23 |
; * |
; * |
24 |
; ***************************************************************************/ |
; ***************************************************************************/ |
25 |
|
|
26 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
%ifdef MARK_FUNCS |
|
|
global _%1:function %1.endfunc-%1 |
|
|
%define %1 _%1:function %1.endfunc-%1 |
|
|
%else |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%endif |
|
|
%else |
|
|
%ifdef MARK_FUNCS |
|
|
global %1:function %1.endfunc-%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endif |
|
|
%endmacro |
|
27 |
|
|
28 |
;============================================================================= |
;============================================================================= |
29 |
; Read only data |
; Read only data |
30 |
;============================================================================= |
;============================================================================= |
31 |
|
|
32 |
%ifdef FORMAT_COFF |
DATA |
|
SECTION .rodata |
|
|
%else |
|
|
SECTION .rodata align=16 |
|
|
%endif |
|
33 |
|
|
34 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
35 |
; yuyv/uyvy mask for extracting yuv components |
; yuyv/uyvy mask for extracting yuv components |
36 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
37 |
; y u y v y u y v |
; y u y v y u y v |
38 |
|
|
39 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
40 |
yuyv_mask: db 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0 |
yuyv_mask: db 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0 |
41 |
mmx_one: dw 1, 1, 1, 1 |
mmx_one: dw 1, 1, 1, 1 |
42 |
|
|
59 |
|
|
60 |
|
|
61 |
%macro YUYV_TO_YV12 2 |
%macro YUYV_TO_YV12 2 |
62 |
movq mm0, [edi] ; x_ptr[0] |
movq mm0, [x_ptr] ; x_ptr[0] |
63 |
movq mm1, [edi + 8] ; x_ptr[8] |
movq mm1, [x_ptr + 8] ; x_ptr[8] |
64 |
movq mm2, [edi + edx] ; x_ptr[x_stride + 0] |
movq mm2, [x_ptr + x_stride] ; x_ptr[x_stride + 0] |
65 |
movq mm3, [edi + edx + 8] ; x_ptr[x_stride + 8] |
movq mm3, [x_ptr + x_stride + 8] ; x_ptr[x_stride + 8] |
66 |
|
|
67 |
; average uv-components |
; average uv-components |
68 |
;---[ plain mmx ]---------------------------------------------------- |
;---[ plain mmx ]---------------------------------------------------- |
128 |
packuswb mm2, mm3 |
packuswb mm2, mm3 |
129 |
|
|
130 |
%ifidn %2,pavgb ; xmm |
%ifidn %2,pavgb ; xmm |
131 |
movntq [esi], mm0 |
movntq [y_ptr], mm0 |
132 |
movntq [esi+eax], mm2 |
movntq [y_ptr+y_stride], mm2 |
133 |
%else ; plain mmx,3dnow |
%else ; plain mmx,3dnow |
134 |
movq [esi], mm0 |
movq [y_ptr], mm0 |
135 |
movq [esi+eax], mm2 |
