|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
synth_sse_s32: SSE optimized synth (s32 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 |
/* real *window; */
|
|
Packit |
c32a2d |
#define WINDOW %ebx
|
|
Packit |
c32a2d |
/* real *b0; */
|
|
Packit |
c32a2d |
#define B0 %edx
|
|
Packit |
c32a2d |
/* real *samples; */
|
|
Packit |
c32a2d |
#define SAMPLES %esi
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#define MMREG_CLIP %mm7
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
/*
|
|
Packit |
c32a2d |
int synth_1to1_s32_sse_asm(real *window, real *b0, int32_t *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 |
scale_s32:
|
|
Packit |
c32a2d |
.long 1199570944 /* 65536.0 */
|
|
Packit |
c32a2d |
.long 1199570944
|
|
Packit |
c32a2d |
.long 1199570944
|
|
Packit |
c32a2d |
.long 1199570944
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
maxmin_s32:
|
|
Packit |
c32a2d |
.long 1191182335 /* 32767.999 */
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long 1191182335
|
|
Packit |
c32a2d |
.long -956301312 /* -32768.0 */
|
|
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_s32_sse_asm)
|
|
Packit |
c32a2d |
ASM_NAME(synth_1to1_s32_sse_asm):
|
|
Packit |
c32a2d |
pushl %ebp
|
|
Packit |
c32a2d |
movl %esp, %ebp
|
|
Packit |
c32a2d |
pushl %ebx
|
|
Packit |
c32a2d |
pushl %esi
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
pxor MMREG_CLIP, MMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movl 8(%ebp), WINDOW
|
|
Packit |
c32a2d |
movl 12(%ebp), B0
|
|
Packit |
c32a2d |
movl 16(%ebp), SAMPLES
|
|
Packit |
c32a2d |
movl 20(%ebp), %eax
|
|
Packit |
c32a2d |
shll $2, %eax
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 64(WINDOW), WINDOW
|
|
Packit |
c32a2d |
subl %eax, WINDOW
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
#undef _EBX_
|
|
Packit |
c32a2d |
#define _EBX_ %eax
|
|
Packit |
c32a2d |
GET_GOT
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movl $4, %ecx
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ALIGN16
|
|
Packit |
c32a2d |
1:
|
|
Packit |
c32a2d |
movups (WINDOW), %xmm0
|
|
Packit |
c32a2d |
movups 16(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 32(WINDOW), %xmm2
|
|
Packit |
c32a2d |
movups 48(WINDOW), %xmm3
|
|
Packit |
c32a2d |
movups 128(WINDOW), %xmm4
|
|
Packit |
c32a2d |
movups 144(WINDOW), %xmm5
|
|
Packit |
c32a2d |
movups 160(WINDOW), %xmm6
|
|
Packit |
c32a2d |
movups 176(WINDOW), %xmm7
|
|
Packit |
c32a2d |
mulps 0(B0), %xmm0
|
|
Packit |
c32a2d |
mulps 16(B0), %xmm1
|
|
Packit |
c32a2d |
mulps 32(B0), %xmm2
|
|
Packit |
c32a2d |
mulps 48(B0), %xmm3
|
|
Packit |
c32a2d |
mulps 64(B0), %xmm4
|
|
Packit |
c32a2d |
mulps 80(B0), %xmm5
|
|
Packit |
c32a2d |
mulps 96(B0), %xmm6
|
|
Packit |
c32a2d |
mulps 112(B0), %xmm7
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm2
|
|
Packit |
c32a2d |
addps %xmm5, %xmm4
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm2, %xmm0
