|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
synth_neon_s32: ARM NEON optimized synth (32-bit output version)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
copyright 1995-2010 by the mpg123 project - free software under the terms of the LGPL 2.1
|
|
Packit |
c32a2d |
see COPYING and AUTHORS files in distribution or http://mpg123.org
|
|
Packit |
c32a2d |
initially written by Taihei Monma
|
|
Packit |
c32a2d |
*/
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#include "mangle.h"
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#define WINDOW r0
|
|
Packit |
c32a2d |
#define B0 r1
|
|
Packit |
c32a2d |
#define SAMPLES r2
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
int synth_1to1_s32_neon_asm(real *window, real *b0, int *samples, int bo1);
|
|
Packit |
c32a2d |
return value: number of clipped samples (0)
|
|
Packit |
c32a2d |
*/
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#ifndef _M_ARM
|
|
Packit |
c32a2d |
.code 32
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
#ifndef __APPLE__
|
|
Packit |
c32a2d |
.fpu neon
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
.text
|
|
Packit |
c32a2d |
GLOBAL_SYMBOL ASM_NAME(synth_1to1_s32_neon_asm)
|
|
Packit |
c32a2d |
#ifdef __ELF__
|
|
Packit |
c32a2d |
.type ASM_NAME(synth_1to1_s32_neon_asm), %function
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
ALIGN4
|
|
Packit |
c32a2d |
ASM_NAME(synth_1to1_s32_neon_asm):
|
|
Packit |
c32a2d |
push {r4-r6, lr}
|
|
Packit |
c32a2d |
vpush {q4-q7}
|
|
Packit |
c32a2d |
mov r6, sp
|
|
Packit |
c32a2d |
sub sp, sp, #16
|
|
Packit |
c32a2d |
mov r5, sp
|
|
Packit |
c32a2d |
and r5, r5, #0xf
|
|
Packit |
c32a2d |
sub sp, sp, r5
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
add WINDOW, WINDOW, #64
|
|
Packit |
c32a2d |
sub WINDOW, WINDOW, r3, lsl #2
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
mov r3, #4
|
|
Packit |
c32a2d |
mov r4, #128
|
|
Packit |
c32a2d |
mov r5, #64
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
vld1.32 {q0,q1}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q2,q3}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q6,q7}, [WINDOW]
|
|
Packit |
c32a2d |
sub WINDOW, WINDOW, #352
|
|
Packit |
c32a2d |
vld1.32 {q8,q9}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q12,q13}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q14,q15}, [B0, :128]
|
|
Packit |
c32a2d |
vswp q1, q4
|
|
Packit |
c32a2d |
vswp q3, q6
|
|
Packit |
c32a2d |
sub B0, B0, #160
|
|
Packit |
c32a2d |
vmul.f32 q0, q0, q8
|
|
Packit |
c32a2d |
vmul.f32 q2, q2, q10
|
|
Packit |
c32a2d |
vmul.f32 q1, q1, q12
|
|
Packit |
c32a2d |
vmul.f32 q3, q3, q14
|
|
Packit |
c32a2d |
vmla.f32 q0, q4, q9
|
|
Packit |
c32a2d |
vmla.f32 q2, q6, q11
|
|
Packit |
c32a2d |
vmla.f32 q1, q5, q13
|
|
Packit |
c32a2d |
vmla.f32 q3, q7, q15
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q6,q7}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q8,q9}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q12,q13}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q14,q15}, [B0, :128], r5
|
|
Packit |
c32a2d |
vswp q5, q6
|
|
Packit |
c32a2d |
vswp q11, q12
|
|
Packit |
c32a2d |
vmla.f32 q0, q4, q10
|
|
Packit |
c32a2d |
vmla.f32 q2, q5, q11
|
|
Packit |
c32a2d |
vmla.f32 q1, q8, q14
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW]
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128]!
|
|
Packit |
c32a2d |
add WINDOW, WINDOW, #96
|
|
Packit |
c32a2d |
vmla.f32 q3, q4, q10
|
|
Packit |
c32a2d |
vmla.f32 q0, q6, q12
|
|
Packit |
c32a2d |
vmla.f32 q2, q7, q13
|
|
Packit |
c32a2d |
vmla.f32 q1, q9, q15
|
|
Packit |
c32a2d |
vmla.f32 q3, q5, q11
|
|
Packit |
c32a2d |
vmvn.i32 q5, #0xb9000000
|
|
Packit |
c32a2d |
vpadd.f32 d0, d0, d1
|
|
Packit |
c32a2d |
vpadd.f32 d4, d4, d5
|
|
Packit |
c32a2d |
vpadd.f32 d2, d2, d3
|
|
Packit |
c32a2d |
vpadd.f32 d6, d6, d7
|
|
Packit |
c32a2d |
vld1.32 {q6}, [sp, :128]
|
|
Packit |
c32a2d |
vpadd.f32 d0, d0, d4
|
|
Packit |
c32a2d |
vpadd.f32 d1, d2, d6
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vcvt.s32.f32 q3, q0, #16
|
|
Packit |
c32a2d |
vacgt.f32 q5, q0, q5
|
|
Packit |
c32a2d |
vld2.32 {q1,q2}, [SAMPLES]
|
|
Packit |
c32a2d |
vshr.u32 q5, q5, #31
|
|
Packit |
c32a2d |
vmov q1, q3
|
|
Packit |
c32a2d |
vst2.32 {q1,q2}, [SAMPLES]!
