]> git.sesse.net Git - ffmpeg/blob - libavfilter/x86/vf_blend_init.c
avfilter/x86/vf_360: add most of >8 depth asm
[ffmpeg] / libavfilter / x86 / vf_blend_init.c
1 /*
2  * Copyright (c) 2015 Paul B Mahol
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg 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.
10  *
11  * FFmpeg 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.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "libavutil/attributes.h"
22 #include "libavutil/cpu.h"
23 #include "libavutil/x86/cpu.h"
24 #include "libavfilter/blend.h"
25
26 #define BLEND_FUNC(name, opt) \
27 void ff_blend_##name##_##opt(const uint8_t *top, ptrdiff_t top_linesize,       \
28                              const uint8_t *bottom, ptrdiff_t bottom_linesize, \
29                              uint8_t *dst, ptrdiff_t dst_linesize,             \
30                              ptrdiff_t width, ptrdiff_t height,                \
31                              struct FilterParams *param, double *values, int starty);
32
33 BLEND_FUNC(addition, sse2)
34 BLEND_FUNC(addition, avx2)
35 BLEND_FUNC(grainmerge, sse2)
36 BLEND_FUNC(grainmerge, avx2)
37 BLEND_FUNC(average, sse2)
38 BLEND_FUNC(average, avx2)
39 BLEND_FUNC(and, sse2)
40 BLEND_FUNC(and, avx2)
41 BLEND_FUNC(darken, sse2)
42 BLEND_FUNC(darken, avx2)
43 BLEND_FUNC(grainextract, sse2)
44 BLEND_FUNC(grainextract, avx2)
45 BLEND_FUNC(multiply, sse2)
46 BLEND_FUNC(multiply, avx2)
47 BLEND_FUNC(screen, sse2)
48 BLEND_FUNC(screen, avx2)
49 BLEND_FUNC(hardmix, sse2)
50 BLEND_FUNC(hardmix, avx2)
51 BLEND_FUNC(divide, sse2)
52 BLEND_FUNC(lighten, sse2)
53 BLEND_FUNC(lighten, avx2)
54 BLEND_FUNC(or, sse2)
55 BLEND_FUNC(or, avx2)
56 BLEND_FUNC(phoenix, sse2)
57 BLEND_FUNC(phoenix, avx2)
58 BLEND_FUNC(subtract, sse2)
59 BLEND_FUNC(subtract, avx2)
60 BLEND_FUNC(xor, sse2)
61 BLEND_FUNC(xor, avx2)
62 BLEND_FUNC(difference, sse2)
63 BLEND_FUNC(difference, ssse3)
64 BLEND_FUNC(difference, avx2)
65 BLEND_FUNC(extremity, sse2)
66 BLEND_FUNC(extremity, ssse3)
67 BLEND_FUNC(extremity, avx2)
68 BLEND_FUNC(negation, sse2)
69 BLEND_FUNC(negation, ssse3)
70 BLEND_FUNC(negation, avx2)
71
72 #if ARCH_X86_64
73 BLEND_FUNC(addition_16, sse2)
74 BLEND_FUNC(addition_16, avx2)
75 BLEND_FUNC(grainmerge_16, sse4)
76 BLEND_FUNC(grainmerge_16, avx2)
77 BLEND_FUNC(average_16, sse2)
78 BLEND_FUNC(average_16, avx2)
79 BLEND_FUNC(and_16, sse2)
80 BLEND_FUNC(and_16, avx2)
81 BLEND_FUNC(darken_16, sse4)
82 BLEND_FUNC(darken_16, avx2)
83 BLEND_FUNC(grainextract_16, sse4)
84 BLEND_FUNC(grainextract_16, avx2)
85 BLEND_FUNC(difference_16, sse4)
86 BLEND_FUNC(difference_16, avx2)
87 BLEND_FUNC(extremity_16, sse4)
88 BLEND_FUNC(extremity_16, avx2)
89 BLEND_FUNC(negation_16, sse4)
90 BLEND_FUNC(negation_16, avx2)
91 BLEND_FUNC(lighten_16, sse4)
92 BLEND_FUNC(lighten_16, avx2)
93 BLEND_FUNC(or_16, sse2)
94 BLEND_FUNC(or_16, avx2)
95 BLEND_FUNC(phoenix_16, sse4)
96 BLEND_FUNC(phoenix_16, avx2)
97 BLEND_FUNC(subtract_16, sse2)
98 BLEND_FUNC(subtract_16, avx2)
99 BLEND_FUNC(xor_16, sse2)
100 BLEND_FUNC(xor_16, avx2)
101 #endif /* ARCH_X86_64 */
102
103 av_cold void ff_blend_init_x86(FilterParams *param, int depth)
104 {
105     int cpu_flags = av_get_cpu_flags();
106
107     if (depth == 8) {
108         if (EXTERNAL_SSE2(cpu_flags) && param->opacity == 1) {
109             switch (param->mode) {
