From c430922a7e51cbb0a612b60413c844116a54e8ac Mon Sep 17 00:00:00 2001 From: Marshall Lochbaum Date: Tue, 27 Sep 2022 17:38:19 -0400 Subject: [PATCH] Use shared functions and shuffle array for sizes 3 to 7 in constant Replicate --- src/singeli/src/slash.singeli | 85 +++++++++++++++++++++++------------ 1 file changed, 56 insertions(+), 29 deletions(-) diff --git a/src/singeli/src/slash.singeli b/src/singeli/src/slash.singeli index ba636a31..123f7320 100644 --- a/src/singeli/src/slash.singeli +++ b/src/singeli/src/slash.singeli @@ -55,26 +55,28 @@ include './avx' include './avx2' include './mask' include 'util/tup' -def idiv{a,b} = (a-a%b)/b +def incl{a,b} = slice{iota{b+1},a} def get_shufs{step, wv} = split{step, replicate{wv, iota{step}}} -def expandsel{ex}{s} = make{[32]i8, join{table{+, ex*s, iota{ex}}}} -def get_rep_iter{V, wv} = { - def step = vcount{V} - def sh = each{expandsel{32 / step}, get_shufs{step, wv}} - def h = wv >> 1 +def get_shuf_data{wv, elbytes} = { + def expand{ex}{s} = join{table{+, ex*s, iota{ex}}} + each{expand{elbytes}, get_shufs{32 / elbytes, wv}} +} +def get_shuf_Vs{V, wv, len} = { + def s = get_shuf_data{wv, width{eltype{V}}/8} + each{bind{make,[32]i8}, slice{s,0,len}} +} +def rep_iter_from_sh{sh} = { + def l = tuplen{sh} + def h = l >> 1 {x, gen} => { def fs{v, s} = gen{sel{[16]i8, v, s}} a := shuf{[4]u64, x, 4b1010}; each{bind{fs,a}, slice{sh,0,h}} - fs{x, tupsel{h, sh}} + if (l%2) fs{x, tupsel{h, sh}} b := shuf{[4]u64, x, 4b3232}; each{bind{fs,b}, slice{sh,-h}} } } -def get_rep_iter{V, wv & wv%2 == 0} = { - def r2 = get_rep_iter{V, 2} - def rk = get_rep_iter{V, wv/2} - {x, gen} => rk{x, {a}=>r2{a,gen}} -} +def get_rep_iter{V, wv} = rep_iter_from_sh{get_shuf_Vs{V, wv, wv}} # Unused def get_rep_iter{V, wv==2}{x, gen} = { def s = shuf{[4]u64, x, 4b3120} each{{q}=>gen{V~~q}, unpackQ{s, s}} @@ -85,8 +87,8 @@ def get_rep_iter{V==[4]u64, wv} = { def sh = each{base4, get_shufs{step, wv}} {x, gen} => each{{s}=>gen{shuf{V, x, s}}, sh} } -def rep_const_shuffle{V, step, wv, xv:*V, rv:*V, n:u64} = { - def onreps = get_rep_iter{V, wv} +def rep_const_shuffle{V, wv, onreps, xv:*V, rv:*V, n:u64} = { + def step = vcount{V} nv := n / step j:u64 = 0 def write{v} = { store{rv, j, v}; ++j } @@ -106,18 +108,43 @@ def rep_const_shuffle{V, step, wv, xv:*V, rv:*V, n:u64} = { if (q) maskstoreF{rv, maskOf{V, q}, e, s} } } -rep_const_sub{T, V}(wv:u64, x:*T, r:*T, n:u64, sh0:[32]i8, sh1:[32]i8, sh2:[32]i8, sh3:[32]i8) : void = { - def sh = tup{sh0,sh1,sh2,sh3} - def h = 4 - def step = vcount{V} - def hs = idiv{h*step, wv} # Actual step size - re := r + n*wv - h*step +def rep_const_shuffle{V, wv, xv:*V, rv:*V, n:u64} = rep_const_shuffle{V, wv, get_rep_iter{V, wv}, xv, rv, n} + +def