3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * - MMX and XMM YV12->YV12 conversion - |
; * - MMX and XMM YV12->YV12 conversion - |
5 |
; * |
; * |
6 |
; * Copyright(C) 2001 Michael Militzer <isibaar@xvid.org> |
; * Copyright(C) 2001-2008 Michael Militzer <michael@xvid.org> |
7 |
; * |
; * |
8 |
; * This program is free software; you can redistribute it and/or modify it |
; * This program is free software; you can redistribute it and/or modify it |
9 |
; * under the terms of the GNU General Public License as published by |
; * under the terms of the GNU General Public License as published by |
23 |
; * |
; * |
24 |
; ***************************************************************************/ |
; ***************************************************************************/ |
25 |
|
|
26 |
BITS 32 |
%include "nasm.inc" |
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
27 |
|
|
28 |
;============================================================================= |
;============================================================================= |
29 |
; Helper macros |
; Helper macros |
30 |
;============================================================================= |
;============================================================================= |
31 |
|
|
32 |
|
%macro _MOVQ 3 |
33 |
|
%if %1 == 1 |
34 |
|
movntq %2, %3 ; xmm |
35 |
|
%else |
36 |
|
movq %2, %3 ; plain mmx |
37 |
|
%endif |
38 |
|
%endmacro |
39 |
|
|
40 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
41 |
; PLANE_COPY ( DST, DST_DIF, SRC, SRC_DIF, WIDTH, HEIGHT, OPT ) |
; PLANE_COPY ( DST, DST_STRIDE, SRC, SRC_STRIDE, WIDTH, HEIGHT, OPT ) |
42 |
; DST dst buffer |
; DST dst buffer |
43 |
; DST_DIF dst stride difference (e.g. stride - width) |
; DST_STRIDE dst stride |
44 |
; SRC src destination buffer |
; SRC src destination buffer |
45 |
; SRC_DIF src stride difference (e.g. stride - width) |
; SRC_STRIDE src stride |
46 |
; WIDTH width |
; WIDTH width |
47 |
; HEIGHT height |
; HEIGHT height |
48 |
; OPT 0=plain mmx, 1=xmm |
; OPT 0=plain mmx, 1=xmm |
49 |
|
; |
50 |
|
; |
51 |
|
; Trashes: DST, SRC, WIDTH, HEIGHT, _EBX, _ECX, _EDX |
52 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
53 |
|
|
54 |
%macro PLANE_COPY 7 |
%macro PLANE_COPY 7 |
55 |
%define DST %1 |
%define DST %1 |
56 |
%define DST_DIF %2 |
%define DST_STRIDE %2 |
57 |
%define SRC %3 |
%define SRC %3 |
58 |
%define SRC_DIF %4 |
%define SRC_STRIDE %4 |
59 |
%define WIDTH %5 |
%define WIDTH %5 |
60 |
%define HEIGHT %6 |
%define HEIGHT %6 |
61 |
%define OPT %7 |
%define OPT %7 |
62 |
|
|
63 |
mov eax, WIDTH |
mov _EBX, WIDTH |
64 |
mov ebp, HEIGHT ; $ebp$ = height |
shr WIDTH, 6 ; $_EAX$ = width / 64 |
65 |
mov esi, SRC |
