• AutoTiler:
  • Added:
    • Improved fused activations in HWC pooling
    • MatMul for DSP processing (float16 and 32)
    • Support for OCTOSPIFlash/Ram in code generation
  • Fixed:
    • NE16 Depthwise mode bug in real HW
    • Real FFT evaluation of last bin spectrogram
• Examples:
  • Removed:
    • remove efuse_burner example
• FreeRTOS:
  • Fixed:
    • fix use of notify in pi_task
• SDK:
  • Changed:
    • change the documentation navigation depth limit to infinite
  • Fixed:
    • add missing PMSIS include directory to documentation
• NNTool:
  • Added:
    • ONNX Upscale operator
    • Batchnorm converted to discrete expressions ops when not fusable
    • Expression compiler with better iteration handling + more flexible quantization + code comments
  • Fixed:
    • NNCF mixed precision import for <8bits weights
• GVSOC:
  • Fixed:
    • Fixed issue with hw loop save/restore

• AutoTiler:
  • Added:
    • Quantizers from/to float32
    • Automatic evaluation of optimal TileOrientation strategy for SQ8 and fp16 generators
    • Improved MatMul SW performance
  • Fixed:
    • Fix in ProcessArgPadding, we need to locate the most inner KER_ITER_i space not the fisrt tileable dim
    • Quantizer generator bug and FP_FL FL_FP basic kernel (bad read of zero/scale infos)
    • NextAddress code generation fix
• FreeRTOS:
  • Fixed:
    • Fix i2s1 configure bug in gap8
• SDK:
  • Added:
    • WSL support when launch run from WSL instance
• NNTool:
  • Added:
    • Concat in tflite on same input
    • Resizer in fp16 mode
    • Add DSP preprocessing in front of the network now working
    • Convert onnx Gather to slice if possible at import
    • FC tflite support for keep_dims
  • Fixed:
    • Fix bug in 1D padding onnx
    • Bug in Padding for onnx

• AutoTiler:
  • Added:
    • Cmake build alternative for Mnist example.
    • Improved MatMul Kernels
    • HSigmoid/HTanh/HSwish support in float16 layers
  • Fixed:
    • Tiling bug fix for vertical tiled 1D kernels
    • Bug in quantizer due to ClipF macro expansion
    • Several warnings fixed for unsigned type of ScaleN pointer
• Examples:
  • Added:
    • add new customer_bsp board example with EPEAS CIS001 camera
• FreeRTOS:
  • Changed:
    • Enable float printf, but can be disabled by adding "PRINTF_FLOAT = DISABLE" in Makefile
  • Fixed:
    • removed some circular inclusions
• SDK:
  • Added:
    • dynamic runner args support to cmake run
    • cmake gaptest support
    • README in hello for cmake and test description
    • Options to compile with cluster enabled
    • cmake top documentation
    • README in bsp ram simple test for cmake and test description
    • Options to compile with chosen ram
    • default values for gvsoc build
    • cmake build for whole sdk
  • Fixed:
    • clean target for gap8, cmake options
• GVSoC:
  • Changed:
    • change -march=native to -mtune=generic for better compatibility with newer architectures
  • Fixed:
    • Fix several problems in ASRC model integration.
• NNTool:
  • Added:
    • Global sum and min/max
    • MatMul performance improvements
    • New generation code for SQ8
    • New qtune syntax
    • Support TinyYolo
    • Split to Concat matcher to simplify network topology
    • Support Onnx NonMaxSuppression Layer
    • Json file to override quantization import in onnx
    • Initial auto compression based on global QSNR setting
    • Sigmoid and Tanh now default non hard also in onnx
    • Import quantization from nncf
  • Fixed:
    • Buf of Match Hsigmoid (https://github.com/GreenWaves-Technologies/gap_sdk/issues/290)
    • Bugs in adjuster
    • Improved quantizer resolver
    • Transpose generation issues
    • Matadd+act generation issues
• Pulp-OS:
  • Added:
    • Move assert implem from PMSIS_API into RTOS.

New Features

• SDK:
  • Re-orgnisation of SDK
  • Tools reorganization
    • Move tools (efuse, gap_configs, gaptest, gapy, json-tools, littlefs, openocd, openocd_tools, plptest, rules) from gap_sdk/tools to gap_sdk/utils
  • Removed old native examples (examples/natives).
  • Removed old applications.
  • Removed old doc, replaced by new one.
• NNTool:
  • Added support for LUT sigmoid/tanh in fp16
  • unsigned sq8 add
  • rsqrt expressions
  • new dsp functions in expressions
  • Added support for LUT sigmoid/tanh in fp16
• Autotiler:
  • Added DSP Lib in autotiler rules
  • Support for float16 LUT based Sigmoid/Tanh
  • map Hx1 Conv onto 1D as well as Wx1
  • Unsigned SQ8 Add
  • rsqrt DSP functions
  • improved performance in DSP functions
• PULP OS:
  • Added preset_fork

