unroll bitwise arith
This commit is contained in:
parent
e3239d29bd
commit
ad73e075e0
@ -130,7 +130,7 @@ def arithAAimpl{vw, mode, F, W, X, R, w, x, r, len} = {
|
|||||||
if (R==u1) {
|
if (R==u1) {
|
||||||
def bulk = vw/64
|
def bulk = vw/64
|
||||||
def VT = [bulk]u64
|
def VT = [bulk]u64
|
||||||
@for_masked{bulk}(r in tup{'g',*u64~~r}, cw in tup{VT,*u64~~w}, cx in tup{VT,*u64~~x} over cdiv{len, 64}) r{F{cw,cx}}
|
@for_mu{bulk, 4}(r in tup{'g',*u64~~r}, cw in tup{VT,*u64~~w}, cx in tup{VT,*u64~~x} over cdiv{len, 64}) r{each{F,cw,cx}}
|
||||||
} else if (same{F,__mul} and W!=u1 and X==u1 and W==R) { # 0‿1‿1‿1‿1‿0‿1‿1×3‿1‿4‿1‿5‿9‿2‿6
|
} else if (same{F,__mul} and W!=u1 and X==u1 and W==R) { # 0‿1‿1‿1‿1‿0‿1‿1×3‿1‿4‿1‿5‿9‿2‿6
|
||||||
def bulk = vw / width{W}
|
def bulk = vw / width{W}
|
||||||
def TU = ty_u{R}
|
def TU = ty_u{R}
|
||||||
|
|||||||
Loading…
Reference in New Issue
Block a user