projects / platform / upstream / tensorflow.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: 50e1888) | patch
[XLA] Accurately measure FLOPs for base-dilated convolutions
authorDavid Majnemer <majnemer@google.com>
Wed, 28 Mar 2018 00:35:14 +0000 (17:35 -0700)
committerTensorFlower Gardener <gardener@tensorflow.org>
Wed, 28 Mar 2018 00:37:54 +0000 (17:37 -0700)
commitec3123979cd622a70fc809ee98b5a7b35772f369
treea1b5d5220066daaec3b3420d538a34b6d13a84dbtree | snapshot
parent50e1888fa89bce621e988a92ede3dc362e37b248commit | diff
[XLA] Accurately measure FLOPs for base-dilated convolutions

We incorrectly counted FLOPs when the output and kernel line up to access the
padding or the dilated area. These should not be accounted as contributing to
the FLOP count.

PiperOrigin-RevId: 190702384
tensorflow/compiler/xla/service/BUILD diff | blob | history
tensorflow/compiler/xla/service/hlo_cost_analysis.cc diff | blob | history
tensorflow/compiler/xla/service/hlo_cost_analysis_test.cc diff | blob | history
Domain: Machine Learning / ML Framework;