More tuning based on dzaima's measurements

This commit is contained in:
Marshall Lochbaum 2025-02-26 22:01:37 -05:00
parent a520ff70a1
commit eda1cd8b50

View File

@ -357,24 +357,6 @@ def loose_mask_gen{V=[vl]T, l} = { # Slow, for ≠` only
def get = loose_mask_gen{T, l}
{} => make{V, @collect (vl) get{}}
}
def loose_mask_gen{V=[vl](u64), l if hasarch{'SSSE3'}} = {
# Shuffles can substitute for variable shifts, in a pinch
assert{l < 256}
def I = re_el{i8, V}; def [il]_ = I
def U = [il]u8
q := make{I, (-128) ^ (-8*iota{il})} # distance to next row boundary, -128
l8:= cast_i{u8, l}; vl:= I**i8~~l8
def q_mod{} = { q+= vl &~ I~~(q < I**0) }
q_mod{}
o:u8 = width{V}; while (o>l8) { o-=l8; q_mod{} }
oo:= I**i8~~(o - 128)
s := make{U, 1<<(iota{il}%8)}
{} => {
m:= shuf{s, q & I**7} & (q < I**(8-128))
q+= (vl & I~~(q < oo)) - I~~U**o
V~~m
}
}
def has_vecshift = hasarch{'AVX2'} or hasarch{'AARCH64'}
def loose_mask_gen{V=[vl](u64), l if has_vecshift} = {
q := -make{V, 64*iota{vl}} # distance to next row boundary
@ -484,15 +466,24 @@ fn scan_rows_andor{id}(src:*u64, dst:*u64, nl:usz, l:usz) : void = {
}}
}
} else if ((hasarch{'SSE4.1'} or hasarch{'AARCH64'}) and l < (if (hasarch{'AVX2'}) 256 else 160)) {
def scan_words{x:V, m:V, _} = {
s:= (if (qand) x &~ ((x+V**1) & (x+m))
else x | ((-x) &~ (x-m)))
p:= (if (qand) x&~m == ~V**0
else x| m == V**0)
k:= s>>63 | p # Carry of 0 or 1, but -1 to propagate previous
tup{s, k}
if (hasarch{'SSE4.1'}) {
def scan_words{x:V, m:V, _} = {
s:= (if (qand) x &~ ((x+V**1) & (x+m))
else x | ((-x) &~ (x-m)))
p:= (if (qand) x&~m == ~V**0
else x| m == V**0)
k:= s>>63 | p # Carry of 0 or 1, but -1 to propagate previous
tup{s, k}
}
vec_loop_with_loose_mask{src, dst, nw, l, id, scan_words, apply_carry}
} else {
def get_m = loose_mask_gen{u64, l}
c:u64 = id # carry
@for (r in dst, x in src over nw) {
r = res_m1{x, c, get_m{}}
c = r >> 63
}
}
vec_loop_with_loose_mask{src, dst, nw, l, id, scan_words, apply_carry}
} else {
i :usz = 0 # row bit index
wn:usz = 0 # starting word of next row
@ -548,7 +539,7 @@ fn scan_rows_neq(x:*u64, r:*u64, nl:usz, l:usz) : void = {
s ^ ((c & f) | (b<<l - b))
}}
}
} else if (has_simd and l < (if (hasarch{'AVX2'}) 320 else 200)) {
} else if (has_simd and l < (if (hasarch{'AVX2'}) 320 else 192)) {
def scan_words{x:V, m:V, ml:V} = {
s:= scan_words{x}
s^= -(s<<1 & m)
@ -557,7 +548,7 @@ fn scan_rows_neq(x:*u64, r:*u64, nl:usz, l:usz) : void = {
}
def propagate{k:V=[vl]_, c:V} = {
def bl{b,a} = b ^ (a &~ -(b>>63))
def bl{b,a if hasarch{'SSE4.1'}} = blend_top{a^b,b, b}
def bl{b,a if hasarch{'AVX2'}} = blend_top{a^b,b, b}
k = bl{k, vec_shift_right_128{k, 1}}
if (vl>2) k = bl{k, shuf{V, blend{V**0, k, 0,1,0,1}, 0,0,1,1}}
bl{k, c}