/*

	synth_stereo_avx: AVX optimized synth for x86-64 (stereo specific version)

	copyright 1995-2013 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 IS_MSABI

/* 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

/*

	int synth_1to1_s_avx_asm(short *window, short *b0l, short *b0r, short *samples, int bo1);

	return value: number of clipped samples

*/

	ALIGN16

.globl ASM_NAME(synth_1to1_s_avx_asm)

ASM_NAME(synth_1to1_s_avx_asm):

#ifdef IS_MSABI /* should save xmm6-15 */

	push		%rbp

	mov			%rsp, %rbp

	sub			$144, %rsp

	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)

	movl		48(%rbp), %eax /* 5th argument; placed after 32-byte shadow space */

#endif

#ifdef IS_MSABI

	shl			$1, %eax

	mov			%rcx, WINDOW

#else

	mov			%r8d, %eax

	shl			$1, %eax

	movq		%rcx, SAMPLES

#endif

	add			$32, WINDOW

	sub			%rax, WINDOW

	mov			$64, %rax

	movl		$4, %ecx

	vpxor		%xmm14, %xmm14, %xmm14

	ALIGN16

1:

	movups		(WINDOW), %xmm8

	movups		16(WINDOW), %xmm9

	movups		(WINDOW,%rax), %xmm10

	movups		16(WINDOW,%rax), %xmm11

	vpmaddwd	(B0L), %xmm8, %xmm0

	vpmaddwd	16(B0L), %xmm9, %xmm1

	vpmaddwd	(B0R), %xmm8, %xmm2

	vpmaddwd	16(B0R), %xmm9, %xmm3

	vpmaddwd	32(B0L), %xmm10, %xmm4

	vpmaddwd	48(B0L), %xmm11, %xmm5

	vpmaddwd	32(B0R), %xmm10, %xmm6

	vpmaddwd	48(B0R), %xmm11, %xmm7

	vpaddd		%xmm1, %xmm0, %xmm8

	vpaddd		%xmm3, %xmm2, %xmm0

	vpaddd		%xmm5, %xmm4, %xmm9

	vpaddd		%xmm7, %xmm6, %xmm1

	lea			(WINDOW,%rax,2), WINDOW

	add			%rax, B0L

	add			%rax, B0R

	movups		(WINDOW), %xmm10

	movups		16(WINDOW), %xmm11

	movups		(WINDOW,%rax), %xmm12

	movups		16(WINDOW,%rax), %xmm13

	vpmaddwd	(B0L), %xmm10, %xmm2

	vpmaddwd	16(B0L), %xmm11, %xmm3

	vpmaddwd	(B0R), %xmm10, %xmm4

	vpmaddwd	16(B0R), %xmm11, %xmm5

	vpmaddwd	32(B0L), %xmm12, %xmm6

	vpmaddwd	48(B0L), %xmm13, %xmm10

	vpmaddwd	32(B0R), %xmm12, %xmm7

	vpmaddwd	48(B0R), %xmm13, %xmm11

	vpaddd		%xmm3, %xmm2, %xmm2

	vpaddd		%xmm5, %xmm4, %xmm3

	vpaddd		%xmm6, %xmm10, %xmm4

	vpaddd		%xmm7, %xmm11, %xmm5

	lea			(WINDOW,%rax,2), WINDOW

	add			%rax, B0L

	add			%rax, B0R

	vpunpckldq	%xmm0, %xmm8, %xmm6

	vpunpckhdq	%xmm0, %xmm8, %xmm0

	vpunpckldq	%xmm1, %xmm9, %xmm7

	vpunpckhdq	%xmm1, %xmm9, %xmm1

	vpaddd		%xmm6, %xmm0, %xmm0

	vpaddd		%xmm7, %xmm1, %xmm1

	vpunpckldq	%xmm3, %xmm2, %xmm6

	vpunpckhdq	%xmm3, %xmm2, %xmm2

	vpunpckldq	%xmm5, %xmm4, %xmm7

	vpunpckhdq	%xmm5, %xmm4, %xmm3

	vpaddd		%xmm6, %xmm2, %xmm2

	vpaddd		%xmm7, %xmm3, %xmm3

	vpunpcklqdq	%xmm1, %xmm0, %xmm4

	vpunpckhqdq	%xmm1, %xmm0, %xmm0

	vpunpcklqdq	%xmm3, %xmm2, %xmm5

	vpunpckhqdq	%xmm3, %xmm2, %xmm1

	vpaddd		%xmm0, %xmm4, %xmm0

	vpaddd		%xmm1, %xmm5, %xmm1

	vpsrad		$13, %xmm0, %xmm0

	vpsrad		$13, %xmm1, %xmm1

	vpackssdw	%xmm1, %xmm0, %xmm2

	vpcmpeqd	%xmm3, %xmm3, %xmm3

	vpslld		$16, %xmm0, %xmm0

	vpslld		$16, %xmm1, %xmm1

	vpsrld		$16, %xmm0, %xmm0

	vpsrld		$16, %xmm1, %xmm1

	vpackusdw	%xmm1, %xmm0, %xmm0

	vpcmpeqw	%xmm2, %xmm0, %xmm0

	vpxor		%xmm3, %xmm0, %xmm0

	vpaddw		%xmm0, %xmm14, %xmm14

	movups		%xmm2, (SAMPLES)

