yading@11
|
1 /*
|
yading@11
|
2 * Copyright (c) 2012 Justin Ruggles <justin.ruggles@gmail.com>
|
yading@11
|
3 *
|
yading@11
|
4 * This file is part of Libav.
|
yading@11
|
5 *
|
yading@11
|
6 * Libav is free software; you can redistribute it and/or
|
yading@11
|
7 * modify it under the terms of the GNU Lesser General Public
|
yading@11
|
8 * License as published by the Free Software Foundation; either
|
yading@11
|
9 * version 2.1 of the License, or (at your option) any later version.
|
yading@11
|
10 *
|
yading@11
|
11 * Libav is distributed in the hope that it will be useful,
|
yading@11
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
yading@11
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
yading@11
|
14 * Lesser General Public License for more details.
|
yading@11
|
15 *
|
yading@11
|
16 * You should have received a copy of the GNU Lesser General Public
|
yading@11
|
17 * License along with Libav; if not, write to the Free Software
|
yading@11
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
yading@11
|
19 */
|
yading@11
|
20
|
yading@11
|
21 #include "config.h"
|
yading@11
|
22 #include "libavutil/cpu.h"
|
yading@11
|
23 #include "libavutil/x86/cpu.h"
|
yading@11
|
24 #include "libavresample/audio_mix.h"
|
yading@11
|
25
|
yading@11
|
26 void ff_mix_2_to_1_fltp_flt_sse(float **src, float **matrix, int len,
|
yading@11
|
27 int out_ch, int in_ch);
|
yading@11
|
28 void ff_mix_2_to_1_fltp_flt_avx(float **src, float **matrix, int len,
|
yading@11
|
29 int out_ch, int in_ch);
|
yading@11
|
30
|
yading@11
|
31 void ff_mix_2_to_1_s16p_flt_sse2(int16_t **src, float **matrix, int len,
|
yading@11
|
32 int out_ch, int in_ch);
|
yading@11
|
33 void ff_mix_2_to_1_s16p_flt_sse4(int16_t **src, float **matrix, int len,
|
yading@11
|
34 int out_ch, int in_ch);
|
yading@11
|
35
|
yading@11
|
36 void ff_mix_2_to_1_s16p_q8_sse2(int16_t **src, int16_t **matrix,
|
yading@11
|
37 int len, int out_ch, int in_ch);
|
yading@11
|
38
|
yading@11
|
39 void ff_mix_1_to_2_fltp_flt_sse(float **src, float **matrix, int len,
|
yading@11
|
40 int out_ch, int in_ch);
|
yading@11
|
41 void ff_mix_1_to_2_fltp_flt_avx(float **src, float **matrix, int len,
|
yading@11
|
42 int out_ch, int in_ch);
|
yading@11
|
43
|
yading@11
|
44 void ff_mix_1_to_2_s16p_flt_sse2(int16_t **src, float **matrix, int len,
|
yading@11
|
45 int out_ch, int in_ch);
|
yading@11
|
46 void ff_mix_1_to_2_s16p_flt_sse4(int16_t **src, float **matrix, int len,
|
yading@11
|
47 int out_ch, int in_ch);
|
yading@11
|
48 void ff_mix_1_to_2_s16p_flt_avx (int16_t **src, float **matrix, int len,
|
yading@11
|
49 int out_ch, int in_ch);
|
yading@11
|
50
|
yading@11
|
51 #define DEFINE_MIX_3_8_TO_1_2(chan) \
|
yading@11
|
52 void ff_mix_ ## chan ## _to_1_fltp_flt_sse(float **src, \
|
yading@11
|
53 float **matrix, int len, \
|
yading@11
|
54 int out_ch, int in_ch); \
|
yading@11
|
55 void ff_mix_ ## chan ## _to_2_fltp_flt_sse(float **src, \
|
yading@11
|
56 float **matrix, int len, \
|
yading@11
|
57 int out_ch, int in_ch); \
|
yading@11
|
58 \
|
yading@11
|
59 void ff_mix_ ## chan ## _to_1_s16p_flt_sse2(int16_t **src, \
|
yading@11
|
60 float **matrix, int len, \
|
yading@11
|
61 int out_ch, int in_ch); \
|
yading@11
|
62 void ff_mix_ ## chan ## _to_2_s16p_flt_sse2(int16_t **src, \
|
yading@11
|
63 float **matrix, int len, \
|
yading@11
|
64 int out_ch, int in_ch); \
|
yading@11
|
65 \
|
yading@11
|
66 void ff_mix_ ## chan ## _to_1_s16p_flt_sse4(int16_t **src, \
|
yading@11
|
67 float **matrix, int len, \
|
yading@11
|
68 int out_ch, int in_ch); \
|
yading@11
|
