...
Integrated MKLDNN for CPU training and inference acceleration acceleration. The high-level design is at The design of MKLDNN integration.
Bug-fixes
Fix I/O multiprocessing for too many open file handles (#8904), race condition (#8995), deadlock (#9126).
Fix image IO integration with OpenCV 3.3 (#8757).
Fix Gluon block printing (#8956).
Fix float16 argmax when there is negative input. (#9149)
Fix random number generator to ensure sufficient randomness. (#9119, #9256, #9300)
Fix custom op multi-GPU scaling (#9283)
Fix gradient of gather_nd when duplicate entries exist in index. (#9200)
...