Singeli shuffle-based code for constant replicates <=7

This commit is contained in:
Marshall Lochbaum 2022-09-26 10:33:54 -04:00
parent 851f50635e
commit 7535d768dd
2 changed files with 81 additions and 1 deletions

View File

@ -720,6 +720,13 @@ B slash_c2(B t, B w, B x) {
u8 xk = xl-3;
void* rv = m_tyarrv(&r, 1<<xk, s, xt);
void* xv = tyany_ptr(x);
#if SINGELI
if (wv<=7) {
#define CASE(L,T) case L: rep_##T(wv, xv, rv, xlen); break;
switch (xk) { default: UD; CASE(0,u8) CASE(1,u16) CASE(2,u32) CASE(3,u64) }
#undef CASE
} else
#endif
#define CASE(L,T) case L: { REP_BY_SCAN(T, wv) break; }
switch (xk) { default: UD; CASE(0,u8) CASE(1,u16) CASE(2,u32) CASE(3,u64) }
#undef CASE

View File

@ -48,4 +48,77 @@ slash1{F, T, iota, add}(w:*u64, r:*T, l:u64) : void = {
'bmipopc_2slash8' = slash2{comp8, i8}
'bmipopc_2slash16' = slash2{comp16, i16}
'bmipopc_1slash8' = slash1{comp8, i8, 0x0706050403020100, 0x0808080808080808}
'bmipopc_1slash16' = slash1{comp16, i16, 0x0003000200010000, 0x0004000400040004}
'bmipopc_1slash16' = slash1{comp16, i16, 0x0003000200010000, 0x0004000400040004}
include './sse3'
include './avx'
include './avx2'
include './mask'
include 'util/tup'
def get_shufs{step, wv} = split{step, replicate{wv, iota{step}}}
def get_rep_iter{V, wv} = {
def step = vcount{V}
def ie = 32 / step # index expansion
def tosel{s} = make{[32]i8, join{table{+, ie*s, iota{ie}}}}
def sh = each{tosel, get_shufs{step, wv}}
def h = wv >> 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}}
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==2}{x, gen} = {
def s = shuf{[4]u64, x, 4b3120}
each{{q}=>gen{V~~q}, unpackQ{s, s}}
}
def get_rep_iter{V==[4]u64, wv} = {
def step = 4
def base4{l} = { if (0==tuplen{l}) 0; else tupsel{0,l}+4*base4{slice{l,1}} }
def sh = each{base4, get_shufs{step, wv}}
{x, gen} => each{{s}=>gen{shuf{V, x, s}}, sh}
}
rep_const{T, wv}(x:*void, r:*void, n:u64) : void = {
def step = 256/width{T}
def V = [step]T
xv := *V~~x
rv := *V~~r
def onreps = get_rep_iter{V, wv}
nv := n / step
j:u64 = 0
def write{v} = { store{rv, j, v}; ++j }
@for (xv over nv) onreps{xv, write}
if (nv*step < n) {
nr := n * wv
e := nr / step
s := broadcast{V, 0}
def end = makelabel{}
onreps{load{xv,nv}, {v} => {
s = v
if (j == e) goto{end}
write{s}
}}
setlabel{end}
q := nr & (step-1)
if (q) maskstoreF{rv, maskOf{V, q}, e, s}
}
}
rep_const{T}(wv:i32, x:*void, r:*void, n:u64) : void = {
assert{wv>=2}
@unroll (k from 2 to 8) {
if (wv==k) return{rep_const{T, k}(x, r, n)}
}
def rc = rep_const{T}
}
'rep_u8' = rep_const{i8 }; 'rep_u16' = rep_const{i16}
'rep_u32' = rep_const{i32}; 'rep_u64' = rep_const{u64}