Structure Types

Enums and Data Structures used in public API.

cmsis_nn_tilestructCMSIS-NN object to contain the width and height of a tile
cmsis_nn_contextstructCMSIS-NN object used for the function context.
cmsis_nn_bias_datastructCMSIS-NN object used to hold bias data for int16 variants.
cmsis_nn_dimsstructCMSIS-NN object to contain the dimensions of the tensors
cmsis_nn_lstm_dimsstructCMSIS-NN object to contain LSTM specific input parameters related to dimensions
cmsis_nn_per_channel_quant_paramsstructCMSIS-NN object for the per-channel quantization parameters
cmsis_nn_per_tensor_quant_paramsstructCMSIS-NN object for the per-tensor quantization parameters
cmsis_nn_quant_paramsstructCMSIS-NN object for quantization parameters.
cmsis_nn_activationstructCMSIS-NN object for the quantized Relu activation
cmsis_nn_conv_paramsstructCMSIS-NN object for the convolution layer parameters
cmsis_nn_transpose_conv_paramsstructCMSIS-NN object for the transpose convolution layer parameters
cmsis_nn_dw_conv_paramsstructCMSIS-NN object for the depthwise convolution layer parameters
cmsis_nn_pool_paramsstructCMSIS-NN object for pooling layer parameters
cmsis_nn_gather_paramsstructCMSIS-NN object for the gather operator
cmsis_nn_gather_nd_paramsstructCMSIS-NN object for the gathernd operator
cmsis_nn_tile_paramsstructCMSIS-NN object for the tile operator
cmsis_nn_broadcast_to_paramsstructCMSIS-NN object for the broadcastto operator
cmsis_nn_scatter_nd_paramsstructCMSIS-NN object for the scatternd operator
cmsis_nn_mirror_pad_paramsstructCMSIS-NN object for the mirrorpad operator
cmsis_nn_where_paramsstructCMSIS-NN object for the WHERE operator
cmsis_nn_select_v2_paramsstructCMSIS-NN object for the selectv2 operator (with broadcast)
cmsis_nn_reverse_sequence_paramsstructCMSIS-NN object for the reversesequence operator
cmsis_nn_dynamic_update_slice_paramsstructCMSIS-NN object for the dynamicupdateslice operator
cmsis_nn_fc_paramsstructCMSIS-NN object for Fully Connected layer parameters
cmsis_nn_bmm_paramsstructCMSIS-NN object for Batch Matmul layer parameters
cmsis_nn_transpose_paramsstructCMSIS-NN object for Transpose layer parameters
cmsis_nn_resize_paramsstructCMSIS-NN object for Resize Nearest Neighbor layer parameters
cmsis_nn_svdf_paramsstructCMSIS-NN object for SVDF layer parameters
cmsis_nn_softmax_lut_s16structCMSIS-NN object for Softmax s16 layer parameters
cmsis_nn_concatenation_paramsstruct
cmsis_nn_scalingstructCMSIS-NN object for quantization parameters
cmsis_nn_lstm_gatestructCMSIS-NN object for LSTM gate parameters
cmsis_nn_lstm_paramsstructCMSIS-NN object for LSTM parameters
cmsis_nn_lstm_contextstructCMSIS-NN object for LSTM scratch buffers.
cmsis_nn_activation_f32structActivation clamp range for floating-point operators.
cmsis_nn_conv_params_f32structConvolution parameters for float32 operators.
cmsis_nn_transpose_conv_params_f32structTranspose convolution parameters for float32 operators.
cmsis_nn_dw_conv_params_f32structDepthwise convolution parameters for float32 operators.
cmsis_nn_pool_params_f32structPooling parameters for float32 operators.
cmsis_nn_fc_params_f32structFully connected layer parameters for float32 operators.
cmsis_nn_bmm_params_f32structBatched matrix multiplication parameters for float32 operators.
cmsis_nn_ew_params_f32structElementwise operator parameters for float32 operators.
cmsis_nn_transpose_params_f32structTranspose parameters for float32 operators.
cmsis_nn_svdf_params_f32structSingular value decomposition filter parameters for float32 operators.
cmsis_nn_lstm_gate_f32structRead-only weights and bias metadata for one float32 LSTM gate.
cmsis_nn_lstm_params_f32structParameters for a unidirectional float32 LSTM layer.
cmsis_nn_lstm_context_f32structScratch and mutable state buffers for a float32 LSTM invocation.
cmsis_nn_gru_gate_f32structWeights and biases for a single float32 GRU gate.
cmsis_nn_gru_params_f32structParameters for a float32 unidirectional GRU invocation.
cmsis_nn_gru_context_f32structScratch buffers for a float32 GRU invocation.
cmsis_nn_activation_f16structActivation clamp range for floating-point operators.
cmsis_nn_conv_params_f16structConvolution parameters for float32 operators.
cmsis_nn_transpose_conv_params_f16structTranspose convolution parameters for float32 operators.
cmsis_nn_dw_conv_params_f16structDepthwise convolution parameters for float32 operators.
cmsis_nn_pool_params_f16structPooling parameters for float32 operators.
cmsis_nn_fc_params_f16structFully connected layer parameters for float32 operators.
cmsis_nn_bmm_params_f16structBatched matrix multiplication parameters for float32 operators.
cmsis_nn_ew_params_f16structElementwise operator parameters for float32 operators.
cmsis_nn_transpose_params_f16structTranspose parameters for float32 operators.
cmsis_nn_svdf_params_f16structSingular value decomposition filter parameters for float32 operators.
cmsis_nn_lstm_gate_f16structRead-only weights and bias metadata for one float32 LSTM gate.
cmsis_nn_lstm_params_f16structParameters for a unidirectional float32 LSTM layer.
cmsis_nn_lstm_context_f16structScratch and mutable state buffers for a float32 LSTM invocation.
cmsis_nn_gru_gate_f16structWeights and biases for a single float16 GRU gate.
cmsis_nn_gru_params_f16structParameters for a float16 unidirectional GRU invocation.
cmsis_nn_gru_context_f16structScratch buffers for a float16 GRU invocation.
arm_nn_tensor_layoutenumTensor layout selector for floating-point APIs.
arm_nn_activation_typeenumEnum for specifying activation function types
arm_nn_activation_type_fltenumActivation selector for floating-point operator APIs.
arm_nn_compare_operationenumEnum for specifying comparison operator
arm_cmsis_nn_statusenumFunction return codes
arm_nn_dw_kernel_layout_f32enumDepthwise kernel storage layout selector for floating-point kernels.
arm_nn_weight_format_fltenumWeight storage format selector for floating-point operators.
arm_nn_dw_kernel_layout_f16type

