yading@11: /*
yading@11:  * Copyright (c) 2012 Justin Ruggles <justin.ruggles@gmail.com>
yading@11:  *
yading@11:  * This file is part of Libav.
yading@11:  *
yading@11:  * Libav is free software; you can redistribute it and/or
yading@11:  * modify it under the terms of the GNU Lesser General Public
yading@11:  * License as published by the Free Software Foundation; either
yading@11:  * version 2.1 of the License, or (at your option) any later version.
yading@11:  *
yading@11:  * Libav is distributed in the hope that it will be useful,
yading@11:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@11:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
yading@11:  * Lesser General Public License for more details.
yading@11:  *
yading@11:  * You should have received a copy of the GNU Lesser General Public
yading@11:  * License along with Libav; if not, write to the Free Software
yading@11:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@11:  */
yading@11: 
yading@11: #include "config.h"
yading@11: #include "libavutil/cpu.h"
yading@11: #include "libavutil/x86/cpu.h"
yading@11: #include "libavresample/audio_mix.h"
yading@11: 
yading@11: void ff_mix_2_to_1_fltp_flt_sse(float **src, float **matrix, int len,
yading@11:                                 int out_ch, int in_ch);
yading@11: void ff_mix_2_to_1_fltp_flt_avx(float **src, float **matrix, int len,
yading@11:                                 int out_ch, int in_ch);
yading@11: 
yading@11: void ff_mix_2_to_1_s16p_flt_sse2(int16_t **src, float **matrix, int len,
yading@11:                                  int out_ch, int in_ch);
yading@11: void ff_mix_2_to_1_s16p_flt_sse4(int16_t **src, float **matrix, int len,
yading@11:                                  int out_ch, int in_ch);
yading@11: 
yading@11: void ff_mix_2_to_1_s16p_q8_sse2(int16_t **src, int16_t **matrix,
yading@11:                                 int len, int out_ch, int in_ch);
yading@11: 
yading@11: void ff_mix_1_to_2_fltp_flt_sse(float **src, float **matrix, int len,
yading@11:                                 int out_ch, int in_ch);
yading@11: void ff_mix_1_to_2_fltp_flt_avx(float **src, float **matrix, int len,
yading@11:                                 int out_ch, int in_ch);
yading@11: 
yading@11: void ff_mix_1_to_2_s16p_flt_sse2(int16_t **src, float **matrix, int len,
yading@11:                                  int out_ch, int in_ch);
yading@11: void ff_mix_1_to_2_s16p_flt_sse4(int16_t **src, float **matrix, int len,
yading@11:                                  int out_ch, int in_ch);
yading@11: void ff_mix_1_to_2_s16p_flt_avx (int16_t **src, float **matrix, int len,
yading@11:                                  int out_ch, int in_ch);
yading@11: 
yading@11: #define DEFINE_MIX_3_8_TO_1_2(chan)                                     \
yading@11: void ff_mix_ ## chan ## _to_1_fltp_flt_sse(float **src,                 \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11: void ff_mix_ ## chan ## _to_2_fltp_flt_sse(float **src,                 \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_s16p_flt_sse2(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11: void ff_mix_ ## chan ## _to_2_s16p_flt_sse2(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_s16p_flt_sse4(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11: void ff_mix_ ## chan ## _to_2_s16p_flt_sse4(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_fltp_flt_avx(float **src,                 \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11: void ff_mix_ ## chan ## _to_2_fltp_flt_avx(float **src,                 \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_s16p_flt_avx(int16_t **src,               \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11: void ff_mix_ ## chan ## _to_2_s16p_flt_avx(int16_t **src,               \
yading@11:                                            float **matrix, int len,     \
yading@11:                                            int out_ch, int in_ch);      \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_fltp_flt_fma4(float **src,                \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11: void ff_mix_ ## chan ## _to_2_fltp_flt_fma4(float **src,                \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11:                                                                         \
yading@11: void ff_mix_ ## chan ## _to_1_s16p_flt_fma4(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);     \
yading@11: void ff_mix_ ## chan ## _to_2_s16p_flt_fma4(int16_t **src,              \
yading@11:                                             float **matrix, int len,    \
yading@11:                                             int out_ch, int in_ch);
yading@11: 
yading@11: DEFINE_MIX_3_8_TO_1_2(3)
yading@11: DEFINE_MIX_3_8_TO_1_2(4)
yading@11: DEFINE_MIX_3_8_TO_1_2(5)
yading@11: DEFINE_MIX_3_8_TO_1_2(6)
yading@11: DEFINE_MIX_3_8_TO_1_2(7)
yading@11: DEFINE_MIX_3_8_TO_1_2(8)
yading@11: 
yading@11: #define SET_MIX_3_8_TO_1_2(chan)                                            \
yading@11:     if (EXTERNAL_SSE(mm_flags)) {                                           \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, 16, 4, "SSE",                        \
yading@11:                               ff_mix_ ## chan ## _to_1_fltp_flt_sse);       \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, 16, 4, "SSE",                        \
yading@11:                               ff_mix_## chan ##_to_2_fltp_flt_sse);         \
yading@11:     }                                                                       \
yading@11:     if (EXTERNAL_SSE2(mm_flags)) {                                          \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, 16, 8, "SSE2",                       \