|
|
Packit |
c32a2d |
vadd.i32 q5, q5, q6
|
|
Packit |
c32a2d |
vst1.32 {q5}, [sp, :128]
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
subs r3, r3, #1
|
|
Packit |
c32a2d |
bne 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
mov r3, #4
|
|
Packit |
c32a2d |
mov r5, #-64
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
vld1.32 {q0,q1}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q2,q3}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q6,q7}, [WINDOW]
|
|
Packit |
c32a2d |
sub WINDOW, WINDOW, #352
|
|
Packit |
c32a2d |
vld1.32 {q8,q9}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q12,q13}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q14,q15}, [B0, :128]
|
|
Packit |
c32a2d |
vswp q1, q4
|
|
Packit |
c32a2d |
vswp q3, q6
|
|
Packit |
c32a2d |
add B0, B0, #224
|
|
Packit |
c32a2d |
vmul.f32 q0, q0, q8
|
|
Packit |
c32a2d |
vmul.f32 q2, q2, q10
|
|
Packit |
c32a2d |
vmul.f32 q1, q1, q12
|
|
Packit |
c32a2d |
vmul.f32 q3, q3, q14
|
|
Packit |
c32a2d |
vmla.f32 q0, q4, q9
|
|
Packit |
c32a2d |
vmla.f32 q2, q6, q11
|
|
Packit |
c32a2d |
vmla.f32 q1, q5, q13
|
|
Packit |
c32a2d |
vmla.f32 q3, q7, q15
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q6,q7}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q8,q9}, [WINDOW], r4
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q12,q13}, [B0, :128], r5
|
|
Packit |
c32a2d |
vld1.32 {q14,q15}, [B0, :128], r5
|
|
Packit |
c32a2d |
vswp q5, q6
|
|
Packit |
c32a2d |
vswp q11, q12
|
|
Packit |
c32a2d |
vmla.f32 q0, q4, q10
|
|
Packit |
c32a2d |
vmla.f32 q2, q5, q11
|
|
Packit |
c32a2d |
vmla.f32 q1, q8, q14
|
|
Packit |
c32a2d |
vld1.32 {q4,q5}, [WINDOW]
|
|
Packit |
c32a2d |
vld1.32 {q10,q11}, [B0, :128]
|
|
Packit |
c32a2d |
add WINDOW, WINDOW, #96
|
|
Packit |
c32a2d |
sub B0, B0, #96
|
|
Packit |
c32a2d |
vmla.f32 q3, q4, q10
|
|
Packit |
c32a2d |
vmla.f32 q0, q6, q12
|
|
Packit |
c32a2d |
vmla.f32 q2, q7, q13
|
|
Packit |
c32a2d |
vmla.f32 q1, q9, q15
|
|
Packit |
c32a2d |
vmla.f32 q3, q5, q11
|
|
Packit |
c32a2d |
vmvn.i32 q5, #0xb9000000
|
|
Packit |
c32a2d |
vpadd.f32 d0, d0, d1
|
|
Packit |
c32a2d |
vpadd.f32 d4, d4, d5
|
|
Packit |
c32a2d |
vpadd.f32 d2, d2, d3
|
|
Packit |
c32a2d |
vpadd.f32 d6, d6, d7
|
|
Packit |
c32a2d |
vld1.32 {q6}, [sp, :128]
|
|
Packit |
c32a2d |
vpadd.f32 d0, d0, d4
|
|
Packit |
c32a2d |
vpadd.f32 d1, d2, d6
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vcvt.s32.f32 q3, q0, #16
|
|
Packit |
c32a2d |
vacgt.f32 q5, q0, q5
|
|
Packit |
c32a2d |
vld2.32 {q1,q2}, [SAMPLES]
|
|
Packit |
c32a2d |
vshr.u32 q5, q5, #31
|
|
Packit |
c32a2d |
vmov q1, q3
|
|
Packit |
c32a2d |
vst2.32 {q1,q2}, [SAMPLES]!
|
|
Packit |
c32a2d |
vadd.i32 q5, q5, q6
|
|
Packit |
c32a2d |
vst1.32 {q5}, [sp, :128]
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
subs r3, r3, #1
|
|
Packit |
c32a2d |
bne 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vld1.32 {q0}, [sp, :128]
|
|
Packit |
c32a2d |
vpadd.i32 d0, d0, d1
|
|
Packit |
c32a2d |
vpadd.i32 d0, d0, d0
|
|
Packit |
c32a2d |
vmov.32 r0, d0[0]
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
mov sp, r6
|
|
Packit |
c32a2d |
vpop {q4-q7}
|
|
Packit |
c32a2d |
pop {r4-r6, pc}
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
NONEXEC_STACK
|