110             case BLEND_ADDITION:     param->blend = ff_blend_addition_sse2;     break;
111             case BLEND_GRAINMERGE:   param->blend = ff_blend_grainmerge_sse2;   break;
112             case BLEND_AND:          param->blend = ff_blend_and_sse2;          break;
113             case BLEND_AVERAGE:      param->blend = ff_blend_average_sse2;      break;
114             case BLEND_DARKEN:       param->blend = ff_blend_darken_sse2;       break;
115             case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_sse2; break;
116             case BLEND_DIVIDE:       param->blend = ff_blend_divide_sse2;       break;
117             case BLEND_HARDMIX:      param->blend = ff_blend_hardmix_sse2;      break;
118             case BLEND_LIGHTEN:      param->blend = ff_blend_lighten_sse2;      break;
119             case BLEND_MULTIPLY:     param->blend = ff_blend_multiply_sse2;     break;
120             case BLEND_OR:           param->blend = ff_blend_or_sse2;           break;
121             case BLEND_PHOENIX:      param->blend = ff_blend_phoenix_sse2;      break;
122             case BLEND_SCREEN:       param->blend = ff_blend_screen_sse2;       break;
123             case BLEND_SUBTRACT:     param->blend = ff_blend_subtract_sse2;     break;
124             case BLEND_XOR:          param->blend = ff_blend_xor_sse2;          break;
125             case BLEND_DIFFERENCE:   param->blend = ff_blend_difference_sse2;   break;
126             case BLEND_EXTREMITY:    param->blend = ff_blend_extremity_sse2;    break;
127             case BLEND_NEGATION:     param->blend = ff_blend_negation_sse2;     break;
128             }
129         }
130         if (EXTERNAL_SSSE3(cpu_flags) && param->opacity == 1) {
131             switch (param->mode) {
132             case BLEND_DIFFERENCE: param->blend = ff_blend_difference_ssse3; break;
133             case BLEND_EXTREMITY:  param->blend = ff_blend_extremity_ssse3;  break;
134             case BLEND_NEGATION:   param->blend = ff_blend_negation_ssse3;   break;
135             }
136         }
137
138         if (EXTERNAL_AVX2_FAST(cpu_flags) && param->opacity == 1) {
139             switch (param->mode) {
140             case BLEND_ADDITION:     param->blend = ff_blend_addition_avx2;     break;
141             case BLEND_GRAINMERGE:   param->blend = ff_blend_grainmerge_avx2;   break;
142             case BLEND_AND:          param->blend = ff_blend_and_avx2;          break;
143             case BLEND_AVERAGE:      param->blend = ff_blend_average_avx2;      break;
144             case BLEND_DARKEN:       param->blend = ff_blend_darken_avx2;       break;
145             case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_avx2; break;
146             case BLEND_HARDMIX:      param->blend = ff_blend_hardmix_avx2;      break;
147             case BLEND_LIGHTEN:      param->blend = ff_blend_lighten_avx2;      break;
148             case BLEND_MULTIPLY:     param->blend = ff_blend_multiply_avx2;     break;
149             case BLEND_OR:           param->blend = ff_blend_or_avx2;           break;
150             case BLEND_PHOENIX:      param->blend = ff_blend_phoenix_avx2;      break;
151             case BLEND_SCREEN:       param->blend = ff_blend_screen_avx2;       break;
152             case BLEND_SUBTRACT:     param->blend = ff_blend_subtract_avx2;     break;
