|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
synth_x86_64_accurate: SSE optimized synth for x86-64 (MPEG-compliant 16bit output version)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
copyright 1995-2009 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 |
/* short *window; */
|
|
Packit |
c32a2d |
#define ARG0 %r10
|
|
Packit |
c32a2d |
/* short *b0; */
|
|
Packit |
c32a2d |
#define ARG1 %rdx
|
|
Packit |
c32a2d |
/* short *samples; */
|
|
Packit |
c32a2d |
#define ARG2 %r8
|
|
Packit |
c32a2d |
/* int bo1; */
|
|
Packit |
c32a2d |
#define ARG3 %r9
|
|
Packit |
c32a2d |
#else
|
|
Packit |
c32a2d |
/* real *window; */
|
|
Packit |
c32a2d |
#define ARG0 %rdi
|
|
Packit |
c32a2d |
/* real *b0; */
|
|
Packit |
c32a2d |
#define ARG1 %rsi
|
|
Packit |
c32a2d |
/* real *samples; */
|
|
Packit |
c32a2d |
#define ARG2 %rdx
|
|
Packit |
c32a2d |
/* int bo1; */
|
|
Packit |
c32a2d |
#define ARG3 %rcx
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#define XMMREG_MAX %xmm15 /* {32767.0, 32767.0, 32767.0, 32767.0} */
|
|
Packit |
c32a2d |
#define XMMREG_MIN %xmm14 /* {-32768.0, -32768.0, -32768.0, -32768.0} */
|
|
Packit |
c32a2d |
#define XMMREG_CLIP %xmm13
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
int synth_1to1_x86_64_accurate_asm(real *window, real *b0, short *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 |
ASM_NAME(maxmin_s16):
|
|
Packit |
c32a2d |
.long 1191181824
|
|
Packit |
c32a2d |
.long 1191181824
|
|
Packit |
c32a2d |
.long 1191181824
|
|
Packit |
c32a2d |
.long 1191181824
|
|
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_x86_64_accurate_asm)
|
|
Packit |
c32a2d |
ASM_NAME(synth_1to1_x86_64_accurate_asm):
|
|
Packit |
c32a2d |
#ifdef IS_MSABI /* should save xmm6-15 */
|
|
Packit |
c32a2d |
movq %rcx, ARG0
|
|
Packit |
c32a2d |
subq $152, %rsp /* stack alignment + 9 xmm registers */
|
|
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 %xmm13, 96(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm14, 112(%rsp)
|
|
Packit |
c32a2d |
movaps %xmm15, 128(%rsp)
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leaq ASM_NAME(maxmin_s16)(%rip), %rax
|
|
Packit |
c32a2d |
movaps (%rax), XMMREG_MAX
|
|
Packit |
c32a2d |
movaps 16(%rax), XMMREG_MIN
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
xorps XMMREG_CLIP, XMMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
andq $0xf, ARG3
|
|
Packit |
c32a2d |
shlq $2, ARG3
|
|
Packit |
c32a2d |
leaq 64(ARG0), ARG0
|
|
Packit |
c32a2d |
subq ARG3, ARG0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movl $4, %ecx
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
movups (ARG0), %xmm8
|
|
Packit |
c32a2d |
movups 16(ARG0), %xmm1
|
|
Packit |
c32a2d |
movups 32(ARG0), %xmm2
|
|
Packit |
c32a2d |
movups 48(ARG0), %xmm3
|
|
Packit |
c32a2d |
movups 128(ARG0), %xmm9
|
|
Packit |
c32a2d |
movups 144(ARG0), %xmm5
|
|
Packit |
c32a2d |
movups 160(ARG0), %xmm6
|
|
Packit |
c32a2d |
movups 176(ARG0), %xmm7
|
|
Packit |
c32a2d |
mulps (ARG1), %xmm8
|
|
Packit |
c32a2d |
mulps 16(ARG1), %xmm1
|
|
Packit |
c32a2d |
mulps 32(ARG1), %xmm2
|
|
Packit |
c32a2d |
mulps 48(ARG1), %xmm3
|
|
Packit |
c32a2d |
mulps 64(ARG1), %xmm9
|
|
Packit |
c32a2d |
mulps 80(ARG1), %xmm5
|
|
Packit |
c32a2d |
mulps 96(ARG1), %xmm6
|
|
Packit |
c32a2d |
mulps 112(ARG1), %xmm7
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
addps %xmm1, %xmm8
|
|
Packit |
c32a2d |
addps %xmm2, %xmm3
|
|
Packit |
c32a2d |
addps %xmm5, %xmm9
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm3, %xmm8
|
|
Packit |
c32a2d |
addps %xmm6, %xmm9
|
|
Packit |
c32a2d |
leaq 256(ARG0), ARG0
|
|
Packit |
c32a2d |
leaq 128(ARG1), ARG1
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (ARG0), %xmm10