BUG Fixes

• Compiler:
  • Fix wrong vector type used in sign inject and in abs for float16 (https://github.com/GreenWaves-Technologies/gap_riscv_toolchain_ubuntu_18)
• NNTool:
  • gen_project bugfixes
• Autotiler:
  • bugfixes for DSP library and generators:
  • Fixed point fft
  • center padding like librosa
  • bugfix in DSP generator and NE16 conv + pool
  • bugfix in USQ8 reduction act
  • DSP: log offset fix for float16 version
  • Code Gen: Fix in Code generation of tiles copie
  • fp16 activation bug in leftover
  • Fixed SSD non max supression num of total BBs
• FreeRTOS:
  • Update timer and perf driver to reduce footprint
  • Calibrated the performance counter

[Be careful]: Due to the reorg of SDK, please clean and rebuild the SDK:

make clean all

GAP Profiler -- V1.0.0

• Profiler - 1st release of gap profiler [version - 1.0.0]: Profiler is a visualization tool for profiling and debugging GAP applications and used with GVSOC, our Full System SoC Simulator. Profiler gives you a visual view of what is happening inside the chip and allows you to control the simulator through a graphic interface

Autotiler -- V4.0.0

• New DSP library with optimised (i)MFCC implementation (with example)
• RFFT kernel

NNTool -- V3.11.x

• Added graph drawing ability - draw command
• Added support for mixed quantization - qtune command reworked
• Improved expression support
• Matmul support extended and added in all quantization schemes. 'channel' scaled matmul support for constant tensor multiplications
• bfloat and ieee16 quantization support
• New gen_project command generates template GAP project automatically
• New performance command creates temporary project, runs it and extracts performance
• several new fusions handle optimising different patterns

BSP

• Updated SPI RAM/FLASH size in gapoc boards
• GAPOC C BSP added camera capture + compression example

Others

• Corrected bug in gvsoc when read/write files bigger then 1 MB
• New Fuser tool for GAP8 (removed old one)
• Add AES encryption for GAP8 flash image
• GAP LIB imageIO added write out image from L3

New Features

NNTool

• New quantizer: initial work for mixed precision quantization
• onnx GRU parameter support
• LSTM quantization: keep different scales for recurrent and input weights + override cell_clip quantization parameter with statistic from inference
• Handling of 1x1 inputs in resizer
• Support for channel padded Matrix addition

Autotiler

• MFCC support for float16 float32 execution
• Channel padded MatAdd kernel and code generation support
• int16 Resizers
• Matrix Multiplication optimization

BSP:

• New camera driver: HM0360

Customer BSP example:

An example about how to customize a BSP for a customer board

• gap_sdk/examples/customer_bsp

Others

• Scripts for supporting vscode

Bug fixes

• Fix NNtool in requirements installation (support also in WSL2)
• Bug fix in 1D onnx convolution and pad operators
• Bug fix in hyperbus driver (Both FreeRTOS and PULP-OS) when TX size = 6
• Fix thermal camera driver for capa load and GSK
• NINA B112 driver fixes
• Modified I2C wait cycles from 0xFF to 0x1 after the STOP command

BUG FIX:

• Regression in GVSOC: fix the bug in event unit
• NNTOOL: fix pad operator for onnx

• NNTool
  - New importer: ONNX graphs support - LSTM/RNN new quantization schemes (use LUT for activations + allow different quantization ranges between input and states) - Bug fixes
• Autotiler - New Tiling engine (more computationally efficient) - Add support for warm Construct/Destruct (only L1 buffer allocation/deallocation) - MFCC generators and basic kernels
• FreeRTOS: - New FLL driver - Add malloc dump feature
• PMSIS_BSP: - Pixart camera driver update for ULPC module - Himax camera driver update for AI-Deck - OTA fixes - Fixes for building AI Deck bsp with freeRTOS - Thermeye IR camera driver updates - Fix for MX25 spi flash erase
• Others: - Support SSBL (Second Stage Boot Loader)

• FreeRTOS - I2S Frequency calculation bug - CPI Slice mode fix - SPI fix for keeping cs
• PulpOS - CPI fix
• NNTool
  - New TFLite operators support: Transpose, ResizeBilinear - Bug fixes (dilated convolutions nntool execution + lstm nntool execution)
• Autotiler (from my side) - Bug fix in SSD PostProcess operator
• BSP - Optimized Himax BSP driver

NEW FEATURES

NNTool

• Expressions: fuse sequences of piecewise TFLite operators into a single Autotiler layer call
• Resizer: insert an image resizer in front of the network (algorithms: Bilinear or Nearest Neighbor)
• Support for Nearest Neighbor tflite node
• Support for LeakyReLU tflite node
• Bug fixes

Autotiler

• Nearest Neighbor algorithm for resizer

PulpOS

• Fixed issue with GPIO callbacks, some registers were corrupted after the execution of the callback

GVSOC

• Cache flushing is now implemented

FreeRTOS

• SPI copy API fixes for SPIRAM
• Important optimizations for Freertos
• Freertos Directory tree reorganisation