[BUG] String to float casting mismatch with cpu in different way from compatibility doc #10037
Assignees
Labels
documentation
Improvements or additions to documentation
Comments
thirtiseven
added
bug
mattahrens
added
documentation
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe the bug
Our compatibility doc is outdated after we switched string to float to new kernel.
The results of current code mismatch slightly with cpu but in different way from compatibility doc. I'm not sure if we are aware of this. I think the cpu did things correctly because the cpu results are also matched with python.
Steps/Code to reproduce bug
Compatibility doc:
My test code
Hack into DoubleGen to make it produce string:
class DoubleGen(DataGen): """Generate doubles, which some built in corner cases.""" def __init__(self, min_exp=DOUBLE_MIN_EXP, max_exp=DOUBLE_MAX_EXP, no_nans=False, nullable=True, special_cases = None): self._min_exp = min_exp self._max_exp = max_exp self._no_nans = no_nans self._use_full_range = (self._min_exp == DOUBLE_MIN_EXP) and (self._max_exp == DOUBLE_MAX_EXP) if special_cases is None: special_cases = [ self.make_from(1, self._max_exp, DOUBLE_MAX_FRACTION), self.make_from(0, self._max_exp, DOUBLE_MAX_FRACTION), self.make_from(1, self._min_exp, DOUBLE_MAX_FRACTION), self.make_from(0, self._min_exp, DOUBLE_MAX_FRACTION) ] if self._min_exp <= 0 and self._max_exp >= 0: special_cases.append(0.0) special_cases.append(-0.0) if self._min_exp <= 3 and self._max_exp >= 3: special_cases.append(1.0) special_cases.append(-1.0) if not no_nans: special_cases.append(float('inf')) special_cases.append(float('-inf')) special_cases.append(float('nan')) special_cases.append(NEG_DOUBLE_NAN_MAX_VALUE) - super().__init__(DoubleType(), nullable=nullable, special_cases=special_cases) + super().__init__(StringType(), nullable=nullable, special_cases=special_cases) def _cache_repr(self): return super()._cache_repr() + '(' + str(self._min_exp) + ',' + str(self._max_exp) + ',' + str(self._no_nans) + ')' @staticmethod def make_from(sign, exp, fraction): sign = sign & 1 # 1 bit exp = (exp + 1023) & 0x7FF # add bias and 11 bits fraction = fraction & DOUBLE_MAX_FRACTION i = (sign << 63) | (exp << 52) | fraction p = struct.pack('L', i) ret = struct.unpack('d', p)[0] return ret def _fixup_nans(self, v): if self._no_nans and (math.isnan(v) or v == math.inf or v == -math.inf): v = None if self.nullable else 0.0 return v def start(self, rand): if self._use_full_range: def gen_double(): i = rand.randint(LONG_MIN, LONG_MAX) p = struct.pack('l', i) - return self._fixup_nans(struct.unpack('d', p)[0]) + return str(self._fixup_nans(struct.unpack('d', p)[0])) self._start(rand, gen_double) else: def gen_part_double(): sign = rand.getrandbits(1) exp = rand.randint(self._min_exp, self._max_exp) fraction = rand.getrandbits(52) - return self._fixup_nans(self.make_from(sign, exp, fraction)) + return str(self._fixup_nans(self.make_from(sign, exp, fraction))) self._start(rand, gen_part_double)and run test:
Results (part)
Expected behavior
At least we can note how it mismatches with cpu in the compatibility doc. We may also want to fix it to fully match cpu one day.
The text was updated successfully, but these errors were encountered: