uCBQN/src/singeli/src/select.singeli

170 lines
5.3 KiB
Plaintext

include './base'
include './cbqnDefs'
include './sse3'
include './avx'
include './avx2'
include './mask'
include './bitops'
include 'util/tup'
oper &~ andnot infix none 35
# def:T - masked original content
# b:B - pointer to data to index; if width{B}<elwidth{T}, padding bytes are garbage read after wanted position
# idx - actual (unscaled) index list
def gather{def:T, b:B, idx:[8]i32, M & w256{T,32}} = {
if (M{0}) T ~~ emit{[8]i32, '_mm256_mask_i32gather_epi32', def, *void~~b, idx, M{T,'to sign bits'}, elwidth{B}/8}
else T ~~ emit{[8]i32, '_mm256_i32gather_epi32', *void~~b, idx, elwidth{B}/8}
}
def gather{def:T, b:B, idx:[4]i32, M & w256{T,64}} = {
if (M{0}) T ~~ emit{[4]i64, '_mm256_mask_i32gather_epi64', def, *void~~b, idx, M{T,'to sign bits'}, elwidth{B}/8}
else T ~~ emit{[4]i64, '_mm256_i32gather_epi64', *void~~b, idx, elwidth{B}/8}
}
def wrapChk{cw0, VI,xlf, M} = {
cw:= cw0 + (xlf & VI~~(cw0<VI**0))
if (homAny{M{ty_u{cw} >= ty_u{xlf}}}) return{0}
cw
}
def storeExp{dst, ind, val, M, ext, rd, wl} = {
def s{M} = storeBatch{dst, ind, val, M}
if (ext==1 or not M{0}) s{M}
else if (ind*rd+rd <= wl) s{maskNone}
else { if (ind*rd < wl) s{maskAfter{wl & (rd-1)}}; return{1} }
}
def shuf_select{ri, rd, TI, w, r, wl, xl, selx} = {
def VI = [ri]TI
def ext = ri/rd
xlf:= VI**cast_i{TI, xl}
maskedLoop{ri, wl, {i, M} => {
cw:= wrapChk{loadBatch{w, i, VI}, VI,xlf, M}
is:= (if (ext>1) i<<lb{ext}; else i)
def se{e, c, o} = {
c2:= shuf{[4]u64, c+c, 4b3120}
each{
{c,o} => se{e*2, VI~~c, o},
unpackQ{c2, c2 + VI**1},
2*o + iota{2}
}
}
def se{e==ext, c, o} = storeExp{r, is+o, selx{c}, M, ext, rd, wl}
se{1, cw, 0}
}}
}
def perm_select{ri, rd, TI, w, r, wl, xl, selx} = {
def VI = [ri]TI
def ext = ri/rd
xlf:= VI**cast_i{TI, xl}
maskedLoop{ri, wl, {i, M} => {
cw:= wrapChk{loadBatch{w, i, VI}, VI,xlf, M}
is:= (if (ext>1) i<<lb{ext}; else i)
def part{o} = cvt{i8, [8]i32, shuf{[4]u64, cw, 4b3210+o}}
def se{o} = storeExp{r, is+o, selx{part{o}}, M, ext, rd, wl}
each{se, iota{ext}}
}}
}
def makeselx{VI, VD, nsel, xd, logv, cshuf} = {
def bblend {m}{ft} = blend{tupsel{0,ft}, tupsel{1,ft}, m}
def bblendn{m}{tf} = bblend{m}{reverse{tf}}
def bb{c}{f, v} = (if (f) bblendn{c<v}; else bblend{(c&v)==v})
def bs{b, c, x} = cshuf{x, c}
def bs{b, c, x & tuplen{b}>0} = {
tupsel{0,b}{each{bind{bs, slice{b,1}, c}, x}}
}
def i = iota{logv}
def vs = each{bind{broadcast,VI}, nsel<<reverse{i}}
{c} => VD~~bs{each{bb{c},i==0,vs}, c, xd}
}
