Allow an 8x8 row for 2-byte transpose

This commit is contained in:
Marshall Lochbaum 2023-03-20 21:43:34 -04:00
parent 239c765998
commit 0e2e315e51
2 changed files with 14 additions and 4 deletions

View File

@ -1340,7 +1340,7 @@ B transp_c1(B t, B x) {
case el_i8: case el_c8: { u8* xp=tyany_ptr(x); u8* rp = m_tyarrp(&r,1,ia,el2t(xe)); PLAINLOOP for(usz y=0;y<h;y++) NOVECTORIZE for(usz x=0;x<w;x++) rp[x*h+y] = xp[xi++]; break; }
case el_i16:case el_c16:
#if SINGELI_X86_64
if (w>=8 && h>=16) { u16* xp=tyany_ptr(x); u16* rp = m_tyarrp(&r,4,ia,el2t(xe)); simd_transpose_i16(rp, xp, w, h); break; }
if (w>=8 && h>=8) { u16* xp=tyany_ptr(x); u16* rp = m_tyarrp(&r,4,ia,el2t(xe)); simd_transpose_i16(rp, xp, w, h); break; }
#endif
{ u16* xp=tyany_ptr(x); u16* rp = m_tyarrp(&r,2,ia,el2t(xe)); PLAINLOOP for(usz y=0;y<h;y++) NOVECTORIZE for(usz x=0;x<w;x++) rp[x*h+y] = xp[xi++]; break; }
case el_i32:case el_c32:

View File

@ -61,7 +61,7 @@ fn transpose{T, k}(r0:*void, x0:*void, w:u64, h:u64) : void = {
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}
}
}
}
} else {
# Result rows are aligned with each other so it's possible to
@ -132,12 +132,22 @@ fn transpose2{T, k & T < i32}(r0:*void, x0:*void, w:u64, h:u64) : void = {
def xvs = each{{i}=>{p:=xpo+i*w; load2{p, p+k*w}}, iota{k}}
def rvs = vtranspose2{xvs}
each{{i,v}=>store{*VT~~(rpo+i*h), 0, v}, iota{k}, rvs}
}
}
}
if ((h & k) != 0) { y := h-h%d
@for_mult{k} (x to w) {
xpo:= xp + y*w + x
rpo:= rp + x*h + 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}
}
}
def base = 'base_transpose_i16'
if (w%k) emit{void, base, rp+h*(w-w%k), xp+ (w-w%k), w%k, h, w, h}
if (h%d) emit{void, base, rp+ (h-h%d), xp+w*(h-h%d), w-w%k, h%d, w, h}
if (h%k) emit{void, base, rp+ (h-h%k), xp+w*(h-h%k), w-w%k, h%k, w, h}
}
export{'simd_transpose_i16', transpose2{i16, 8}}