movq [y_ptr+y_stride], mm2 |
136 |
%endif |
%endif |
137 |
|
|
138 |
; write uv-components |
; write uv-components |
143 |
pand mm4, mm7 |
pand mm4, mm7 |
144 |
packuswb mm5,mm5 |
packuswb mm5,mm5 |
145 |
packuswb mm4,mm4 |
packuswb mm4,mm4 |
146 |
movd [ebx],mm5 |
movd [u_ptr],mm5 |
147 |
movd [ecx],mm4 |
movd [v_ptr],mm4 |
148 |
%endmacro |
%endmacro |
149 |
|
|
150 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
151 |
; YV12_TO_YUYV( TYPE ) |
; YV12_TO_YUYV( TYPE ) |
152 |
; |
; |
153 |
; TYPE 0=yuyv, 1=uyvy |
; bytes=2, pixels = 16, vpixels=2 |
|
; |
|
|
; bytes=2, pixels = 8, vpixels=2 |
|
154 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
155 |
|
|
156 |
%macro YV12_TO_YUYV_INIT 2 |
%macro YV12_TO_YUYV_INIT 2 |
158 |
|
|
159 |
|
|
160 |
%macro YV12_TO_YUYV 2 |
%macro YV12_TO_YUYV 2 |
161 |
movd mm4, [ebx] ; [ |uuuu] |
movq mm6, [u_ptr] ; [ |uuuu] |
162 |
movd mm5, [ecx] ; [ |vvvv] |
movq mm2, [v_ptr] ; [ |vvvv] |
163 |
movq mm0, [esi] ; [yyyy|yyyy] ; y row 0 |
movq mm0, [y_ptr ] ; [yyyy|yyyy] ; y row 0 |
164 |
movq mm1, [esi+eax] ; [yyyy|yyyy] ; y row 1 |
movq mm1, [y_ptr+y_stride] ; [yyyy|yyyy] ; y row 1 |
165 |
punpcklbw mm4, mm5 ; [vuvu|vuvu] ; uv row 0 |
movq mm7, mm6 |
166 |
|
punpcklbw mm6, mm2 ; [vuvu|vuvu] ; uv[0..3] |
167 |
|
punpckhbw mm7, mm2 ; [vuvu|vuvu] ; uv[4..7] |
168 |
|
|
169 |
%if %1 == 0 ; YUYV |
%if %1 == 0 ; YUYV |
170 |
movq mm2, mm0 |
movq mm2, mm0 |
171 |
movq mm3, mm1 |
movq mm3, mm1 |
172 |
punpcklbw mm0, mm4 ; [vyuy|vyuy] ; y row 0 + 0 |
movq mm4, [y_ptr +8] ; [yyyy|yyyy] ; y[8..15] row 0 |
173 |
punpckhbw mm2, mm4 ; [vyuy|vyuy] ; y row 0 + 8 |
movq mm5, [y_ptr+y_stride+8] ; [yyyy|yyyy] ; y[8..15] row 1 |
174 |
punpcklbw mm1, mm4 ; [vyuy|vyuy] ; y row 1 + 0 |
punpcklbw mm0, mm6 ; [vyuy|vyuy] ; y row 0 + 0 |
175 |
punpckhbw mm3, mm4 ; [vyuy|vyuy] ; y row 1 + 8 |
punpckhbw mm2, mm6 ; [vyuy|vyuy] ; y row 0 + 8 |
176 |
movq [edi], mm0 |
punpcklbw mm1, mm6 ; [vyuy|vyuy] ; y row 1 + 0 |
177 |
movq [edi+8], mm2 |
punpckhbw mm3, mm6 ; [vyuy|vyuy] ; y row 1 + 8 |
178 |
movq [edi+edx], mm1 |
movq [x_ptr ], mm0 |
179 |
movq [edi+edx+8], mm3 |
movq [x_ptr+8 ], mm2 |
180 |
|
movq [x_ptr+x_stride ], mm1 |
181 |
|
movq [x_ptr+x_stride+8], mm3 |
182 |
|
movq mm0, mm4 |
183 |
|
movq mm2, mm5 |
184 |
|
punpcklbw mm0, mm7 ; [vyuy|vyuy] ; y row 0 + 16 |
185 |
|
punpckhbw mm4, mm7 ; [vyuy|vyuy] ; y row 0 + 24 |
186 |
|
punpcklbw mm2, mm7 ; [vyuy|vyuy] ; y row 1 + 16 |
187 |
|
punpckhbw mm5, mm7 ; [vyuy|vyuy] ; y row 1 + 24 |
188 |
|
movq [x_ptr +16], mm0 |
189 |
|
movq [x_ptr +24], mm4 |
190 |
|
movq [x_ptr+x_stride+16], mm2 |
191 |
|
movq [x_ptr+x_stride+24], mm5 |
