diff --git a/src/singeli/src/transpose.singeli b/src/singeli/src/transpose.singeli index c77cbba9..82843b76 100644 --- a/src/singeli/src/transpose.singeli +++ b/src/singeli/src/transpose.singeli @@ -40,6 +40,11 @@ def vtranspose2{x & ktest{'X86_64',8,[16]i16}{x}} = { def r = unpack_pass{4, unpack_pass{2, unpack_pass{1, x}}} each{bind{~~,[16]i16}, r} } +# Transpose 16x16 packed as halves +def vtranspose{x & ktest{'X86_64',8,[32]i8}{x}} = { + def r = unpack_pass{4, unpack_pass{2, unpack_pass{1, x}}} + each{{v}=>[32]i8~~shuf{[4]i64, v, 4b3120}, r} +} def load2{a:T, b:T & w128i{eltype{T}}} = { def V = eltype{T} emit{[2*vcount{V}](eltype{V}), '_mm256_loadu2_m128i', b, a} @@ -48,6 +53,35 @@ def store2{a:T, b:T, v:T2 & w128i{eltype{T}} & w256{T2}} = { each{{p, i} => store{p, 0, half{v, i}}, tup{a,b}, iota{2}} } +def kernel{src:P, dst:P, k, k, w, h & k*width{eltype{P}}==256} = { + def VT = [k](eltype{P}) + def xvs = each{{i}=>load{*VT~~(src+i*w), 0}, iota{k}} + def rvs = vtranspose{xvs} + each{{i,v}=>store{*VT~~(dst+i*h), 0, v}, iota{k}, rvs} +} +def kernel{src:P, dst:P, k, k, w, h & k*width{eltype{P}}==128} = { + def VT = [k](eltype{P}) + def s = k/2 + def xvs = each{{i}=>{p:=src+i*w; load2{*VT~~p, *VT~~(p+s*w)}}, iota{s}} + def rvs = vtranspose{xvs} + each{{i,v}=>{p:=dst+2*i*h; store2{*VT~~p, *VT~~(p+h), v}}, iota{s}, rvs} +} +def kernel{src:P, dst:P, k, d, w, h & d==2*k & d*width{eltype{P}}==256} = { + def HT = [k](eltype{P}) + def VT = [d](eltype{P}) + def xvs = each{{i}=>{p:=src+i*w; load2{*HT~~p, *HT~~(p+k*w)}}, iota{k}} + def rvs = vtranspose2{xvs} + each{{i,v}=>store{*VT~~(dst+i*h), 0, v}, iota{k}, rvs} +} +# 8x8 i16 with a wasted half: would be better to store as 4x16 +def kernel{src:P, dst:P, k==8, k, w, h & eltype{P}==i16} = { + def T = i16 + def VT = [2*k]i16 + def lw = k*width{T} + def xvs = each{{i}=>loadLow{*VT~~(src+i*w), lw}, iota{k}} + def rvs = vtranspose2{xvs} + each{{i,v}=>storeLow{*VT~~(dst+i*h), lw, v}, iota{k}, rvs} +} def for_mult{k}{vars,begin,end,block} = { @@ -57,6 +91,14 @@ def for_mult{k}{vars,begin,end,block} = { def mat_at{rp,xp,w,h}{x,y} = tup{xp + y*w + x, rp + x*h + y} +def transpose_kernels{kw, kh, rp, xp, w, h} = { + @for_mult{kh} (y to h) { + @for_mult{kw} (x to w) { + kernel{...mat_at{rp,xp,w,h}{x,y}, kw, kh, w, h} + } + } +} + # Scalar transpose defined in C def call_base{T} = { def ts = if (T==i8) 'i8' else if (T==i16) 'i16' else if (T==i32) 'i32' else 'i64' @@ -75,7 +117,6 @@ fn transpose{T, k}(r0:*void, x0:*void, w:u64, h:u64) : void = { rp:*T = *T~~r0 