|
|
Packit |
c32a2d |
movups 16(ARG0), %xmm1
|
|
Packit |
c32a2d |
movups 32(ARG0), %xmm2
|
|
Packit |
c32a2d |
movups 48(ARG0), %xmm3
|
|
Packit |
c32a2d |
movups 128(ARG0), %xmm11
|
|
Packit |
c32a2d |
movups 144(ARG0), %xmm5
|
|
Packit |
c32a2d |
movups 160(ARG0), %xmm6
|
|
Packit |
c32a2d |
movups 176(ARG0), %xmm7
|
|
Packit |
c32a2d |
mulps (ARG1), %xmm10
|
|
Packit |
c32a2d |
mulps 16(ARG1), %xmm1
|
|
Packit |
c32a2d |
mulps 32(ARG1), %xmm2
|
|
Packit |
c32a2d |
mulps 48(ARG1), %xmm3
|
|
Packit |
c32a2d |
mulps 64(ARG1), %xmm11
|
|
Packit |
c32a2d |
mulps 80(ARG1), %xmm5
|
|
Packit |
c32a2d |
mulps 96(ARG1), %xmm6
|
|
Packit |
c32a2d |
mulps 112(ARG1), %xmm7
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
addps %xmm1, %xmm10
|
|
Packit |
c32a2d |
addps %xmm2, %xmm3
|
|
Packit |
c32a2d |
addps %xmm5, %xmm11
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm3, %xmm10
|
|
Packit |
c32a2d |
addps %xmm6, %xmm11
|
|
Packit |
c32a2d |
leaq 256(ARG0), ARG0
|
|
Packit |
c32a2d |
leaq 128(ARG1), ARG1
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm8, %xmm0
|
|
Packit |
c32a2d |
movaps %xmm10, %xmm1
|
|
Packit |
c32a2d |
unpcklps %xmm9, %xmm8
|
|
Packit |
c32a2d |
unpcklps %xmm11, %xmm10
|
|
Packit |
c32a2d |
unpckhps %xmm9, %xmm0
|
|
Packit |
c32a2d |
unpckhps %xmm11, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm8, %xmm2
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm3
|
|
Packit |
c32a2d |
movlhps %xmm10, %xmm8
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm10
|
|
Packit |
c32a2d |
movlhps %xmm1, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm1
|
|
Packit |
c32a2d |
subps %xmm10, %xmm8
|
|
Packit |
c32a2d |
subps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm8, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (ARG2), %xmm3
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm2
|
|
Packit |
c32a2d |
cmpnleps XMMREG_MAX, %xmm1
|
|
Packit |
c32a2d |
cmpltps XMMREG_MIN, %xmm2
|
|
Packit |
c32a2d |
cvtps2dq %xmm0, %xmm0
|
|
Packit |
c32a2d |
packssdw %xmm0, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm4
|
|
Packit |
c32a2d |
pshuflw $0xdd, %xmm3, %xmm3
|
|
Packit |
c32a2d |
pshuflw $0xdd, %xmm4, %xmm4
|
|
Packit |
c32a2d |
psrlq $32, %xmm3
|
|
Packit |
c32a2d |
psllq $32, %xmm4
|
|
Packit |
c32a2d |
por %xmm4, %xmm3
|
|
Packit |
c32a2d |
punpcklwd %xmm3, %xmm0
|
|
Packit |
c32a2d |
movups %xmm0, (ARG2)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
psrld $31, %xmm1
|
|
Packit |
c32a2d |
psrld $31, %xmm2
|
|
Packit |
c32a2d |
paddd %xmm2, %xmm1
|
|
Packit |
c32a2d |
paddd %xmm1, XMMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leaq 16(ARG2), ARG2
|
|
Packit |
c32a2d |
decl %ecx
|
|
Packit |
c32a2d |
jnz 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movl $4, %ecx
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
movups (ARG0), %xmm8
|
|
Packit |
c32a2d |
movups 16(ARG0), %xmm1
|
|
Packit |
c32a2d |
movups 32(ARG0), %xmm2
|
|
Packit |
c32a2d |
movups 48(ARG0), %xmm3
|
|
Packit |
c32a2d |
movups 128(ARG0), %xmm9
|
|
Packit |
c32a2d |
movups 144(ARG0), %xmm5
|
|
Packit |
c32a2d |
movups 160(ARG0), %xmm6
|
|
Packit |
c32a2d |
movups 176(ARG0), %xmm7
|
|
Packit |
c32a2d |
mulps (ARG1), %xmm8
|
|
Packit |
c32a2d |
mulps 16(ARG1), %xmm1
|
|
Packit |
c32a2d |
mulps 32(ARG1), %xmm2
|
|
Packit |
c32a2d |
mulps 48(ARG1), %xmm3
|
|
Packit |
c32a2d |
mulps -64(ARG1), %xmm9
|
|
Packit |
c32a2d |
mulps -48(ARG1), %xmm5
|
|
Packit |
c32a2d |
mulps -32(ARG1), %xmm6
|
|
Packit |
c32a2d |
mulps -16(ARG1), %xmm7
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
addps %xmm1, %xmm8
|
|
Packit |
c32a2d |
addps %xmm2, %xmm3
|
|
Packit |