and _EBX, 63 ; remainder = width % 64 |
66 |
mov edi, DST |
mov _EDX, _EBX |
67 |
|
shr _EBX, 4 ; $_EBX$ = remainder / 16 |
68 |
mov ebx, eax |
and _EDX, 15 ; $_EDX$ = remainder % 16 |
69 |
shr eax, 6 ; $eax$ = width / 64 |
|
70 |
and ebx, 63 ; remainder = width % 64 |
%%loop64_start_pc: |
71 |
mov edx, ebx |
push DST |
72 |
shr ebx, 4 ; $ebx$ = remainder / 16 |
push SRC |
73 |
and edx, 15 ; $edx$ = remainder % 16 |
|
74 |
|
mov _ECX, WIDTH ; width64 |
75 |
%%loop64_start |
test WIDTH, WIDTH |
76 |
or eax, eax |
jz %%loop16_start_pc |
77 |
jz %%loop16_start |
|
78 |
mov ecx, eax ; width64 |
%%loop64_pc: |
|
%%loop64: |
|
79 |
%if OPT == 1 ; xmm |
%if OPT == 1 ; xmm |
80 |
prefetchnta [esi + 64] ; non temporal prefetch |
prefetchnta [SRC + 64] ; non temporal prefetch |
81 |
prefetchnta [esi + 96] |
prefetchnta [SRC + 96] |
|
%endif |
|
|
movq mm1, [esi] ; read from src |
|
|
movq mm2, [esi + 8] |
|
|
movq mm3, [esi + 16] |
|
|
movq mm4, [esi + 24] |
|
|
movq mm5, [esi + 32] |
|
|
movq mm6, [esi + 40] |
|
|
movq mm7, [esi + 48] |
|
|
movq mm0, [esi + 56] |
|
|
|
|
|
%if OPT == 0 ; plain mmx |
|
|
movq [edi], mm1 ; write to y_out |
|
|
movq [edi + 8], mm2 |
|
|
movq [edi + 16], mm3 |
|
|
movq [edi + 24], mm4 |
|
|
movq [edi + 32], mm5 |
|
|
movq [edi + 40], mm6 |
|
|
movq [edi + 48], mm7 |
|
|
movq [edi + 56], mm0 |
|
|
%else |
|
|
movntq [edi], mm1 ; write to y_out |
|
|
movntq [edi + 8], mm2 |
|
|
movntq [edi + 16], mm3 |
|
|
movntq [edi + 24], mm4 |
|
|
movntq [edi + 32], mm5 |
|
|
movntq [edi + 40], mm6 |
|
|
movntq [edi + 48], mm7 |
|
|
movntq [edi + 56], mm0 |
|
82 |
%endif |
%endif |
83 |
|
movq mm1, [SRC ] ; read from src |
84 |
|
movq mm2, [SRC + 8] |
85 |
|
movq mm3, [SRC + 16] |
86 |
|
movq mm4, [SRC + 24] |
87 |
|
movq mm5, [SRC + 32] |
88 |
|
movq mm6, [SRC + 40] |
89 |
|
movq mm7, [SRC + 48] |
90 |
|
movq mm0, [SRC + 56] |
91 |
|
|
92 |
|
_MOVQ OPT, [DST ], mm1 ; write to y_out |
93 |
|
_MOVQ OPT, [DST + 8], mm2 |
94 |
|
_MOVQ OPT, [DST + 16], mm3 |
95 |
|
_MOVQ OPT, [DST + 24], mm4 |
96 |
|
_MOVQ OPT, [DST + 32], mm5 |
97 |
|
_MOVQ OPT, [DST + 40], mm6 |
98 |
|
_MOVQ OPT, [DST + 48], mm7 |
99 |
|
_MOVQ OPT, [DST + 56], mm0 |
100 |
|
|
101 |
|
add SRC, 64 |
102 |
|
add DST, 64 |
103 |
|
loop %%loop64_pc |
104 |
|
|
105 |
|
%%loop16_start_pc: |
106 |
|
mov _ECX, _EBX ; width16 |
107 |
|
test _EBX, _EBX |
108 |
|
jz %%loop1_start_pc |
109 |
|
|
110 |
|
%%loop16_pc: |
111 |
|
movq mm1, [SRC] |
112 |
|
movq mm2, [SRC + 8] |
113 |
|
|
114 |
|
_MOVQ OPT, [DST], mm1 |
115 |
|
_MOVQ OPT, [DST + 8], mm2 |
116 |
|
|
117 |
|
add SRC, 16 |
118 |
|
add DST, 16 |
119 |
|
loop %%loop16_pc |
120 |
|
|
121 |
|
%%loop1_start_pc: |
122 |
|
mov _ECX, _EDX |
123 |
|
rep movsb |
