libavresample: NEON optimized FIR audio resampling

modelled after aarch64 code

on Cortex-A8, s16 and s32 code is about 2x faster,
float code about 7x faster

Signed-off-by: Peter Meerwald <pmeerw@pmeerw.net>
Signed-off-by: Martin Storsjö <martin@martin.st>
This commit is contained in:
Peter Meerwald 2015-02-19 23:28:26 +01:00 committed by Martin Storsjö
parent cbdd1806ea
commit 12655c4804
6 changed files with 470 additions and 2 deletions

View File

@ -1,5 +1,7 @@
OBJS += arm/audio_convert_init.o
OBJS += arm/audio_convert_init.o \
arm/resample_init.o
OBJS-$(CONFIG_NEON_CLOBBER_TEST) += arm/neontest.o
NEON-OBJS += arm/audio_convert_neon.o
NEON-OBJS += arm/audio_convert_neon.o \
arm/resample_neon.o

View File

@ -0,0 +1,29 @@
/*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVRESAMPLE_ARM_ASM_OFFSETS_H
#define AVRESAMPLE_ARM_ASM_OFFSETS_H
/* struct ResampleContext */
#define FILTER_BANK 0x08
#define FILTER_LENGTH 0x0c
#define SRC_INCR 0x20
#define PHASE_SHIFT 0x28
#define PHASE_MASK (PHASE_SHIFT + 0x04)
#endif /* AVRESAMPLE_ARM_ASM_OFFSETS_H */

View File

@ -0,0 +1,74 @@
/*
* Copyright (c) 2014 Peter Meerwald <pmeerw@pmeerw.net>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/cpu.h"
#include "libavutil/arm/cpu.h"
#include "libavutil/internal.h"
#include "libavutil/samplefmt.h"
#include "libavresample/resample.h"
#include "asm-offsets.h"
AV_CHECK_OFFSET(struct ResampleContext, filter_bank, FILTER_BANK);
AV_CHECK_OFFSET(struct ResampleContext, filter_length, FILTER_LENGTH);
AV_CHECK_OFFSET(struct ResampleContext, src_incr, SRC_INCR);
AV_CHECK_OFFSET(struct ResampleContext, phase_shift, PHASE_SHIFT);
AV_CHECK_OFFSET(struct ResampleContext, phase_mask, PHASE_MASK);
void ff_resample_one_flt_neon(struct ResampleContext *c, void *dst0,
int dst_index, const void *src0,
unsigned int index, int frac);
void ff_resample_one_s16_neon(struct ResampleContext *c, void *dst0,
int dst_index, const void *src0,
unsigned int index, int frac);
void ff_resample_one_s32_neon(struct ResampleContext *c, void *dst0,
int dst_index, const void *src0,
unsigned int index, int frac);
void ff_resample_linear_flt_neon(struct ResampleContext *c, void *dst0,
int dst_index, const void *src0,
unsigned int index, int frac);
av_cold void ff_audio_resample_init_arm(ResampleContext *c,
enum AVSampleFormat sample_fmt)
{
int cpu_flags = av_get_cpu_flags();
if (have_neon(cpu_flags)) {
switch (sample_fmt) {
case AV_SAMPLE_FMT_FLTP:
if (c->linear)
c->resample_one = ff_resample_linear_flt_neon;
else
c->resample_one = ff_resample_one_flt_neon;
break;
case AV_SAMPLE_FMT_S16P:
if (!c->linear)
c->resample_one = ff_resample_one_s16_neon;
break;
case AV_SAMPLE_FMT_S32P:
if (!c->linear)
c->resample_one = ff_resample_one_s32_neon;
break;
}
}
}