rcsh_min = 3 +def rcsh_max = 7 +def rcsh_vals = incl{rcsh_min, rcsh_max} +def rcsh_offs = shiftright{0, scan{+,rcsh_vals}} +def rcsh_table = table{get_shuf_data, rcsh_vals, tup{1,2,4}} +rcsh_data:*i8 = join{join{join{rcsh_table}}} +rcsh_sub{wv}(x:*i8, r:*i8, n:u64, sh:*[32]i8) : void = { + def V = [32]i8 + def st = each{bind{load,sh}, iota{wv}} + rep_const_shuffle{V, wv, rep_iter_from_sh{st}, *V~~x, *V~~r, n} +} +def rcsh_subs = each{rcsh_sub, rcsh_vals} +def rep_const_shuffle{V, wv, xv:*V, rv:*V, n:u64 & width{eltype{V}}<=32 & rcsh_min<=wv & wv<=rcsh_max} = { + def ri = wv - tupsel{0,rcsh_vals} + def eb = width{eltype{V}}/8 + def ti = eb-1 - (eb==4) + tupsel{ri, rcsh_subs}( + *i8~~xv, *i8~~rv, n*(width{eltype{V}}/8), + *[32]i8~~rcsh_data + (tupsel{ri,rcsh_offs}*3 + ti*wv) + ) +} + +rep_const_part4_sub(wv:u64, elbytes:u64, x:*i8, r:*i8, n:u64, sh0:[32]i8, sh1:[32]i8, sh2:[32]i8, sh3:[32]i8) : void = { + def h = 4; def sh = tup{sh0,sh1,sh2,sh3} + def V = [32]i8 + def step = vcount{V} # Bytes written + def wvb = wv * elbytes + def hs = (h*step) / wvb # Actual step size in argument elements + re := r + n*wvb - h*step i:u64 = 0 while (r <= re) { a := shuf{[4]u64, load{*V~~(x+i),0}, 4b1010} @unroll (j to h) store{*V~~r, j, sel{[16]i8, a, tupsel{j,sh}}} - i += hs - r += hs*wv + i += hs*elbytes + r += hs*wvb } re += (h-1)*step a := shuf{[4]u64, load{*V~~(x+i),0}, 4b1010} @@ -133,10 +160,10 @@ rep_const_sub{T, V}(wv:u64, x:*T, r:*T, n:u64, sh0:[32]i8, sh1:[32]i8, sh2:[32]i q := (re+step) - r if (q) maskstoreF{*V~~r, maskOf{V, q}, 0, s} } -def rep_const_shuffle{V, step, wv, xv:*V, rv:*V, n:u64 & wv >= 11} = { +def rep_const_shuffle{V, wv, xv:*V, rv:*V, n:u64 & wv >= 11} = { def T = eltype{V} - def sh = each{expandsel{32/step}, slice{get_shufs{step, wv}, 0, 4}} - apply{bind{call, rep_const_sub{T, V}, wv, *T~~xv, *T~~rv, n}, sh} + def sh = get_shuf_Vs{V, wv, 4} + apply{bind{call, rep_const_part4_sub, wv, width{T}/8, *i8~~xv, *i8~~rv, n}, sh} } rep_const_broadcast{T, V, kv}(wv:u64, x:*T, r:*T, n:u64) : void = { @@ -174,19 +201,19 @@ def rep_const{T, wv, x, r, n} = { def step = 256/width{T} def V = [step]T if (wv <= max{2*step, 6}) { - rep_const_shuffle{V, step, wv, *V~~x, *V~~r, n} + rep_const_shuffle{V, wv, *V~~x, *V~~r, n} } else { - def kv = idiv{wv, step} + def kv = (wv-wv%step) / step rep_const_broadcast{T, V, kv}(wv, *T~~x, *T~~r, n) } } -# 1+˝∨`⌾⌽(2+↕6)(0=|⌜)8+↕25 +# 1+˝∨`⌾⌽0=div|⌜range def makefact{divisor, range} = { def t = table{{a,b}=>0==b%a, divisor, range} fold{+, 1, reverse{scan{|, reverse{t}}}} } -def basic_rep = 2+iota{6} +def basic_rep = incl{2,7} def fact_size = 128 def fact_inds = slice{iota{fact_size},8} def fact_tab = makefact{basic_rep, fact_inds}