struct cmsis_nn_tile

CMSIS-NN object to contain the width and height of a tile

int32_t w

Width

int32_t h

Height

struct cmsis_nn_context

CMSIS-NN object used for the function context.

void * buf

Pointer to a buffer needed for the optimization

int32_t size

Buffer size

struct cmsis_nn_bias_data

CMSIS-NN object used to hold bias data for int16 variants.

const void * data

Pointer to bias data

const bool is_int32_bias

Indicate type of bias data. True means int32 else int64

struct cmsis_nn_dims

CMSIS-NN object to contain the dimensions of the tensors

int32_t n

Generic dimension to contain either the batch size or output channels. Please refer to the function documentation for more information

int32_t h

Height

int32_t w

Width

int32_t c

Input channels

struct cmsis_nn_lstm_dims

CMSIS-NN object to contain LSTM specific input parameters related to dimensions

int32_t max_time

int32_t num_inputs

int32_t num_batches

int32_t num_outputs

struct cmsis_nn_per_channel_quant_params

CMSIS-NN object for the per-channel quantization parameters

int32_t * multiplier

Multiplier values

int32_t * shift

Shift values

struct cmsis_nn_per_tensor_quant_params

CMSIS-NN object for the per-tensor quantization parameters

int32_t multiplier

Multiplier value

int32_t shift

Shift value

struct cmsis_nn_quant_params

CMSIS-NN object for quantization parameters. This struct supports both per-tensor and per-channels requantization and is recommended for new operators.

