Add armclang support for target based on Arm Corstone-300 software
Armclang need to use -ffp-mode=full because quantization unit test is using NaN and without this use of NaN is undefined behavior. Because of a compiler defect which only occurs with -ffp-mode=full a workaround is done in the lstm unit test. The workaround prevents implicit conversion from double to float. Since many of those conversions would be inexact, the compiler would not consider those conversions as constant expressions with -ffp-mode=full. BUG=https://github.com/tensorflow/tflite-micro/issues/1305
Showing
因为 它太大了无法显示 source diff 。你可以改为 查看blob。
想要评论请 注册 或 登录