This project is no longer maintained and has been moved to https://github.com/airockchip/rknn-toolkit2/tree/master/rknpu2
RKNPU2 provides an advanced interface to access Rockchip NPU.
- RK3566/RK3568
- RK3588/RK3588S
- RV1103/RV1106
- RK3562
Note: The rknn model must be generated using RKNN Toolkit 2: https://github.com/rockchip-linux/rknn-toolkit2
For RK1808/RV1109/RV1126/RK3399Pro, please use:
https://github.com/rockchip-linux/rknn-toolkit
https://github.com/rockchip-linux/rknpu
https://github.com/airockchip/RK3399Pro_npu
- Improved dynamic shape support
- Improved matmul api support
- Add GPU back-end implementations for some operators such as matmul
- Improve transformer support
- Reduce rknn_init memory usage
- Optimize rknn_init time-consuming
- Support RK3562
- Support more NPU operator fuse, such as Conv-Silu/Conv-Swish/Conv-Hardswish/Conv-sigmoid/Conv-HardSwish/Conv-Gelu ..
- Improve support for NHWC output layout
- RK3568/RK3588:The maximum input resolution up to 8192
- Improve support for Swish/DataConvert/Softmax/Lstm/LayerNorm/Gather/Transpose/Mul/Maxpool/Sigmoid/Pad
- Improve support for CPU operators (Cast, Sin, Cos, RMSNorm, ScalerND, GRU)
- Limited support for dynamic resolution
- Provide MATMUL API
- Add RV1103/RV1106 rknn_server application as proxy between PC and board
- Add more examples such as rknn_dynamic_shape_input_demo and video demo for yolov5
- Bug fix
- Support more NPU operators, such as Reshape、Transpose、MatMul、 Max、Min、exGelu、exSoftmax13、Resize etc.
- Add Weight Share function, reduce memory usage.
- Add Weight Compression function, reduce memory and bandwidth usage.(RK3588/RV1103/RV1106)
- RK3588 supports storing weights or feature maps on SRAM, reducing system bandwidth consumption.
- RK3588 adds the function of running a single model on multiple cores at the same time.
- Add new output layout NHWC (C has alignment restrictions) .
- Improve support for non-4D input.
- Add more examples such as rknn_yolov5_android_apk_demo and rknn_internal_mem_reuse_demo.
- Bug fix.
- Support RV1103/RV1106(Beta SDK)
- rknn_tensor_attr support w_stride(rename from stride) and h_stride
- Rename rknn_destroy_mem()
- Support more NPU operators, such as Where, Resize, Pad, Reshape, Transpose etc.
- RK3588 support multi-batch multi-core mode
- When RKNN_LOG_LEVEL=4, it supports to display the MACs utilization and bandwidth occupation of each layer.
- Bug fix
- Support RK3588
- Support more operators, such as GRU、Swish、LayerNorm etc.
- Reduce memory usage
- Improve zero-copy interface implementation
- Bug fix
- Support INT8+FP16 mixed quantization to improve model accuracy
- Support specifying input and output dtype, which can be solidified into the model
- Support multiple inputs of the model with different channel mean/std
- Improve the stability of multi-thread + multi-process runtime
- Support flashing cache for fd pointed to internal tensor memory which are allocated by users
- Improve dumping internal layer results of the model
- Add rknn_server application as proxy between PC and board
- Support more operators, such as HardSigmoid、HardSwish、Gather、ReduceMax、Elu
- Add LSTM support (structure cifg and peephole are not supported, function: layernormal, clip is not supported)
- Bug fix
- Optimize the performance of rknn_inputs_set()
- Add more functions for zero-copy
- Add new OP support, see OP support list document for details.
- Add multi-process support
- Support per-channel quantitative model
- Bug fix
- Optimize the performance of rknn_inputs_set(), especially for models whose input width is 8-byte aligned.
- Add new OP support, see OP support list document for details.
- Bug fix
- Initial version