v1.2.1

Added

• Support more graph optimizations : Convolution+Convolution, LayerNorm
• Support more operators: ROIAlign, GenerateProposals, Reciprocal, Not, Log, ReductionL2, InstanceNorm, Expand, Gather, Scatter
• Support more operators(PReLU) process NCHW input data.
• Support ONNX share weight between Linear, MatMul, Gemm and Gather
• Support more networks on CPU: vision transformer(ViT, TNT), recommendation networks
• Support more networks on GPU : ASR, Faster_RCNN
• Support Armv7 int8 to accelerate NLP network(50%+ speed-up)
• Support X86 AVX512 int8 to accelerate NLP network(3x+ speed-up)
• Support using image on Qualcomm GPU, add GPU image manage methods
• Improve inference performance on Qualcomm GPU
• Add more kit android/iOS demos : Chinese ASR, Face Detection, Sentiment Analysis
• Try to bind core when using GPU

Changed

• Replace mali option with gpu in install shell script, and remove default target option setting
• Change data format NCWHC4 TO NCHWC4 for GPU
• Simplified tensor padding method with OclMemory for GPU
• Tool preprocess_ocl produces algofile and xxxlib.so before, for now algofile has been packaged into this xxxlib.so
• Add BNN_FP16 option in X2bolt tool to convert ONNX 1-bit model
• Replace original INT8 option with INT8_FP16 in post_training_quantization tool to convert int8+float16 hybrid inference model, and add INT8_FP32 option to convert int8+float32 hybrid inference model.
• Add shell environment variable BOLT_INT8_STORAGE_ERROR_THRESHOLD to control post_training_quantization convert int8 model, default value is 0.002. post_training_quantization will use int8 storage when when quantization error lower than BOLT_INT8_STORAGE_ERROR_THRESHOLD.

Fixed

• Fix PReLU 2d, 3d support
• Fix Resize bug on some mode
• Fix ONNX converter read Squeeze, UnSqueeze, Deconv parameter bug
• Fix Arm Sigmoid precision
• Fix ONNX RNN optimizer, and add support for NCHWC8 input data
• Fix Concat with weight tensor in onnx converter
• Simplify C API example