Sum counts in groups of 4 when possible

This commit is contained in:
Marshall Lochbaum 2023-03-02 10:42:06 -05:00
parent b655dd4771
commit e6a3e846d9

View File

@ -23,6 +23,8 @@ def fold{F, x:T & width{T}==128 & hasarch{'X86_64'}} = {
}
def fold{F, x:T & width{T}==256 & hasarch{'X86_64'}} = fold{F, F{half{x, 0}, half{x, 1}}}
fn sum_vec{T==[32]i8}(v:T) = fold{+, [16]i16~~fold{+, unpackQ{v, T**0}}}
def inc{ptr, ind, v} = store{ptr, ind, v + load{ptr, ind}}
def inc{ptr, ind} = inc{ptr, ind, 1}
@ -39,25 +41,30 @@ fn count{T}(tab:*usz, x:*ty_u{T}, n:u64) : u1 = {
b := r / vec
xv := *V~~x
used_eq:u1 = 0
if (r >= 256) {
if (r >= 128) {
b = block; if (r < vec*b) b = r / vec
mv := V**0
@for (xv over b) mv = max{mv, xv}
mi := iV~~mv
if (homAny{mi < iV**0}) return{1}
if (homAll{mi <= iV**32}) {
if (homAll{mi <= iV**48}) {
used_eq = 1
r = b * vec
m := fold{max, mv}
total := b*vec
@for (j to promote{u64,m}) {
c := V**0
e := V**trunc{uT, j}
@for (xv over b) c -= xv == e
s := fold{+, [vec/2]i16~~fold{+, unpackQ{iV~~c, iV**0}}}
total -= promote{u64, s}
inc{tab, j, promote{usz, s}}
m := promote{u64, fold{max, mv}}
total := trunc{usz, b*vec}
def count_each{j} = {
c := copy{tuplen{j}, V**0}
e := each{{j}=>V**trunc{uT, j}, j}
@for (xv over b) each{{c,e} => c -= xv == e, c, e}
def add_sum{c, j} = {
s := promote{usz, sum_vec{iV}(iV~~c)}
total -= s; inc{tab, j, s}
}
each{add_sum, c, j}
}
m4 := m / 4
@for (j4 to m4) count_each{each{{k}=>4*j4 + k, iota{4}}}
@for (j from 4*m4 to m) count_each{tup{j}}
inc{tab, m, trunc{usz,total}}
}
}