|
|
Packit |
c32a2d |
addps %xmm6, %xmm4
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm5
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm4
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 256(WINDOW), WINDOW
|
|
Packit |
c32a2d |
leal 128(B0), B0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (WINDOW), %xmm0
|
|
Packit |
c32a2d |
movups 16(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 32(WINDOW), %xmm2
|
|
Packit |
c32a2d |
movups 48(WINDOW), %xmm3
|
|
Packit |
c32a2d |
movups 128(WINDOW), %xmm6
|
|
Packit |
c32a2d |
movups 144(WINDOW), %xmm7
|
|
Packit |
c32a2d |
mulps (B0), %xmm0
|
|
Packit |
c32a2d |
mulps 16(B0), %xmm1
|
|
Packit |
c32a2d |
mulps 32(B0), %xmm2
|
|
Packit |
c32a2d |
mulps 48(B0), %xmm3
|
|
Packit |
c32a2d |
mulps 64(B0), %xmm6
|
|
Packit |
c32a2d |
mulps 80(B0), %xmm7
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm2
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
movups 160(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 176(WINDOW), %xmm3
|
|
Packit |
c32a2d |
mulps 96(B0), %xmm1
|
|
Packit |
c32a2d |
mulps 112(B0), %xmm3
|
|
Packit |
c32a2d |
addps %xmm2, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm1
|
|
Packit |
c32a2d |
addps %xmm1, %xmm6
|
|
Packit |
c32a2d |
movaps %xmm6, %xmm7
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm6
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 256(WINDOW), WINDOW
|
|
Packit |
c32a2d |
leal 128(B0), B0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm0
|
|
Packit |
c32a2d |
movaps %xmm6, %xmm1
|
|
Packit |
c32a2d |
unpcklps %xmm5, %xmm4
|
|
Packit |
c32a2d |
unpcklps %xmm7, %xmm6
|
|
Packit |
c32a2d |
unpckhps %xmm5, %xmm0
|
|
Packit |
c32a2d |
unpckhps %xmm7, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm2
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm3
|
|
Packit |
c32a2d |
movlhps %xmm6, %xmm4
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm6
|
|
Packit |
c32a2d |
movlhps %xmm1, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm1
|
|
Packit |
c32a2d |
subps %xmm6, %xmm4
|
|
Packit |
c32a2d |
subps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm4, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm2
|
|
Packit |
c32a2d |
mulps LOCAL_VAR(scale_s32), %xmm0
|
|
Packit |
c32a2d |
cmpnleps LOCAL_VAR(maxmin_s32), %xmm1
|
|
Packit |
c32a2d |
cmpltps 16+LOCAL_VAR(maxmin_s32), %xmm2
|
|
Packit |
c32a2d |
cvtps2pi %xmm0, %mm0
|
|
Packit |
c32a2d |
movhlps %xmm0, %xmm0
|
|
Packit |
c32a2d |
cvtps2pi %xmm0, %mm1
|
|
Packit |
c32a2d |
cvtps2pi %xmm1, %mm2
|
|
Packit |
c32a2d |
movhlps %xmm1, %xmm1
|
|
Packit |
c32a2d |
cvtps2pi %xmm1, %mm3
|
|
Packit |
c32a2d |
psrad $31, %mm2
|
|
Packit |
c32a2d |
psrad $31, %mm3
|
|
Packit |
c32a2d |
pxor %mm2, %mm0
|
|
Packit |
c32a2d |
pxor %mm3, %mm1
|
|
Packit |
c32a2d |
movd %mm0, (SAMPLES)
|
|
Packit |
c32a2d |
psrlq $32, %mm0
|
|
Packit |
c32a2d |