yading@11:                               ff_mix_ ## chan ## _to_1_s16p_flt_sse2);      \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, 16, 8, "SSE2",                       \
yading@11:                               ff_mix_ ## chan ## _to_2_s16p_flt_sse2);      \
yading@11:     }                                                                       \
yading@11:     if (EXTERNAL_SSE4(mm_flags)) {                                          \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, 16, 8, "SSE4",                       \
yading@11:                               ff_mix_ ## chan ## _to_1_s16p_flt_sse4);      \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, 16, 8, "SSE4",                       \
yading@11:                               ff_mix_ ## chan ## _to_2_s16p_flt_sse4);      \
yading@11:     }                                                                       \
yading@11:     if (EXTERNAL_AVX(mm_flags)) {                                           \
yading@11:         int ptr_align = 32;                                                 \
yading@11:         int smp_align = 8;                                                  \
yading@11:         if (ARCH_X86_32 || chan >= 6) {                                     \
yading@11:             ptr_align = 16;                                                 \
yading@11:             smp_align = 4;                                                  \
yading@11:         }                                                                   \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, ptr_align, smp_align, "AVX",         \
yading@11:                               ff_mix_ ## chan ## _to_1_fltp_flt_avx);       \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, ptr_align, smp_align, "AVX",         \
yading@11:                               ff_mix_ ## chan ## _to_2_fltp_flt_avx);       \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, 16, 8, "AVX",                        \
yading@11:                               ff_mix_ ## chan ## _to_1_s16p_flt_avx);       \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, 16, 8, "AVX",                        \
yading@11:                               ff_mix_ ## chan ## _to_2_s16p_flt_avx);       \
yading@11:     }                                                                       \
yading@11:     if (EXTERNAL_FMA4(mm_flags)) {                                          \
yading@11:         int ptr_align = 32;                                                 \
yading@11:         int smp_align = 8;                                                  \
yading@11:         if (ARCH_X86_32 || chan >= 6) {                                     \
yading@11:             ptr_align = 16;                                                 \
yading@11:             smp_align = 4;                                                  \
yading@11:         }                                                                   \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, ptr_align, smp_align, "FMA4",        \
yading@11:                               ff_mix_ ## chan ## _to_1_fltp_flt_fma4);      \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, ptr_align, smp_align, "FMA4",        \
yading@11:                               ff_mix_ ## chan ## _to_2_fltp_flt_fma4);      \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 1, 16, 8, "FMA4",                       \
yading@11:                               ff_mix_ ## chan ## _to_1_s16p_flt_fma4);      \
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,\
yading@11:                               chan, 2, 16, 8, "FMA4",                       \
yading@11:                               ff_mix_ ## chan ## _to_2_s16p_flt_fma4);      \
yading@11:     }
yading@11: 
yading@11: av_cold void ff_audio_mix_init_x86(AudioMix *am)
yading@11: {
yading@11: #if HAVE_YASM
yading@11:     int mm_flags = av_get_cpu_flags();
yading@11: 
yading@11:     if (EXTERNAL_SSE(mm_flags)) {
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               2, 1, 16, 8, "SSE", ff_mix_2_to_1_fltp_flt_sse);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               1, 2, 16, 4, "SSE", ff_mix_1_to_2_fltp_flt_sse);
yading@11:     }
yading@11:     if (EXTERNAL_SSE2(mm_flags)) {
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_flt_sse2);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_Q8,
yading@11:                               2, 1, 16, 8, "SSE2", ff_mix_2_to_1_s16p_q8_sse2);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               1, 2, 16, 8, "SSE2", ff_mix_1_to_2_s16p_flt_sse2);
yading@11:     }
yading@11:     if (EXTERNAL_SSE4(mm_flags)) {
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               2, 1, 16, 8, "SSE4", ff_mix_2_to_1_s16p_flt_sse4);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               1, 2, 16, 8, "SSE4", ff_mix_1_to_2_s16p_flt_sse4);
yading@11:     }
yading@11:     if (EXTERNAL_AVX(mm_flags)) {
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               2, 1, 32, 16, "AVX", ff_mix_2_to_1_fltp_flt_avx);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               1, 2, 32, 8, "AVX", ff_mix_1_to_2_fltp_flt_avx);
yading@11:         ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
yading@11:                               1, 2, 16, 8, "AVX", ff_mix_1_to_2_s16p_flt_avx);
yading@11:     }
yading@11: 
yading@11:     SET_MIX_3_8_TO_1_2(3)
yading@11:     SET_MIX_3_8_TO_1_2(4)
yading@11:     SET_MIX_3_8_TO_1_2(5)
yading@11:     SET_MIX_3_8_TO_1_2(6)
yading@11:     SET_MIX_3_8_TO_1_2(7)
yading@11:     SET_MIX_3_8_TO_1_2(8)
yading@11: #endif /* HAVE_YASM */
yading@11: }