experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA] Be slightly more relaxed about denormals in exhaustive_op_test.

Browse Source 
Given subnormal x previously the test would accept any of the following as a
"correct" value for f(x).

 - f(x) evaluated at full precision
 - f(copysign(0, x)), i.e. flush denormals to zero

Now we add a third option:

 - f(copysign(minfloat, x)), i.e. flush denormals to the closest *normal* value

This matches the behavior of XLA:CPU's log implementation when it's run with
ftz disabled on the CPU.  It also seems like a reasonable behavior, especially
for log, where this behavior avoids an infinity, which may avoid a nan down the
road.

PiperOrigin-RevId: 247766690
This commit is contained in:
Justin Lebar 2019-05-11 11:13:41 -07:00 committed by TensorFlower Gardener
parent 7bbd0940e8
commit 1cf2df7f98
1 changed files with 22 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
tensorflow/compiler/xla/tests/exhaustive_op_test.cc
View File

@ -338,6 +338,10 @@ class ExhaustiveOpTest
// denormals.
const T expected_at_pos_zero = static_cast<T>(evaluate_op(0));
const T expected_at_neg_zero = static_cast<T>(evaluate_op(-0.0));
const T expected_at_pos_min_normal_float =
static_cast<T>(evaluate_op(std::numeric_limits<float>::min()));
const T expected_at_neg_min_normal_float =
static_cast<T>(evaluate_op(-std::numeric_limits<float>::min()));
for (int64 i = 0; i < input_arr.size(); ++i) {
T input = input_arr[i];
float input_f32 = static_cast<float>(input);
@ -369,13 +373,23 @@ class ExhaustiveOpTest
// - evaluate_op(input)
// - evaluate_op(+/-0), where the sign of 0 equal to the sign of
// `input`,
// - evaluate_op(+/-min_normal_float), where the sign of
// min_normal_float matches `input`.
// - if relaxed_denormal_signs_, evaluate_op(-/+0), where the sign of
// 0 is the opposite of `input`.
//
// (In particular, the XLA:CPU implementation of log flushes positive
// denormals to min-normal-float. This seems kind of reasonable if our
// goal is to avoid infinities because they cause nans?)
T sign_preserving_ftz_expected =
std::signbit(input_f32) ? expected_at_neg_zero : expected_at_pos_zero;
T flush_to_normal_expected = std::signbit(input_f32)
? expected_at_neg_min_normal_float
: expected_at_pos_min_normal_float;
T sign_nonpreserving_ftz_expected =
std::signbit(input_f32) ? expected_at_pos_zero : expected_at_neg_zero;
if (IsClose(sign_preserving_ftz_expected, actual) ||
IsClose(flush_to_normal_expected, actual) ||
(relaxed_denormal_signs_ &&
IsClose(sign_nonpreserving_ftz_expected, actual))) {
continue;
@ -386,11 +400,13 @@ class ExhaustiveOpTest
return absl::StrFormat(
"Mismatch on denormal value %s. Expected one of:\n"
" %10s (evaluated at full-precision value)\n"
" %10s (evaluated at sign-preserving min-normal-float)\n"
" %10s (evaluated after flushing to sign-preserving zero)\n"
" %10s (evaluated after flushing to non-sign-preserving "
"zero)\n"
"but got %s.",
StringifyNum(input), StringifyNum(expected),
StringifyNum(input), //
StringifyNum(expected), StringifyNum(flush_to_normal_expected),
StringifyNum(sign_preserving_ftz_expected),
StringifyNum(sign_nonpreserving_ftz_expected),
StringifyNum(actual));
@ -400,10 +416,13 @@ class ExhaustiveOpTest
return absl::StrFormat(
"Mismatch on denormal value %s. Expected one of:\n"
" %10s (evaluated at full-precision value)\n"
" %10s (evaluated at sign-preserving min-normal-float)\n"
" %10s (evaluated after flushing to sign-preserving zero)\n"
"but got %s.",
StringifyNum(input), StringifyNum(expected),
StringifyNum(sign_preserving_ftz_expected), StringifyNum(actual));
StringifyNum(input), //
StringifyNum(expected), StringifyNum(flush_to_normal_expected),
StringifyNum(sign_preserving_ftz_expected), //
StringifyNum(actual));
});
}
}