69 void ff_mix_ ## chan ## _to_2_s16p_flt_sse4(int16_t **src, \
|
yading@11
|
70 float **matrix, int len, \
|
yading@11
|
71 int out_ch, int in_ch); \
|
yading@11
|
72 \
|
yading@11
|
73 void ff_mix_ ## chan ## _to_1_fltp_flt_avx(float **src, \
|
yading@11
|
74 float **matrix, int len, \
|
yading@11
|
75 int out_ch, int in_ch); \
|
yading@11
|
76 void ff_mix_ ## chan ## _to_2_fltp_flt_avx(float **src, \
|
yading@11
|
77 float **matrix, int len, \
|
yading@11
|
78 int out_ch, int in_ch); \
|
yading@11
|
79 \
|
yading@11
|
80 void ff_mix_ ## chan ## _to_1_s16p_flt_avx(int16_t **src, \
|
yading@11
|
81 float **matrix, int len, \
|
yading@11
|
82 int out_ch, int in_ch); \
|
yading@11
|
83 void ff_mix_ ## chan ## _to_2_s16p_flt_avx(int16_t **src, \
|
yading@11
|
84 float **matrix, int len, \
|
yading@11
|
85 int out_ch, int in_ch); \
|
yading@11
|
86 \
|
yading@11
|
87 void ff_mix_ ## chan ## _to_1_fltp_flt_fma4(float **src, \
|
yading@11
|
88 float **matrix, int len, \
|
yading@11
|
89 int out_ch, int in_ch); \
|
yading@11
|
90 void ff_mix_ ## chan ## _to_2_fltp_flt_fma4(float **src, \
|
yading@11
|
91 float **matrix, int len, \
|
yading@11
|
92 int out_ch, int in_ch); \
|
yading@11
|
93 \
|
yading@11
|
94 void ff_mix_ ## chan ## _to_1_s16p_flt_fma4(int16_t **src, \
|
yading@11
|
95 float **matrix, int len, \
|
yading@11
|
96 int out_ch, int in_ch); \
|
yading@11
|
97 void ff_mix_ ## chan ## _to_2_s16p_flt_fma4(int16_t **src, \
|
yading@11
|
98 float **matrix, int len, \
|
yading@11
|
99 int out_ch, int in_ch);
|
yading@11
|
100
|
yading@11
|
101 DEFINE_MIX_3_8_TO_1_2(3)
|
yading@11
|
102 DEFINE_MIX_3_8_TO_1_2(4)
|
yading@11
|
103 DEFINE_MIX_3_8_TO_1_2(5)
|
yading@11
|
104 DEFINE_MIX_3_8_TO_1_2(6)
|
yading@11
|
105 DEFINE_MIX_3_8_TO_1_2(7)
|
yading@11
|
106 DEFINE_MIX_3_8_TO_1_2(8)
|
yading@11
|
107
|
yading@11
|
108 #define SET_MIX_3_8_TO_1_2(chan) \
|
yading@11
|
109 if (EXTERNAL_SSE(mm_flags)) { \
|
yading@11
|
110 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
111 chan, 1, 16, 4, "SSE", \
|
yading@11
|
112 ff_mix_ ## chan ## _to_1_fltp_flt_sse); \
|
yading@11
|
113 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
114 chan, 2, 16, 4, "SSE", \
|
yading@11
|
115 ff_mix_## chan ##_to_2_fltp_flt_sse); \
|
yading@11
|
116 } \
|
yading@11
|
117 if (EXTERNAL_SSE2(mm_flags)) { \
|
yading@11
|
118 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
119 chan, 1, 16, 8, "SSE2", \
|
yading@11
|
120 ff_mix_ ## chan ## _to_1_s16p_flt_sse2); \
|
yading@11
|
121 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
122 chan, 2, 16, 8, "SSE2", \
|
yading@11
|
123 ff_mix_ ## chan ## _to_2_s16p_flt_sse2); \
|
yading@11
|
124 } \
|
yading@11
|
125 if (EXTERNAL_SSE4(mm_flags)) { \
|
yading@11
|
126 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
127 chan, 1, 16, 8, "SSE4", \
|
yading@11
|
128 ff_mix_ ## chan ## _to_1_s16p_flt_sse4); \
|
yading@11
|
129 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
130 chan, 2, 16, 8, "SSE4", \
|
yading@11
|
131 ff_mix_ ## chan ## _to_2_s16p_flt_sse4); \
|
yading@11
|
132 } \
|
yading@11
|
133 if (EXTERNAL_AVX(mm_flags)) { \
|
yading@11
|
134 int ptr_align = 32; \
|
yading@11
|
135 int smp_align = 8; \
|
yading@11
|
136 if (ARCH_X86_32 || chan >= 6) { \
|
yading@11
|
137 ptr_align = 16; \
|
yading@11
|
138 smp_align = 4; \
|
yading@11
|
139 } \
|
yading@11
|
140 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
141 chan, 1, ptr_align, smp_align, "AVX", \
|
yading@11
|
142 ff_mix_ ## chan ## _to_1_fltp_flt_avx); \
|
yading@11
|
143 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
