Overlap SIMD transpose to handle uneven heights (not widths yet)
This commit is contained in:
parent
fad7f3aa8b
commit
8ff40cb2e5
@ -65,31 +65,43 @@ def for_mult{k}{vars,begin,end,block} = {
|
||||
fn transpose{T, k, kh}(r0:*void, x0:*void, w:u64, h:u64) : void = {
|
||||
# Scalar transpose defined in C
|
||||
def ts = if (T==i8) 'i8' else if (T==i16) 'i16' else if (T==i32) 'i32' else 'i64'
|
||||
def call_base{...a} = emit{void, merge{'base_transpose_',ts}, ...a}
|
||||
def call_base{...a} = emit{void, merge{'base_transpose_',ts}, ...a, w, h}
|
||||
|
||||
rp:*T = *T~~r0
|
||||
xp:*T = *T~~x0
|
||||
if (w<k or h<k) { call_base{rp, xp, w, h, w, h}; return{} }
|
||||
if (w<k or h<k) { call_base{rp, xp, w, h}; return{} }
|
||||
|
||||
def at{x,y} = tup{xp + y*w + x, rp + x*h + y}
|
||||
|
||||
# Cache line info
|
||||
def line_bytes = 64
|
||||
def line_elts = line_bytes / (width{T}/8)
|
||||
|
||||
def use_overlap{o} = o >= 2 # For overlapped SIMD instead of scalar
|
||||
|
||||
# Handle uneven height (extra rows) here, but not uneven width
|
||||
if (line_elts > 2*k or h&(line_elts-1) != 0) {
|
||||
ho := h%k
|
||||
# Effective height: number of rows read, counting overlap twice
|
||||
# Just use base transpose for short overhang; otherwise round up
|
||||
he := h; if (use_overlap{ho}) he += k - ho
|
||||
# Main transpose
|
||||
@for_mult{kh} (y to h) {
|
||||
hm := h - kh
|
||||
@for_mult{kh} (y0 to he) { y:=y0; if (y>hm) y = hm
|
||||
@for_mult{k} (x to w) {
|
||||
kernel{...at{x,y}, k, kh, w, h}
|
||||
}
|
||||
}
|
||||
# Extra row for uneven i16 case
|
||||
if (2*k == kh and (h & k) != 0) { y := h-h%kh
|
||||
# Half-row for non-square i16 case
|
||||
if (2*k == kh and (he & k) != 0) {
|
||||
e := h%kh; if (e<k) e = k
|
||||
y := h - e
|
||||
@for_mult{k} (x to w) {
|
||||
kernel{...at{x,y}, k, k, w, h}
|
||||
}
|
||||
}
|
||||
# Base transpose used if overlap wasn't
|
||||
if (ho!=0 and he==h) { hs := h-ho; call_base{rp+hs, xp+w*hs, w-w%k, ho} }
|
||||
} else {
|
||||
# Result rows are aligned with each other so it's possible to
|
||||
# write a full cache line at a time
|
||||
@ -137,9 +149,7 @@ fn transpose{T, k, kh}(r0:*void, x0:*void, w:u64, h:u64) : void = {
|
||||
}
|
||||
}
|
||||
|
||||
def edge_tr{...a} = call_base{...a, w, h}
|
||||
wo := w%k; ws := w-wo; if (wo) edge_tr{rp+h*ws, xp+ ws, wo, h }
|
||||
ho := h%k; hs := h-ho; if (ho) edge_tr{rp+ hs, xp+w*hs, ws, ho}
|
||||
wo := w%k; if (wo!=0) { ws := w-wo; call_base{rp+h*ws, xp+ws, wo, h } }
|
||||
}
|
||||
|
||||
def transpose{T, k} = transpose{T, k, k}
|
||||
|
||||
Loading…
Reference in New Issue
Block a user