Skip to content
heliaRT
User guide
HELIA HUB

Operator explorer

Search the HELIA adapter inventory by operator, family or input data type. Each result lists important constraints and links to the adapter at this documentation’s source revision.

A listed type can use an optimized, reference or storage-only path. Shape, weight/output types and build settings still determine whether your model is supported.

How to read type support
  • FP16 arithmetic needs the matching heliaCORE feature and target support. Storage operations such as DEQUANTIZE are different.
  • int4 weights appear in compute-operator notes. They do not imply int4 activations.
  • Fallback is operator-specific. Check both allocation and invocation status. See Model compatibility and Floating point.
  • Operators outside this inventory may come from upstream Reference sources. Registering an operator does not prove that all its types and options are supported.

59 operators

ABS

Arithmetic
  • int8
  • int16
  • float32

Reference implementation in the HELIA adapter source. Quantized scaling requirements apply.

Adapter source

ADD

Arithmetic
  • int8
  • int16
  • float32
  • float16
  • int32

Float identical-shape inputs support arbitrary rank. Optimized broadcasting supports rank <= 4; higher-rank FP32 uses Reference and FP16 is rejected. int32 follows the adapter reference path.

Adapter source

ARG_MAX

Reduction
  • int8
  • float32
  • float16

FP16 requires CORE >= 7.35.0, rank 1..4, INT32 axis/output and positive reduced extent. FP32/int8 use Reference.

Adapter source

ARG_MIN

Reduction
  • int8
  • float32
  • float16

FP16 requires CORE >= 7.35.0, rank 1..4, INT32 axis/output and positive reduced extent. FP32/int8 use Reference.

Adapter source

AVERAGE_POOL_2D

Compute
  • int8
  • int16
  • float32
  • float16

Four-dimensional tensors; check filter/stride/padding and output dimensions. Kernel failure is returned to the application.

Adapter source

BATCH_MATMUL

Compute
  • int8
  • int16
  • float32
  • float16

Check batch broadcasting, adjoint flags and quantized weight/activation combinations against the adapter.

Adapter source

CONCATENATION

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64
  • bool

Optimized float paths support rank <= 4; higher-rank float inputs use Reference. Input/output type and quantization must agree.

Adapter source

CONV_2D

Compute
  • int8
  • int16
  • float32
  • float16

Quantized weight/bias combinations are constrained. int4 weights use int8 activations. Float grouped convolution uses Reference for FP32 and is rejected for FP16.

Adapter source

COS

Arithmetic
  • float32

Reference implementation in the HELIA adapter source; Float32 input/output.

Adapter source

DEPTHWISE_CONV_2D

Compute
  • int8
  • int16
  • float32
  • float16

Quantized weight/bias combinations and depth multiplier are constrained. int4 is a weight type with int8 activations; it is not a general activation type.

Adapter source

DEQUANTIZE

Quantization
  • int8
  • int16
  • uint8
  • float16

Output is FP32. FP16 means weight-storage widening, not FP16 computation; it works without ARM_NN_ENABLE_F16.

Adapter source

EQUAL

Comparison
  • bool
  • float32
  • int8
  • int16
  • int32
  • int64

Boolean output. Supported input types depend on the comparison; input shapes and broadcast rules must match.

Adapter source

FILL

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64

Constant dimensions and scalar fill value. Float APIs require CORE >= 7.33.0 and matching enabled feature.

Adapter source

FULLY_CONNECTED

Compute
  • int8
  • int16
  • float32
  • float16

Quantized activation/weight/bias types must match a supported combination; includes int4 weights with int8 activations. Float paths require matching CORE features.

Adapter source

GATHER

Data movement
  • int8
  • int16
  • float32
  • float16

CORE >= 7.34.0. Native rank <= 4. Scalar indices use float native paths or int8 Reference; int16 rejects them. Higher-rank FP32/int8 use Reference; FP16/int16 reject.