124 |
|
|
125 |
|
pop SRC |
126 |
|
pop DST |
127 |
|
|
128 |
add esi, 64 |
%ifdef ARCH_IS_X86_64 |
129 |
add edi, 64 |
XVID_MOVSXD _ECX, SRC_STRIDE |
130 |
dec ecx |
add SRC, _ECX |
131 |
jnz %%loop64 |
mov ecx, DST_STRIDE |
132 |
|
add DST, _ECX |
|
|
|
|
%%loop16_start |
|
|
or ebx, ebx |
|
|
jz %%loop1_start |
|
|
mov ecx, ebx ; width16 |
|
|
%%loop16: |
|
|
movq mm1, [esi] |
|
|
movq mm2, [esi + 8] |
|
|
%if OPT == 0 ; plain mmx |
|
|
movq [edi], mm1 |
|
|
movq [edi + 8], mm2 |
|
133 |
%else |
%else |
134 |
movntq [edi], mm1 |
add SRC, SRC_STRIDE |
135 |
movntq [edi + 8], mm2 |
add DST, DST_STRIDE |
136 |
%endif |
%endif |
137 |
|
|
138 |
add esi, 16 |
dec HEIGHT |
139 |
add edi, 16 |
jg near %%loop64_start_pc |
|
dec ecx |
|
|
jnz %%loop16 |
|
140 |
|
|
141 |
|
%undef DST |
142 |
|
%undef DST_STRIDE |
143 |
|
%undef SRC |
144 |
|
%undef SRC_STRIDE |
145 |
|
%undef WIDTH |
146 |
|
%undef HEIGHT |
147 |
|
%undef OPT |
148 |
|
%endmacro |
149 |
|
|
150 |
%%loop1_start |
;------------------------------------------------------------------------------ |
151 |
mov ecx, edx |
; PLANE_FILL ( DST, DST_STRIDE, WIDTH, HEIGHT, OPT ) |
152 |
rep movsb |
; DST dst buffer |
153 |
|
; DST_STRIDE dst stride |
154 |
|
; WIDTH width |
155 |
|
; HEIGHT height |
156 |
|
; OPT 0=plain mmx, 1=xmm |
157 |
|
; |
158 |
|
; Trashes: DST, WIDTH, HEIGHT, _EBX, _ECX, _EDX, _EAX |
159 |
|
;------------------------------------------------------------------------------ |
160 |
|
|
161 |
|
%macro PLANE_FILL 5 |
162 |
|
%define DST %1 |
163 |
|
%define DST_STRIDE %2 |
164 |
|
%define WIDTH %3 |
165 |
|
%define HEIGHT %4 |
166 |
|
%define OPT %5 |
167 |
|
|
168 |
|
mov _EAX, 0x80808080 |
169 |
|
mov _EBX, WIDTH |
170 |
|
shr WIDTH, 6 ; $_ESI$ = width / 64 |
171 |
|
and _EBX, 63 ; _EBX = remainder = width % 64 |
172 |
|
movd mm0, eax |
173 |
|
mov _EDX, _EBX |
174 |
|
shr _EBX, 4 ; $_EBX$ = remainder / 16 |
175 |
|
and _EDX, 15 ; $_EDX$ = remainder % 16 |
176 |
|
punpckldq mm0, mm0 |
177 |
|
|
178 |
|
%%loop64_start_pf: |
179 |
|
push DST |
180 |
|
mov _ECX, WIDTH ; width64 |
181 |
|
test WIDTH, WIDTH |
182 |
|
jz %%loop16_start_pf |
183 |
|
|
184 |
|
%%loop64_pf: |
185 |
|
|
186 |
|
_MOVQ OPT, [DST ], mm0 ; write to y_out |
187 |
|
_MOVQ OPT, [DST + 8], mm0 |
188 |
|
_MOVQ OPT, [DST + 16], mm0 |
189 |
|
_MOVQ OPT, [DST + 24], mm0 |
190 |
|
_MOVQ OPT, [DST + 32], mm0 |
191 |
|
_MOVQ OPT, [DST + 40], mm0 |
192 |
|
_MOVQ OPT, [DST + 48], mm0 |
193 |
|
_MOVQ OPT, [DST + 56], mm0 |
194 |
|
|
195 |
|
add DST, 64 |
196 |
|
loop %%loop64_pf |
197 |
|
|
198 |
|
%%loop16_start_pf: |
199 |
|
mov _ECX, _EBX ; width16 |
200 |
|
test _EBX, _EBX |
201 |
|
jz %%loop1_start_pf |
202 |
|
|
203 |
|
%%loop16_pf: |
204 |