int32_t * multiplier

Multiplier values

int32_t * shift

Shift values

int32_t is_per_channel

struct cmsis_nn_activation

CMSIS-NN object for the quantized Relu activation

int32_t min

Min value used to clamp the result

int32_t max

Max value used to clamp the result

struct cmsis_nn_conv_params

CMSIS-NN object for the convolution layer parameters

int32_t input_offset

The negative of the zero value for the input tensor

int32_t output_offset

The negative of the zero value for the output tensor

cmsis_nn_tile stride

cmsis_nn_tile padding

cmsis_nn_tile dilation

cmsis_nn_activation activation

struct cmsis_nn_transpose_conv_params

CMSIS-NN object for the transpose convolution layer parameters

int32_t input_offset

The negative of the zero value for the input tensor

int32_t output_offset

The negative of the zero value for the output tensor

cmsis_nn_tile stride

cmsis_nn_tile padding

cmsis_nn_tile padding_offsets

cmsis_nn_tile dilation

cmsis_nn_activation activation

struct cmsis_nn_dw_conv_params

CMSIS-NN object for the depthwise convolution layer parameters

int32_t input_offset

The negative of the zero value for the input tensor

int32_t output_offset

The negative of the zero value for the output tensor

int32_t ch_mult

Channel Multiplier. ch_mult * in_ch = out_ch

cmsis_nn_tile stride

cmsis_nn_tile padding

cmsis_nn_tile dilation

cmsis_nn_activation activation

struct cmsis_nn_pool_params

CMSIS-NN object for pooling layer parameters

cmsis_nn_tile stride

cmsis_nn_tile padding

cmsis_nn_activation activation

struct cmsis_nn_gather_params

CMSIS-NN object for the gather operator

int32_t axis

Axis to gather from. Supports negative indexing.

int32_t batch_dims

Number of leading batch dimensions

int32_t input_rank

Rank of the input tensor (range: [1, 4])

int32_t coords_rank

Rank of the coordinate tensor ([1, 4]; float gather also accepts scalar rank 0)

struct cmsis_nn_gather_nd_params

CMSIS-NN object for the gather_nd operator

int32_t params_rank

Rank of the params tensor (range: [1, 4])

int32_t indices_rank

Rank of the indices tensor (range: [1, 4])

int32_t batch_dims

Number of batch dimensions

struct cmsis_nn_tile_params

CMSIS-NN object for the tile operator

int32_t rank

Rank of the input tensor (range: [1, 8])

const int32_t * input_shape

Input shape array (length = rank)

const int32_t * multiples

Multiples array (length = rank)

struct cmsis_nn_broadcast_to_params

CMSIS-NN object for the broadcast_to operator

int32_t rank

Rank of input/output tensors (range: [1, 8])

const int32_t * input_shape

Input shape array (length = rank)

const int32_t * output_shape

Output (broadcast target) shape array (length = rank)

struct cmsis_nn_scatter_nd_params

CMSIS-NN object for the scatter_nd operator

int32_t num_updates

Number of update slices

int32_t index_depth

Depth of each index vector

int32_t slice_size

Size of each update slice

int32_t output_size

Total number of elements in output

const int32_t * output_strides

Strides of the output tensor (length = index_depth)

struct cmsis_nn_mirror_pad_params

CMSIS-NN object for the mirror_pad operator

int32_t rank

Rank of the input tensor (range: [1, 8])

const int32_t * input_shape

Input shape array (length = rank)

const int32_t * output_shape

Output shape array (length = rank)

const int32_t * pad_before

Padding before each dimension (length = rank)

int32_t mode

0 = REFLECT, 1 = SYMMETRIC

struct cmsis_nn_where_params

CMSIS-NN object for the WHERE operator

int32_t rank

Rank of the condition tensor (range: [1, 8])

const int32_t * shape

Condition tensor shape array (length = rank)

struct cmsis_nn_select_v2_params

CMSIS-NN object for the select_v2 operator (with broadcast)

int32_t rank

Rank of the output tensor (range: [1, 8])

const int32_t * output_shape

Output shape array (length = rank)

