Use minimum in addition to maximum for AVX2 counting
This commit is contained in:
parent
e6a3e846d9
commit
2ddcc14852
@ -34,7 +34,8 @@ fn count{T}(tab:*usz, x:*ty_u{T}, n:u64) : u1 = {
|
||||
def uT = ty_u{T}
|
||||
def V = [vec]uT
|
||||
def iV = [vec]T
|
||||
def block = (1024*8) / vbits # Desired vectors per block
|
||||
def block = (2048*8) / vbits # Target vectors per block
|
||||
assert{block < 1<<width{T}} # Don't overflow count in vector section
|
||||
i:u64 = 0
|
||||
while (i < n) {
|
||||
r:u64 = n - i
|
||||
@ -43,16 +44,19 @@ fn count{T}(tab:*usz, x:*ty_u{T}, n:u64) : u1 = {
|
||||
used_eq:u1 = 0
|
||||
if (r >= 128) {
|
||||
b = block; if (r < vec*b) b = r / vec
|
||||
mv := V**0
|
||||
@for (xv over b) mv = max{mv, xv}
|
||||
jv := load{xv}; mv := jv
|
||||
@for (xv over _ from 1 to b) { jv = min{jv, xv}; mv = max{mv, xv} }
|
||||
mi := iV~~mv
|
||||
if (homAny{mi < iV**0}) return{1}
|
||||
if (homAll{mi <= iV**48}) {
|
||||
jt := fold{min, jv}
|
||||
if (homAll{mi <= iV**(48 + i8~~jt)}) {
|
||||
used_eq = 1
|
||||
r = b * vec
|
||||
m := promote{u64, fold{max, mv}}
|
||||
j0 := promote{u64, jt}
|
||||
m := promote{u64, fold{max, mv}} - j0
|
||||
total := trunc{usz, b*vec}
|
||||
def count_each{j} = {
|
||||
def count_each{js, num} = {
|
||||
j := (@collect (k to num) js+k)
|
||||
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}
|
||||
@ -63,9 +67,9 @@ fn count{T}(tab:*usz, x:*ty_u{T}, n:u64) : u1 = {
|
||||
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}}
|
||||
@for (j4 to m4) count_each{j0 + 4*j4, 4}
|
||||
@for (j from 4*m4 to m) count_each{j0 + j, 1}
|
||||
inc{tab, j0 + m, trunc{usz,total}}
|
||||
}
|
||||
}
|
||||
if (not used_eq) @for (x over r) inc{tab, x}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user