xp:*T = *T~~x0 small_transpose_out{T, k, rp, xp, w, h} - def at = mat_at{rp,xp,w,h} def VT = [k]T # Cache line info @@ -83,15 +124,9 @@ fn transpose{T, k}(r0:*void, x0:*void, w:u64, h:u64) : void = { def line_elts = line_bytes / (width{T}/8) def line_vecs = line_bytes / (width{VT}/8) - if (h&(line_elts-1) != 0) { - @for_mult{k} (y to h) { - @for_mult{k} (x to w) { - {xpo,rpo} := at{x, y} - def xvs = each{{i}=>load{*VT~~(xpo+i*w), 0}, iota{k}} - def rvs = vtranspose{xvs} - each{{i,v}=>store{*VT~~(rpo+i*h), 0, v}, iota{k}, rvs} - } - } + if (line_elts > 2*k or h&(line_elts-1) != 0) { + # Main transpose + transpose_kernels{k, k, rp, xp, w, h} } else { # Result rows are aligned with each other so it's possible to # write a full cache line at a time @@ -129,7 +164,7 @@ fn transpose{T, k}(r0:*void, x0:*void, w:u64, h:u64) : void = { } @for_mult{line_elts} (y0 to yn) { y := y0 + ro @for_mult{k} (x to w) { - {xpo,rpo} := at{x, y} + {xpo,rpo} := mat_at{rp,xp,w,h}{x, 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} } @@ -143,50 +178,13 @@ fn transpose{T, k, m==2}(r0:*void, x0:*void, w:u64, h:u64) : void = { rp:*T = *T~~r0 xp:*T = *T~~x0 small_transpose_out{T, k, rp, xp, w, h} - def at = mat_at{rp,xp,w,h} def d = m*k - def VT = [d]T - def HT = [k]T - @for_mult{d} (y to h) { - @for_mult{k} (x to w) { - {xpo, rpo} := at{x, y} - def xvs = each{{i}=>{p:=xpo+i*w; load2{*HT~~p, *HT~~(p+k*w)}}, iota{k}} - def rvs = vtranspose2{xvs} - each{{i,v}=>store{*VT~~(rpo+i*h), 0, v}, iota{k}, rvs} - } - } + transpose_kernels{k, d, rp, xp, w, h} + if ((h & k) != 0) { y := h-h%d @for_mult{k} (x to w) { - {xpo, rpo} := at{x, y} - def lw = k*width{T} - def xvs = each{{i}=>loadLow{*VT~~(xpo+i*w), lw}, iota{k}} - def rvs = vtranspose2{xvs} - each{{i,v}=>storeLow{*VT~~(rpo+i*h), lw, v}, iota{k}, rvs} - } - } - - edge_transpose{T, k, rp, xp, w, h} -} - -def vtranspose{x & ktest{'X86_64',8,[32]i8}{x}} = { - def r = unpack_pass{4, unpack_pass{2, unpack_pass{1, x}}} - each{{v}=>[32]i8~~shuf{[4]i64, v, 4b3120}, r} -} -fn transpose{T==i8, k}(r0:*void, x0:*void, w:u64, h:u64) : void = { - rp:*T = *T~~r0 - xp:*T = *T~~x0 - small_transpose_out{T, k, rp, xp, w, h} - def at = mat_at{rp,xp,w,h} - def VT = [k]T - - @for_mult{k} (y to h) { - @for_mult{k} (x to w) { - {xpo, rpo} := at{x, y} - def s = k/2 - def xvs = each{{i}=>{p:=xpo+i*w; load2{*VT~~p, *VT~~(p+s*w)}}, iota{s}} - def rvs = vtranspose{xvs} - each{{i,v}=>{p:=rpo+2*i*h; store2{*VT~~p, *VT~~(p+h), v}}, iota{s}, rvs} + kernel{...mat_at{rp,xp,w,h}{x,y}, k, k, w, h} } }