2 * Copyright (C) 2012 Ronald S. Bultje
4 * This file is part of Libav.
6 * Libav is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * Libav is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with Libav; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "libavutil/attributes.h"
23 #include "libavutil/common.h"
24 #include "libavutil/cpu.h"
25 #include "libavutil/mem.h"
26 #include "libavutil/x86/asm.h"
27 #include "libavutil/x86/cpu.h"
28 #include "libavcodec/videodsp.h"
31 typedef void emu_edge_vfix_func(uint8_t *dst
, const uint8_t *src
,
32 x86_reg dst_stride
, x86_reg src_stride
,
33 x86_reg start_y
, x86_reg end_y
, x86_reg bh
);
34 typedef void emu_edge_vvar_func(uint8_t *dst
, const uint8_t *src
,
35 x86_reg dst_stride
, x86_reg src_stride
,
36 x86_reg start_y
, x86_reg end_y
, x86_reg bh
,
39 extern emu_edge_vfix_func ff_emu_edge_vfix1_mmx
;
40 extern emu_edge_vfix_func ff_emu_edge_vfix2_mmx
;
41 extern emu_edge_vfix_func ff_emu_edge_vfix3_mmx
;
42 extern emu_edge_vfix_func ff_emu_edge_vfix4_mmx
;
43 extern emu_edge_vfix_func ff_emu_edge_vfix5_mmx
;
44 extern emu_edge_vfix_func ff_emu_edge_vfix6_mmx
;
45 extern emu_edge_vfix_func ff_emu_edge_vfix7_mmx
;
46 extern emu_edge_vfix_func ff_emu_edge_vfix8_mmx
;
47 extern emu_edge_vfix_func ff_emu_edge_vfix9_mmx
;
48 extern emu_edge_vfix_func ff_emu_edge_vfix10_mmx
;
49 extern emu_edge_vfix_func ff_emu_edge_vfix11_mmx
;
50 extern emu_edge_vfix_func ff_emu_edge_vfix12_mmx
;
51 extern emu_edge_vfix_func ff_emu_edge_vfix13_mmx
;
52 extern emu_edge_vfix_func ff_emu_edge_vfix14_mmx
;
53 extern emu_edge_vfix_func ff_emu_edge_vfix15_mmx
;
54 extern emu_edge_vfix_func ff_emu_edge_vfix16_mmx
;
55 extern emu_edge_vfix_func ff_emu_edge_vfix17_mmx
;
56 extern emu_edge_vfix_func ff_emu_edge_vfix18_mmx
;
57 extern emu_edge_vfix_func ff_emu_edge_vfix19_mmx
;
58 extern emu_edge_vfix_func ff_emu_edge_vfix20_mmx
;
59 extern emu_edge_vfix_func ff_emu_edge_vfix21_mmx
;
60 extern emu_edge_vfix_func ff_emu_edge_vfix22_mmx
;
62 static emu_edge_vfix_func
* const vfixtbl_mmx
[22] = {
63 &ff_emu_edge_vfix1_mmx
, &ff_emu_edge_vfix2_mmx
, &ff_emu_edge_vfix3_mmx
,
64 &ff_emu_edge_vfix4_mmx
, &ff_emu_edge_vfix5_mmx
, &ff_emu_edge_vfix6_mmx
,
65 &ff_emu_edge_vfix7_mmx
, &ff_emu_edge_vfix8_mmx
, &ff_emu_edge_vfix9_mmx
,
66 &ff_emu_edge_vfix10_mmx
, &ff_emu_edge_vfix11_mmx
, &ff_emu_edge_vfix12_mmx
,
67 &ff_emu_edge_vfix13_mmx
, &ff_emu_edge_vfix14_mmx
, &ff_emu_edge_vfix15_mmx
,
68 &ff_emu_edge_vfix16_mmx
, &ff_emu_edge_vfix17_mmx
, &ff_emu_edge_vfix18_mmx
,
69 &ff_emu_edge_vfix19_mmx
, &ff_emu_edge_vfix20_mmx
, &ff_emu_edge_vfix21_mmx
,
70 &ff_emu_edge_vfix22_mmx
73 extern emu_edge_vvar_func ff_emu_edge_vvar_mmx
;
74 extern emu_edge_vfix_func ff_emu_edge_vfix16_sse
;
75 extern emu_edge_vfix_func ff_emu_edge_vfix17_sse
;
76 extern emu_edge_vfix_func ff_emu_edge_vfix18_sse
;
77 extern emu_edge_vfix_func ff_emu_edge_vfix19_sse
;
78 extern emu_edge_vfix_func ff_emu_edge_vfix20_sse
;
79 extern emu_edge_vfix_func ff_emu_edge_vfix21_sse
;
80 extern emu_edge_vfix_func ff_emu_edge_vfix22_sse
;