movd %mm0, 8(SAMPLES)
|
|
Packit |
c32a2d |
movd %mm1, 16(SAMPLES)
|
|
Packit |
c32a2d |
psrlq $32, %mm1
|
|
Packit |
c32a2d |
movd %mm1, 24(SAMPLES)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
cvtps2pi %xmm2, %mm0
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm2
|
|
Packit |
c32a2d |
cvtps2pi %xmm2, %mm1
|
|
Packit |
c32a2d |
packssdw %mm3, %mm2
|
|
Packit |
c32a2d |
packssdw %mm1, %mm0
|
|
Packit |
c32a2d |
psrlw $15, %mm2
|
|
Packit |
c32a2d |
psrlw $15, %mm0
|
|
Packit |
c32a2d |
paddw %mm2, %mm0
|
|
Packit |
c32a2d |
paddw %mm0, MMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 32(SAMPLES), SAMPLES
|
|
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 (WINDOW), %xmm0
|
|
Packit |
c32a2d |
movups 16(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 32(WINDOW), %xmm2
|
|
Packit |
c32a2d |
movups 48(WINDOW), %xmm3
|
|
Packit |
c32a2d |
movups 128(WINDOW), %xmm4
|
|
Packit |
c32a2d |
movups 144(WINDOW), %xmm5
|
|
Packit |
c32a2d |
movups 160(WINDOW), %xmm6
|
|
Packit |
c32a2d |
movups 176(WINDOW), %xmm7
|
|
Packit |
c32a2d |
mulps 0(B0), %xmm0
|
|
Packit |
c32a2d |
mulps 16(B0), %xmm1
|
|
Packit |
c32a2d |
mulps 32(B0), %xmm2
|
|
Packit |
c32a2d |
mulps 48(B0), %xmm3
|
|
Packit |
c32a2d |
mulps -64(B0), %xmm4
|
|
Packit |
c32a2d |
mulps -48(B0), %xmm5
|
|
Packit |
c32a2d |
mulps -32(B0), %xmm6
|
|
Packit |
c32a2d |
mulps -16(B0), %xmm7
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm2
|
|
Packit |
c32a2d |
addps %xmm5, %xmm4
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
addps %xmm2, %xmm0
|
|
Packit |
c32a2d |
addps %xmm6, %xmm4
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm5
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm4
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 256(WINDOW), WINDOW
|
|
Packit |
c32a2d |
leal -128(B0), B0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movups (WINDOW), %xmm0
|
|
Packit |
c32a2d |
movups 16(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 32(WINDOW), %xmm2
|
|
Packit |
c32a2d |
movups 48(WINDOW), %xmm3
|
|
Packit |
c32a2d |
movups 128(WINDOW), %xmm6
|
|
Packit |
c32a2d |
movups 144(WINDOW), %xmm7
|
|
Packit |
c32a2d |
mulps (B0), %xmm0
|
|
Packit |
c32a2d |
mulps 16(B0), %xmm1
|
|
Packit |
c32a2d |
mulps 32(B0), %xmm2
|
|
Packit |
c32a2d |
mulps 48(B0), %xmm3
|
|
Packit |
c32a2d |
mulps -64(B0), %xmm6
|
|
Packit |
c32a2d |
mulps -48(B0), %xmm7
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm2
|
|
Packit |
c32a2d |
addps %xmm7, %xmm6
|
|
Packit |
c32a2d |
movups 160(WINDOW), %xmm1
|
|
Packit |
c32a2d |
movups 176(WINDOW), %xmm3
|
|
Packit |
c32a2d |
mulps -32(B0), %xmm1
|
|
Packit |
c32a2d |
mulps -16(B0), %xmm3
|
|
Packit |
c32a2d |
addps %xmm2, %xmm0
|
|
Packit |
c32a2d |
addps %xmm3, %xmm1
|
|
Packit |
c32a2d |
addps %xmm1, %xmm6
|
|
Packit |
c32a2d |
movaps %xmm6, %xmm7
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm6