192 |
%else ; UYVY |
%else ; UYVY |
193 |
movq mm5, mm4 |
movq mm2, mm6 |
194 |
movq mm6, mm4 |
movq mm3, mm6 |
195 |
movq mm7, mm4 |
movq mm4, mm6 |
196 |
punpcklbw mm4, mm0 ; [yvyu|yvyu] ; y row 0 + 0 |
punpcklbw mm2, mm0 ; [yvyu|yvyu] ; y row 0 + 0 |
197 |
punpckhbw mm5, mm0 ; [yvyu|yvyu] ; y row 0 + 8 |
punpckhbw mm3, mm0 ; [yvyu|yvyu] ; y row 0 + 8 |
198 |
punpcklbw mm6, mm1 ; [yvyu|yvyu] ; y row 1 + 0 |
movq mm0, [y_ptr +8] ; [yyyy|yyyy] ; y[8..15] row 0 |
199 |
punpckhbw mm7, mm1 ; [yvyu|yvyu] ; y row 1 + 8 |
movq mm5, [y_ptr+y_stride+8] ; [yyyy|yyyy] ; y[8..15] row 1 |
200 |
movq [edi], mm4 |
punpcklbw mm4, mm1 ; [yvyu|yvyu] ; y row 1 + 0 |
201 |
movq [edi+8], mm5 |
punpckhbw mm6, mm1 ; [yvyu|yvyu] ; y row 1 + 8 |
202 |
movq [edi+edx], mm6 |
movq [x_ptr ], mm2 |
203 |
movq [edi+edx+8], mm7 |
movq [x_ptr +8], mm3 |
204 |
|
movq [x_ptr+x_stride ], mm4 |
205 |
|
movq [x_ptr+x_stride+8], mm6 |
206 |
|
movq mm2, mm7 |
207 |
|
movq mm3, mm7 |
208 |
|
movq mm6, mm7 |
209 |
|
punpcklbw mm2, mm0 ; [yvyu|yvyu] ; y row 0 + 0 |
210 |
|
punpckhbw mm3, mm0 ; [yvyu|yvyu] ; y row 0 + 8 |
211 |
|
punpcklbw mm6, mm5 ; [yvyu|yvyu] ; y row 1 + 0 |
212 |
|
punpckhbw mm7, mm5 ; [yvyu|yvyu] ; y row 1 + 8 |
213 |
|
movq [x_ptr +16], mm2 |
214 |
|
movq [x_ptr +24], mm3 |
215 |
|
movq [x_ptr+x_stride+16], mm6 |
216 |
|
movq [x_ptr+x_stride+24], mm7 |
217 |
%endif |
%endif |
218 |
%endmacro |
%endmacro |
219 |
|
|
229 |
%endmacro |
%endmacro |
230 |
|
|
231 |
%macro YV12_TO_YUYVI 2 |
%macro YV12_TO_YUYVI 2 |
232 |
xchg ebp, [uv_stride] |
%ifdef ARCH_IS_X86_64 |
233 |
movd mm0, [ebx] ; [ |uuuu] |
mov TMP1d, prm_uv_stride |
234 |
movd mm1, [ebx+ebp] ; [ |uuuu] |
movd mm0, [u_ptr] ; [ |uuuu] |
235 |
punpcklbw mm0, [ecx] ; [vuvu|vuvu] ; uv row 0 |
movd mm1, [u_ptr+TMP1] ; [ |uuuu] |
236 |
punpcklbw mm1, [ecx+ebp] ; [vuvu|vuvu] ; uv row 1 |
punpcklbw mm0, [v_ptr] ; [vuvu|vuvu] ; uv row 0 |
237 |
xchg ebp, [uv_stride] |
punpcklbw mm1, [v_ptr+TMP1] ; [vuvu|vuvu] ; uv row 1 |
238 |
|
%else |
239 |
|
xchg width, prm_uv_stride |
240 |
|
movd mm0, [u_ptr] ; [ |uuuu] |
241 |
|
movd mm1, [u_ptr+width] ; [ |uuuu] |
242 |
|
punpcklbw mm0, [v_ptr] ; [vuvu|vuvu] ; uv row 0 |
243 |
|
punpcklbw mm1, [v_ptr+width] ; [vuvu|vuvu] ; uv row 1 |
244 |
|
xchg width, prm_uv_stride |
245 |
|
%endif |
246 |
|
|
247 |
%if %1 == 0 ; YUYV |
%if %1 == 0 ; YUYV |
248 |
movq mm4, [esi] ; [yyyy|yyyy] ; y row 0 |
movq mm4, [y_ptr] ; [yyyy|yyyy] ; y row 0 |
249 |
movq mm6, [esi+eax] ; [yyyy|yyyy] ; y row 1 |
movq mm6, [y_ptr+y_stride] ; [yyyy|yyyy] ; y row 1 |
250 |
movq mm5, mm4 |
movq mm5, mm4 |
251 |
movq mm7, mm6 |
