AVX2 boolean Compress using variable shifts

This commit is contained in:
Marshall Lochbaum 2023-08-05 21:50:22 -04:00
parent 66c0fe041c
commit 371aa6f2ef

View File

@ -288,7 +288,8 @@ export{'si_2slash64', slash{1, i64}}; export{'si_thresh_2slash64', u64~~thresh{1
# pext, or boolean compress
def pext_popc{x:T, m:T} = {
def w = width{T}
def w = if (isvec{T}) elwidth{T} else width{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}
# At each step, x and z are split into groups of length k
@ -298,11 +299,18 @@ 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 = lowbits{k} # Low bit in each new group
def low_s = lowbits{k} # Low bit in each new group
def low = scal{low_s}
if (k == 2) {
z0 := z & low
zm := z>>1 & low
tup{ x - (x>>1 & z0), zm + z0 }
} else if (hasarch{'AVX2'} and isvec{T} and k >= 32) {
# We have variable shifts at these sizes
lh := scal{low_s*(1<<h - 1)}
zl := z & lh
def S = re_el{ty_u{k}, T}
tup{T~~(S~~(x&~lh) >> S~~zl) | (x&lh), T~~(S~~z >> h) + zl}
} else {
# SWAR shifter: shift x by sh*o, in length-k groups
def shift{sh, o, x} = {
@ -311,9 +319,9 @@ 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 = mod{low*(1<<k - 1<<h)} # Top half
odd:T = scal{mod{low_s*(1<<k - 1<<h)}} # Top half
ze := z&~odd
z1 := ze + low*(1<<(k-1) - 1) # z-1, as signed k-bit
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
tup{
(x&~move) | shift{1, z1, x&move}>>1,
@ -323,12 +331,16 @@ def pext_popc{x:T, m:T} = {
}
# Finally, compose groups with regular shifts
def g = 8
{b,z} := build{g}
o := z*lowbits{g} # Offsets by prefix sum
def s = 1<<g - 1
def gr{sh} = (b & mod{s<<sh}) >> (o>>(sh-g) & s)
pe := fold{|, b&s, each{gr, g*slice{iota{cdiv{w,g}},1}}}
tup{pe, w - o>>(w-g)}
def build{k & ~isvec{T} & k > g} = {
{x,z} := build{g}
o := z*lowbits{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 {pe, z} = build{w}
tup{pe, scal{w} - z}
}
def pext_popc{xs:T, ms:T & hasarch{'PCLMUL'} & T==u64} = {
@ -350,13 +362,12 @@ def pext_popc{xs:T, ms:T & hasarch{'PCLMUL'} & T==u64} = {
tup{extract{x, 0}, popc{ms}}
}
def pext_popc{x:T, m:T & hasarch{'BMI2'}} = tup{pext{x, m}, popc{m}}
def pext_popc{x:T, m:T & hasarch{'BMI2'} & T==u64} = tup{pext{x, m}, popc{m}}
fn compress_bool(w:*u64, x:*u64, r:*u64, n:u64) : void = {
cw:u64 = 0; # current word
ro:u64 = 0; # offset in word where next bit should be written; never 64
@for (w, x over i to cdiv{n,64}) {
{v, c} := pext_popc{x, w}
def add_bits{{v, c}} = {
cw |= v<<ro
ro2 := ro+c
if (ro2 >= 64) {
@ -365,6 +376,17 @@ fn compress_bool(w:*u64, x:*u64, r:*u64, n:u64) : void = {
}
ro = ro2%64
}
if (hasarch{'AVX2'}) {
def V = [4]u64
nv := n/256
@for (w in *V~~w, x in *V~~x over i to nv) {
vc := pext_popc{x, w}
@unroll (j to 4) add_bits{each{extract{., j}, vc}}
}
@for (w, x over i from nv*4 to cdiv{n,64}) add_bits{pext_popc{x, w}}
} else {
@for (w, x over i to cdiv{n,64}) add_bits{pext_popc{x, w}}
}
if (ro > 0) store{r, 0, cw}
}