Faster i32 transpose when height is a multiple of 16 by writing cache lines
This commit is contained in:
parent
bfb2519aed
commit
ca44d41daa
@ -22,16 +22,74 @@ def vtranspose{x & tuplen{x}==8 & type{tupsel{0,x}}==[8]i32 & hasarch{'X86_64'}}
|
||||
fn transpose{T}(r0:*void, x0:*void, w:u64, h:u64) : void = {
|
||||
rp:*T = *T~~r0
|
||||
xp:*T = *T~~x0
|
||||
|
||||
def for_mult{k}{vars,begin,end,block} = {
|
||||
assert{begin == 0}
|
||||
@for (i to end/k) exec{k*i, vars, block}
|
||||
}
|
||||
|
||||
# Kernel size
|
||||
def k = 8
|
||||
def VT = [k]i32
|
||||
|
||||
# Cache line info
|
||||
def line_bytes = 64
|
||||
def line_elts = line_bytes / (width{T}/8)
|
||||
def line_vecs = line_bytes / (width{VT}/8)
|
||||
|
||||
@for (y to h/8) {
|
||||
@for (x to w/8) {
|
||||
def VT = [8]i32
|
||||
xpo:= xp + y*8*w + x*8
|
||||
rpo:= rp + x*8*h + y*8
|
||||
def xvs = each{{i}=>load{*VT~~(xpo+i*w), 0}, iota{vcount{VT}}}
|
||||
def rvs = vtranspose{xvs}
|
||||
each{{i,v}=>store{*VT~~(rpo+i*h), 0, v}, iota{vcount{VT}}, rvs}
|
||||
}
|
||||
if (h&(line_elts-1) != 0) {
|
||||
@for_mult{k} (y to h) {
|
||||
@for_mult{k} (x to w) {
|
||||
xpo:= xp + y*w + x
|
||||
rpo:= rp + x*h + y
|
||||
def xvs = each{{i}=>load{*VT~~(xpo+i*w), 0}, iota{vcount{VT}}}
|
||||
def rvs = vtranspose{xvs}
|
||||
each{{i,v}=>store{*VT~~(rpo+i*h), 0, v}, iota{vcount{VT}}, rvs}
|
||||
}
|
||||
}
|
||||
} else {
|
||||
# Result rows are aligned with each other so it's possible to
|
||||
# write a full cache line at a time
|
||||
# This case is here to mitigate cache associativity problems at
|
||||
# at multiples of 256 or so, but it's faster whenever it applies
|
||||
def store_line{p, vs} = each{bind{store,p}, iota{line_vecs}, vs}
|
||||
def get_lines{loadx} = {
|
||||
def vt{i} = vtranspose{each{loadx, k*i + iota{k}}}
|
||||
each{tup, ...each{vt, iota{line_vecs}}}
|
||||
}
|
||||
ro := tail{6, -u64~~r0} / 4 # Offset to align within cache line; assume elt-aligned
|
||||
wh := w*h
|
||||
yn := h
|
||||
if (ro != 0) {
|
||||
ra := line_elts - ro
|
||||
y := h - ra
|
||||
rpo := rp + y # Cache aligned
|
||||
rpe := rpo + (w-1)*h
|
||||
# Part of first and last result row aren't covered by the split loop
|
||||
def trtail{dst, src, len} = @for (i to len) store{dst, i, load{src, w*i}}
|
||||
trtail{rp, xp, ro}
|
||||
trtail{rpe, xp + y*w + w-1, ra}
|
||||
# Transpose first few rows and last few rows together
|
||||
@for_mult{k} (x to w) {
|
||||
o := w*y + x
|
||||
def loadx{_} = {
|
||||
l:=load{*VT~~(xp+o)}
|
||||
o+=w; if (o>wh-k) o -= wh-1 # Jump from last source row to first, shifting right 1
|
||||
l
|
||||
}
|
||||
def rls = get_lines{loadx} # 4 rows of 2 vectors each
|
||||
each{{i,v} => {if (i<3 or rpo<rpe) store_line{*VT~~rpo, v}; rpo+=h}, iota{k}, rls}
|
||||
}
|
||||
--yn # One strip handled
|
||||
}
|
||||
@for_mult{line_elts} (y0 to yn) { y := y0 + ro
|
||||
@for_mult{k} (x to w) {
|
||||
xpo:= xp + y*w + x
|
||||
rpo:= rp + x*h + y
|
||||
def rls = get_lines{{i} => load{*VT~~(xpo+i*w), 0}}
|
||||
each{{i,v} => store_line{*VT~~(rpo+i*h), v}, iota{k}, rls}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (w%8) emit{void, 'base_transpose_u32', rp+h*(w-w%8), xp+ (w-w%8), w%8, h, w, h}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user