const int32_t * cond_strides

Condition tensor broadcast strides (length = rank)

const int32_t * x_strides

X tensor broadcast strides (length = rank)

const int32_t * y_strides

Y tensor broadcast strides (length = rank)

struct cmsis_nn_reverse_sequence_params

CMSIS-NN object for the reverse_sequence operator

int32_t rank

Rank of the input tensor (range: [1, 8])

const int32_t * shape

Input shape array (length = rank)

int32_t seq_dim

Dimension along which to reverse

int32_t batch_dim

Batch dimension

struct cmsis_nn_dynamic_update_slice_params

CMSIS-NN object for the dynamic_update_slice operator

int32_t rank

Rank of the operand tensor (range: [1, 8])

const int32_t * operand_shape

Operand shape array (length = rank)

const int32_t * update_shape

Update shape array (length = rank)

int32_t operand_size

Total number of elements in operand

int32_t update_size

Total number of elements in update

const int32_t * operand_strides

Strides of the operand tensor (length = rank)

struct cmsis_nn_fc_params

CMSIS-NN object for Fully Connected layer parameters

int32_t input_offset

The negative of the zero value for the input tensor

int32_t filter_offset

The negative of the zero value for the filter tensor

int32_t output_offset

The negative of the zero value for the output tensor

cmsis_nn_activation activation

struct cmsis_nn_bmm_params

CMSIS-NN object for Batch Matmul layer parameters

const bool adj_x

const bool adj_y

cmsis_nn_fc_params fc_params

struct cmsis_nn_transpose_params

CMSIS-NN object for Transpose layer parameters

const int32_t num_dims

const uint32_t * permutations

The dimensions applied to the input dimensions

struct cmsis_nn_resize_params

CMSIS-NN object for Resize Nearest Neighbor layer parameters

bool align_corners

Align corners when calculating interpolation

bool half_pixel_centers

Use half pixel centers when calculating interpolation

struct cmsis_nn_svdf_params

CMSIS-NN object for SVDF layer parameters

int32_t rank

int32_t input_offset

The negative of the zero value for the input tensor

int32_t output_offset

The negative of the zero value for the output tensor

cmsis_nn_activation input_activation

cmsis_nn_activation output_activation

struct cmsis_nn_softmax_lut_s16

CMSIS-NN object for Softmax s16 layer parameters

const int16_t * exp_lut

const int16_t * one_by_one_lut

struct cmsis_nn_concatenation_params

const int32_t axis

struct cmsis_nn_scaling

CMSIS-NN object for quantization parameters

int32_t multiplier

Multiplier value

int32_t shift

Shift value

struct cmsis_nn_lstm_gate

CMSIS-NN object for LSTM gate parameters

int32_t input_multiplier

int32_t input_shift

const void * input_weights

const void * input_effective_bias

Bias added with precomputed kernel_sum * lhs_offset

int32_t hidden_multiplier

int32_t hidden_shift

const void * hidden_weights

const void * hidden_effective_bias

Precomputed kernel_sum * lhs_offset

const void * bias

arm_nn_activation_type activation_type

struct cmsis_nn_lstm_params

CMSIS-NN object for LSTM parameters

int32_t time_major

0 if first dimension is batch, else first dimension is time

int32_t batch_size

int32_t time_steps

int32_t input_size

Size of new data input into the LSTM cell

int32_t hidden_size

Size of output from the LSTM cell, used as output and recursively into the next time step

int32_t input_offset

int32_t forget_to_cell_multiplier

int32_t forget_to_cell_shift

int32_t input_to_cell_multiplier

int32_t input_to_cell_shift

int32_t cell_clip

Min/max value of cell output

int32_t cell_scale_power

int32_t output_multiplier

int32_t output_shift

int32_t output_offset

cmsis_nn_lstm_gate forget_gate

cmsis_nn_lstm_gate input_gate

cmsis_nn_lstm_gate cell_gate

cmsis_nn_lstm_gate output_gate

struct cmsis_nn_lstm_context

CMSIS-NN object for LSTM scratch buffers.

