NRF_AXON_GENERATE_VERSION

#define NRF_AXON_GENERATE_VERSION	(		major,
			minor,
			patch )

Value:

( ((major)<<16) | ((minor) << 8) | (patch))

The APIs in this header file are for the pre-compiled driver library. Users should avoid these low-level APIs in favor of higher level APIs in nrf_axon_nn_infer.h and nrf_axon_infer_test.h NRF_AXON_VERSION applies to the entire axon software tool chain. version history 1.3.0

• batch support added for SpaceToBatch, BatchToSpace, and convolutions. This enables dilation on convolutions.

1.2.1 05/04/2026

• Remove RRAM work-around.

1.2.0

• Multiple outputs supported
• Resize_Nearest_Neighbor operater supported.
• Rev'ed bin file version to 1.2.0 1.1.0 03/19/2026:
• TFLite 2.19 is supported version (was 2.15)
• More fixes to per-channel quantized dense layers: maximum input length increased to 2048 from 2046. maximum output length increased to 1024 from 512. signmoid and tanh actviation functions after per-channel quantized, fully-connected fixed. 1.0.1 03/04/2026:
• softmax after fully-connected with per channel quantization fixed. 1.0.0 03/02/2026:
• softmax fixed to report packed output.
• reshape implemented as a CPU op. 0.2.0 02/17/2026:
• Compiler catches unsupported dilation settings.
• Renamed broadcast add op function. 0.1.4 02/03/2026:
• Constant inputs for add and multiply operations.
• Axis broadcast for add operation.
• Width axis broadcast disabled (temporarily) for multiply operation.
• Maximum input channels increased from 512 to 1023 for many operations.
• Passlist functionality added.
• Optimization for 1D convolutions whose channel count is <= 16. 0.1.2 12/18/2025 :
• Fix to SplitV for bug experienced on Linux (not Windows).
• Average pool operations that are "mean-like" in that have an output width of 1 on height and/or width axis but whose filter size on that axis is less than the input size are now implemented with mean operation, allowing a maximum axis size of 1024 (vs 32). For example, input 49x20x64, filter 48x20x64, output 1x1x64, can now be handled.
• (INTERNAL) 1x1 pointwise output optimized with matrix mult instead of conv. Allows output channels up to 512 insteand of just 16.
• (INTERNAL) Places packing conv output in scratch mem. 0.1.1 Internal development 0.1.0 12/11/2025 :
• 1st versioned release