Notes:
- Training in fp16 (half precision) including mixed-precision is now fully supported
- Batch Normalization in fp16 (half precision) including mixed-precision are now available
- Performance improvements for 3x3 and 1x1 single-precision convolutions
- Layer fusions for BatchNorm+Activation are now available
- Layer fusions with convolutions now support varying strides and padding configurations
Changes:
- rocBLAS is now used as the default BLAS library for the HIP backend (minimum version 14.3.0)
- Fixed various bugs in convolution kernels
- Fixed issues with bad references in layer fusion
- Fixed gfx803 assembily issues
- Added support fp16 Winograd convolutions
- Added support for fp16 pooling
- Improved error reporting for convolutions and layer fusions
- Improved documentation
Known Issues:
- RNNs do not support fp16
- OpenCL backend does not have full fp16 support
- Layer fusions for convolution 1x1 fp16 are not supported