1.1 DNN Operator Library Development: Batch-norm/Convert/Broadcast Design the Parallel split scheme and vectorization parallel scheme, Cache Management: Set pingpong buffer in L3/L2/L1 data transfer, Pipeline Setup: Unroll and packet instructions, reduce the slot stall 1.2 Aten Operator Library Development: Pooling/Resize Generalize Dataflow template to Pooling/Resize: Thread parallel split + L3->L1 slice + Cache pipeline Refinement Work on ECC problem and Scalar calculation part to reach 60-80% hardware bandwidth peak 1.3 Model Performance Tuning Support: Element-wise operator fusion work: for LSTM structure, fuse the rest activation and other element-wise ops in L1 level Reduce the L3-L3 copy demand: fuse the the DMA operators (slice broadcast gather) into L2/L1 level 1.4 Support an outer-source team(8 member) Skills/knowledge Training for Operator development Technique support for operator implementation design, performance analysis and mismatch problem debug work 1.5 Generalized operator accuracy validation tool: Use Statistical The Central Limit Theorem to calculate the tolerance limit for convolution/matmul/reduce op accuracy validation according to data type, input data distribution, input shape and statistical tolerance feature.

participating in establising the GPU cluster monitor platform for CV training.

transplant the CPU codes with CUDA onto Nvidia V100 cards. Detailed refinement on the Share memory usage and Key for-loop rewrite work.