SIMD table intialization for Index-of

This commit is contained in:
Marshall Lochbaum 2023-04-29 10:50:51 -04:00
parent 844b537593
commit 464dd27a37
2 changed files with 42 additions and 17 deletions

View File

@ -226,6 +226,18 @@ B indexOf_c2(B t, B w, B x) {
}
if (xia+wia>20 && we<=el_i16 && xe<=el_i16) {
#if SINGELI
if (wia>256 && we==el_i8 && xe==el_i8) {
TALLOC(u8, tab, 256*(1+sizeof(usz))); usz* ind = (usz*)(tab+256);
void* fp = tyany_ptr(x);
simd_index_tab_u8(tyany_ptr(w), wia, fp, xia, tab, ind);
decG(w);
i32* rp; B r = m_i32arrc(&rp, x);
for (usz i=0; i<xia; i++) rp[i]=ind[((u8*)fp)[i]];
TFREE(tab); decG(x);
return reduceI32Width(r, wia);
}
#endif
B r;
TABLE(w, x, i32, wia, i)
return reduceI32Width(r, wia);

View File

@ -138,25 +138,28 @@ def bittab_lookup{x0:*void, n:u64, r0:*void, tab:*void & hasarch{'AVX2'}} = {
# are equal: by the time it's checked at least one has been set to t,
# so they're all t
# Fill r0 depending on mode:
# - 'none': ignore
# - 'mask': Mark Firsts of x0
# - 'none': ignore
# - 'mask': Mark Firsts of x0
# - 'unique': Deduplicate of x0
# - 'index': First index of value x at r0+x
def do_bittab{x0:*void, n:u64, tab:*void, u:u8, t, mode, r0} = {
def rbit = mode == 'mask'
def storebit{i, v:T} = if (rbit) store{*T~~r0, i, v}
def rval = mode == 'unique'
def rind = mode == 'index'
def storebit{i, v:T} = if (rbit) store{*T~~r0, i, v}
if (rbit or rval) assert{match{t,0}}
btab:= *i8~~tab
def settab_sub{x, v} = {
def settab_sub{x, v, i} = {
if (rval) store{*u8~~r0, u, x}
u+= u8~~(t - v) # u tracks the total of btab
if (rind and v!=t) store{r0, x, i}
u+= u8~~i8~~(t - v) # u tracks the total of btab
store{btab, x, t}
v
}
def settab1{x} = settab_sub{x, -1 - t} # Known new
def settab{x} = settab_sub{x, load{btab, x}} # General case
def settab{T, x} = T~~promote{ty_s{T}, settab{x}}
def settab1{x, i} = settab_sub{x, -1 - t, i} # Known new
def settab{x, i} = settab_sub{x, load{btab, x}, i} # General case
def settab{T, x, i} = T~~promote{ty_s{T}, settab{x, i}}
x:= *u8~~x0
if (not hasarch{'AVX2'}) {
@ -165,7 +168,7 @@ def do_bittab{x0:*void, n:u64, tab:*void, u:u8, t, mode, r0} = {
k:= rem; if (k>64) k=64
rw:u64 = 0
@for (x over j to k) {
new:= settab{u64, x}
new:= settab{u64, x, i*64+j} # Index usually unused
if (rbit) rw|= new & ((u64~~1)<<j)
}
storebit{i, rw}
@ -178,11 +181,11 @@ def do_bittab{x0:*void, n:u64, tab:*void, u:u8, t, mode, r0} = {
{rw,rv} := undef{tup{u64,u32}} # Bit results, used if rbit
if (rbit) rw = 0
@for (x over j to i) {
new:= settab{u64, x}
new:= settab{u64, x, j}
if (rbit) rw|= new & ((u64~~1)<<j)
}
storebit{0, rw}
if (mode == 'none' and u == 0) return{u} # Won't ever trigger (m != 0)!
if ((mode == 'none' or rind) and u == 0) return{u} # Won't ever trigger (m != 0)!
def done = makelabel{}
def {bitsel, reload_tab} = bittab_selector{readbytes{*VI~~tab}}
@ -200,8 +203,9 @@ def do_bittab{x0:*void, n:u64, tab:*void, u:u8, t, mode, r0} = {
} else {
# Add values to the table and filter m
if (rbit) rv = m
xi:= load{x, i0 + ctzi{m}}
settab1{xi}
im:= i0 + ctzi{m}
xi:= load{x, im}
settab1{xi, im}
if ((m&(m-1)) != 0) { # More bits than one
# Filter out values equal to the previous, or first new
def pind = (iota{32}&15) - 1
@ -211,7 +215,8 @@ def do_bittab{x0:*void, n:u64, tab:*void, u:u8, t, mode, r0} = {
if (rbit) rv&= e | -m # Don't remove first bit
m&= e
while (m != 0) {
new:= settab{u32, load{x, i0 + ctzi{m}}}
im:= i0 + ctzi{m}
new:= settab{u32, load{x, im}, im}
m1:= m-1; m&= m1 # Clear low bit
if (rbit) rv&= m1 | new # Clear if not new
}
@ -250,16 +255,13 @@ fn fill_bittab(x0:*void, n:u64, tab:*void, u:u8, t:i8) : u8 = {
fn simd_member_u8(w0:*void, nw:u64, x0:*void, nx:u64, r0:*void, tab:*void) : void = {
assert{nw > 0}
# Initialize
rev:u1 = nx < nw/4 # Reverse lookup
bittab_init{tab, -promote{i8,rev}}
u:u8 = 0 # Sum of table, either 0 or 256
if (rev) u = fill_bittab(x0, nx, tab, u, 0)
# Fill table
u = fill_bittab(w0, nw, tab, u, -1)
# Read result
if (u == 0) { # All found!
@for (r in *u64~~r0 over cdiv{nx,64}) r = maxvalue{u64}
} else {
@ -267,6 +269,17 @@ fn simd_member_u8(w0:*void, nw:u64, x0:*void, nx:u64, r0:*void, tab:*void) : voi
}
}
fn simd_index_tab_u8{I}(w0:*void, nw:u64, x0:*void, nx:u64, tab:*void, i0:*void) : void = {
rev:u1 = nx < nw/4
bittab_init{tab, -promote{i8,rev}}
ind:= *I~~i0
@for (ind over 256) ind = cast_i{I, nw}
u:u8 = 0
if (rev) u = fill_bittab(x0, nx, tab, u, 0)
do_bittab{w0, nw, tab, u, -1, 'index', ind}
}
export{'simd_mark_firsts_u8', simd_mark_firsts_u8}
export{'simd_deduplicate_u8', simd_deduplicate_u8}
export{'simd_member_u8', simd_member_u8}
export{'simd_index_tab_u8', simd_index_tab_u8{usz}}