153             case BLEND_XOR:          param->blend = ff_blend_xor_avx2;          break;
154             case BLEND_DIFFERENCE:   param->blend = ff_blend_difference_avx2;   break;
155             case BLEND_EXTREMITY:    param->blend = ff_blend_extremity_avx2;    break;
156             case BLEND_NEGATION:     param->blend = ff_blend_negation_avx2;     break;
157             }
158         }
159     } else if (depth == 16) {
160 #if ARCH_X86_64
161         if (EXTERNAL_SSE2(cpu_flags) && param->opacity == 1) {
162             switch (param->mode) {
163             case BLEND_ADDITION: param->blend = ff_blend_addition_16_sse2; break;
164             case BLEND_AND:      param->blend = ff_blend_and_16_sse2;      break;
165             case BLEND_AVERAGE:  param->blend = ff_blend_average_16_sse2;  break;
166             case BLEND_OR:       param->blend = ff_blend_or_16_sse2;       break;
167             case BLEND_SUBTRACT: param->blend = ff_blend_subtract_16_sse2; break;
168             case BLEND_XOR:      param->blend = ff_blend_xor_16_sse2;      break;
169             }
170         }
171         if (EXTERNAL_SSE4(cpu_flags) && param->opacity == 1) {
172             switch (param->mode) {
173             case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_16_sse4; break;
174             case BLEND_DARKEN:   param->blend = ff_blend_darken_16_sse4;     break;
175             case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_16_sse4; break;
176             case BLEND_DIFFERENCE: param->blend = ff_blend_difference_16_sse4; break;
177             case BLEND_EXTREMITY:  param->blend = ff_blend_extremity_16_sse4;    break;
178             case BLEND_NEGATION:  param->blend = ff_blend_negation_16_sse4;     break;
179             case BLEND_LIGHTEN:  param->blend = ff_blend_lighten_16_sse4;    break;
180             case BLEND_PHOENIX:  param->blend = ff_blend_phoenix_16_sse4;    break;
181             }
182         }
183         if (EXTERNAL_AVX2_FAST(cpu_flags) && param->opacity == 1) {
184             switch (param->mode) {
185             case BLEND_ADDITION: param->blend = ff_blend_addition_16_avx2; break;
186             case BLEND_GRAINMERGE: param->blend = ff_blend_grainmerge_16_avx2;   break;
187             case BLEND_AND:      param->blend = ff_blend_and_16_avx2;      break;
188             case BLEND_AVERAGE:  param->blend = ff_blend_average_16_avx2;  break;
189             case BLEND_DARKEN:   param->blend = ff_blend_darken_16_avx2;   break;
190             case BLEND_GRAINEXTRACT: param->blend = ff_blend_grainextract_16_avx2; break;
191             case BLEND_DIFFERENCE: param->blend = ff_blend_difference_16_avx2; break;
192             case BLEND_EXTREMITY:  param->blend = ff_blend_extremity_16_avx2;    break;
193             case BLEND_NEGATION:  param->blend = ff_blend_negation_16_avx2;     break;
194             case BLEND_LIGHTEN:  param->blend = ff_blend_lighten_16_avx2;  break;
195             case BLEND_OR:       param->blend = ff_blend_or_16_avx2;       break;
196             case BLEND_PHOENIX:  param->blend = ff_blend_phoenix_16_avx2;  break;
197             case BLEND_SUBTRACT: param->blend = ff_blend_subtract_16_avx2; break;
198             case BLEND_XOR:      param->blend = ff_blend_xor_16_avx2;      break;
199             }
200         }
201 #endif /* ARCH_X86_64 */
202     }
203 }