def test_fp_mul():
# Regression test for https://github.com/StanfordAHA/lassen/issues/111
inst = asm.fp_mul()
data0 = Data(0x4040)
data1 = Data(0x4049)
res, res_p, _ = pe(inst, data0, data1)
if CAD_ENV:
rtl_tester(inst, data0, data1, res=res)
else:
pytest.skip("Skipping since DW not available")
def __call__(self, in0: Data, in1: Data) -> Data:
inst1 = asm.fgetmant()
inst2 = asm.fsubexp()
inst3 = asm.fp_mul()
rom_instr = mem_asm.rom([TLUT.div_lut(i) for i in range(0, 128)] +
[0x0000] * (depth - 128))
op_a = in0
op_b = in1
mant, _, _ = self.pe_get_mant(inst1, op_b, Data(0))
lookup_result = self.rom(rom_instr, mant, Data(0))
scaled_result, _, _ = self.pe_scale_res(inst2, lookup_result, op_b)
result, _, _ = self.pe_mult(inst3, scaled_result, op_a)
return result
def __call__(self, in0: Data) -> Data:
inst1 = asm.fgetmant()
inst2 = asm.fcnvexp2f()
inst3 = asm.fp_mul()
inst4 = asm.fp_add()
rom_instr = mem_asm.rom([TLUT.ln_lut(i) for i in range(0, 128)] +
[0x0000] * (depth - 128))
op_a = in0
ln2 = math.log(2)
ln2_bf = int(float2bfbin(ln2), 2)
const_ln2 = Data(ln2_bf)
mant, _, _ = self.pe_get_mant(inst1, op_a, Data(0))
fexp, _, _ = self.pe_get_exp(inst2, op_a, Data(0))
lookup_result = self.rom(rom_instr, mant, Data(0))
mult, _, _ = self.pe_mult(inst3, fexp, const_ln2)
result, _, _ = self.pe_mult(inst4, lookup_result, mult)
return result
def __call__(self, in0: Data) -> Data:
# Perform op_a/ln(2)
inst1 = asm.fp_mul()
# Compute 2**op_a
inst2 = asm.fgetfint()
inst3 = asm.fgetffrac()
inst4 = asm.and_()
inst5 = asm.faddiexp()
rom_instr = mem_asm.rom([TLUT.exp_lut(i) for i in range(0, 128)] +
[TLUT.exp_lut(i) for i in range(-128, 0)] +
[0x0000] * (depth - 256))
op_a = in0
ln2_inv = 1.0 / math.log(2)
ln2_inv_bf = int(float2bfbin(ln2_inv), 2)
const_ln2_inv = Data(ln2_inv_bf)
div_res, _, _ = self.pe_div_mult(inst1, const_ln2_inv, op_a)
fint, _, _ = self.pe_get_int(inst2, div_res, Data(0))
ffrac, _, _ = self.pe_get_frac(inst3, div_res, Data(0))
idx, _, _ = self.pe_rom_idx(inst4, ffrac, Data(0xFF))
lookup_result = self.rom(rom_instr, idx, Data(0))
result, _, _ = self.pe_incr_exp(inst5, lookup_result, fint)
return result
def _make_random(cls):
if issubclass(cls, hwtypes.BitVector):
return cls.random(len(cls))
if issubclass(cls, hwtypes.FPVector):
while True:
val = cls.random()
if val.fp_is_normal():
return val.reinterpret_as_bv()
return NotImplemented
_CAD_DIR = "/cad/synopsys/syn/P-2019.03/dw/sim_ver/"
_EXPENSIVE = {
"bits32.mul": ((umult0(),), "magma_Bits_32_mul_inst0", hwtypes.UIntVector[16]), # noqa
"bfloat16.mul": ((fp_mul(),), "magma_BFloat_16_mul_inst0", BFloat16_fc(PyFamily())), # noqa
"bfloat16.add": ((fp_add(),), "magma_BFloat_16_add_inst0", BFloat16_fc(PyFamily())), # noqa
}
@pytest.mark.parametrize("op", list(_EXPENSIVE.keys()))
def test_pe_data_gate(op, dw_files):
instrs, fu, BV = _EXPENSIVE[op]
is_float = issubclass(BV, hwtypes.FPVector)
if not irun_available() and is_float:
pytest.skip("Need irun to test fp ops")
core = PeakCore(PE_fc)
core.name = lambda: "PECore"
circuit = core.circuit()
def _make_random(cls):
if issubclass(cls, hwtypes.BitVector):
return cls.random(len(cls))
if issubclass(cls, hwtypes.FPVector):
while True:
val = cls.random()
if val.fp_is_normal():
return val.reinterpret_as_bv()
return NotImplemented
_EXPENSIVE = {
"bits32.mul":
((umult0(), ), "magma_Bits_32_mul_inst0", hwtypes.UIntVector[16]), # noqa
"bfloat16.mul": ((fp_mul(), ), "magma_BFloat_16_mul_inst0",
BFloat16_fc(PyFamily())), # noqa
"bfloat16.add": ((fp_add(), ), "magma_BFloat_16_add_inst0",
BFloat16_fc(PyFamily())), # noqa
}
@pytest.mark.parametrize("op", list(_EXPENSIVE.keys()))
def test_pe_data_gate(op, run_tb):
instrs, fu, BV = _EXPENSIVE[op]
is_float = issubclass(BV, hwtypes.FPVector)
if not irun_available() and is_float:
pytest.skip("Need irun to test fp ops")
# note to skip mul since CW BFloat is faulty
#
def BV(val):
return BFloat16(val)
fp_sign_vec = [BV(2.0), BV(-2.0), BV(3.0), BV(-3.0)]
fp_zero_vec = [BV(0.0), BV('-0.0')]
fp_inf_vec = [BV('inf'), BV('-inf')]
@pytest.mark.parametrize("op", [
op(asm.fp_add(), lambda x, y: x + y),
op(asm.fp_sub(), lambda x, y: x - y),
op(asm.fp_mul(), lambda x, y: x * y)
])
@pytest.mark.parametrize(
"args", [(BFloat16.random(), BFloat16.random()) for _ in range(NTESTS)] +
list(product(fp_sign_vec + fp_zero_vec, fp_sign_vec + fp_zero_vec)))
def test_fp_binary_op(op, args):
inst = op.inst
in0 = args[0]
in1 = args[1]
out = op.func(in0, in1)
data0 = BFloat16.reinterpret_as_bv(in0)
data1 = BFloat16.reinterpret_as_bv(in1)
res, res_p, _ = pe(inst, data0, data1)
assert res == BFloat16.reinterpret_as_bv(out)
if CAD_ENV:
rtl_tester(op, data0, data1, res=res)