81 static emu_edge_vfix_func
* const vfixtbl_sse
[22] = {
82 ff_emu_edge_vfix1_mmx
, ff_emu_edge_vfix2_mmx
, ff_emu_edge_vfix3_mmx
,
83 ff_emu_edge_vfix4_mmx
, ff_emu_edge_vfix5_mmx
, ff_emu_edge_vfix6_mmx
,
84 ff_emu_edge_vfix7_mmx
, ff_emu_edge_vfix8_mmx
, ff_emu_edge_vfix9_mmx
,
85 ff_emu_edge_vfix10_mmx
, ff_emu_edge_vfix11_mmx
, ff_emu_edge_vfix12_mmx
,
86 ff_emu_edge_vfix13_mmx
, ff_emu_edge_vfix14_mmx
, ff_emu_edge_vfix15_mmx
,
87 ff_emu_edge_vfix16_sse
, ff_emu_edge_vfix17_sse
, ff_emu_edge_vfix18_sse
,
88 ff_emu_edge_vfix19_sse
, ff_emu_edge_vfix20_sse
, ff_emu_edge_vfix21_sse
,
89 ff_emu_edge_vfix22_sse
91 extern emu_edge_vvar_func ff_emu_edge_vvar_sse
;
93 typedef void emu_edge_hfix_func(uint8_t *dst
, x86_reg dst_stride
,
94 x86_reg start_x
, x86_reg bh
);
95 typedef void emu_edge_hvar_func(uint8_t *dst
, x86_reg dst_stride
,
96 x86_reg start_x
, x86_reg n_words
, x86_reg bh
);
98 extern emu_edge_hfix_func ff_emu_edge_hfix2_mmx
;
99 extern emu_edge_hfix_func ff_emu_edge_hfix4_mmx
;
100 extern emu_edge_hfix_func ff_emu_edge_hfix6_mmx
;
101 extern emu_edge_hfix_func ff_emu_edge_hfix8_mmx
;
102 extern emu_edge_hfix_func ff_emu_edge_hfix10_mmx
;
103 extern emu_edge_hfix_func ff_emu_edge_hfix12_mmx
;
104 extern emu_edge_hfix_func ff_emu_edge_hfix14_mmx
;
105 extern emu_edge_hfix_func ff_emu_edge_hfix16_mmx
;
106 extern emu_edge_hfix_func ff_emu_edge_hfix18_mmx
;
107 extern emu_edge_hfix_func ff_emu_edge_hfix20_mmx
;
108 extern emu_edge_hfix_func ff_emu_edge_hfix22_mmx
;
110 static emu_edge_hfix_func
* const hfixtbl_mmx
[11] = {
111 ff_emu_edge_hfix2_mmx
, ff_emu_edge_hfix4_mmx
, ff_emu_edge_hfix6_mmx
,
112 ff_emu_edge_hfix8_mmx
, ff_emu_edge_hfix10_mmx
, ff_emu_edge_hfix12_mmx
,
113 ff_emu_edge_hfix14_mmx
, ff_emu_edge_hfix16_mmx
, ff_emu_edge_hfix18_mmx
,
114 ff_emu_edge_hfix20_mmx
, ff_emu_edge_hfix22_mmx
117 extern emu_edge_hvar_func ff_emu_edge_hvar_mmx
;
118 extern emu_edge_hfix_func ff_emu_edge_hfix16_sse2
;
119 extern emu_edge_hfix_func ff_emu_edge_hfix18_sse2
;
120 extern emu_edge_hfix_func ff_emu_edge_hfix20_sse2
;
121 extern emu_edge_hfix_func ff_emu_edge_hfix22_sse2
;
122 static emu_edge_hfix_func
* const hfixtbl_sse2
[11] = {
123 ff_emu_edge_hfix2_mmx
, ff_emu_edge_hfix4_mmx
, ff_emu_edge_hfix6_mmx
,
124 ff_emu_edge_hfix8_mmx
, ff_emu_edge_hfix10_mmx
, ff_emu_edge_hfix12_mmx
,
125 ff_emu_edge_hfix14_mmx
, ff_emu_edge_hfix16_sse2
, ff_emu_edge_hfix18_sse2
,
126 ff_emu_edge_hfix20_sse2
, ff_emu_edge_hfix22_sse2
128 extern emu_edge_hvar_func ff_emu_edge_hvar_sse2
;
130 static av_always_inline
void emulated_edge_mc(uint8_t *dst
, const uint8_t *src
,
131 ptrdiff_t dst_stride
,
132 ptrdiff_t src_stride
,
133 x86_reg block_w
, x86_reg block_h
,
134 x86_reg src_x
, x86_reg src_y
,
135 x86_reg w
, x86_reg h
,
136 emu_edge_vfix_func
* const *vfix_tbl
,
137 emu_edge_vvar_func
*v_extend_var
,
138 emu_edge_hfix_func
* const *hfix_tbl
,
139 emu_edge_hvar_func
*h_extend_var
)
141 x86_reg start_y
, start_x
, end_y
, end_x
, src_y_add
= 0, p
;
147 src
-= src_y
* src_stride
;
148 src_y
= src_y_add
= h
- 1;
149 } else if (src_y
<= -block_h
