# CMSIS-NN foundation and extensions

Arm CMSIS-NN provides the foundation for heliaCORE. For production models on
Ambiq devices, heliaCORE builds on its convolution,
fully connected, pooling, and recurrent kernels with additional operations spanning tensor
indexing, graph updates, reductions, comparisons, and floating-point math.

## Additional graph operations

The following examples are declared in heliaCORE's public kernel headers and
are absent from the upstream public headers at the comparison revision below.
They illustrate the expansion; they are not the complete list of additions.

| Operation | heliaCORE data types | Upstream public API |
|---|---|---|
| Gather and GatherND | s8, s16, FP16, FP32 | Not present |
| ScatterND and Tile | s8, s16 | Not present |
| Select and DynamicUpdateSlice | s8, s16 | Not present |
| ReverseSequence | s8, s16 | Not present |
| ArgMin and ArgMax | s8, s16, FP16, FP32 | Not present |
| Reduce minimum and maximum | s8, s16, FP16, FP32 | Not present |
| Reduce sum | FP16, FP32 | Not present |
| Broadcast add, subtract, multiply | FP16, FP32 | Not present |
| GRU | FP16, FP32 | Not present |

This broader coverage lets model integrations use dedicated library kernels
for more of the graph. Support for a data type does not imply that every shape
has a DSP or MVE implementation. Use the [kernel index](https://ambiqai.github.io/ns-cmsis-nn/reference/kernel-index/)
for individual functions and the [data-type matrix](https://ambiqai.github.io/ns-cmsis-nn/guide/coverage/data-types-by-family/)
for family-level coverage.

## Quantized and floating-point models

heliaCORE supports A8W8 and A16W8 quantized kernels as well as opt-in
FP16 and FP32 APIs. Upstream also provides experimental FP16/FP32;
the distinction is the additional operations and variants, not the existence
of floating-point support alone.

## Integration with the Arm ecosystem

heliaCORE retains inherited CMSIS-NN-style interfaces where supported, adds
Ambiq-targeted implementations, and integrates through CMake, CMSIS-Pack,
Zephyr, and neuralSPOT-X. Upstream-derived files retain their Arm attribution
and Apache-2.0 licensing. See [About and licenses](https://ambiqai.github.io/ns-cmsis-nn/about/).

## Comparison sources

The comparison uses the public integer and floating-point headers, with the
source trees checked for the named implementations:

- heliaCORE revision `af724c78`: [integer header](https://github.com/AmbiqAI/ns-cmsis-nn/blob/af724c78778b545cc5a86bd74c7c830ea6df3899/Include/arm_nnfunctions.h),
  [float header](https://github.com/AmbiqAI/ns-cmsis-nn/blob/af724c78778b545cc5a86bd74c7c830ea6df3899/Include/arm_nnfunctions_flt.h).
- Arm CMSIS-NN revision `9e1b4768`: [integer header](https://github.com/ARM-software/CMSIS-NN/blob/9e1b4768c640606817f0d8f8e53a3d39be817ab4/Include/arm_nnfunctions.h),
  [float header](https://github.com/ARM-software/CMSIS-NN/blob/9e1b4768c640606817f0d8f8e53a3d39be817ab4/Include/arm_nnfunctions_flt.h).

This is an operator/API comparison, not an upstream performance benchmark.
[Kernel benchmarks](https://ambiqai.github.io/ns-cmsis-nn/guide/performance/kernel-benchmarks/) compare
heliaCORE's own execution paths under the documented measurement conditions.