View File

@ -0,0 +1,358 @@
/*
* Copyright (c) 2014 Peter Meerwald <pmeerw@pmeerw.net>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* Libav is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with Libav; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/arm/asm.S"
#include "asm-offsets.h"
.macro resample_one fmt, es=2
function ff_resample_one_\fmt\()_neon, export=1
push {r4, r5}
add r1, r1, r2, lsl #\es
ldr r2, [r0, #PHASE_SHIFT+4] /* phase_mask */
ldr ip, [sp, #8] /* index */
ldr r5, [r0, #FILTER_LENGTH]
and r2, ip, r2 /* (index & phase_mask) */
ldr r4, [r0, #PHASE_SHIFT]
lsr r4, ip, r4 /* compute sample_index */
mul r2, r2, r5
ldr ip, [r0, #FILTER_BANK]
add r3, r3, r4, lsl #\es /* &src[sample_index] */
cmp r5, #8
add r0, ip, r2, lsl #\es /* filter = &filter_bank[...] */
blt 5f
8:
subs r5, r5, #8
LOAD4
MUL4
7:
LOAD4
beq 6f
cmp r5, #8
MLA4
blt 4f
subs r5, r5, #8
LOAD4
MLA4
b 7b
6:
MLA4
STORE
pop {r4, r5}
bx lr
5:
INIT4
4: /* remaining filter_length 1 to 7 */
cmp r5, #4
blt 2f
subs r5, r5, #4
LOAD4
MLA4
beq 0f
2: /* remaining filter_length 1 to 3 */
cmp r5, #2
blt 1f
subs r5, r5, #2
LOAD2
MLA2
beq 0f
1: /* remaining filter_length 1 */
LOAD1
MLA1
0:
STORE
pop {r4, r5}
bx lr
endfunc
.purgem LOAD1
.purgem LOAD2
.purgem LOAD4
.purgem MLA1
.purgem MLA2
.purgem MLA4
.purgem MUL4
.purgem INIT4
.purgem STORE
.endm
/* float32 */
.macro LOAD1
veor.32 d0, d0
vld1.32 {d0[0]}, [r0]! /* load filter */
vld1.32 {d4[0]}, [r3]! /* load src */
.endm
.macro LOAD2
vld1.32 {d0}, [r0]! /* load filter */
vld1.32 {d4}, [r3]! /* load src */
.endm
.macro LOAD4
vld1.32 {d0,d1}, [r0]! /* load filter */
vld1.32 {d4,d5}, [r3]! /* load src */
.endm
.macro MLA1
vmla.f32 d16, d0, d4[0]
.endm
.macro MLA2
vmla.f32 d16, d0, d4
.endm
.macro MLA4
vmla.f32 d16, d0, d4
vmla.f32 d17, d1, d5
.endm
.macro MUL4
vmul.f32 d16, d0, d4
vmul.f32 d17, d1, d5
.endm
.macro INIT4
veor.f32 q8, q8
.endm
.macro STORE
vpadd.f32 d16, d16, d17
vpadd.f32 d16, d16, d16
vst1.32 d16[0], [r1]
.endm
resample_one flt, 2
/* s32 */
.macro LOAD1
veor.32 d0, d0
vld1.32 {d0[0]}, [r0]! /* load filter */
vld1.32 {d4[0]}, [r3]! /* load src */
.endm
.macro LOAD2
vld1.32 {d0}, [r0]! /* load filter */
vld1.32 {d4}, [r3]! /* load src */
.endm
.macro LOAD4
vld1.32 {d0,d1}, [r0]! /* load filter */
vld1.32 {d4,d5}, [r3]! /* load src */
.endm
.macro MLA1
vmlal.s32 q8, d0, d4[0]
.endm
.macro MLA2
vmlal.s32 q8, d0, d4
.endm
.macro MLA4
vmlal.s32 q8, d0, d4
vmlal.s32 q9, d1, d5
.endm
.macro MUL4
vmull.s32 q8, d0, d4
vmull.s32 q9, d1, d5
.endm
.macro INIT4
veor.s64 q8, q8
veor.s64 q9, q9
.endm
.macro STORE
vadd.s64 q8, q8, q9
vadd.s64 d16, d16, d17
vqrshrn.s64 d16, q8, #30
vst1.32 d16[0], [r1]
.endm
resample_one s32, 2
/* s16 */
.macro LOAD1
veor.16 d0, d0
vld1.16 {d0[0]}, [r0]! /* load filter */
vld1.16 {d4[0]}, [r3]! /* load src */
.endm
.macro LOAD2
veor.16 d0, d0
vld1.32 {d0[0]}, [r0]! /* load filter */
veor.16 d4, d4
vld1.32 {d4[0]}, [r3]! /* load src */
.endm
.macro LOAD4
vld1.16 {d0}, [r0]! /* load filter */
vld1.16 {d4}, [r3]! /* load src */
.endm
.macro MLA1
vmlal.s16 q8, d0, d4[0]
.endm
.macro MLA2
vmlal.s16 q8, d0, d4
.endm
.macro MLA4
