|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
synth_stereo_avx_accurate: AVX optimized synth for x86-64 (stereo specific, MPEG-compliant 16bit output version)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
copyright 1995-2013 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 |
#ifdef IS_MSABI
|
|
Packit |
c32a2d |
/* real *window; */
|
|
Packit |
c32a2d |
#define WINDOW %r10
|
|
Packit |
c32a2d |
/* real *b0l; */
|
|
Packit |
c32a2d |
#define B0L %rdx
|
|
Packit |
c32a2d |
/* real *b0r; */
|
|
Packit |
c32a2d |
#define B0R %r8
|
|
Packit |
c32a2d |
/* real *samples; */
|
|
Packit |
c32a2d |
#define SAMPLES %r9
|
|
Packit |
c32a2d |
#else
|
|
Packit |
c32a2d |
/* real *window; */
|
|
Packit |
c32a2d |
#define WINDOW %rdi
|
|
Packit |
c32a2d |
/* real *b0l; */
|
|
Packit |
c32a2d |
#define B0L %rsi
|
|
Packit |
c32a2d |
/* real *b0r; */
|
|
Packit |
c32a2d |
#define B0R %rdx
|
|
Packit |
c32a2d |
/* real *samples; */
|
|
Packit |
c32a2d |
#define SAMPLES %r9
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
int synth_1to1_s_avx_accurate_asm(real *window, real *b0l, real *b0r, real *samples, int bo1);
|
|
Packit |
c32a2d |
return value: number of clipped samples
|
|
Packit |
c32a2d |
*/
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#ifndef __APPLE__
|
|
Packit |
c32a2d |
.section .rodata
|
|
Packit |
c32a2d |
#else
|
|
Packit |
c32a2d |
.data
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
ALIGN32
|
|
Packit |
c32a2d |
maxmin_avx:
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.long -956301312
|
|
Packit |
c32a2d |
.text
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
.globl ASM_NAME(synth_1to1_s_avx_accurate_asm)
|
|
Packit |
c32a2d |
ASM_NAME(synth_1to1_s_avx_accurate_asm):
|
|
Packit |
c32a2d |
#ifdef IS_MSABI /* should save xmm6-15 */
|
|
Packit |
c32a2d |
push %rbp
|
|
Packit |
c32a2d |
mov %rsp, %rbp
|
|
Packit |
c32a2d |
sub $144, %rsp
|
|
Packit |
c32a2d |
movaps %xmm6, (%rsp)
|
|
Packit |
c32a2d |
movaps %xmm7, 16(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm8, 32(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm9, 48(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm10, 64(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm11, 80(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm12, 96(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm13, 112(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm14, 128(%rsp)
|
|
Packit |
c32a2d |
movl 48(%rbp), %eax /* 5th argument; placed after 32-byte shadow space */
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#ifdef IS_MSABI
|
|
Packit |
c32a2d |
shl $2, %eax
|
|
Packit |
c32a2d |
mov %rcx, WINDOW
|
|
Packit |
c32a2d |
#else
|
|
Packit |
c32a2d |
mov %r8d, %eax
|
|
Packit |
c32a2d |
shl $2, %eax
|
|
Packit |
c32a2d |
mov %rcx, SAMPLES
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
add $64, WINDOW
|
|
Packit |
c32a2d |
sub %rax, WINDOW
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
mov $128, %rax
|
|
Packit |
c32a2d |
mov $4, %ecx
|
|
Packit |
