commit | 8e53547454aa8e8a335b25d027246977d2041fe3 | [log] [tgz] |
---|---|---|
author | Przemyslaw Szczepaniak <pszczepaniak@google.com> | Thu May 02 10:03:44 2019 +0100 |
committer | Przemyslaw Szczepaniak <pszczepaniak@google.com> | Thu May 02 16:04:04 2019 +0100 |
tree | f7b70099ddb51c0df959c39225c10135efa3332e | |
parent | 33c1e0537b438a6cfc83ccd25c0df42d89764fa4 [diff] |
Calculate quantized convolution multiplier using double type. NNAPI has been using the float type for quantized convolution multiplier calculation, which was causing an occasional off-by-one differences for convolution operation results between TFLite/CPU and NNAPI/CPU implementations. Test: MLTS com.android.nn.benchmark.app.NNTest#testNNAPI[mobilenet_*_quant] error rates Test: NeuralNetworksTest_static Bug: 115377572 Change-Id: If0bb9ff7229f5b3864a69c66566a4e4008f94c37