/* Intel SIMD (SSE2) implementations of Viterbi ACS butterflies
   for 64-state (k=7) convolutional code
   Copyright 2001 Phil Karn, KA9Q
   This code may be used under the terms of the GNU Public License

   int update_viterbi27(struct v27 *vp,unsigned char sym1,unsigned char sym2); 
*/
	# SSE2 (128-bit integer SIMD) version
	# Requires Pentium 4 or better

	# These are offsets into struct v27, defined in viterbi27.h
	.set DP,128
	.set OLDMETRICS,132
	.set NEWMETRICS,136
	.text	
	.global update_viterbi27,Branchtab27_1,Branchtab27_2
	.type update_viterbi27,@function
	.align 16
	
update_viterbi27:
	pushl %ebp
	movl %esp,%ebp
	pushl %esi
	pushl %edi
	pushl %edx
	pushl %ebx
	movl 8(%ebp),%edx	# edx = vp
	movl OLDMETRICS(%edx),%esi	# esi -> old metrics
	movl NEWMETRICS(%edx),%edi	# edi -> new metrics
	movl DP(%edx),%edx	# edx -> decisions

	movd 12(%ebp),%xmm6	# xmm6[0] = first symbol
	movd 16(%ebp),%xmm5	# xmm5[0] = second symbol
	punpcklbw %xmm6,%xmm6	# xmm6[1] = xmm6[0]
	punpcklbw %xmm5,%xmm5
	movdqa fifteens,%xmm7
	pshuflw $0,%xmm6,%xmm6	# copy low word to low 3
	pshuflw $0,%xmm5,%xmm5
	punpcklqdq %xmm6,%xmm6  # propagate to all 16
	punpcklqdq %xmm5,%xmm5
	pand %xmm7,%xmm6
	pand %xmm7,%xmm5
	# xmm6 now contains first symbol in each byte, xmm5 the second

	# each invocation of this macro does 16 butterflies in parallel
	.MACRO butterfly GROUP
	# compute branch metrics
	movdqa Branchtab27_1+(16*\GROUP),%xmm4
	movdqa Branchtab27_2+(16*\GROUP),%xmm3
	pxor %xmm6,%xmm4
	pxor %xmm5,%xmm3
	pavgb %xmm3,%xmm4	# xmm4 contains branch metrics

	movdqa (16*\GROUP)(%esi),%xmm0	# Incoming path metric, high bit = 0
	movdqa ((16*\GROUP)+32)(%esi),%xmm3	# Incoming path metric, high bit = 1
	movdqa %xmm0,%xmm2
	movdqa %xmm3,%xmm1
	paddusb %xmm4,%xmm0
	paddusb %xmm4,%xmm3
	
	# invert branch metrics. This works only because they're 4 bits
	pxor %xmm7,%xmm4
	
	paddusb %xmm4,%xmm1
	paddusb %xmm4,%xmm2
	
	# Find survivors, leave in mm0,2
	pminub %xmm1,%xmm0
	pminub %xmm3,%xmm2
	# get decisions, leave in mm1,3
	pcmpeqb %xmm0,%xmm1
	pcmpeqb %xmm2,%xmm3
	
	# interleave and store new branch metrics in mm0,2
	movdqa %xmm0,%xmm4
	punpckhbw %xmm2,%xmm0	# interleave second 16 new metrics
	punpcklbw %xmm2,%xmm4	# interleave first 16 new metrics
	movdqa %xmm0,(32*\GROUP+16)(%edi)
	movdqa %xmm4,(32*\GROUP)(%edi)

	# interleave decisions & store
	movdqa %xmm1,%xmm4
	punpckhbw %xmm3,%xmm1
	punpcklbw %xmm3,%xmm4
	# work around bug in gas due to Intel doc error
	.byte 0x66,0x0f,0xd7,0xd9	# pmovmskb %xmm1,%ebx
	shll $16,%ebx
	.byte 0x66,0x0f,0xd7,0xc4	# pmovmskb %xmm4,%eax
	orl %eax,%ebx
	movl %ebx,(4*\GROUP)(%edx)
	.endm

	# invoke macro 2 times for a total of 32 butterflies
	butterfly GROUP=0
	butterfly GROUP=1

	addl $8,%edx		# bump decision pointer
		
	# see if we have to normalize
	movl (%edi),%eax	# extract first output metric
	orl $255,%eax
	cmp $150,%eax		# is it greater than 150?
	movl $0,%eax
	jge done		# No, no need to normalize

	# Normalize by finding smallest metric and subtracting it
	# from all metrics
	movdqa (%edi),%xmm0
	pminub 16(%edi),%xmm0
	pminub 32(%edi),%xmm0
	pminub 48(%edi),%xmm0

	# crunch down to single lowest metric
	movdqa %xmm0,%xmm1
	psrldq $8,%xmm0     # the count to psrldq is bytes, not bits!
	pminub %xmm1,%xmm0
	movdqa %xmm0,%xmm1
	psrlq $32,%xmm0
	pminub %xmm1,%xmm0
	movdqa %xmm0,%xmm1
	psrlq $16,%xmm0
	pminub %xmm1,%xmm0
	movdqa %xmm0,%xmm1
	psrlq $8,%xmm0
	pminub %xmm1,%xmm0

	punpcklbw %xmm0,%xmm0	# lowest 2 bytes
	pshuflw $0,%xmm0,%xmm0  # lowest 8 bytes
	punpcklqdq %xmm0,%xmm0	# all 16 bytes

	# xmm0 now contains lowest metric in all 16 bytes
	# subtract it from every output metric
	movdqa (%edi),%xmm1
	psubusb %xmm0,%xmm1
	movdqa %xmm1,(%edi)
	movdqa 16(%edi),%xmm1
	psubusb %xmm0,%xmm1
	movdqa %xmm1,16(%edi)	
	movdqa 32(%edi),%xmm1
	psubusb %xmm0,%xmm1	
	movdqa %xmm1,32(%edi)	
	movdqa 48(%edi),%xmm1
	psubusb %xmm0,%xmm1	
	movdqa %xmm1,48(%edi)	
	
	movd %xmm1,%eax
	and $0xff,%eax

done:	movl 8(%ebp),%ebx	# ebx = vp
	movl %edx,DP(%ebx)	# stash incremented value of vp->dp
	
	# swap metrics
	movl %edi,OLDMETRICS(%ebx)
	movl %esi,NEWMETRICS(%ebx)

	popl %ebx
	popl %edx
	popl %edi
	popl %esi
	popl %ebp
	ret

	.data
	.align 16
fifteens:	
	.byte 15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15