movq mm7, mm6 |
252 |
punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 0 + 0 |
punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 0 + 0 |
253 |
punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 0 + 8 |
punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 0 + 8 |
254 |
punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 1 + 0 |
punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 1 + 0 |
255 |
punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 1 + 8 |
punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 1 + 8 |
256 |
movq [edi], mm4 |
movq [x_ptr], mm4 |
257 |
movq [edi+8], mm5 |
movq [x_ptr+8], mm5 |
258 |
movq [edi+edx], mm6 |
movq [x_ptr+x_stride], mm6 |
259 |
movq [edi+edx+8], mm7 |
movq [x_ptr+x_stride+8], mm7 |
260 |
|
|
261 |
push esi |
push y_ptr |
262 |
push edi |
push x_ptr |
263 |
add esi, eax |
add y_ptr, y_stride |
264 |
add edi, edx |
add x_ptr, x_stride |
265 |
movq mm4, [esi+eax] ; [yyyy|yyyy] ; y row 2 |
movq mm4, [y_ptr+y_stride] ; [yyyy|yyyy] ; y row 2 |
266 |
movq mm6, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 |
movq mm6, [y_ptr+2*y_stride] ; [yyyy|yyyy] ; y row 3 |
267 |
movq mm5, mm4 |
movq mm5, mm4 |
268 |
movq mm7, mm6 |
movq mm7, mm6 |
269 |
punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 2 + 0 |
punpcklbw mm4, mm0 ; [yuyv|yuyv] ; y row 2 + 0 |
270 |
punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 2 + 8 |
punpckhbw mm5, mm0 ; [yuyv|yuyv] ; y row 2 + 8 |
271 |
punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 3 + 0 |
punpcklbw mm6, mm1 ; [yuyv|yuyv] ; y row 3 + 0 |
272 |
punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 3 + 8 |
punpckhbw mm7, mm1 ; [yuyv|yuyv] ; y row 3 + 8 |
273 |
movq [edi+edx], mm4 |
movq [x_ptr+x_stride], mm4 |
274 |
movq [edi+edx+8], mm5 |
movq [x_ptr+x_stride+8], mm5 |
275 |
movq [edi+2*edx], mm6 |
movq [x_ptr+2*x_stride], mm6 |
276 |
movq [edi+2*edx+8], mm7 |
movq [x_ptr+2*x_stride+8], mm7 |
277 |
pop edi |
pop x_ptr |
278 |
pop esi |
pop y_ptr |
279 |
%else ; UYVY |
%else ; UYVY |
280 |
movq mm2, [esi] ; [yyyy|yyyy] ; y row 0 |
movq mm2, [y_ptr] ; [yyyy|yyyy] ; y row 0 |
281 |
movq mm3, [esi+eax] ; [yyyy|yyyy] ; y row 1 |
movq mm3, [y_ptr+y_stride] ; [yyyy|yyyy] ; y row 1 |
282 |
movq mm4, mm0 |
movq mm4, mm0 |
283 |
movq mm5, mm0 |
movq mm5, mm0 |
284 |
movq mm6, mm1 |
movq mm6, mm1 |
287 |
punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 0 + 8 |
punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 0 + 8 |
288 |
punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 1 + 0 |
punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 1 + 0 |