|
_MOVQ OPT, [DST ], mm0 |
205 |
|
_MOVQ OPT, [DST + 8], mm0 |
206 |
|
|
207 |
|
add DST, 16 |
208 |
|
loop %%loop16_pf |
209 |
|
|
210 |
|
%%loop1_start_pf: |
211 |
|
mov _ECX, _EDX |
212 |
|
rep stosb |
213 |
|
|
214 |
|
pop DST |
215 |
|
|
216 |
|
%ifdef ARCH_IS_X86_64 |
217 |
|
mov ecx, DST_STRIDE |
218 |
|
add DST, _ECX |
219 |
|
%else |
220 |
|
add DST, DST_STRIDE |
221 |
|
%endif |
222 |
|
|
223 |
|
dec HEIGHT |
224 |
|
jg near %%loop64_start_pf |
225 |
|
|
226 |
add esi, SRC_DIF |
%undef DST |
227 |
add edi, DST_DIF |
%undef DST_STRIDE |
228 |
dec ebp |
%undef WIDTH |
229 |
jnz near %%loop64_start |
%undef HEIGHT |
230 |
|
%undef OPT |
231 |
%endmacro |
%endmacro |
232 |
|
|
233 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
243 |
;------------------------------------------------------------------------------ |
;------------------------------------------------------------------------------ |
244 |
%macro MAKE_YV12_TO_YV12 2 |
%macro MAKE_YV12_TO_YV12 2 |
245 |
%define NAME %1 |
%define NAME %1 |
246 |
%define OPT %2 |
%define XMM_OPT %2 |
247 |
ALIGN 16 |
ALIGN SECTION_ALIGN |
248 |
cglobal NAME |
cglobal NAME |
249 |
NAME: |
NAME: |
|
%define pushsize 16 |
|
|
%define localsize 24 |
|
250 |
|
|
251 |
%define vflip esp + localsize + pushsize + 52 |
push _EBX ; _ESP + localsize + 3*PTR_SIZE |
252 |
%define height esp + localsize + pushsize + 48 |
|
253 |
%define width esp + localsize + pushsize + 44 |
%define localsize 2*4 |
254 |
%define uv_src_stride esp + localsize + pushsize + 40 |
|
255 |
%define y_src_stride esp + localsize + pushsize + 36 |
%ifdef ARCH_IS_X86_64 |
|
%define v_src esp + localsize + pushsize + 32 |
|
|
%define u_src esp + localsize + pushsize + 28 |
|
|
%define y_src esp + localsize + pushsize + 24 |
|
|
%define uv_dst_stride esp + localsize + pushsize + 20 |
|
|
%define y_dst_stride esp + localsize + pushsize + 16 |
|
|
%define v_dst esp + localsize + pushsize + 12 |
|
|
%define u_dst esp + localsize + pushsize + 8 |
|
|
%define y_dst esp + localsize + pushsize + 4 |
|
|
%define _ip esp + localsize + pushsize + 0 |
|
|
|
|
|
push ebx ; esp + localsize + 16 |
|
|
push esi ; esp + localsize + 8 |
|
|
push edi ; esp + localsize + 4 |
|
|
push ebp ; esp + localsize + 0 |
|
|
|
|
|
%define width2 esp + localsize - 4 |
|
|
%define height2 esp + localsize - 8 |
|
|
%define y_src_dif esp + localsize - 12 |
|
|
%define y_dst_dif esp + localsize - 16 |
|
|
%define uv_src_dif esp + localsize - 20 |
|
|
%define uv_dst_dif esp + localsize - 24 |
|
256 |
|
|
257 |
sub esp, localsize |
%ifndef WINDOWS |
258 |
|
%define pushsize 2*PTR_SIZE |
259 |
|
%define shadow 0 |
260 |
|
%else |
261 |
|
%define pushsize 4*PTR_SIZE |
262 |
|
%define shadow 32 + 2*PTR_SIZE |