def makeshuf{VI, VD, x0, logv} = {
x:= *VD~~x0
def halves{v} = each{bind{shuf, [4]u64, v}, tup{4b1010, 4b3232}}
def readx{l,o} = each{bind{readx,l-1}, o + iota{2}<<(l-2)}
def readx{l==0,o} = shuf{[4]u64, load{x}, 4b1010}
def readx{l==1,o} = halves{load{x, o}}
xd:= readx{logv, 0}
makeselx{VI,VD,16,xd,logv, bind{sel,[16]i8}}
}
def makeperm{VI, VD, x0, logv} = {
x:= *VD~~x0
def readx{l,o} = each{bind{readx,l-1}, o + iota{2}<<(l-1)}
def readx{l==0,o} = load{x, o}
makeselx{[8]i32,VD,8, readx{logv, 0}, logv, bind{sel,[8]i32}}
}
select{rw, TI, TD}(w0:*void, x0:*void, r0:*void, wl:u64, xl:u64) : u1 = {
w:= *TI ~~ w0
x:= *TD ~~ x0
r:= *TD ~~ r0
def wd = width{TD}; def rd = rw/wd
def wi = width{TI}; def ri = rw/wi
def reg_select{sel,make}{l} = {
sel{ri, rd, TI, w, r, wl, xl, make{[ri]TI,[rd]TD, x,l}}
}
def shuf_select = reg_select{shuf_select, makeshuf}
def perm_select = reg_select{perm_select, makeperm}
if (wi==8 and wd==32 and xl*wd<=256 ) perm_select{0}
else if (wi==8 and wd==32 and xl*wd<=256<<1) perm_select{1}
else if (wi==8 and wd<=16 and xl*wd<=128 ) shuf_select{0}
else if (wi==8 and wd<=16 and xl*wd<=128<<1) shuf_select{1}
else if (wi==8 and wd<=16 and xl*wd<=128<<2) shuf_select{2}
else if (wi==8 and wd<= 8 and xl*wd<=128<<3) shuf_select{3}
else {
def TIE = i32
def TDE = tern{wd<32, u32, TD}
def bulk = rw / width{TDE}
def VI = [bulk]TIE
def VD = [bulk]TDE
def xlf = VI**cast_i{TIE, xl}
maskedLoop{bulk, wl, {i, M} => {
cw:= wrapChk{loadBatch{w, i, VI}, VI,xlf, M}
got:= gather{VD**0, x, cw, M}
if (TDE!=TD) got&= VD**((1<<wd)-1)
storeBatch{r, i, got, M}
}}
}
1
}
def select{TI, TD} = select{256, TI, TD}
def selects = join{table{select, tup{i8, i16, i32}, # indices
tup{u8, u16, u32, u64}}} # values
avx2_select_tab:*type{tupsel{0,selects}} = selects
'avx2_select_tab' = avx2_select_tab
avx2_select_bool128(w0:*void, x0:*void, r0:*void, wl:u64, xl:u64) : u1 = {
def TI = i8
def VI = [32]TI
w:= *VI ~~ w0
r:= *u32 ~~ r0
xlf:= VI**cast_i{TI, xl}
if (wl>32 and xl<=16) {
xb:= shuf{[4]u64, spreadBits{[32]u8, load{*u32~~x0}}, 4b1010}
maskedLoop{32, wl, {i, M} => {
cw:= wrapChk{load{w, i}, VI,xlf, M}
store{r, i, getmask{sel{[16]i8, xb, cw}}}
}}
} else {
x:= shuf{[4]u64, load{*VI ~~ x0}, 4b1010}
low:= VI**7
b := VI~~make{[32]u8, 1 << (iota{32} & 7)}
maskedLoop{32, wl, {i, M} => {
cw:= wrapChk{load{w, i}, VI,xlf, M}
byte:= sel{[16]i8, x, VI~~(([8]u32~~andnot{cw, low})>>3)}
mask:= sel{[16]i8, b, cw & low}
store{r, i, getmask{(mask & byte) == mask}}
}}
}
1
}
'avx2_select_bool128' = avx2_select_bool128