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

[XLA] Avoid fp division by 0 in literal_comparison.cc.

Browse Source 
Found by running hlo_evaluator_test with ubsan.

PiperOrigin-RevId: 226937627
This commit is contained in:
Justin Lebar 2018-12-26 11:32:32 -08:00 committed by TensorFlower Gardener
parent c220247fe3
commit 79126ece73
1 changed files with 8 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tensorflow/compiler/xla/literal_comparison.cc
View File

@ -387,7 +387,14 @@ class NearComparator {
rel_error = std::numeric_limits<float>::infinity();
} else {
abs_error = FpAbsoluteValue(actual - expected);
rel_error = abs_error / FpAbsoluteValue(expected);
// Avoid division by 0 even though it's well-defined because ubsan can be
// configured to treat this as a fatal error.
if (expected != T{0}) {
rel_error = abs_error / FpAbsoluteValue(expected);
} else {
rel_error = std::numeric_limits<float>::infinity();
}
}
const bool is_abs_mismatch = abs_error > error_.abs;
const bool is_rel_mismatch = rel_error > error_.rel;