263 |
|
%endif |
264 |
|
|
265 |
|
%define prm_vflip dword [_ESP + localsize + pushsize + shadow + 7*PTR_SIZE] |
266 |
|
%define prm_height dword [_ESP + localsize + pushsize + shadow + 6*PTR_SIZE] |
267 |
|
%define prm_width dword [_ESP + localsize + pushsize + shadow + 5*PTR_SIZE] |
268 |
|
%define prm_uv_src_stride dword [_ESP + localsize + pushsize + shadow + 4*PTR_SIZE] |
269 |
|
%define prm_y_src_stride dword [_ESP + localsize + pushsize + shadow + 3*PTR_SIZE] |
270 |
|
%define prm_v_src [_ESP + localsize + pushsize + shadow + 2*PTR_SIZE] |
271 |
|
%define prm_u_src [_ESP + localsize + pushsize + shadow + 1*PTR_SIZE] |
272 |
|
|
273 |
|
%ifdef WINDOWS |
274 |
|
push _ESI ; _ESP + localsize + 2*PTR_SIZE |
275 |
|
push _EDI ; _ESP + localsize + 1*PTR_SIZE |
276 |
|
push _EBP ; _ESP + localsize + 0*PTR_SIZE |
277 |
|
|
278 |
|
sub _ESP, localsize |
279 |
|
|
280 |
|
%define prm_y_src _ESI |
281 |
|
%define prm_uv_dst_stride TMP0d |
282 |
|
%define prm_y_dst_stride prm4d |
283 |
|
%define prm_v_dst prm3 |
284 |
|
%define prm_u_dst TMP1 |
285 |
|
%define prm_y_dst _EDI |
286 |
|
|
287 |
|
mov _EDI, prm1 |
288 |
|
mov TMP1, prm2 |
289 |
|
|
290 |
|
mov _ESI, [_ESP + localsize + pushsize + shadow + 0*PTR_SIZE] |
291 |
|
mov TMP0d, dword [_ESP + localsize + pushsize + shadow - 1*PTR_SIZE] |
292 |
|
|
293 |
|
%else |
294 |
|
push _EBP ; _ESP + localsize + 0*PTR_SIZE |
295 |
|
|
296 |
|
sub _ESP, localsize |
297 |
|
|
298 |
|
%define prm_y_src _ESI |
299 |
|
%define prm_uv_dst_stride prm5d |
300 |
|
%define prm_y_dst_stride TMP1d |
301 |
|
%define prm_v_dst prm6 |
302 |
|
%define prm_u_dst TMP0 |
303 |
|
%define prm_y_dst _EDI |
304 |
|
|
305 |
|
mov TMP0, prm2 |
306 |
|
mov _ESI, prm6 |
307 |
|
|
308 |
|
mov prm6, prm3 |
309 |
|
mov TMP1d, prm4d |
310 |
|
%endif |
311 |
|
|
312 |
|
%define _ip _ESP + localsize + pushsize + 0 |
313 |
|
|
314 |
|
%else |
315 |
|
|
316 |
|
%define pushsize 4*PTR_SIZE |
317 |
|
|
318 |
mov eax, [width] |
%define prm_vflip [_ESP + localsize + pushsize + 13*PTR_SIZE] |
319 |
mov ebx, [height] |
%define prm_height [_ESP + localsize + pushsize + 12*PTR_SIZE] |
320 |
|
%define prm_width [_ESP + localsize + pushsize + 11*PTR_SIZE] |
321 |
|
%define prm_uv_src_stride [_ESP + localsize + pushsize + 10*PTR_SIZE] |
322 |
|
%define prm_y_src_stride [_ESP + localsize + pushsize + 9*PTR_SIZE] |
323 |
|
%define prm_v_src [_ESP + localsize + pushsize + 8*PTR_SIZE] |
324 |
|
%define prm_u_src [_ESP + localsize + pushsize + 7*PTR_SIZE] |
325 |
|
|
326 |
|
%define prm_y_src _ESI |
327 |
|
%define prm_uv_dst_stride [_ESP + localsize + pushsize + 5*PTR_SIZE] |
328 |
|
%define prm_y_dst_stride [_ESP + localsize + pushsize + 4*PTR_SIZE] |
329 |
|