Adapter source

GATHER_ND

Data movement
  • int8
  • int16
  • float32
  • float16

CORE >= 7.34.0. Native ranks <= 4 and positive index tuple width. Higher ranks and zero-width tuples use FP32/int8 Reference; FP16/int16 reject.

Adapter source

GREATER

Comparison
  • float32
  • int8
  • int16
  • int32
  • int64

Boolean output. Check broadcast and operand types; no FP16 comparison branch.

Adapter source

HARD_SWISH

Activation
  • int8
  • int16
  • float32
  • float16

FP16 requires the CORE float feature; without it preparation fails. FP32 can use Reference.

Adapter source

LESS

Comparison
  • float32
  • int8
  • int16
  • int32
  • int64

Boolean output. Check broadcast and operand types; no FP16 comparison branch.

Adapter source

LESS_EQUAL

Comparison
  • float32
  • int8
  • int16
  • int32
  • int64

Boolean output. Check broadcast and operand types; no FP16 comparison branch.

Adapter source

LOG

Arithmetic
  • float32

Reference implementation in the HELIA adapter source; Float32 input/output.

Adapter source

LOGISTIC

Activation
  • int8
  • int16
  • float32
  • float16

int8 uses a lookup table; int16 and enabled floats have HELIA paths. NaN is not supported by the optimized float activation path.

Adapter source

MAXIMUM

Arithmetic
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64

Optimized paths support rank <= 4. Higher-rank FP32/int8/int16 use Reference; FP16 is rejected. int32/int64 use Reference.

Adapter source

MAX_POOL_2D

Compute
  • int8
  • int16
  • float32
  • float16

Four-dimensional tensors; check filter/stride/padding and output dimensions. Kernel failure is returned to the application.

Adapter source

MEAN

Reduction
  • int8
  • int16
  • float32
  • float16

Float optimized paths support rank <= 4 and axis sets; higher-rank FP32 uses Reference, FP16 is rejected. See reduce_common.cc.

Adapter source

MINIMUM

Arithmetic
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64

Optimized paths support rank <= 4. Higher-rank FP32/int8/int16 use Reference; FP16 is rejected. int32/int64 use Reference.

Adapter source

MUL

Arithmetic
  • int8
  • int16
  • float32
  • float16
  • int32

Float identical-shape inputs support arbitrary rank. Optimized broadcasting supports rank <= 4; higher-rank FP32 uses Reference and FP16 is rejected. int32 follows the adapter reference path.

Adapter source

NOT_EQUAL

Comparison
  • bool
  • float32
  • int8
  • int16
  • int32
  • int64

Boolean output. Supported input types depend on the comparison; input shapes and broadcast rules must match.

Adapter source

PACK

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64

Equal input shapes, including scalars. Float APIs require CORE >= 7.33.0 and matching enabled feature.

Adapter source

PAD

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32

FP16 requires four-dimensional tensors and the float feature. Padding values and output shape must match.

Adapter source

PADV2

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32

FP16 requires four-dimensional tensors and the float feature. Padding values and output shape must match.

Adapter source

QUANTIZE

Quantization
  • int8
  • int16
  • int32
  • uint8
  • float32

Supported input/output pairs are specific: this type list does not imply all pairings. Registration is supplied by kernels/quantize.cc.

Adapter source

REDUCE_MAX

Reduction
  • int8
  • int16
  • float32

No FP16 execution path. Axis, output shape and quantization checks apply; implementation is in reduce_common.cc.

Adapter source

REDUCE_MIN

Reduction
  • int8
  • int16
  • float32

No FP16 execution path. Axis, output shape and quantization checks apply; implementation is in reduce_common.cc.

Adapter source

RELU

Activation
  • int8
  • int16
  • float32
  • float16

Check activation bounds, tensor shape and quantization. Optimized float paths require the matching CORE feature.

