Skip to content

Kernel Selection

heliaRT supports three kernel backends. The backend is chosen at build time — not at runtime.

Backends

Backend Kconfig OPTIMIZED_KERNEL_DIR Requires
Reference HELIA_RT_BACKEND_REFERENCE (empty) Nothing extra
CMSIS-NN HELIA_RT_BACKEND_CMSIS_NN cmsis_nn Arm CMSIS-NN module
HELIA HELIA_RT_BACKEND_HELIA helia Ambiq ns-cmsis-nn module

How It Works

flowchart TD
    A[Build system] --> B{OPTIMIZED_KERNEL_DIR?}
    B -->|helia| C[Look in kernels/helia/]
    B -->|cmsis_nn| D[Look in kernels/cmsis_nn/]
    B -->|empty| E[Use kernels/ reference]
    C --> F{Kernel file exists?}
    F -->|Yes| G[Use HELIA kernel]
    F -->|No| E
    D --> H{Kernel file exists?}
    H -->|Yes| I[Use CMSIS-NN kernel]
    H -->|No| E

For each operator, the build system checks whether an optimized implementation exists in the selected backend directory. If it does, that implementation is compiled instead of the Reference one. If not, the Reference kernel is used automatically.

Selecting a Backend

# prj.conf
CONFIG_HELIA_RT=y

The Kconfig default depends on what modules are available:

  • If NS_CMSIS_NN module is present → defaults to HELIA
  • If only CMSIS_NN module is present → defaults to CMSIS-NN
  • Otherwise → defaults to Reference
make -f tensorflow/lite/micro/tools/make/Makefile \
    OPTIMIZED_KERNEL_DIR=helia \
    ...

HELIA Kernel Coverage

The HELIA backend currently provides optimized implementations for 36 operators — expanding to 230+ kernel variants when counting per-dtype paths (int8 / int16 / float):

Full list

activations · add · batch_matmul · comparisons · concatenation · conv · depthwise_conv · dequantize · fill · fully_connected · hard_swish · leaky_relu · logistic · maximum_minimum · mul · pack · pad · pooling · quantize_common · reduce · reshape · softmax · split · split_v · squeeze · strided_slice · sub · svdf · tanh · transpose · transpose_conv · unidirectional_sequence_lstm · zeros_like

Full operator coverage matrix

HELIA Kernel Optimization Profile

The HELIA backend supports a SPEED/SIZE kernel profile in source builds. The default is SPEED.

CONFIG_HELIA_RT=y
CONFIG_HELIA_RT_KERNEL_OPTIMIZE_SPEED=y
# or
CONFIG_HELIA_RT_KERNEL_OPTIMIZE_SIZE=y

cmake -S . -B build \
  -DHELIA_RT_ENABLE_HELIA=ON \
  -DHELIA_RT_GLOBAL_KERNEL_OPTIMIZE=SIZE

make ... OPTIMIZED_KERNEL_DIR=helia GLOBAL_KERNEL_OPTIMIZE=SIZE microlite

Advanced CMake and Makefile builds can also override selected kernel families:

CONV_OPT=SPEED    # optimize Conv2D for latency
FC_OPT=SIZE       # optimize FullyConnected for code size

For CMake, the equivalent cache variables are HELIA_RT_CONV_OPT and HELIA_RT_FC_OPT. These default to the global profile when unset.

Next Steps