Vector Member-of search up to an offset of 4

This commit is contained in:
Marshall Lochbaum 2023-11-05 15:04:30 -05:00
parent af81da50e4
commit ec31543604
2 changed files with 38 additions and 1 deletions

View File

@ -371,6 +371,7 @@ fn hashtab{T, name}(rp:*rty{name}, ip:*T, m:usz, fp:*T, n:usz, links:ity{name})
def setm{found, i} = has_maxh |= found
end := maxh>>sh # Clip trailing maxh if it shouldn't be in the table
def usem{} = {
try_vec_memb{T, hash, sz+ext, sh, maxh, has_maxh, rp, fp, n, abort}
if (has_maxh) { while (load{hash,end}!=maxh) ++end }
else end += promote{T,ext} # Unreachable
}
@ -496,6 +497,43 @@ fn hashtab{T, name}(rp:*rty{name}, ip:*T, m:usz, fp:*T, n:usz, links:ity{name})
i == m # Whether it finished
}
def try_vec_memb{} = {}
def try_vec_memb{T, hash, hashlen, sh, maxh, has_maxh, rp, fp, n, done
& hasarch{'SSE4.2'} & T==u32} = {
# Hash h wants bin h>>sh, so the offset for h in slot i is (in infite-precision ints)
# i-h>>sh = i+((1<<sh-1)-h)>>sh = (((i+1)<<sh-1)-h)>>sh
# We maintain io = (i+1)<<sh-1
def vl = 4; def V = [vl]T
nh := hashlen/vl; o := hashlen-vl*nh
io := make{V, each{{k} => (o+k)<<sh - 1, 1+iota{vl}}}
id := V**(T~~vl<<sh)
mv := V**0
@for (h in *V~~(hash+o) over nh) { mv=max{mv, io-min{h,io}}; io+=id }
max_off := vfold{max, mv} >> sh
def memb{test} = {
def R = i8; def rw = width{R}; def u = width{T}/rw
l := n/u
@for (r in *T~~rp over i to l) {
c := V**0 # Will combine u results to avoid folding too much
@unroll (f in fp+u*i over a to u) c |= V**(1<<(rw*a)) & test{f}
r = vfold{|, c}
}
@for (rp, fp over _ from u*l to n) rp = promote{R, homAny{test{fp}}}
goto{done}
}
if (max_off < vl) {
hp := *V~~(hash + maxh>>sh)
hv := load{hp}
store{hp, 0, hv &~ (hv == V**(maxh &- ~has_maxh))}
def test{x} = {
h := hash_val{x}; j := h>>sh
V**h == load{*V~~(hash+j), 0}
}
memb{test}
}
}
def exp_hash{T, name} = {
export{merge{name,'_c2_hash',fmtnat{width{T}}}, hashtab{T, name}}
}

View File

@ -1,5 +1,4 @@
# Fold associative/commutative operation across a register
# Used by squeeze.singeli, count.singeli
def vfold{F, x:T & w128{T} & hasarch{'X86_64'}} = {
c:= x