In AVX2 boolean Compress, jump from 8 to 32 with 32-bit shifts

This commit is contained in:
Marshall Lochbaum 2023-08-07 11:32:02 -04:00
parent 120df26814
commit bbad98bce1

View File

@ -293,8 +293,7 @@ def pext_width{} = if (hasarch{'AVX2'}) 4 else 1
def pext_popc{x:T, m:T} = {
def w = scalwidth{T}
def scal{v} = if (isvec{T}) T**v else v
def mod{a} = a % (1<<w)
def lowbits{k} = base{1<<k, cdiv{w,k}**1}
def lowbits{w,k} = base{1<<k, cdiv{w,k}**1}
# At each step, x and z are split into groups of length k
# - z tells how many bits in the group are NOT used
# - x contains the bits, with z zeros above
@ -302,7 +301,7 @@ def pext_popc{x:T, m:T} = {
def build{k & k > 1} = {
def h = k>>1 # Increase size from h to k
{x,z} := build{h}
def low_s = lowbits{k} # Low bit in each new group
def low_s = lowbits{w,k} # Low bit in each new group
def low = scal{low_s}
if (k == 2) {
z0 := z & low
@ -322,7 +321,7 @@ def pext_popc{x:T, m:T} = {
if (2*sh<k/2) shift{2*sh, o>>1, s} else s
}
# Shift high x group down by low z, then add halves of z
odd:T = scal{mod{low_s*(1<<k - 1<<h)}} # Top half
odd:T = scal{low_s*(1<<k - 1<<h)} # Top half
ze := z&~odd
z1 := ze + scal{low_s*(1<<(k-1) - 1)} # z-1, as signed k-bit
move := odd &~ (z1<<1) # Only groups where z>0 move
@ -332,16 +331,25 @@ def pext_popc{x:T, m:T} = {
}
}
}
# Finally, compose groups with regular shifts
def g = 8
def build{k & ~isvec{T} & k > g} = {
{x,z} := build{g}
o := z*lowbits{g} # Offsets by prefix sum
# Compose k/g groups with k/g-1 regular shifts
def multi_shift{x, z, g, k, sc} = {
o := z * sc{lowbits{k,g}} # Offsets by prefix sum
def s = 1<<g - 1
def gr{sh} = (x & mod{s<<sh}) >> (o>>(sh-g) & s)
pe := fold{|, x&s, each{gr, g*slice{iota{cdiv{w,g}},1}}}
tup{pe, o>>(w-g)}
def s0 = sc{s}
def oo{sh} = if (sh==g) z else o>>(sh-g) # Offset for group
def gr{sh} = (x & sc{s<<sh}) >> (oo{sh} & s0) # Shifted group
pe := fold{|, x&s0, each{gr, g*slice{iota{k/g},1}}}
tup{pe, o>>(k-g)}
}
def build{k==32 & hasarch{'AVX2'} & isvec{T}} = {
def S = re_el{ty_u{k}, T}
def c{T,vs} = each{{v}=>T~~v, vs}
c{T, multi_shift{...c{S, build{8}}, 8, k, {s}=>S**s}}
}
def build{k & ~isvec{T} & k > 8} = {
multi_shift{...build{8}, 8, k, {s}=>s}
}
# Final result
def {pe, z} = build{w}
tup{pe, scal{w} - z}
}