c32a2d |
addps %xmm5, %xmm9
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm3, %xmm8
|
|
Packit |
c32a2d |
addps %xmm6, %xmm9
|
|
Packit |
c32a2d |
leaq 256(ARG0), ARG0
|
|
Packit |
c32a2d |
leaq -128(ARG1), ARG1
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (ARG0), %xmm10
|
|
Packit |
c32a2d |
movups 16(ARG0), %xmm1
|
|
Packit |
c32a2d |
movups 32(ARG0), %xmm2
|
|
Packit |
c32a2d |
movups 48(ARG0), %xmm3
|
|
Packit |
c32a2d |
movups 128(ARG0), %xmm11
|
|
Packit |
c32a2d |
movups 144(ARG0), %xmm5
|
|
Packit |
c32a2d |
movups 160(ARG0), %xmm6
|
|
Packit |
c32a2d |
movups 176(ARG0), %xmm7
|
|
Packit |
c32a2d |
mulps (ARG1), %xmm10
|
|
Packit |
c32a2d |
mulps 16(ARG1), %xmm1
|
|
Packit |
c32a2d |
mulps 32(ARG1), %xmm2
|
|
Packit |
c32a2d |
mulps 48(ARG1), %xmm3
|
|
Packit |
c32a2d |
mulps -64(ARG1), %xmm11
|
|
Packit |
c32a2d |
mulps -48(ARG1), %xmm5
|
|
Packit |
c32a2d |
mulps -32(ARG1), %xmm6
|
|
Packit |
c32a2d |
mulps -16(ARG1), %xmm7
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
addps %xmm1, %xmm10
|
|
Packit |
c32a2d |
addps %xmm2, %xmm3
|
|
Packit |
c32a2d |
addps %xmm5, %xmm11
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm3, %xmm10
|
|
Packit |
c32a2d |
addps %xmm6, %xmm11
|
|
Packit |
c32a2d |
leaq 256(ARG0), ARG0
|
|
Packit |
c32a2d |
leaq -128(ARG1), ARG1
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm8, %xmm0
|
|
Packit |
c32a2d |
movaps %xmm10, %xmm1
|
|
Packit |
c32a2d |
unpcklps %xmm9, %xmm8
|
|
Packit |
c32a2d |
unpcklps %xmm11, %xmm10
|
|
Packit |
c32a2d |
unpckhps %xmm9, %xmm0
|
|
Packit |
c32a2d |
unpckhps %xmm11, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm8, %xmm2
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm3
|
|
Packit |
c32a2d |
movlhps %xmm10, %xmm8
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm10
|
|
Packit |
c32a2d |
movlhps %xmm1, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm1
|
|
Packit |
c32a2d |
addps %xmm10, %xmm8
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm8, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (ARG2), %xmm3
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm2
|
|
Packit |
c32a2d |
cmpnleps XMMREG_MAX, %xmm1
|
|
Packit |
c32a2d |
cmpltps XMMREG_MIN, %xmm2
|
|
Packit |
c32a2d |
cvtps2dq %xmm0, %xmm0
|
|
Packit |
c32a2d |
packssdw %xmm0, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm4
|
|
Packit |
c32a2d |
pshuflw $0xdd, %xmm3, %xmm3
|
|
Packit |
c32a2d |
pshuflw $0xdd, %xmm4, %xmm4
|
|
Packit |
c32a2d |
psrlq $32, %xmm3
|
|
Packit |
c32a2d |
psllq $32, %xmm4
|
|
Packit |
c32a2d |
por %xmm4, %xmm3
|
|
Packit |
c32a2d |
punpcklwd %xmm3, %xmm0
|
|
Packit |
c32a2d |
movups %xmm0, (ARG2)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
psrld $31, %xmm1
|
|
Packit |
c32a2d |
psrld $31, %xmm2
|
|
Packit |
c32a2d |
paddd %xmm2, %xmm1
|
|
Packit |
c32a2d |
paddd %xmm1, XMMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leaq 16(ARG2), ARG2
|
|
Packit |
c32a2d |
decl %ecx
|
|
Packit |
c32a2d |
jnz 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
pshuflw $0xee, XMMREG_CLIP, %xmm0
|
|
Packit |
c32a2d |
movhlps XMMREG_CLIP, %xmm1
|
|
Packit |
c32a2d |
pshuflw $0xee, %xmm1, %xmm2
|
|
Packit |
c32a2d |
paddd %xmm0, XMMREG_CLIP
|
|
Packit |
c32a2d |
paddd %xmm1, XMMREG_CLIP
|
|
Packit |
c32a2d |
paddd %xmm2, XMMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movd XMMREG_CLIP, %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), %xmm13
|
|
Packit |
c32a2d |
movaps 112(%rsp), %xmm14
|
|
Packit |
c32a2d |
movaps 128(%rsp), %xmm15
|
|
Packit |
c32a2d |
addq $152, %rsp
|
|
Packit |
c32a2d |
#endif
|
|
Packit |
c32a2d |
ret
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
NONEXEC_STACK
|