NVDLA 1.2.0
Release 1.2.0
NVDLA 1.2.0 release is focused on INT8 precision and nv_large/nv_small configurations support along with performance optimizations
New features and improvements
- INT8 precision support
- Per-tensor and per-kernel scale factors
- Symmetric scaling
- nv_small and nv_large configuration support
- FP16 winograd support
- SDP fusion performance feature
Fixes and other changes
- Failures in deconvolution for number of groups greater than one
- Disable scale+bias fusion to BN as it is not required with SDP fusion enabled
- Disable winograd for output layers of network as it causes size mismatch due to special requirements of winograd
- Return error for pooling layers with pad size greater than kernel size
Known limitations
- An INT8 calibration tool is not provided
- More details about supplying INT8 scale factors to the NVDLA Compiler can be found in Low precision support in NVDLA