) {
150 src
-= src_y
*src_stride
;
151 src_y
= src_y_add
= 1 - block_h
;
154 src
+= w
- 1 - src_x
;
156 } else if (src_x
<= -block_w
) {
157 src
+= 1 - block_w
- src_x
;
161 start_y
= FFMAX(0, -src_y
);
162 start_x
= FFMAX(0, -src_x
);
163 end_y
= FFMIN(block_h
, h
-src_y
);
164 end_x
= FFMIN(block_w
, w
-src_x
);
165 assert(start_x
< end_x
&& block_w
> 0);
166 assert(start_y
< end_y
&& block_h
> 0);
168 // fill in the to-be-copied part plus all above/below
169 src
+= (src_y_add
+ start_y
) * src_stride
+ start_x
;
172 vfix_tbl
[w
- 1](dst
+ start_x
, src
,
173 dst_stride
, src_stride
,
174 start_y
, end_y
, block_h
);
176 v_extend_var(dst
+ start_x
, src
, dst_stride
, src_stride
,
177 start_y
, end_y
, block_h
, w
);
183 hfix_tbl
[(start_x
- 1) >> 1](dst
, dst_stride
, start_x
, block_h
);
185 h_extend_var(dst
, dst_stride
,
186 start_x
, (start_x
+ 1) >> 1, block_h
);
194 hfix_tbl
[(p
- 1) >> 1](dst
+ end_x
- (p
& 1), dst_stride
,
197 h_extend_var(dst
+ end_x
- (p
& 1), dst_stride
,
198 -!(p
& 1), (p
+ 1) >> 1, block_h
);
204 static av_noinline
void emulated_edge_mc_mmx(uint8_t *buf
, const uint8_t *src
,
205 ptrdiff_t buf_stride
,
206 ptrdiff_t src_stride
,
207 int block_w
, int block_h
,
208 int src_x
, int src_y
, int w
, int h
)
210 emulated_edge_mc(buf
, src
, buf_stride
, src_stride
, block_w
, block_h
,
211 src_x
, src_y
, w
, h
, vfixtbl_mmx
, &ff_emu_edge_vvar_mmx
,
212 hfixtbl_mmx
, &ff_emu_edge_hvar_mmx
);
215 static av_noinline
void emulated_edge_mc_sse(uint8_t * buf
,const uint8_t *src
,
216 ptrdiff_t buf_stride
,
217 ptrdiff_t src_stride
,
218 int block_w
, int block_h
,
219 int src_x
, int src_y
, int w
, int h
)
221 emulated_edge_mc(buf
, src
, buf_stride
, src_stride
, block_w
, block_h
,
222 src_x
, src_y
, w
, h
, vfixtbl_sse
, &ff_emu_edge_vvar_sse
,
223 hfixtbl_mmx
, &ff_emu_edge_hvar_mmx
);
227 static av_noinline
void emulated_edge_mc_sse2(uint8_t *buf
, const uint8_t *src
,
228 ptrdiff_t buf_stride
,
229 ptrdiff_t src_stride
,
230 int block_w
, int block_h
,
231 int src_x
, int src_y
, int w
,
234 emulated_edge_mc(buf
, src
, buf_stride
, src_stride
, block_w
, block_h
, src_x
,
235 src_y
, w
, h
, vfixtbl_sse
, &ff_emu_edge_vvar_sse
,
236 hfixtbl_sse2
, &ff_emu_edge_hvar_sse2
);
238 #endif /* HAVE_X86ASM */
240 void ff_prefetch_mmxext(uint8_t *buf
, ptrdiff_t stride
, int h
);
241 void ff_prefetch_3dnow(uint8_t *buf
, ptrdiff_t stride
, int h
);
243 av_cold
void ff_videodsp_init_x86(VideoDSPContext
*ctx
, int bpc
)
246 int cpu_flags
= av_get_cpu_flags();
249 if (EXTERNAL_MMX(cpu_flags
) && bpc
<= 8) {
250 ctx
->emulated_edge_mc
= emulated_edge_mc_mmx
;
252 if (EXTERNAL_AMD3DNOW(cpu_flags
)) {
253 ctx
->prefetch
= ff_prefetch_3dnow
;
255 #endif /* ARCH_X86_32 */
256 if (EXTERNAL_MMXEXT(cpu_flags
)) {
257 ctx
->prefetch
= ff_prefetch_mmxext
;
260 if (EXTERNAL_SSE(cpu_flags
) && bpc
<= 8) {
261 ctx
->emulated_edge_mc
= emulated_edge_mc_sse
;
263 #endif /* ARCH_X86_32 */
264 if (EXTERNAL_SSE2(cpu_flags
) && bpc
<= 8) {
265 ctx
->emulated_edge_mc
= emulated_edge_mc_sse2
;
267 #endif /* HAVE_X86ASM */