	add			$16, SAMPLES

	dec			%ecx

	jnz			1b

	movl		$4, %ecx

	ALIGN16

1:

	movups		(WINDOW), %xmm8

	movups		16(WINDOW), %xmm9

	movups		(WINDOW,%rax), %xmm10

	movups		16(WINDOW,%rax), %xmm11

	vpmaddwd	(B0L), %xmm8, %xmm0

	vpmaddwd	16(B0L), %xmm9, %xmm1

	vpmaddwd	(B0R), %xmm8, %xmm2

	vpmaddwd	16(B0R), %xmm9, %xmm3

	vpmaddwd	-32(B0L), %xmm10, %xmm4

	vpmaddwd	-16(B0L), %xmm11, %xmm5

	vpmaddwd	-32(B0R), %xmm10, %xmm6

	vpmaddwd	-16(B0R), %xmm11, %xmm7

	vpaddd		%xmm1, %xmm0, %xmm8

	vpaddd		%xmm3, %xmm2, %xmm0

	vpaddd		%xmm5, %xmm4, %xmm9

	vpaddd		%xmm7, %xmm6, %xmm1

	lea			(WINDOW,%rax,2), WINDOW

	sub			%rax, B0L

	sub			%rax, B0R

	movups		(WINDOW), %xmm10

	movups		16(WINDOW), %xmm11

	movups		(WINDOW,%rax), %xmm12

	movups		16(WINDOW,%rax), %xmm13

	vpmaddwd	(B0L), %xmm10, %xmm2

	vpmaddwd	16(B0L), %xmm11, %xmm3

	vpmaddwd	(B0R), %xmm10, %xmm4

	vpmaddwd	16(B0R), %xmm11, %xmm5

	vpmaddwd	-32(B0L), %xmm12, %xmm6

	vpmaddwd	-16(B0L), %xmm13, %xmm10

	vpmaddwd	-32(B0R), %xmm12, %xmm7

	vpmaddwd	-16(B0R), %xmm13, %xmm11

	vpaddd		%xmm3, %xmm2, %xmm2

	vpaddd		%xmm5, %xmm4, %xmm3

	vpaddd		%xmm6, %xmm10, %xmm4

	vpaddd		%xmm7, %xmm11, %xmm5

	lea			(WINDOW,%rax,2), WINDOW

	sub			%rax, B0L

	sub			%rax, B0R

	vpunpckldq	%xmm0, %xmm8, %xmm6

	vpunpckhdq	%xmm0, %xmm8, %xmm0

	vpunpckldq	%xmm1, %xmm9, %xmm7

	vpunpckhdq	%xmm1, %xmm9, %xmm1

	vpaddd		%xmm6, %xmm0, %xmm0

	vpaddd		%xmm7, %xmm1, %xmm1

	vpunpckldq	%xmm3, %xmm2, %xmm6

	vpunpckhdq	%xmm3, %xmm2, %xmm2

	vpunpckldq	%xmm5, %xmm4, %xmm7

	vpunpckhdq	%xmm5, %xmm4, %xmm3

	vpaddd		%xmm6, %xmm2, %xmm2

	vpaddd		%xmm7, %xmm3, %xmm3

	vpunpcklqdq	%xmm1, %xmm0, %xmm4

	vpunpckhqdq	%xmm1, %xmm0, %xmm0

	vpunpcklqdq	%xmm3, %xmm2, %xmm5

	vpunpckhqdq	%xmm3, %xmm2, %xmm1

	vpaddd		%xmm0, %xmm4, %xmm0

	vpaddd		%xmm1, %xmm5, %xmm1

	vpsrad		$13, %xmm0, %xmm0

	vpsrad		$13, %xmm1, %xmm1

	vpackssdw	%xmm1, %xmm0, %xmm2

	vpcmpeqd	%xmm3, %xmm3, %xmm3

	vpslld		$16, %xmm0, %xmm0

	vpslld		$16, %xmm1, %xmm1

	vpsrld		$16, %xmm0, %xmm0

	vpsrld		$16, %xmm1, %xmm1

	vpackusdw	%xmm1, %xmm0, %xmm0

	vpcmpeqw	%xmm2, %xmm0, %xmm0

	vpxor		%xmm3, %xmm0, %xmm0

	vpaddw		%xmm0, %xmm14, %xmm14

	movups		%xmm2, (SAMPLES)

	add			$16, SAMPLES

	dec			%ecx

	jnz			1b

	pxor		%xmm1, %xmm1

	psubw		%xmm14, %xmm1

	pshufd		$0x4e, %xmm1, %xmm0

	paddw		%xmm1, %xmm0

	pshuflw		$0x4e, %xmm0, %xmm1

	paddw		%xmm1, %xmm0

	pshuflw		$0x11, %xmm0, %xmm1

	paddw		%xmm1, %xmm0

	movd		%xmm0, %eax

	and			$0x7f, %eax

#ifdef IS_MSABI

	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

	mov			%rbp, %rsp

	pop			%rbp

#endif

	ret

NONEXEC_STACK