vmlal.s16 q8, d0, d4
.endm
.macro MUL4
vmull.s16 q8, d0, d4
.endm
.macro INIT4
veor.s32 q8, q8
.endm
.macro STORE
vpadd.s32 d16, d16, d17
vpadd.s32 d16, d16, d16
vqrshrn.s32 d16, q8, #15
vst1.16 d16[0], [r1]
.endm
resample_one s16, 1
.macro resample_linear fmt, es=2
function ff_resample_linear_\fmt\()_neon, export=1
push {r4, r5}
add r1, r1, r2, lsl #\es
ldr r2, [r0, #PHASE_SHIFT+4] /* phase_mask */
ldr ip, [sp, #8] /* index */
ldr r5, [r0, #FILTER_LENGTH]
and r2, ip, r2 /* (index & phase_mask) */
ldr r4, [r0, #PHASE_SHIFT]
lsr r4, ip, r4 /* compute sample_index */
mul r2, r2, r5
ldr ip, [r0, #FILTER_BANK]
add r3, r3, r4, lsl #\es /* &src[sample_index] */
cmp r5, #8
ldr r4, [r0, #SRC_INCR]
add r0, ip, r2, lsl #\es /* filter = &filter_bank[...] */
add r2, r0, r5, lsl #\es /* filter[... + c->filter_length] */
blt 5f
8:
subs r5, r5, #8
LOAD4
MUL4
7:
LOAD4
beq 6f
cmp r5, #8
MLA4
blt 4f
subs r5, r5, #8
LOAD4
MLA4
b 7b
6:
MLA4
STORE
pop {r4, r5}
bx lr
5:
INIT4
4: /* remaining filter_length 1 to 7 */
cmp r5, #4
blt 2f
subs r5, r5, #4
LOAD4
MLA4
beq 0f
2: /* remaining filter_length 1 to 3 */
cmp r5, #2
blt 1f
subs r5, r5, #2
LOAD2
MLA2
beq 0f
1: /* remaining filter_length 1 */
LOAD1
MLA1
0:
STORE
pop {r4, r5}
bx lr
endfunc
.purgem LOAD1
.purgem LOAD2
.purgem LOAD4
.purgem MLA1
.purgem MLA2
.purgem MLA4
.purgem MUL4
.purgem INIT4
.purgem STORE
.endm
/* float32 linear */
.macro LOAD1
veor.32 d0, d0
veor.32 d2, d2
vld1.32 {d0[0]}, [r0]! /* load filter */
vld1.32 {d2[0]}, [r2]! /* load filter */
vld1.32 {d4[0]}, [r3]! /* load src */
.endm
.macro LOAD2
vld1.32 {d0}, [r0]! /* load filter */
vld1.32 {d2}, [r2]! /* load filter */
vld1.32 {d4}, [r3]! /* load src */
.endm
.macro LOAD4
vld1.32 {d0,d1}, [r0]! /* load filter */
vld1.32 {d2,d3}, [r2]! /* load filter */
vld1.32 {d4,d5}, [r3]! /* load src */
.endm
.macro MLA1
vmla.f32 d18, d0, d4[0]
vmla.f32 d16, d2, d4[0]
.endm
.macro MLA2
vmla.f32 d18, d0, d4
vmla.f32 d16, d2, d4
.endm
.macro MLA4
vmla.f32 q9, q0, q2
vmla.f32 q8, q1, q2
.endm
.macro MUL4
vmul.f32 q9, q0, q2
vmul.f32 q8, q1, q2
.endm
.macro INIT4
veor.f32 q9, q9
veor.f32 q8, q8
.endm
.macro STORE
vldr s0, [sp, #12] /* frac */
vmov s1, r4
vcvt.f32.s32 d0, d0
vsub.f32 q8, q8, q9 /* v2 - val */
vpadd.f32 d18, d18, d19
vpadd.f32 d16, d16, d17
vpadd.f32 d2, d18, d18
vpadd.f32 d1, d16, d16
vmul.f32 s2, s2, s0 /* (v2 - val) * frac */
vdiv.f32 s2, s2, s1 /* / c->src_incr */
vadd.f32 s4, s4, s2
vstr s4, [r1]
.endm
resample_linear flt, 2

View File

@ -110,4 +110,7 @@ struct AVAudioResampleContext {
void ff_audio_resample_init_aarch64(ResampleContext *c,
enum AVSampleFormat sample_fmt);
void ff_audio_resample_init_arm(ResampleContext *c,
enum AVSampleFormat sample_fmt);
#endif /* AVRESAMPLE_INTERNAL_H */

View File

@ -172,6 +172,8 @@ ResampleContext *ff_audio_resample_init(AVAudioResampleContext *avr)
if (ARCH_AARCH64)
ff_audio_resample_init_aarch64(c, avr->internal_sample_fmt);
if (ARCH_ARM)
ff_audio_resample_init_arm(c, avr->internal_sample_fmt);
felem_size = av_get_bytes_per_sample(avr->internal_sample_fmt);
c->filter_bank = av_mallocz(c->filter_length * (phase_count + 1) * felem_size);