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 256(WINDOW), WINDOW
|
|
Packit |
c32a2d |
leal -128(B0), B0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm0
|
|
Packit |
c32a2d |
movaps %xmm6, %xmm1
|
|
Packit |
c32a2d |
unpcklps %xmm5, %xmm4
|
|
Packit |
c32a2d |
unpcklps %xmm7, %xmm6
|
|
Packit |
c32a2d |
unpckhps %xmm5, %xmm0
|
|
Packit |
c32a2d |
unpckhps %xmm7, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm4, %xmm2
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm3
|
|
Packit |
c32a2d |
movlhps %xmm6, %xmm4
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm6
|
|
Packit |
c32a2d |
movlhps %xmm1, %xmm0
|
|
Packit |
c32a2d |
movhlps %xmm3, %xmm1
|
|
Packit |
c32a2d |
addps %xmm6, %xmm4
|
|
Packit |
c32a2d |
addps %xmm1, %xmm0
|
|
Packit |
c32a2d |
addps %xmm4, %xmm0
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm1
|
|
Packit |
c32a2d |
movaps %xmm0, %xmm2
|
|
Packit |
c32a2d |
mulps LOCAL_VAR(scale_s32), %xmm0
|
|
Packit |
c32a2d |
cmpnleps LOCAL_VAR(maxmin_s32), %xmm1
|
|
Packit |
c32a2d |
cmpltps 16+LOCAL_VAR(maxmin_s32), %xmm2
|
|
Packit |
c32a2d |
cvtps2pi %xmm0, %mm0
|
|
Packit |
c32a2d |
movhlps %xmm0, %xmm0
|
|
Packit |
c32a2d |
cvtps2pi %xmm0, %mm1
|
|
Packit |
c32a2d |
cvtps2pi %xmm1, %mm2
|
|
Packit |
c32a2d |
movhlps %xmm1, %xmm1
|
|
Packit |
c32a2d |
cvtps2pi %xmm1, %mm3
|
|
Packit |
c32a2d |
psrad $31, %mm2
|
|
Packit |
c32a2d |
psrad $31, %mm3
|
|
Packit |
c32a2d |
pxor %mm2, %mm0
|
|
Packit |
c32a2d |
pxor %mm3, %mm1
|
|
Packit |
c32a2d |
movd %mm0, (SAMPLES)
|
|
Packit |
c32a2d |
psrlq $32, %mm0
|
|
Packit |
c32a2d |
movd %mm0, 8(SAMPLES)
|
|
Packit |
c32a2d |
movd %mm1, 16(SAMPLES)
|
|
Packit |
c32a2d |
psrlq $32, %mm1
|
|
Packit |
c32a2d |
movd %mm1, 24(SAMPLES)
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
cvtps2pi %xmm2, %mm0
|
|
Packit |
c32a2d |
movhlps %xmm2, %xmm2
|
|
Packit |
c32a2d |
cvtps2pi %xmm2, %mm1
|
|
Packit |
c32a2d |
packssdw %mm3, %mm2
|
|
Packit |
c32a2d |
packssdw %mm1, %mm0
|
|
Packit |
c32a2d |
psrlw $15, %mm2
|
|
Packit |
c32a2d |
psrlw $15, %mm0
|
|
Packit |
c32a2d |
paddw %mm2, %mm0
|
|
Packit |
c32a2d |
paddw %mm0, MMREG_CLIP
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
leal 32(SAMPLES), SAMPLES
|
|
Packit |
c32a2d |
decl %ecx
|
|
Packit |
c32a2d |
jnz 1b
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
pshufw $0xee, MMREG_CLIP, %mm0
|
|
Packit |
c32a2d |
paddw MMREG_CLIP, %mm0
|
|
Packit |
c32a2d |
pshufw $0x55, %mm0, %mm1
|
|
Packit |
c32a2d |
paddw %mm1, %mm0
|
|
Packit |
c32a2d |
movd %mm0, %eax
|
|
Packit |
c32a2d |
andl $0xffff, %eax
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
popl %esi
|
|
Packit |
c32a2d |
popl %ebx
|
|
Packit |
c32a2d |
movl %ebp, %esp
|
|
Packit |
c32a2d |
popl %ebp
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
emms
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
ret
|
|
Packit |
c32a2d |
|
|
Packit |
c32a2d |
NONEXEC_STACK
|