#

# Copyright 2002,2005 Free Software Foundation, Inc.

# 

# This file is part of GNU Radio

# 

# GNU Radio is free software; you can redistribute it and/or modify

# it under the terms of the GNU General Public License as published by

# the Free Software Foundation; either version 3, or (at your option)

# any later version.

# 

# GNU Radio is distributed in the hope that it will be useful,

# but WITHOUT ANY WARRANTY; without even the implied warranty of

# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the

# GNU General Public License for more details.

# 

# You should have received a copy of the GNU General Public License

# along with GNU Radio; see the file COPYING.  If not, write to

# the Free Software Foundation, Inc., 51 Franklin Street,

# Boston, MA 02110-1301, USA.

# 

# input and taps are guarenteed to be 16 byte aligned.

# n_2_ccomplex_blocks is != 0

#	

#

#  ccomplex_dotprod_generic (const float *input,

#                         const float *taps, unsigned n_2_ccomplex_blocks, float *result)

#  {

#    float sum0 = 0;

#    float sum1 = 0;

#    float sum2 = 0;

#    float sum3 = 0;

#  

#    do {

#  

#      sum0 += input[0] * taps[0] - input[1] * taps[1];

#      sum1 += input[0] * taps[1] + input[1] * taps[0];

#      sum2 += input[2] * taps[2] - input[3] * taps[3];

#      sum3 += input[2] * taps[3] + input[3] * taps[2];

#  

#      input += 4;

#      taps += 4;

#  

#    } while (--n_2_ccomplex_blocks != 0);

#  

#  

#    result[0] = sum0 + sum2;

#    result[1] = sum1 + sum3;

#  }

#  		

# TODO: prefetch and better scheduling

#include "assembly.h"

	.file	"ccomplex_dotprod_3dnowext64.S"

	.version	"01.01"

.text

	.p2align 4

.globl GLOB_SYMB(ccomplex_dotprod_3dnowext)

	DEF_FUNC_HEAD(ccomplex_dotprod_3dnowext)

GLOB_SYMB(ccomplex_dotprod_3dnowext):

	# intput: rdi, taps: rsi, n_2_ccomplex_blocks: rdx, result: rcx

	mov	%rdx, %rax

	# zero accumulators

	pxor	%mm6, %mm6		# mm6 = 0 0 

	pxor	%mm7, %mm7		# mm7 = 0 0

	movq	0(%rdi), %mm0

	movq	0(%rsi), %mm2

	shr	$1, %rax		# rax = n_2_ccomplex_blocks / 2

	movq	8(%rdi), %mm1

	movq	8(%rsi), %mm3

	jmp	.L1_test

	#

	# 4 taps / loop

	# something like ?? cycles / loop

	#

	.p2align 4

.Loop1:	

# complex prod: C += A * B,  w/ temp Z

#

#	movq	0(%rdi), %mmA

#	movq	0(%rsi), %mmB

#	pswapd	%mmA, %mmZ

#	pfmul	%mmB, %mmA

#	pfmul	%mmZ, %mmB

#	pfpnacc	%mmB, %mmA

#	pfadd	%mmA, %mmC

# A=mm0, B=mm2, Z=mm4

# A'=mm1, B'=mm3, Z'=mm5

	pswapd	%mm0, %mm4

	pfmul	%mm2, %mm0

	pswapd	%mm1, %mm5

	pfmul	%mm4, %mm2

	pfmul	%mm3, %mm1

	pfpnacc	%mm2, %mm0

	pfmul	%mm5, %mm3

	movq	16(%rsi), %mm2

	pfpnacc	%mm3, %mm1

	movq	24(%rsi), %mm3

	pfadd	%mm0, %mm6

	movq	16(%rdi), %mm0

	pfadd	%mm1, %mm7

	movq	24(%rdi), %mm1

# unroll

	pswapd	%mm0, %mm4

	pfmul	%mm2, %mm0

	pswapd	%mm1, %mm5

	pfmul	%mm4, %mm2

	pfmul	%mm3, %mm1

	pfpnacc	%mm2, %mm0

	pfmul	%mm5, %mm3

	movq	32(%rsi), %mm2

	pfpnacc	%mm3, %mm1

	movq	40(%rsi), %mm3

	pfadd	%mm0, %mm6

	movq	32(%rdi), %mm0

	pfadd	%mm1, %mm7

	movq	40(%rdi), %mm1

	add	$32, %rsi

	add	$32, %rdi

.L1_test:

	dec	%rax

	jge	.Loop1

	# We've handled the bulk of multiplies up to here.

	# Let's see if original n_2_ccomplex_blocks was odd.

	# If so, we've got 2 more taps to do.

	and	$1, %rdx

	je	.Leven

	# The count was odd, do 2 more taps.

	# Note that we've already got mm0/mm2 & mm1/mm3 preloaded

	# from the main loop.

# A=mm0, B=mm2, Z=mm4

# A'=mm1, B'=mm3, Z'=mm5

	pswapd	%mm0, %mm4

	pfmul	%mm2, %mm0

	pswapd	%mm1, %mm5

	pfmul	%mm4, %mm2

	pfmul	%mm3, %mm1

	pfpnacc	%mm2, %mm0

	pfmul	%mm5, %mm3

	pfpnacc	%mm3, %mm1

	pfadd	%mm0, %mm6

	pfadd	%mm1, %mm7

.Leven:

	# at this point mm6 and mm7 contain partial sums

	pfadd	%mm7, %mm6

	movq	%mm6, (%rcx)	 	# result

	femms

	retq

FUNC_TAIL(ccomplex_dotprod_3dnowext)

	.ident	"Hand coded x86_64 3DNow!Ext assembly"

#if defined(__linux__) && defined(__ELF__)

.section .note.GNU-stack,"",%progbits

#endif