v1.4.0

New features and changes in this release:

• Reworked user interface, including settings and dependency management (#118)
• Reworked PYNQ driver (#100)
• Integration of experimental ONNX Passes front-end (#116)
• Improved FPGA-in-the-Loop ("Live") FIFO-Sizing (#158)
• Rework of Transformer example models and their build flows (#129, #160)
• Numerous small fixes (#147, #154, #163, #167)
• Cleanup of numerous dependencies, including a switch to the latest Brevitas (v0.12.1) and QONNX (v1.0.0) releases
• Marked some redundant CustomOps as deprecated (#166)

Notable features pulled in from the dev branch of AMD's upstream FINN repository:

• FP16 and fixed-point support for thresholding and elementwise ops (Xilinx#1422, Xilinx#1444, Xilinx#1445)
• Support for multiple weight sets for the memstreamer component (Xilinx#1441, Xilinx#1443)
• Generalized transpose and reshape support (Xilinx#1419)
• Support for QONNX' new operator versioning scheme, specifically Trunc v2 (Xilinx#1468, Xilinx#1480)
• New HLS Softmax operator (Xilinx#1439)
• New HLS Crop operator (Xilinx#1501)
• New RTL + HLS LayerNorm operators (Xilinx#1498, Xilinx#1506)
• Support for Relu activation as elementwise operator (Xilinx#1479)

v1.3.0

New features and improvements:

• Improved code quality by extending documentation and type hinting (#106)
• Partially refactored project structure (#113)
• Removed unneeded dependencies and bump dependency versions (#113)
• Packaged tests and notebook in separate sub-packages (#113)
• Added support for arbitrary Reshape operations (#128)
• Verification: Added configurable tolerance and logging of detailed tensor statistics (#119)

Notable bug fixes:

• Fix RTL simulation timeout due to small max cycle estimation (#115)
• Fix OOC synthesis failure for floating point IPs (#122)
• Fix PE operations assuming identical input and output shapes leading to wrong folding (#128)

v1.2.0

This release incorporates numerous fixes and improvements from the dev branch of AMD's FINN repository.
These are the most notable features:

• Improved streamlining transformations to support Split/Concat nodes and the YOLOv8 topology (Xilinx#1189)
• Added support for MatMul operations with two dynamic inputs to the existing RTL/HLS MVAU operator (Xilinx#1321)
• Switch to finnxsi as new simulation back-end (Xilinx#1365)
• Consolidation of the pooling operators into the pool_hls operator (Xilinx#1328)
• Integration of a new, more generic HLS Upsample operator (Xilinx#1408)
• Support decoupled mem_mode for elementwise ops (Xilinx#1413)
• Support FP32 inputs for the Thresholding operator (Xilinx#1418)

v1.1.0

New features and improvements:

• ZYNQ backend: Integration of instrumentation wrapper shell as alternative to I/O DMAs (#45)
• Introduction of Live (on-board) FIFO-Sizing (#46)
• Introduction of a custom step library (#34)
• New optional builder step for Vivado power estimation after OOC synthesis (#82)
• Alveo C++ driver improvements (#80)
• Improved error handling and logging
  • Introduce custom exceptions (e.g., "FINNUserError") (#69)
  • Overhauled handling of FINN-launched subprocesses (e.g., Vitis/Vivado) (#34)
  • Dedicated check for unsupported layers after CustomOp mapping (#81, #83)
• CI-related improvements
  • Infrastructure for on-board regression testing and micro-benchmarking (#34, #82)
  • Stability improvements related to parallel Vitis HLS instances (#79)
  • Build flows for all available finn-examples are now included in FINN+ (#83)

Potentially breaking changes:

• Removal of oh-my-xilinx dependency, select scripts have been moved to qnn-data/vivado_scripts (#82)
• Switched synthesis strategies/directives to default for OOC-synth and ZynqBuild (#82)
• Step for FIFO-sizing/insertion is performed earlier in the flow
• Changes to data layout inference, which is now performed more often (#87)

Notable updates pulled from upstream FINN:

• Different input datatypes support for AddStream operator (Xilinx#1195)
• Introduce FIFO "gauge" mode for depth monitoring during XSI simulation (#75)

v1.0.0

Initial FINN+ release. On top of many recent changes from the dev branch of the upstream repository, these are the most notable extensions of FINN included in this release:

• Completely new containerless setup (replacing the Docker-based infrastructure)
• Transformer/Attention support
• Generation of an optimized C++ driver for Alveo platforms
• Support for host memory access on Alveo platforms
• Improved streamlining and introduction of a new "StreamlinePlus" transformation
• Generic elementwise binary operations
• YAML-based build configuration
• Extended opset support for the resize operator
• Improved logging
• Improved CI and unit test infrastructure