%define prm_v_dst [_ESP + localsize + pushsize + 3*PTR_SIZE] |
330 |
|
%define prm_u_dst [_ESP + localsize + pushsize + 2*PTR_SIZE] |
331 |
|
%define prm_y_dst _EDI |
332 |
|
|
333 |
|
%define _ip _ESP + localsize + pushsize + 0 |
334 |
|
|
335 |
|
push _ESI ; _ESP + localsize + 2*PTR_SIZE |
336 |
|
push _EDI ; _ESP + localsize + 1*PTR_SIZE |
337 |
|
push _EBP ; _ESP + localsize + 0*PTR_SIZE |
338 |
|
|
339 |
|
sub _ESP, localsize |
340 |
|
|
341 |
|
mov _ESI, [_ESP + localsize + pushsize + 6*PTR_SIZE] |
342 |
|
mov _EDI, [_ESP + localsize + pushsize + 1*PTR_SIZE] |
343 |
|
|
344 |
|
%endif |
345 |
|
|
346 |
|
%define width2 dword [_ESP + localsize - 1*4] |
347 |
|
%define height2 dword [_ESP + localsize - 2*4] |
348 |
|
|
349 |
|
mov eax, prm_width |
350 |
|
mov ebx, prm_height |
351 |
shr eax, 1 ; calculate widht/2, heigh/2 |
shr eax, 1 ; calculate widht/2, heigh/2 |
352 |
shr ebx, 1 |
shr ebx, 1 |
353 |
mov [width2], eax |
mov width2, eax |
354 |
mov [height2], ebx |
mov height2, ebx |
355 |
|
|
356 |
mov ebp, [vflip] |
mov eax, prm_vflip |
357 |
or ebp, ebp |
test eax, eax |
358 |
jz near .dont_flip |
jz near .go |
359 |
|
|
360 |
; flipping support |
; flipping support |
361 |
mov eax, [height] |
mov eax, prm_height |
362 |
mov esi, [y_src] |
mov ecx, prm_y_src_stride |
363 |
mov edx, [y_src_stride] |
sub eax, 1 |
364 |
push edx |
imul eax, ecx |
365 |
mul edx |
add _ESI, _EAX ; y_src += (height-1) * y_src_stride |
366 |
pop edx |
neg ecx |
367 |
add esi, eax ; y_src += (height-1) * y_src_stride |
mov prm_y_src_stride, ecx ; y_src_stride = -y_src_stride |
368 |
neg edx |
|
369 |
mov [y_src], esi |
mov eax, height2 |
370 |
mov [y_src_stride], edx ; y_src_stride = -y_src_stride |
mov _EDX, prm_u_src |
371 |
|
mov _EBP, prm_v_src |
372 |
mov eax, [height2] |
mov ecx, prm_uv_src_stride |
373 |
mov esi, [u_src] |
test _EDX, _EDX |
374 |
mov edi, [v_src] |
jz .go |
375 |
mov edx, [uv_src_stride] |
test _EBP, _EBP |
376 |
sub eax, 1 ; ebp = height2 - 1 |
jz .go |
377 |
push edx |
sub eax, 1 ; _EAX = height2 - 1 |
378 |
mul edx |
imul eax, ecx |
379 |
pop edx |
add _EDX, _EAX ; u_src += (height2-1) * uv_src_stride |
380 |
add esi, eax ; u_src += (height2-1) * uv_src_stride |
add _EBP, _EAX ; v_src += (height2-1) * uv_src_stride |
381 |
add edi, eax ; v_src += (height2-1) * uv_src_stride |
neg ecx |
382 |
neg edx |
mov prm_u_src, _EDX |
383 |
mov [u_src], esi |
mov prm_v_src, _EBP |
384 |
mov [v_src], edi |
mov prm_uv_src_stride, ecx ; uv_src_stride = -uv_src_stride |
385 |
mov [uv_src_stride], edx ; uv_src_stride = -uv_src_stride |
|
386 |
|
.go: |
387 |
.dont_flip |
mov eax, prm_width |
388 |
|
mov ebp, prm_height |
389 |
mov eax, [y_src_stride] |
