diff options
Diffstat (limited to 'src/libmpg123/synth_stereo_x86_64.S')
| -rw-r--r-- | src/libmpg123/synth_stereo_x86_64.S | 338 |
1 files changed, 338 insertions, 0 deletions
diff --git a/src/libmpg123/synth_stereo_x86_64.S b/src/libmpg123/synth_stereo_x86_64.S new file mode 100644 index 0000000..b858efa --- /dev/null +++ b/src/libmpg123/synth_stereo_x86_64.S @@ -0,0 +1,338 @@ +/* + synth_stereo_x86_64: SSE optimized synth for x86-64 (stereo specific version) + + copyright 1995-2009 by the mpg123 project - free software under the terms of the LGPL 2.1 + see COPYING and AUTHORS files in distribution or http://mpg123.org + initially written by Taihei Monma +*/ + +#include "mangle.h" + +#ifdef _WIN64 +/* short *window; */ +#define WINDOW %r10 +/* short *b0l; */ +#define B0L %rdx +/* short *b0r; */ +#define B0R %r8 +/* short *samples; */ +#define SAMPLES %r9 +#else +/* short *window; */ +#define WINDOW %rdi +/* short *b0l; */ +#define B0L %rsi +/* short *b0r; */ +#define B0R %rdx +/* short *samples; */ +#define SAMPLES %r9 +#endif + +#define XMMREG_CLIP %xmm15 +#define XMMREG_MAX %xmm14 /* {32767, 32767, 32767, 32767} */ +#define XMMREG_MIN %xmm13 /* {-32769, -32769, -32769, -32769} : not -32768 because SSE doesn't have "less than" comparison... */ +#define XMMREG_FULL %xmm12 /* {0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF} */ + +/* + int synth_1to1_stereo_x86_64_asm(short *window, short *b0l, short *b0r, short *samples, int bo1); + return value: number of clipped samples +*/ + +#ifndef __APPLE__ + .section .rodata +#else + .data +#endif + ALIGN32 +ASM_NAME(maxmin_x86_64): + .long 32767 + .long 32767 + .long 32767 + .long 32767 + .long -32769 + .long -32769 + .long -32769 + .long -32769 + .text + ALIGN16,,15 +.globl ASM_NAME(synth_1to1_stereo_x86_64_asm) +ASM_NAME(synth_1to1_stereo_x86_64_asm): +#ifdef _WIN64 /* should save xmm6-15 */ + movl 40(%rsp), %eax /* 5th argument; placed after 32-byte shadow space */ + subq $168, %rsp /* stack alignment + 10 xmm registers */ + movaps %xmm6, (%rsp) + movaps %xmm7, 16(%rsp) + movaps %xmm8, 32(%rsp) + movaps %xmm9, 48(%rsp) + movaps %xmm10, 64(%rsp) + movaps %xmm11, 80(%rsp) + movaps %xmm12, 96(%rsp) + movaps %xmm13, 112(%rsp) + movaps %xmm14, 128(%rsp) + movaps %xmm15, 144(%rsp) +#endif + +#ifdef _WIN64 + shlq $32, %rax + shrq $31, %rax + movq %rcx, %r10 +#else + movq %r8, %rax + shlq $32, %rax + shrq $31, %rax + movq %rcx, %r9 +#endif + leaq 32(WINDOW), WINDOW + subq %rax, WINDOW + + leaq ASM_NAME(maxmin_x86_64)(%rip), %rax + movaps (%rax), XMMREG_MAX + movaps 16(%rax), XMMREG_MIN + pxor XMMREG_CLIP, XMMREG_CLIP + pcmpeqd XMMREG_FULL, XMMREG_FULL + + movl $4, %ecx + + ALIGN16 +Loop_start_1: + movups (WINDOW), %xmm0 + movups 16(WINDOW), %xmm1 + movups 64(WINDOW), %xmm2 + movups 80(WINDOW), %xmm3 + movups 128(WINDOW), %xmm4 + movups 144(WINDOW), %xmm5 + movups 192(WINDOW), %xmm6 + movups 208(WINDOW), %xmm7 + movaps %xmm0, %xmm8 + movaps %xmm1, %xmm9 + movaps %xmm2, %xmm10 + movaps %xmm3, %xmm11 + pmaddwd (B0L), %xmm0 + pmaddwd 16(B0L), %xmm1 + pmaddwd 32(B0L), %xmm2 + pmaddwd 48(B0L), %xmm3 + pmaddwd (B0R), %xmm8 + pmaddwd 16(B0R), %xmm9 + pmaddwd 32(B0R), %xmm10 + pmaddwd 48(B0R), %xmm11 + paddd %xmm1, %xmm0 + paddd %xmm3, %xmm2 + paddd %xmm9, %xmm8 + paddd %xmm11, %xmm10 + movaps %xmm4, %xmm1 + movaps %xmm5, %xmm9 + movaps %xmm6, %xmm3 + movaps %xmm7, %xmm11 + pmaddwd 64(B0L), %xmm4 + pmaddwd 80(B0L), %xmm5 + pmaddwd 96(B0L), %xmm6 + pmaddwd 112(B0L), %xmm7 + pmaddwd 64(B0R), %xmm1 + pmaddwd 80(B0R), %xmm9 + pmaddwd 96(B0R), %xmm3 + pmaddwd 112(B0R), %xmm11 + paddd %xmm5, %xmm4 + paddd %xmm7, %xmm6 + paddd %xmm1, %xmm9 + paddd %xmm3, %xmm11 + + movaps %xmm0, %xmm1 + movaps %xmm4, %xmm3 + movaps %xmm8, %xmm5 + movaps %xmm9, %xmm7 + punpckldq %xmm2, %xmm0 + punpckldq %xmm6, %xmm4 + punpckhdq %xmm2, %xmm1 + punpckhdq %xmm6, %xmm3 + punpckldq %xmm10, %xmm8 + punpckldq %xmm11, %xmm9 + punpckhdq %xmm10, %xmm5 + punpckhdq %xmm11, %xmm7 + movaps %xmm0, %xmm2 + movaps %xmm1, %xmm6 + movaps %xmm8, %xmm10 + movaps %xmm5, %xmm11 + movlhps %xmm4, %xmm0 + movhlps %xmm2, %xmm4 + movlhps %xmm3, %xmm1 + movhlps %xmm6, %xmm3 + movlhps %xmm9, %xmm8 + movhlps %xmm10, %xmm9 + movlhps %xmm7, %xmm5 + movhlps %xmm11, %xmm7 + paddd %xmm4, %xmm0 + paddd %xmm3, %xmm1 + paddd %xmm9, %xmm8 + paddd %xmm7, %xmm5 + paddd %xmm1, %xmm0 + paddd %xmm5, %xmm8 + psrad $13, %xmm0 + psrad $13, %xmm8 + + movaps %xmm0, %xmm1 + movaps %xmm0, %xmm2 + movaps %xmm0, %xmm3 + movaps %xmm8, %xmm4 + punpckldq %xmm8, %xmm0 + punpckhdq %xmm8, %xmm1 + packssdw %xmm1, %xmm0 + movups %xmm0, (SAMPLES) + + pcmpgtd XMMREG_MAX, %xmm2 + pcmpgtd XMMREG_MIN, %xmm3 + pcmpgtd XMMREG_MAX, %xmm4 + pcmpgtd XMMREG_MIN, %xmm8 + packssdw %xmm4, %xmm2 + packssdw %xmm8, %xmm3 + pxor XMMREG_FULL, %xmm3 + psrlw $15, %xmm2 + psrlw $15, %xmm3 + paddw %xmm3, %xmm2 + paddw %xmm2, XMMREG_CLIP + + leaq 256(WINDOW), WINDOW + leaq 128(B0L), B0L + leaq 128(B0R), B0R + leaq 16(SAMPLES), SAMPLES + + decl %ecx + jnz Loop_start_1 + + movl $4, %ecx + + ALIGN16 +Loop_start_2: + movups (WINDOW), %xmm0 + movups 16(WINDOW), %xmm1 + movups 64(WINDOW), %xmm2 + movups 80(WINDOW), %xmm3 + movups 128(WINDOW), %xmm4 + movups 144(WINDOW), %xmm5 + movups 192(WINDOW), %xmm6 + movups 208(WINDOW), %xmm7 + movaps %xmm0, %xmm8 + movaps %xmm1, %xmm9 + movaps %xmm2, %xmm10 + movaps %xmm3, %xmm11 + pmaddwd (B0L), %xmm0 + pmaddwd 16(B0L), %xmm1 + pmaddwd -32(B0L), %xmm2 + pmaddwd -16(B0L), %xmm3 + pmaddwd (B0R), %xmm8 + pmaddwd 16(B0R), %xmm9 + pmaddwd -32(B0R), %xmm10 + pmaddwd -16(B0R), %xmm11 + paddd %xmm1, %xmm0 + paddd %xmm3, %xmm2 + paddd %xmm9, %xmm8 + paddd %xmm11, %xmm10 + movaps %xmm4, %xmm1 + movaps %xmm5, %xmm9 + movaps %xmm6, %xmm3 + movaps %xmm7, %xmm11 + pmaddwd -64(B0L), %xmm4 + pmaddwd -48(B0L), %xmm5 + pmaddwd -96(B0L), %xmm6 + pmaddwd -80(B0L), %xmm7 + pmaddwd -64(B0R), %xmm1 + pmaddwd -48(B0R), %xmm9 + pmaddwd -96(B0R), %xmm3 + pmaddwd -80(B0R), %xmm11 + paddd %xmm5, %xmm4 + paddd %xmm7, %xmm6 + paddd %xmm1, %xmm9 + paddd %xmm3, %xmm11 + + movaps %xmm0, %xmm1 + movaps %xmm4, %xmm3 + movaps %xmm8, %xmm5 + movaps %xmm9, %xmm7 + punpckldq %xmm2, %xmm0 + punpckldq %xmm6, %xmm4 + punpckhdq %xmm2, %xmm1 + punpckhdq %xmm6, %xmm3 + punpckldq %xmm10, %xmm8 + punpckldq %xmm11, %xmm9 + punpckhdq %xmm10, %xmm5 + punpckhdq %xmm11, %xmm7 + movaps %xmm0, %xmm2 + movaps %xmm1, %xmm6 + movaps %xmm8, %xmm10 + movaps %xmm5, %xmm11 + movlhps %xmm4, %xmm0 + movhlps %xmm2, %xmm4 + movlhps %xmm3, %xmm1 + movhlps %xmm6, %xmm3 + movlhps %xmm9, %xmm8 + movhlps %xmm10, %xmm9 + movlhps %xmm7, %xmm5 + movhlps %xmm11, %xmm7 + paddd %xmm4, %xmm0 + paddd %xmm3, %xmm1 + paddd %xmm9, %xmm8 + paddd %xmm7, %xmm5 + paddd %xmm1, %xmm0 + paddd %xmm5, %xmm8 + psrad $13, %xmm0 + psrad $13, %xmm8 + + movaps %xmm0, %xmm1 + movaps %xmm0, %xmm2 + movaps %xmm0, %xmm3 + movaps %xmm8, %xmm4 + punpckldq %xmm8, %xmm0 + punpckhdq %xmm8, %xmm1 + packssdw %xmm1, %xmm0 + movups %xmm0, (SAMPLES) + + pcmpgtd XMMREG_MAX, %xmm2 + pcmpgtd XMMREG_MIN, %xmm3 + pcmpgtd XMMREG_MAX, %xmm4 + pcmpgtd XMMREG_MIN, %xmm8 + packssdw %xmm4, %xmm2 + packssdw %xmm8, %xmm3 + pxor XMMREG_FULL, %xmm3 + psrlw $15, %xmm2 + psrlw $15, %xmm3 + paddw %xmm3, %xmm2 + paddw %xmm2, XMMREG_CLIP + + leaq 256(WINDOW), WINDOW + leaq -128(B0L), B0L + leaq -128(B0R), B0R + leaq 16(SAMPLES), SAMPLES + + decl %ecx + jnz Loop_start_2 + + movhlps XMMREG_CLIP, %xmm0 + paddw XMMREG_CLIP, %xmm0 + pshuflw $0x55, %xmm0, %xmm1 + pshuflw $0xaa, %xmm0, %xmm2 + pshuflw $0xff, %xmm0, %xmm3 + paddw %xmm1, %xmm0 + paddw %xmm2, %xmm0 + paddw %xmm3, %xmm0 + + movd %xmm0, %eax + andl $0xffff, %eax + +#ifdef _WIN64 + movaps (%rsp), %xmm6 + movaps 16(%rsp), %xmm7 + movaps 32(%rsp), %xmm8 + movaps 48(%rsp), %xmm9 + movaps 64(%rsp), %xmm10 + movaps 80(%rsp), %xmm11 + movaps 96(%rsp), %xmm12 + movaps 112(%rsp), %xmm13 + movaps 128(%rsp), %xmm14 + movaps 144(%rsp), %xmm15 + addq $168, %rsp +#endif + ret + +/* Mark non-executable stack. */ +#if defined(__linux__) && defined(__ELF__) +.section .note.GNU-stack,"",%progbits +#endif |