def test_alloc():
c = get_core()
(r21,s21,w01,w2x) = c.acquire_fpregisters(range(4))
(i0,i1,ir0,is0,sixteen) = map(IntRegister,range(5))
c.name_registers(a21=FPRegister(4),b21=FPRegister(5))
istream = [
# weights
isa.fpset2(w01,1/9,2/9),
isa.fpset2(w2x,1/9,-1),
# preamble
isa.lfpd('a21',i0,0), # A[0],A[1]
isa.lfdu('a21',i0,16), # A[2],A[1]
isa.lfpd('b21',i1,0), # A[0],A[1]
isa.lfdu('b21',i1,16), # A[2],A[1]
# start loads
isa.fxcpmadd(r21,w01,'a21',r21),
isa.fxcpmadd(s21,w01,'b21',s21),
isa.lfpd('a23',i0,16),
isa.lfpd('b23',i1,16),
isa.fxcxma(r21,w01,'a23',r21),
isa.fxcxma(s21,w01,'b23',s21),
isa.lfdu('a23',i0,16), # a43
isa.lfdu('b23',i1,16), # b43
isa.intset(ir0,8*8),
isa.intset(is0,18*8),
isa.intset(sixteen,16),
isa.stfxdux(r21,ir0,sixteen),
isa.stfxdux(s21,is0,sixteen),
]
#for instr in istream: print(instr); #c.trace = c.trace_none
c.schedule(istream)
c.execute([isa.inspect()])
def stencil():
def label(c,i,j,kp,ks):
return '%s_%d_%d_%d%d' % (c,i,j,kp,ks)
def stream(i,j):
def a(kp,ks): return label('a',i,j,kp,ks)
p = 'p_%d_%d' % (i,j)
yield isa.lfpd(a(0,1),p,0)
for k in (2,4,6):
yield isa.lfdu(a(k,k-1),p,16)
yield isa.lfpd(a(k,k+1),p,0)
def jam(i,j):
def r(kp,ks): return label('r',i,j,kp,ks)
for k in (2,4,6):
rr = r(k,k-1)
for ii in (-1,0,1):
for jj in (-1,0,1):
def a(kp,ks): return label('a',i+ii,j+jj,kp,ks)
yield isa.fxcpmadd(rr,'w01',a(k,k-1),rr)
yield isa.fxcxma(rr,'w01',a(k,k+1),rr)
yield isa.fxcpmadd(rr,'w2x',a(k+2,k+1),rr)
instrs = []
for i in (0,1,2):
for j in (0,1,2):
instrs.append(stream(i,j))
instrs.append(jam(1,1))
yield isa.fpset2('w01',1/9,2/9)
yield isa.fpset2('w2x',1/9,9)
while True:
for ins in instrs:
yield next(ins)
def s_weights(w01,w2x):
return [isa.fpset2(w01,1/9,2/9), isa.fpset2(w2x,1/9,-1)]