/* Intel SIMD (SSE2) implementations of Viterbi ACS butterflies for 64-state (k=7) convolutional code Copyright 2003 Phil Karn, KA9Q This code may be used under the terms of the GNU Lesser General Public License (LGPL) Modifications for x86_64, 2012 Matthias P. Braendli, HB9EGM: - changed registers to x86-64 equivalents - changed instructions accordingly - %rip indirect addressing needed for position independent code, which is required because x86-64 needs dynamic libs to be PIC void update_viterbi27_blk_sse2(struct v27 *vp,unsigned char syms[],int nbits) ; */ # SSE2 (128-bit integer SIMD) version # All X86-64 CPUs include SSE2 # These are offsets into struct v27, defined in viterbi27_av.c .set DP,128 .set OLDMETRICS,132 .set NEWMETRICS,136 .text .global update_viterbi27_blk_sse2,Branchtab27_sse2 .type update_viterbi27_blk_sse2,@function .align 16 update_viterbi27_blk_sse2: pushq %rbp movq %rsp,%rbp /* convention different between i386 and x86_64: rsi and rdi belong to called function, not caller */ /* Let's say we don't care (yet) */ pushq %rsi pushq %rdi pushq %rdx pushq %rbx movq 8(%rbp),%rdx # edx = vp testq %rdx,%rdx jnz 0f movq -1,%rax jmp err 0: movq OLDMETRICS(%rdx),%rsi # esi -> old metrics movq NEWMETRICS(%rdx),%rdi # edi -> new metrics movq DP(%rdx),%rdx # edx -> decisions 1: movq 16(%rbp),%rax # eax = nbits decq %rax jl 2f # passed zero, we're done movq %rax,16(%rbp) xorq %rax,%rax movq 12(%rbp),%rbx # ebx = syms movb (%rbx),%al movd %rax,%xmm6 # xmm6[0] = first symbol movb 1(%rbx),%al movd %rax,%xmm5 # xmm5[0] = second symbol addq $2,%rbx movq %rbx,12(%rbp) punpcklbw %xmm6,%xmm6 # xmm6[1] = xmm6[0] punpcklbw %xmm5,%xmm5 pshuflw $0,%xmm6,%xmm6 # copy low word to low 3 pshuflw $0,%xmm5,%xmm5 punpcklqdq %xmm6,%xmm6 # propagate to all 16 punpcklqdq %xmm5,%xmm5 # xmm6 now contains first symbol in each byte, xmm5 the second movdqa thirtyones(%rip),%xmm7 # each invocation of this macro does 16 butterflies in parallel .MACRO butterfly GROUP # compute branch metrics movdqa (Branchtab27_sse2+(16*\GROUP))(%rip),%xmm4 movdqa (Branchtab27_sse2+32+(16*\GROUP))(%rip),%xmm3 pxor %xmm6,%xmm4 pxor %xmm5,%xmm3 # compute 5-bit branch metric in xmm4 by adding the individual symbol metrics # This is okay for this # code because the worst-case metric spread (at high Eb/No) is only 120, # well within the range of our unsigned 8-bit path metrics, and even within # the range of signed 8-bit path metrics pavgb %xmm3,%xmm4 psrlw $3,%xmm4 pand %xmm7,%xmm4 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 # note use of saturating arithmetic paddusb %xmm4,%xmm3 # this shouldn't be necessary, but why not? # negate branch metrics 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)(%rdi) movdqa %xmm4,(32*\GROUP)(%rdi) # 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 shlq $16,%rbx .byte 0x66,0x0f,0xd7,0xc4 # pmovmskb %xmm4,%eax orq %rax,%rbx movq %rbx,(4*\GROUP)(%rdx) .endm # invoke macro 2 times for a total of 32 butterflies butterfly GROUP=0 butterfly GROUP=1 addq $8,%rdx # bump decision pointer # See if we have to normalize. This requires an explanation. We don't want # our path metrics to exceed 255 on the *next* iteration. Since the # largest branch metric is 30, that means we don't want any to exceed 225 # on *this* iteration. Rather than look them all, we just pick an arbitrary one # (the first) and see if it exceeds 225-120=105, where 120 is the experimentally- # determined worst-case metric spread for this code and branch metrics in the range 0-30. # This is extremely conservative, and empirical testing at a variety of Eb/Nos might # show that a higher threshold could be used without affecting BER performance movq (%rdi),%rax # extract first output metric andq $255,%rax cmp $105,%rax jle done # No, no need to normalize # Normalize by finding smallest metric and subtracting it # from all metrics. We can't just pick an arbitrary small constant because # the minimum metric might be zero! movdqa (%rdi),%xmm0 movdqa %xmm0,%xmm4 movdqa 16(%rdi),%xmm1 pminub %xmm1,%xmm4 movdqa 32(%rdi),%xmm2 pminub %xmm2,%xmm4 movdqa 48(%rdi),%xmm3 pminub %xmm3,%xmm4 # crunch down to single lowest metric movdqa %xmm4,%xmm5 psrldq $8,%xmm5 # the count to psrldq is bytes, not bits! pminub %xmm5,%xmm4 movdqa %xmm4,%xmm5 psrlq $32,%xmm5 pminub %xmm5,%xmm4 movdqa %xmm4,%xmm5 psrlq $16,%xmm5 pminub %xmm5,%xmm4 movdqa %xmm4,%xmm5 psrlq $8,%xmm5 pminub %xmm5,%xmm4 # now in lowest byte of %xmm4 punpcklbw %xmm4,%xmm4 # lowest 2 bytes pshuflw $0,%xmm4,%xmm4 # lowest 8 bytes punpcklqdq %xmm4,%xmm4 # all 16 bytes # xmm4 now contains lowest metric in all 16 bytes # subtract it from every output metric psubusb %xmm4,%xmm0 psubusb %xmm4,%xmm1 psubusb %xmm4,%xmm2 psubusb %xmm4,%xmm3 movdqa %xmm0,(%rdi) movdqa %xmm1,16(%rdi) movdqa %xmm2,32(%rdi) movdqa %xmm3,48(%rdi) done: # swap metrics movq %rsi,%rax movq %rdi,%rsi movq %rax,%rdi jmp 1b 2: movq 8(%rbp),%rbx # ebx = vp # stash metric pointers movq %rsi,OLDMETRICS(%rbx) movq %rdi,NEWMETRICS(%rbx) movq %rdx,DP(%rbx) # stash incremented value of vp->dp xorq %rax,%rax err: popq %rbx popq %rdx popq %rdi popq %rsi popq %rbp ret .data .align 16 thirtyones: .byte 31,31,31,31,31,31,31,31,31,31,31,31,31,31,31,31