def TestExtended():
class Add10(spe.ExtendedInstruction):
isa_module = ppc
def __init__(self, d, value):
self.d = d
self.value = value
spe.ExtendedInstruction.__init__(self)
return
def block(self):
for i in range(10):
ppc.addi(self.d, self.d, self.value)
return
code = InstructionStream()
proc = Processor()
# Using code.add
code.add(ppc.addi(code.gp_return, 0, 0))
code.add(Add10(code.gp_return, 1))
Add10.ex(1).eval(code, reg=code.gp_return)
code.print_code()
r = proc.execute(code)
print r
assert r == 20
# Using active code
code.reset()
ppc.set_active_code(code)
ppc.addi(code.gp_return, 0, 0)
Add10(code.gp_return, 1)
Add10.ex(1).eval(ppc.get_active_code(), reg=code.gp_return)
code.print_code()
r = proc.execute(code)
print r
assert r == 20
return
def TestExtended():
class Add10(spe.ExtendedInstruction):
isa_module = ppc
def __init__(self, d, value):
self.d = d
self.value = value
spe.ExtendedInstruction.__init__(self)
return
def block(self):
for i in range(10):
ppc.addi(self.d, self.d, self.value)
return
code = InstructionStream()
proc = Processor()
# Using code.add
code.add(ppc.addi(code.gp_return, 0, 0))
code.add(Add10(code.gp_return, 1))
Add10.ex(1).eval(code, reg=code.gp_return)
code.print_code()
r = proc.execute(code)
print r
assert (r == 20)
# Using active code
code.reset()
ppc.set_active_code(code)
ppc.addi(code.gp_return, 0, 0)
Add10(code.gp_return, 1)
Add10.ex(1).eval(ppc.get_active_code(), reg=code.gp_return)
code.print_code()
r = proc.execute(code)
print r
assert (r == 20)
return
def synthesize(self, code, tB, N):
"""
Extract a block from B and pack it for fast access.
tB is transposed.
"""
old_code = ppc.get_active_code()
ppc.set_active_code(code)
code.add_storage(tB)
self._init_constants(code, tB, N)
self._init_vars()
self._load_params()
self._pack_b(code)
ppc.set_active_code(old_code)
return
def synthesize(self, prgm, tB, M, K, N, kc, nc, mr = 1, nr = 1):
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
gepb = SynGEPB(self.gepb_mode)
packb = SynPackB()
gepb._init_constants(M, K, N, kc, nc, mr, nr, True)
packb._init_constants(prgm, tB, N)
gepb._init_vars()
# Reuse the C/C_aux registers for B. They are set in init pointers.
packb._init_vars2(gepb.p_C, gepb.c[0][0], gepb.r_tB_addr)
gepb._load_params()
packb._load_params(pvB = 7)
# kN = k * N * 8
# for j in range(0, N * 8, nc * 8):
for j in syn_iter(code, N, nc):
# # Pack B into tB -- tB1.transpose(B[k:k+kc, j:j+nc])
# pack_params.p1 = B_addr + kN + j # (k * N + j) * 8
packb.vN.v = N
packb._pack_b(code)
# proc.execute(cgepb, params = pm)
gepb._init_pointers()
gepb._gepb(code)
# pm.p3 += nc8
gepb.r_C_addr.v = gepb.r_C_addr + nc * 8
packb.vB.v = packb.vB + nc * 8
# /end for j
ppc.set_active_code(old_code)
return
def synthesize(self, prgm, tB, M, K, N, kc, nc, mr=1, nr=1):
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
gepb = SynGEPB(self.gepb_mode)
packb = SynPackB()
gepb._init_constants(M, K, N, kc, nc, mr, nr, True)
packb._init_constants(prgm, tB, N)
gepb._init_vars()
# Reuse the C/C_aux registers for B. They are set in init pointers.
packb._init_vars2(gepb.p_C, gepb.c[0][0], gepb.r_tB_addr)
gepb._load_params()
packb._load_params(pvB=7)
# kN = k * N * 8
# for j in range(0, N * 8, nc * 8):
for j in syn_iter(code, N, nc):
# # Pack B into tB -- tB1.transpose(B[k:k+kc, j:j+nc])
# pack_params.p1 = B_addr + kN + j # (k * N + j) * 8
packb.vN.v = N
packb._pack_b(code)
# proc.execute(cgepb, params = pm)
gepb._init_pointers()
gepb._gepb(code)
# pm.p3 += nc8
gepb.r_C_addr.v = gepb.r_C_addr + nc * 8
packb.vB.v = packb.vB + nc * 8
# /end for j
ppc.set_active_code(old_code)
return
def synthesize(self, prgm, tB, N):
"""
Extract a block from B and pack it for fast access.
tB is transposed.
"""
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
prgm.add_storage(tB)
self._init_constants(prgm, tB, N)
self._init_vars()
self._load_params()
self._pack_b(code)
ppc.set_active_code(old_code)
return
def synthesize(self, prgm, M, K, N, kc, nc, mr = 1, nr = 1, _transpose = False):
"""
tA is M x nc
tB is nc x kc
C is M x nc
I is the current block column in C
"""
code = prgm.get_stream()
old_code = ppc.get_active_code()
ppc.set_active_code(code)
self._init_constants(M, K, N, kc, nc, mr, nr, _transpose)
self._init_vars()
self._load_params()
self._init_pointers()
self._gepb(code)
ppc.set_active_code(old_code)
return
def synthesize(self, code, M, K, N, kc, nc, mr = 1, nr = 1, _transpose = False):
"""
tA is M x nc
tB is nc x kc
C is M x nc
I is the current block column in C
"""
old_code = ppc.get_active_code()
ppc.set_active_code(code)
self._init_constants(M, K, N, kc, nc, mr, nr, _transpose)
self._init_vars()
self._load_params()
self._init_pointers()
self._gepb(code)
ppc.set_active_code(old_code)
return
def _get_active_code(self):
return ppc.get_active_code()
def _get_active_code(self): return ppc.get_active_code()