Adapter source

RELU6

Activation
  • int8
  • int16
  • float32
  • float16

Check activation bounds, tensor shape and quantization. Optimized float paths require the matching CORE feature.

Adapter source

RESHAPE

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64
  • uint8
  • bool

Preserves storage and element count. FP16 storage works without float arithmetic enabled. Types listed here are representative fixed-width storage types.

Adapter source

RSQRT

Arithmetic
  • int8
  • int16
  • float32
  • float16

FP32 and integer paths use Reference. FP16 requires CORE >= 7.33.0 and ARM_NN_ENABLE_F16; input/output shapes and types must match.

Adapter source

SIN

Arithmetic
  • float32

Reference implementation in the HELIA adapter source; Float32 input/output.

Adapter source

SOFTMAX

Activation
  • int8
  • int16
  • float32
  • float16

Optimized float path requires beta = 1. Other beta values use FP32 Reference or fail FP16 preparation. int8-to-int16 output has a dedicated registration.

Adapter source

SPLIT

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32

Constant axis and equal output extents. Float APIs require CORE >= 7.33.0 and corresponding enabled feature.

Adapter source

SPLIT_V

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32

Constant axis and split sizes; one inferred -1 size and zero-length pieces are allowed. Shapes/types/counts are checked during preparation. Float APIs require CORE >= 7.33.0.

Adapter source

SQRT

Arithmetic
  • float32
  • float16

FP32 uses Reference. FP16 requires CORE >= 7.33.0 and ARM_NN_ENABLE_F16; input/output shapes and types must match.

Adapter source

SQUEEZE

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32
  • int64
  • uint8
  • bool

Bitwise storage copy; removes unit dimensions, checks matching byte counts and rejects strings. Input rank <= 8.

Adapter source

STRIDED_SLICE

Data movement
  • int8
  • int16
  • int32
  • float32
  • bool

Check slicing masks, strides and shape constraints in the adapter; no FP16 execution branch.

Adapter source

SUB

Arithmetic
  • int8
  • int16
  • float32
  • float16
  • int32

Float identical-shape inputs support arbitrary rank. Optimized broadcasting supports rank <= 4; higher-rank FP32 uses Reference and FP16 is rejected. int32 follows the adapter reference path.

Adapter source

SUM

Reduction
  • int8
  • int16
  • float32
  • float16

SUM is LiteRT REDUCE_SUM. Float optimized paths support rank <= 4 and axis sets; higher-rank FP32 uses Reference, FP16 is rejected. See reduce_common.cc.

Adapter source

SVDF

Recurrent
  • int8
  • float32
  • float16

Data input is int8, FP32 or enabled FP16. int16 is a time-weight/state type, not a supported data input type. Variable state persists across invocations; weight types, rank and activation configuration are constrained.

Adapter source

TANH

Activation
  • int8
  • int16
  • float32
  • float16

NaN is not supported by the optimized float activation path. Float support is build-gated.

Adapter source

TRANSPOSE

Data movement
  • int8
  • int16
  • float32
  • float16

Optimized rank <= 4. Higher ranks use Reference; FP16 has a bitwise storage fallback even with its float feature disabled.

Adapter source

TRANSPOSE_CONV

Compute
  • int8
  • int16
  • float32
  • float16

Check output shape, padding and quantized weight/bias combinations. Float features are build-gated.

Adapter source

UNIDIRECTIONAL_SEQUENCE_LSTM

Recurrent
  • int8
  • int16
  • float32
  • float16

Standard four-gate LSTM only. Preparation rejects peephole, projection, layer normalization and missing gates (CIFG), including FP32. Quantized inputs use int8 weights and int16 cell state. Preserve variable state across invocations.

Adapter source

UNPACK

Data movement
  • int8
  • int16
  • float32
  • float16
  • int32

One output per selected axis element. Float APIs require CORE >= 7.33.0 and matching enabled feature.

Adapter source

Download the complete operator inventory as Markdown, including input types, constraints and source links.