144 chan, 2, ptr_align, smp_align, "AVX", \
|
yading@11
|
145 ff_mix_ ## chan ## _to_2_fltp_flt_avx); \
|
yading@11
|
146 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
147 chan, 1, 16, 8, "AVX", \
|
yading@11
|
148 ff_mix_ ## chan ## _to_1_s16p_flt_avx); \
|
yading@11
|
149 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
150 chan, 2, 16, 8, "AVX", \
|
yading@11
|
151 ff_mix_ ## chan ## _to_2_s16p_flt_avx); \
|
yading@11
|
152 } \
|
yading@11
|
153 if (EXTERNAL_FMA4(mm_flags)) { \
|
yading@11
|
154 int ptr_align = 32; \
|
yading@11
|
155 int smp_align = 8; \
|
yading@11
|
156 if (ARCH_X86_32 || chan >= 6) { \
|
yading@11
|
157 ptr_align = 16; \
|
yading@11
|
158 smp_align = 4; \
|
yading@11
|
159 } \
|
yading@11
|
160 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
161 chan, 1, ptr_align, smp_align, "FMA4", \
|
yading@11
|
162 ff_mix_ ## chan ## _to_1_fltp_flt_fma4); \
|
yading@11
|
163 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
164 chan, 2, ptr_align, smp_align, "FMA4", \
|
yading@11
|
165 ff_mix_ ## chan ## _to_2_fltp_flt_fma4); \
|
yading@11
|
166 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
167 chan, 1, 16, 8, "FMA4", \
|
yading@11
|
168 ff_mix_ ## chan ## _to_1_s16p_flt_fma4); \
|
yading@11
|
169 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
|
yading@11
|
170 chan, 2, 16, 8, "FMA4", \
|
yading@11
|
171 ff_mix_ ## chan ## _to_2_s16p_flt_fma4); \
|
yading@11
|
172 }
|
yading@11
|
173
|
yading@11
|
174 av_cold void ff_audio_mix_init_x86(AudioMix *am)
|
yading@11
|
175 {
|
yading@11
|
176 #if HAVE_YASM
|
yading@11
|
177 int mm_flags = av_get_cpu_flags();
|
yading@11
|
178
|
yading@11
|
179 if (EXTERNAL_SSE(mm_flags)) {
|
yading@11
|
180 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
181 2, 1, 16, 8, "SSE", ff_mix_2_to_1_fltp_flt_sse);
|
yading@11
|
182 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
183 1, 2, 16, 4, "SSE", ff_mix_1_to_2_fltp_flt_sse);
|
yading@11
|
184 }
|
yading@11
|
185 if (EXTERNAL_SSE2(mm_flags)) {
|
yading@11
|
186 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
187 2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_flt_sse2);
|
yading@11
|
188 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_Q8,
|
yading@11
|
189 2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_q8_sse2);
|
yading@11
|
190 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
191 1, 2, 16, 8, "SSE2", ff_mix_1_to_2_s16p_flt_sse2);
|
yading@11
|
192 }
|
yading@11
|
193 if (EXTERNAL_SSE4(mm_flags)) {
|
yading@11
|
194 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
195 2, 1, 16, 8, "SSE4", ff_mix_2_to_1_s16p_flt_sse4);
|
yading@11
|
196 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
197 1, 2, 16, 8, "SSE4", ff_mix_1_to_2_s16p_flt_sse4);
|
yading@11
|
198 }
|
yading@11
|
199 if (EXTERNAL_AVX(mm_flags)) {
|
yading@11
|
200 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
201 2, 1, 32, 16, "AVX", ff_mix_2_to_1_fltp_flt_avx);
|
yading@11
|
202 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
203 1, 2, 32, 8, "AVX", ff_mix_1_to_2_fltp_flt_avx);
|
yading@11
|
204 ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
|
yading@11
|
205 1, 2, 16, 8, "AVX", ff_mix_1_to_2_s16p_flt_avx);
|
yading@11
|
206 }
|
yading@11
|
207
|
yading@11
|
208 SET_MIX_3_8_TO_1_2(3)
|
yading@11
|
209 SET_MIX_3_8_TO_1_2(4)
|
yading@11
|
210 SET_MIX_3_8_TO_1_2(5)
|
yading@11
|
211 SET_MIX_3_8_TO_1_2(6)
|
yading@11
|
212 SET_MIX_3_8_TO_1_2(7)
|
yading@11
|
213 SET_MIX_3_8_TO_1_2(8)
|
yading@11
|
214 #endif /* HAVE_YASM */
|
yading@11
|
215 }
|