There is no size field and no runtime enforcement: an undersized temp1 or temp2 is written past on every build target, so size them from the queries below, not by transcribing a formula.

void * temp1

Gate-vector scratch (int16_t elements for both the s8 and s16 layers). Sized by arm_lstm_unidirectional_s8_temp1_get_buffer_size() / arm_lstm_unidirectional_s16_temp1_get_buffer_size().

void * temp2

Cell-gate and tanh(cell_state) scratch (int16_t elements for both layers). Sized by arm_lstm_unidirectional_s8_temp2_get_buffer_size() / arm_lstm_unidirectional_s16_temp2_get_buffer_size().

void * cell_state

Cell-state buffer, batch_size * hidden_size int16_t elements for both layers.

void * hidden_state

Optional in/out hidden state for streaming; NULL selects stateless.

struct cmsis_nn_activation_f32

Activation clamp range for floating-point operators.

float32_t min

Minimum value used to clamp the result.

float32_t max

Maximum value used to clamp the result.

struct cmsis_nn_conv_params_f32

Convolution parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f32 activation

Output activation clamp range.

arm_nn_weight_format_flt weight_format

Filter storage format.

struct cmsis_nn_transpose_conv_params_f32

Transpose convolution parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile padding_offsets

Output padding adjustment for transpose convolution.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f32 activation

Output activation clamp range.

struct cmsis_nn_dw_conv_params_f32

Depthwise convolution parameters for float32 operators.

int32_t ch_mult

Channel multiplier. ch_mult * in_ch = out_ch.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f32 activation

Output activation clamp range.

struct cmsis_nn_pool_params_f32

Pooling parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_activation_f32 activation

Output activation clamp range.

struct cmsis_nn_fc_params_f32

Fully connected layer parameters for float32 operators.

cmsis_nn_activation_f32 activation

Output activation clamp range.

arm_nn_weight_format_flt weight_format

Weight storage format.

struct cmsis_nn_bmm_params_f32

Batched matrix multiplication parameters for float32 operators.

const bool adj_x

True when the left-hand-side operand is stored transposed.

const bool adj_y

True when the right-hand-side operand is stored transposed.

cmsis_nn_activation_f32 activation

Output activation clamp range.

arm_nn_weight_format_flt rhs_format

Right-hand-side operand storage format. ARM_NN_WEIGHT_FORMAT_NT_N_PACKED is currently supported only when adj_x == false and adj_y == false.

struct cmsis_nn_ew_params_f32

Elementwise operator parameters for float32 operators.

cmsis_nn_activation_f32 activation

Output activation clamp range.

struct cmsis_nn_transpose_params_f32

Transpose parameters for float32 operators.

int32_t num_dims

Number of active dimensions in the permutation.

int32_t perm[4]

Permutation indices.

arm_nn_tensor_layout layout

Layout convention used to interpret tensor dimensions.

struct cmsis_nn_svdf_params_f32

Singular value decomposition filter parameters for float32 operators.

int32_t rank

SVDF rank.

cmsis_nn_activation_f32 input_activation

Clamp range applied after the input projection.

cmsis_nn_activation_f32 output_activation

Clamp range applied to the final output.

struct cmsis_nn_lstm_gate_f32

Read-only weights and bias metadata for one float32 LSTM gate.

const float32_t * input_weights

Input-to-gate weight matrix.

const float32_t * hidden_weights

Hidden-state-to-gate weight matrix.

const float32_t * bias

Optional gate bias vector.

arm_nn_activation_type_flt activation_type

Gate activation selector.

struct cmsis_nn_lstm_params_f32

Parameters for a unidirectional float32 LSTM layer.

int32_t time_major

Non-zero when input/output tensors are time-major.

int32_t batch_size

Batch size processed per invocation.

int32_t time_steps

Number of time steps processed per invocation.

int32_t input_size

Input feature size per time step.

int32_t hidden_size

Hidden-state size.

float32_t cell_clip

Optional cell-state clip value.

cmsis_nn_lstm_gate_f32 forget_gate

Forget gate weights and activation.

cmsis_nn_lstm_gate_f32 input_gate

Input gate weights and activation.

