Skip to content

Latest commit

 

History

History
159 lines (103 loc) · 4.85 KB

CHANGELOG.md

File metadata and controls

159 lines (103 loc) · 4.85 KB
Raw

CHANGELOG

v2.2.0

  • Support installation via pip
  • Optimize transformer model performance
  • Support Python 3.12
  • Operator optimization, such as softmax, hardmax, MatMul, etc.

v2.1.0

  • Support RV1103B (Beta)
  • Support RK2118 (Beta)
  • Support Flash Attention (Only RK3562 and RK3576)
  • Improve MatMul API
  • Improve support for int32 and int64
  • Support more operators and operator fusion

v2.0.0-beta0

  • Support RK3576 (Beta)
  • Support RK2118 (Beta)
  • Support SDPA (Scaled Dot Product Attention) to improve transformer performance
  • Improve custom operators support
  • Improve MatMul API
  • Improve support for Reshape,Transpose,BatchLayernorm,Softmax,Deconv,Matmul,ScatterND etc.
  • Support pytorch 2.1
  • Improve support for QAT models of pytorch and onnx
  • Optimize automatic generation of C++ code

v1.6.0

  • Support ONNX model of OPSET 12~19
  • Support custom operators (including CPU and GPU)
  • Improve support for dynamic weight convolution, Layernorm, RoiAlign, Softmax, ReduceL2, Gelu, GLU, etc.
  • Added support for python3.7/3.9/3.11
  • Add rknn_convert function
  • Improve transformer support
  • Improve MatMul API, such as increasing the K limit length, RK3588 adding int4 * int4 -> int16 support, etc.
  • Reduce RV1106 rknn_init initialization time, memory consumption, etc.
  • RV1106 adds int16 support for some operators
  • Fixed the problem that the convolution operator of RV1106 platform may make random errors in some cases.
  • Improve user manual
  • Reconstruct the rknn model zoo and add support for multiple models such as detection, segmentation, OCR, and license plate recognition.

v1.5.2

  • Improve dynamic shape support
  • Improve 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

v1.5.0

  • 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

v1.4.0

  • 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.

v1.3.0

  • 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

v1.2.0

  • Support RK3588
  • Support more operators, such as GRU、Swish、LayerNorm etc.
  • Reduce memory usage
  • Improve zero-copy interface implementation
  • Bug fix

v1.1.0

  • 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

v1.0

  • 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

v0.7

  • 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

v0.6

  • Initial version