289 |
punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 1 + 8 |
punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 1 + 8 |
290 |
movq [edi], mm4 |
movq [x_ptr], mm4 |
291 |
movq [edi+8], mm5 |
movq [x_ptr+8], mm5 |
292 |
movq [edi+edx], mm6 |
movq [x_ptr+x_stride], mm6 |
293 |
movq [edi+edx+8], mm7 |
movq [x_ptr+x_stride+8], mm7 |
294 |
|
|
295 |
push esi |
push y_ptr |
296 |
push edi |
push x_ptr |
297 |
add esi, eax |
add y_ptr, y_stride |
298 |
add edi, edx |
add x_ptr, x_stride |
299 |
movq mm2, [esi+eax] ; [yyyy|yyyy] ; y row 2 |
movq mm2, [y_ptr+y_stride] ; [yyyy|yyyy] ; y row 2 |
300 |
movq mm3, [esi+2*eax] ; [yyyy|yyyy] ; y row 3 |
movq mm3, [y_ptr+2*y_stride] ; [yyyy|yyyy] ; y row 3 |
301 |
movq mm4, mm0 |
movq mm4, mm0 |
302 |
movq mm5, mm0 |
movq mm5, mm0 |
303 |
movq mm6, mm1 |
movq mm6, mm1 |
306 |
punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 2 + 8 |
punpckhbw mm5, mm2 ; [uyvy|uyvy] ; y row 2 + 8 |
307 |
punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 3 + 0 |
punpcklbw mm6, mm3 ; [uyvy|uyvy] ; y row 3 + 0 |
308 |
punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 3 + 8 |
punpckhbw mm7, mm3 ; [uyvy|uyvy] ; y row 3 + 8 |
309 |
movq [edi+edx], mm4 |
movq [x_ptr+x_stride], mm4 |
310 |
movq [edi+edx+8], mm5 |
movq [x_ptr+x_stride+8], mm5 |
311 |
movq [edi+2*edx], mm6 |
movq [x_ptr+2*x_stride], mm6 |
312 |
movq [edi+2*edx+8], mm7 |
movq [x_ptr+2*x_stride+8], mm7 |
313 |
pop edi |
pop x_ptr |
314 |
pop esi |
pop y_ptr |
315 |
%endif |
%endif |
316 |
%endmacro |
%endmacro |
317 |
|
|
319 |
; Code |
; Code |
320 |
;============================================================================= |
;============================================================================= |
321 |
|
|
322 |
SECTION .text |
TEXT |
323 |
|
|
324 |
%include "colorspace_mmx.inc" |
%include "colorspace_mmx.inc" |
325 |
|
|
335 |
|
|
336 |
; output |
; output |
337 |
|
|
338 |
MAKE_COLORSPACE yv12_to_yuyv_mmx,0, 2,8,2, YV12_TO_YUYV, 0, -1 |
MAKE_COLORSPACE yv12_to_yuyv_mmx,0, 2,16,2, YV12_TO_YUYV, 0, -1 |
339 |
MAKE_COLORSPACE yv12_to_uyvy_mmx,0, 2,8,2, YV12_TO_YUYV, 1, -1 |
MAKE_COLORSPACE yv12_to_uyvy_mmx,0, 2,16,2, YV12_TO_YUYV, 1, -1 |
340 |
|
|
341 |
MAKE_COLORSPACE yv12_to_yuyvi_mmx,0, 2,8,4, YV12_TO_YUYVI, 0, -1 |
MAKE_COLORSPACE yv12_to_yuyvi_mmx,0, 2,8,4, YV12_TO_YUYVI, 0, -1 |
342 |
MAKE_COLORSPACE yv12_to_uyvyi_mmx,0, 2,8,4, YV12_TO_YUYVI, 1, -1 |
MAKE_COLORSPACE yv12_to_uyvyi_mmx,0, 2,8,4, YV12_TO_YUYVI, 1, -1 |
343 |
|
|
344 |
|
%ifidn __OUTPUT_FORMAT__,elf |
345 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
346 |
|
%endif |
347 |
|
|