
movq mm1, [TEMP_G1] |