c32a2d |
vpxor %xmm14, %xmm14, %xmm14
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
vmovups (WINDOW), %ymm8
|
|
Packit |
c32a2d |
vmovups 32(WINDOW), %ymm9
|
|
Packit |
c32a2d |
vmovups (WINDOW,%rax), %ymm10
|
|
Packit |
c32a2d |
vmovups 32(WINDOW,%rax), %ymm11
|
|
Packit |
c32a2d |
vmulps (B0L), %ymm8, %ymm0
|
|
Packit |
c32a2d |
vmulps 32(B0L), %ymm9, %ymm1
|
|
Packit |
c32a2d |
vmulps (B0R), %ymm8, %ymm2
|
|
Packit |
c32a2d |
vmulps 32(B0R), %ymm9, %ymm3
|
|
Packit |
c32a2d |
vmulps 64(B0L), %ymm10, %ymm4
|
|
Packit |
c32a2d |
vmulps 96(B0L), %ymm11, %ymm5
|
|
Packit |
c32a2d |
vmulps 64(B0R), %ymm10, %ymm6
|
|
Packit |
c32a2d |
vmulps 96(B0R), %ymm11, %ymm7
|
|
Packit |
c32a2d |
vaddps %ymm1, %ymm0, %ymm8
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm0
|
|
Packit |
c32a2d |
vaddps %ymm5, %ymm4, %ymm9
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm6, %ymm1
|
|
Packit |
c32a2d |
lea (WINDOW,%rax,2), WINDOW
|
|
Packit |
c32a2d |
add %rax, B0L
|
|
Packit |
c32a2d |
add %rax, B0R
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vmovups (WINDOW), %ymm10
|
|
Packit |
c32a2d |
vmovups 32(WINDOW), %ymm11
|
|
Packit |
c32a2d |
vmovups (WINDOW,%rax), %ymm12
|
|
Packit |
c32a2d |
vmovups 32(WINDOW,%rax), %ymm13
|
|
Packit |
c32a2d |
vmulps (B0L), %ymm10, %ymm2
|
|
Packit |
c32a2d |
vmulps 32(B0L), %ymm11, %ymm3
|
|
Packit |
c32a2d |
vmulps (B0R), %ymm10, %ymm4
|
|
Packit |
c32a2d |
vmulps 32(B0R), %ymm11, %ymm5
|
|
Packit |
c32a2d |
vmulps 64(B0L), %ymm12, %ymm6
|
|
Packit |
c32a2d |
vmulps 96(B0L), %ymm13, %ymm10
|
|
Packit |
c32a2d |
vmulps 64(B0R), %ymm12, %ymm7
|
|
Packit |
c32a2d |
vmulps 96(B0R), %ymm13, %ymm11
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vaddps %ymm5, %ymm4, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm10, %ymm4
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm11, %ymm5
|
|
Packit |
c32a2d |
lea (WINDOW,%rax,2), WINDOW
|
|
Packit |
c32a2d |
add %rax, B0L
|
|
Packit |
c32a2d |
add %rax, B0R
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vunpcklps %ymm0, %ymm8, %ymm6
|
|
Packit |
c32a2d |
vunpckhps %ymm0, %ymm8, %ymm0
|
|
Packit |
c32a2d |
vunpcklps %ymm1, %ymm9, %ymm7
|
|
Packit |
c32a2d |
vunpckhps %ymm1, %ymm9, %ymm1
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm0, %ymm0
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm1, %ymm1
|
|
Packit |
c32a2d |
vunpcklps %ymm3, %ymm2, %ymm6
|
|
Packit |
c32a2d |
vunpckhps %ymm3, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vunpcklps %ymm5, %ymm4, %ymm7
|
|
Packit |
c32a2d |
vunpckhps %ymm5, %ymm4, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm3, %ymm3
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vunpcklpd %ymm1, %ymm0, %ymm4
|
|
Packit |
c32a2d |
vunpckhpd %ymm1, %ymm0, %ymm0
|
|
Packit |
c32a2d |
vunpcklpd %ymm3, %ymm2, %ymm5
|
|
Packit |
c32a2d |
vunpckhpd %ymm3, %ymm2, %ymm1
|
|
Packit |
c32a2d |
vsubps %ymm0, %ymm4, %ymm0
|
|
Packit |
c32a2d |
vsubps %ymm1, %ymm5, %ymm1
|
|
Packit |
c32a2d |
vperm2f128 $0x20, %ymm1, %ymm0, %ymm2
|
|
Packit |
