Avoid using popcount primitive in Singeli if POPCNT isn't present
This commit is contained in:
parent
6df283c379
commit
93e1262864
@ -35,12 +35,19 @@ def loadu{p:T & *u64==T} = emit{eltype{T}, 'loadu_u64', p}
|
||||
def popcRand{x:T & isint{T} & width{T}==64} = emit{u8, 'rand_popc64', x} # under valgrind, return a random result in the range of possible ones
|
||||
def popcRand{x:T & isint{T} & width{T}<=32} = emit{u8, 'rand_popc64', x}
|
||||
|
||||
def maketab{l,w} = { # Table from l bits to w-bit indices
|
||||
def lw = l*w
|
||||
fold{{t,k} => join{each{tup,t,k+(t<<w)%(1<<lw)}}, tup{base{1<<w,l**l}}, reverse{iota{l}}}
|
||||
# Table from l bits to w-bit indices
|
||||
def maketab{l,w} = {
|
||||
def bot = fold{
|
||||
{t,k} => join{each{tup, t, k + t<<w}},
|
||||
tup{0},
|
||||
reverse{iota{l}}
|
||||
}
|
||||
# Store popcnt-1 in the high element
|
||||
def top = (fold{bind{flat_table,+}, l**iota{2}} - 1)%(1<<w)
|
||||
top<<(l*w-w) | bot # Overlaps for all-1 value only
|
||||
}
|
||||
# 16-element tables
|
||||
i64tab :*u32 = maketab{4,8}*2
|
||||
i64tab :*u32 = (maketab{4,8}*2)%(1<<32)
|
||||
tab_4_16:*u64 = maketab{4,16}
|
||||
if (1) {
|
||||
def use_table = 1
|
||||
@ -48,6 +55,10 @@ if (1) {
|
||||
} else {
|
||||
def use_table = 0
|
||||
}
|
||||
# Recover popcount, for when POPCNT isn't there
|
||||
def has_popc = hasarch{'POPCNT'}
|
||||
def tab_popc{i, w} = (i>>(64-w) + 1) & (1<<w - 1)
|
||||
def popc_alt{v, i, w} = if (has_popc) popc{v} else tab_popc{i, w}
|
||||
|
||||
def arg{c,T} = if (c) *T else if (T==i32) T else tup{}
|
||||
|
||||
@ -207,7 +218,6 @@ fn slash{c, T==i8 & hasarch{'AVX2'}}(w:*u64, x:arg{c,T}, r:*T, l:u64, sum:u64) :
|
||||
}
|
||||
}
|
||||
|
||||
# TODO avoid calling popcnt if it's not a single instruction
|
||||
def thresh{c==0, T==i8 & use_table} = 32
|
||||
def thresh{c==0, T==i16} = 16
|
||||
fn slash{c==0, T & (if (T==i8) use_table else T==i16)}(w:*u64, x:arg{c,T}, r:*T, l:u64, sum:u64) : void = {
|
||||
@ -217,17 +227,16 @@ fn slash{c==0, T & (if (T==i8) use_table else T==i16)}(w:*u64, x:arg{c,T}, r:*T,
|
||||
j:u64 = 0
|
||||
def inc = base{1<<tw, n**n}
|
||||
@for_special_buffered{r,8} (w in *u8~~w over sum) {
|
||||
pc:= popc{w}
|
||||
def step{r, w} = storeu{*u64~~r, 0, j + load{tab, w}}
|
||||
if (tw==8) {
|
||||
step{r,w}
|
||||
} else {
|
||||
h := w&0xf
|
||||
step{r, h}; j += inc
|
||||
step{r+popcRand{h}, w>>4}
|
||||
def rn = if (has_popc) r+popc{w} else 0
|
||||
def step{w} = {
|
||||
i := load{tab, w}
|
||||
storeu{*u64~~r, 0, j + i}
|
||||
r += popc_alt{w, i, tw}
|
||||
j += inc
|
||||
}
|
||||
j += inc
|
||||
r += pc
|
||||
if (tw==8) { step{w} }
|
||||
else { step{w&0xf}; step{w>>4} }
|
||||
if (has_popc) r = rn # Shorter dependency chain
|
||||
}
|
||||
}
|
||||
|
||||
@ -237,9 +246,9 @@ fn slash{c==1, T & T<=i16 & hasarch{'SSSE3'} & use_table}(wp:*u64, x:arg{c,T}, r
|
||||
def tw = width{T}
|
||||
def V = [16]i8
|
||||
@for_special_buffered{r,8} (w in *u8~~wp over i to sum) {
|
||||
pc := popc{w}
|
||||
ind := load{itab, w}; def I = type{ind}
|
||||
s := V~~make{[2]I, ind,0}
|
||||
ind := load{itab, w}
|
||||
pc := popc_alt{w, ind, 8}
|
||||
s := V~~make{[2]u64, ind,0}
|
||||
if (T==i16) { s+=s; s = V~~unpackLo{s, s+V**1} }
|
||||
res := sel{V, load{*V~~(x+8*i)}, s}
|
||||
if (T==i8) store{*u64~~r, 0, extract{[2]u64~~res, 0}}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user