PLANE_COPY _EDI, prm_y_dst_stride, _ESI, prm_y_src_stride, _EAX, _EBP, XMM_OPT |
390 |
mov ebx, [y_dst_stride] |
|
391 |
mov ecx, [uv_src_stride] |
mov _EAX, prm_u_src |
392 |
mov edx, [uv_dst_stride] |
or _EAX, prm_v_src |
393 |
sub eax, [width] |
jz near .UVFill_0x80 |
394 |
sub ebx, [width] |
|
395 |
sub ecx, [width2] |
mov eax, width2 |
396 |
sub edx, [width2] |
mov ebp, height2 |
397 |
mov [y_src_dif], eax ; y_src_dif = y_src_stride - width |
mov _ESI, prm_u_src |
398 |
mov [y_dst_dif], ebx ; y_dst_dif = y_dst_stride - width |
mov _EDI, prm_u_dst |
399 |
mov [uv_src_dif], ecx ; uv_src_dif = uv_src_stride - width2 |
PLANE_COPY _EDI, prm_uv_dst_stride, _ESI, prm_uv_src_stride, _EAX, _EBP, XMM_OPT |
400 |
mov [uv_dst_dif], edx ; uv_dst_dif = uv_dst_stride - width2 |
|
401 |
|
mov eax, width2 |
402 |
PLANE_COPY [y_dst], [y_dst_dif], [y_src], [y_src_dif], [width], [height], OPT |
mov ebp, height2 |
403 |
PLANE_COPY [u_dst], [uv_dst_dif], [u_src], [uv_src_dif], [width2], [height2], OPT |
mov _ESI, prm_v_src |
404 |
PLANE_COPY [v_dst], [uv_dst_dif], [v_src], [uv_src_dif], [width2], [height2], OPT |
mov _EDI, prm_v_dst |
405 |
|
PLANE_COPY _EDI, prm_uv_dst_stride, _ESI, prm_uv_src_stride, _EAX, _EBP, XMM_OPT |
406 |
add esp, localsize |
|
407 |
pop ebp |
.Done_UVPlane: |
408 |
pop edi |
add _ESP, localsize |
409 |
pop esi |
|
410 |
pop ebx |
pop _EBP |
411 |
|
%ifndef ARCH_IS_X86_64 |
412 |
|
pop _EDI |
413 |
|
pop _ESI |
414 |
|
%else |
415 |
|
%ifdef WINDOWS |
416 |
|
pop _EDI |
417 |
|
pop _ESI |
418 |
|
%endif |
419 |
|
%endif |
420 |
|
pop _EBX |
421 |
|
|
422 |
ret |
ret |
423 |
|
|
424 |
|
.UVFill_0x80: |
425 |
|
|
426 |
|
mov esi, width2 |
427 |
|
mov ebp, height2 |
428 |
|
mov _EDI, prm_u_dst |
429 |
|
PLANE_FILL _EDI, prm_uv_dst_stride, _ESI, _EBP, XMM_OPT |
430 |
|
|
431 |
|
mov esi, width2 |
432 |
|
mov ebp, height2 |
433 |
|
mov _EDI, prm_v_dst |
434 |
|
PLANE_FILL _EDI, prm_uv_dst_stride, _ESI, _EBP, XMM_OPT |
435 |
|
|
436 |
|
jmp near .Done_UVPlane |
437 |
|
|
438 |
|
ENDFUNC |
439 |
|
|
440 |
|
%undef NAME |
441 |
|
%undef XMM_OPT |
442 |
%endmacro |
%endmacro |
443 |
|
|
444 |
;============================================================================= |
;============================================================================= |
445 |
; Code |
; Code |
446 |
;============================================================================= |
;============================================================================= |
447 |
|
|
448 |
SECTION .text |
TEXT |
449 |
|
|
450 |
MAKE_YV12_TO_YV12 yv12_to_yv12_mmx, 0 |
MAKE_YV12_TO_YV12 yv12_to_yv12_mmx, 0 |
451 |
|
|
452 |
MAKE_YV12_TO_YV12 yv12_to_yv12_xmm, 1 |
MAKE_YV12_TO_YV12 yv12_to_yv12_xmm, 1 |
453 |
|
|
454 |
|
NON_EXEC_STACK |