c32a2d |
vperm2f128 $0x31, %ymm1, %ymm0, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm0
|
|
Packit |
c32a2d |
vcmpnleps maxmin_avx(%rip), %ymm0, %ymm1
|
|
Packit |
c32a2d |
vcmpltps 32+maxmin_avx(%rip), %ymm0, %ymm2
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm1, %xmm3
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm2, %xmm4
|
|
Packit |
c32a2d |
vpackssdw %xmm2, %xmm1, %xmm1
|
|
Packit |
c32a2d |
vpackssdw %xmm4, %xmm3, %xmm3
|
|
Packit |
c32a2d |
vpaddw %xmm3, %xmm1, %xmm1
|
|
Packit |
c32a2d |
vpaddw %xmm1, %xmm14, %xmm14
|
|
Packit |
c32a2d |
vcvtps2dq %ymm0, %ymm0
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm0, %xmm1
|
|
Packit |
c32a2d |
vpackssdw %xmm1, %xmm0, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vmovups %xmm0, (SAMPLES)
|
|
Packit |
c32a2d |
add $16, SAMPLES
|
|
Packit |
c32a2d |
dec %ecx
|
|
Packit |
c32a2d |
jnz 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
mov $4, %ecx
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
vmovups (WINDOW), %ymm8
|
|
Packit |
c32a2d |
vmovups 32(WINDOW), %ymm9
|
|
Packit |
c32a2d |
vmovups (WINDOW,%rax), %ymm10
|
|
Packit |
c32a2d |
vmovups 32(WINDOW,%rax), %ymm11
|
|
Packit |
c32a2d |
vmulps (B0L), %ymm8, %ymm0
|
|
Packit |
c32a2d |
vmulps 32(B0L), %ymm9, %ymm1
|
|
Packit |
c32a2d |
vmulps (B0R), %ymm8, %ymm2
|
|
Packit |
c32a2d |
vmulps 32(B0R), %ymm9, %ymm3
|
|
Packit |
c32a2d |
vmulps -64(B0L), %ymm10, %ymm4
|
|
Packit |
c32a2d |
vmulps -32(B0L), %ymm11, %ymm5
|
|
Packit |
c32a2d |
vmulps -64(B0R), %ymm10, %ymm6
|
|
Packit |
c32a2d |
vmulps -32(B0R), %ymm11, %ymm7
|
|
Packit |
c32a2d |
vaddps %ymm1, %ymm0, %ymm8
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm0
|
|
Packit |
c32a2d |
vaddps %ymm5, %ymm4, %ymm9
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm6, %ymm1
|
|
Packit |
c32a2d |
lea (WINDOW,%rax,2), WINDOW
|
|
Packit |
c32a2d |
sub %rax, B0L
|
|
Packit |
c32a2d |
sub %rax, B0R
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vmovups (WINDOW), %ymm10
|
|
Packit |
c32a2d |
vmovups 32(WINDOW), %ymm11
|
|
Packit |
c32a2d |
vmovups (WINDOW,%rax), %ymm12
|
|
Packit |
c32a2d |
vmovups 32(WINDOW,%rax), %ymm13
|
|
Packit |
c32a2d |
vmulps (B0L), %ymm10, %ymm2
|
|
Packit |
c32a2d |
vmulps 32(B0L), %ymm11, %ymm3
|
|
Packit |
c32a2d |
vmulps (B0R), %ymm10, %ymm4
|
|
Packit |
c32a2d |
vmulps 32(B0R), %ymm11, %ymm5
|
|
Packit |
c32a2d |
vmulps -64(B0L), %ymm12, %ymm6
|
|
Packit |
c32a2d |
vmulps -32(B0L), %ymm13, %ymm10
|
|
Packit |
c32a2d |
vmulps -64(B0R), %ymm12, %ymm7
|
|
Packit |
c32a2d |
vmulps -32(B0R), %ymm13, %ymm11
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vaddps %ymm5, %ymm4, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm10, %ymm4
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm11, %ymm5
|
|
Packit |
c32a2d |
lea (WINDOW,%rax,2), WINDOW
|
|
Packit |
c32a2d |
sub %rax, B0L
|
|
Packit |
c32a2d |
sub %rax, B0R
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vunpcklps %ymm0, %ymm8, %ymm6
|
|
Packit |
c32a2d |
vunpckhps %ymm0, %ymm8, %ymm0
|
|
