Tuning; drop the scalar loose_mask branch for ∧∨ as it never seems faster

This commit is contained in:
Marshall Lochbaum 2025-02-25 19:33:26 -05:00
parent b5ec2f5fa8
commit 709712f03c

View File

@ -458,25 +458,16 @@ fn scan_rows_andor{id}(src:*u64, dst:*u64, nl:usz, l:usz) : void = {
a >> 63} # new c a >> 63} # new c
}} }}
} }
} else if (l < (if (hasarch{'AVX2'}) 256 else 160)) { } else if ((hasarch{'SSE4.1'} or hasarch{'AARCH64'}) and l < (if (hasarch{'AVX2'}) 256 else 160)) {
if (hasarch{'SSE4.1'} or hasarch{'AARCH64'}) { def scan_words{x:V, m:V, _} = {
def scan_words{x:V, m:V, _} = { s:= (if (qand) x &~ ((x+V**1) & (x+m))
s:= (if (qand) x &~ ((x+V**1) & (x+m)) else x | ((-x) &~ (x-m)))
else x | ((-x) &~ (x-m))) p:= (if (qand) x&~m == ~V**0
p:= (if (qand) x&~m == ~V**0 else x| m == V**0)
else x| m == V**0) k:= s>>63 | p # Carry of 0 or 1, but -1 to propagate previous
k:= s>>63 | p # Carry of 0 or 1, but -1 to propagate previous tup{s, k}
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 { } else {
i :usz = 0 # row bit index i :usz = 0 # row bit index
wn:usz = 0 # starting word of next row wn:usz = 0 # starting word of next row
@ -538,7 +529,7 @@ fn scan_rows_neq(x:*u64, r:*u64, nl:usz, l:usz) : void = {
s ^ ((c & f) | (b<<l - b)) s ^ ((c & f) | (b<<l - b))
}} }}
} }
} else if (has_simd and l < (if (hasarch{'SSSE3'} and not hasarch{'AVX2'}) 256 else 320)) { } else if (has_simd and l < (if (hasarch{'AVX2'}) 320 else 256)) {
def scan_words{x:V, m:V, ml:V} = { def scan_words{x:V, m:V, ml:V} = {
s:= scan_words{x} s:= scan_words{x}
s^= -(s<<1 & m) s^= -(s<<1 & m)
@ -588,7 +579,7 @@ fn scan_rows_left(x:*u64, r:*u64, nl:usz, l:usz) : void = {
(c & f) | apply_mask{x, m} (c & f) | apply_mask{x, m}
}} }}
} }
} else if ((hasarch{'SSE4.1'} or hasarch{'AARCH64'}) and l < 176) { } else if ((hasarch{'SSE4.1'} or hasarch{'AARCH64'}) and l < (if (hasarch{'SSE4.1'} and not hasarch{'AVX2'}) 72 else 176)) {
def scan_words{x:V, m:V, _} = { def scan_words{x:V, m:V, _} = {
s:= -(x & m) s:= -(x & m)
tup{s, s>>63 | (m == V**0)} tup{s, s>>63 | (m == V**0)}