Packit |
c32a2d |
vunpcklps %ymm1, %ymm9, %ymm7
|
|
Packit |
c32a2d |
vunpckhps %ymm1, %ymm9, %ymm1
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm0, %ymm0
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm1, %ymm1
|
|
Packit |
c32a2d |
vunpcklps %ymm3, %ymm2, %ymm6
|
|
Packit |
c32a2d |
vunpckhps %ymm3, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vunpcklps %ymm5, %ymm4, %ymm7
|
|
Packit |
c32a2d |
vunpckhps %ymm5, %ymm4, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm6, %ymm2, %ymm2
|
|
Packit |
c32a2d |
vaddps %ymm7, %ymm3, %ymm3
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vunpcklpd %ymm1, %ymm0, %ymm4
|
|
Packit |
c32a2d |
vunpckhpd %ymm1, %ymm0, %ymm0
|
|
Packit |
c32a2d |
vunpcklpd %ymm3, %ymm2, %ymm5
|
|
Packit |
c32a2d |
vunpckhpd %ymm3, %ymm2, %ymm1
|
|
Packit |
c32a2d |
vaddps %ymm0, %ymm4, %ymm0
|
|
Packit |
c32a2d |
vaddps %ymm1, %ymm5, %ymm1
|
|
Packit |
c32a2d |
vperm2f128 $0x20, %ymm1, %ymm0, %ymm2
|
|
Packit |
c32a2d |
vperm2f128 $0x31, %ymm1, %ymm0, %ymm3
|
|
Packit |
c32a2d |
vaddps %ymm3, %ymm2, %ymm0
|
|
Packit |
c32a2d |
vcmpnleps maxmin_avx(%rip), %ymm0, %ymm1
|
|
Packit |
c32a2d |
vcmpltps 32+maxmin_avx(%rip), %ymm0, %ymm2
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm1, %xmm3
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm2, %xmm4
|
|
Packit |
c32a2d |
vpackssdw %xmm2, %xmm1, %xmm1
|
|
Packit |
c32a2d |
vpackssdw %xmm4, %xmm3, %xmm3
|
|
Packit |
c32a2d |
vpaddw %xmm3, %xmm1, %xmm1
|
|
Packit |
c32a2d |
vpaddw %xmm1, %xmm14, %xmm14
|
|
Packit |
c32a2d |
vcvtps2dq %ymm0, %ymm0
|
|
Packit |
c32a2d |
vextractf128 $0x1, %ymm0, %xmm1
|
|
Packit |
c32a2d |
vpackssdw %xmm1, %xmm0, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vmovups %xmm0, (SAMPLES)
|
|
Packit |
c32a2d |
add $16, SAMPLES
|
|
Packit |
c32a2d |
dec %ecx
|
|
Packit |
c32a2d |
jnz 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
vzeroupper
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
pxor %xmm1, %xmm1
|
|
Packit |
c32a2d |
psubw %xmm14, %xmm1
|
|
Packit |
c32a2d |
pshufd $0x4e, %xmm1, %xmm0
|
|
Packit |
c32a2d |
paddw %xmm1, %xmm0
|
|
Packit |
c32a2d |
pshuflw $0x4e, %xmm0, %xmm1
|
|
Packit |
c32a2d |
paddw %xmm1, %xmm0
|
|
Packit |
c32a2d |
pshuflw $0x11, %xmm0, %xmm1
|
|
Packit |
c32a2d |
paddw %xmm1, %xmm0
|
|
Packit |
c32a2d |
movd %xmm0, %eax
|
|
Packit |
c32a2d |
and $0x7f, %eax
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#ifdef IS_MSABI
|
|
Packit |
c32a2d |
movaps (%rsp), %xmm6
|
|
Packit |
c32a2d |
movaps 16(%rsp), %xmm7
|
|
Packit |
c32a2d |
movaps 32(%rsp), %xmm8
|
|
Packit |
c32a2d |
movaps 48(%rsp), %xmm9
|
|
Packit |
c32a2d |
movaps 64(%rsp), %xmm10
|
|
Packit |
c32a2d |
movaps 80(%rsp), %xmm11
|
|
Packit |
c32a2d |
movaps 96(%rsp), %xmm12
|
|
Packit |
c32a2d |
movaps 112(%rsp), %xmm13
|
|
Packit |
c32a2d |
movaps 128(%rsp), %xmm14
|
|
Packit |
c32a2d |
mov %rbp, %rsp
|
|
Packit |
c32a2d |
pop %rbp
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
ret
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
NONEXEC_STACK
|