cmsis_nn_lstm_gate_f32 cell_gate

Cell-update gate weights and activation.

cmsis_nn_lstm_gate_f32 output_gate

Output gate weights and activation.

struct cmsis_nn_lstm_context_f32

Scratch and mutable state buffers for a float32 LSTM invocation.

float32_t * temp1

Unused by the current implementation and may be NULL. Sized by arm_lstm_unidirectional_f32_temp1_get_buffer_size(), which reports 0; size from the query rather than hard-coding NULL if the buffer is arena-allocated.

float32_t * temp2

Unused by the current implementation and may be NULL. Sized by arm_lstm_unidirectional_f32_temp2_get_buffer_size().

float32_t * cell_state

Mutable cell-state buffer (in/out when streaming).

float32_t * hidden_state

Optional in/out hidden state for streaming; NULL selects stateless. Streaming is NULL-gated, so zero-initialise the context (e.g. designated initialisers) to keep legacy 3-field callers stateless. Matches the quantized cmsis_nn_lstm_context contract.

struct cmsis_nn_gru_gate_f32

Weights and biases for a single float32 GRU gate.

The activation (sigmoid for update/reset, tanh for candidate) is implied by the gate’s role and is not stored here. The reset-after formulation keeps the input-projection bias and the recurrent-projection bias separate, because the reset gate multiplies the recurrent projection (including its bias) after the matmul.

const float32_t * input_weights

Input-to-gate weight matrix [hidden_size, input_size].

const float32_t * hidden_weights

Hidden-to-gate weight matrix [hidden_size, hidden_size].

const float32_t * input_bias

Optional input-projection bias [hidden_size]. May be NULL.

const float32_t * hidden_bias

Optional recurrent-projection bias [hidden_size]. May be NULL.

struct cmsis_nn_gru_params_f32

Parameters for a float32 unidirectional GRU invocation.

GRU has three gates (update, reset, candidate) and, unlike LSTM, no cell state. The hidden state is the layer output.

int32_t time_major

Non-zero when input/output tensors are time-major.

int32_t batch_size

Batch size processed per invocation.

int32_t time_steps

Number of time steps processed per invocation.

int32_t input_size

Input feature size per time step.

int32_t hidden_size

Hidden-state size.

int32_t reset_after

Non-zero: reset gate applied after the recurrent matmul (Keras/TFLite default).

cmsis_nn_gru_gate_f32 update_gate

Update gate (z), sigmoid activation.

cmsis_nn_gru_gate_f32 reset_gate

Reset gate (r), sigmoid activation.

cmsis_nn_gru_gate_f32 candidate_gate

Candidate/new gate (n), tanh activation.

struct cmsis_nn_gru_context_f32

Scratch buffers for a float32 GRU invocation.

float32_t * temp1

Scratch required when reset_after == 0; sized by arm_gru_unidirectional_f32_temp1_get_buffer_size().

float32_t * hidden_state

Optional in/out persistent hidden state [hidden_size] for streaming (batch_size == 1).

struct cmsis_nn_activation_f16

Activation clamp range for floating-point operators.

float16_t min

Minimum value used to clamp the result.

float16_t max

Maximum value used to clamp the result.

struct cmsis_nn_conv_params_f16

Convolution parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f16 activation

Output activation clamp range.

arm_nn_weight_format_flt weight_format

Filter storage format.

struct cmsis_nn_transpose_conv_params_f16

Transpose convolution parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile padding_offsets

Output padding adjustment for transpose convolution.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f16 activation

Output activation clamp range.

struct cmsis_nn_dw_conv_params_f16

Depthwise convolution parameters for float32 operators.

int32_t ch_mult

Channel multiplier. ch_mult * in_ch = out_ch.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_tile dilation

Spatial dilation.

cmsis_nn_activation_f16 activation

Output activation clamp range.

struct cmsis_nn_pool_params_f16

Pooling parameters for float32 operators.

cmsis_nn_tile stride

Spatial stride.

cmsis_nn_tile padding

Spatial zero-padding.

cmsis_nn_activation_f16 activation

Output activation clamp range.

struct cmsis_nn_fc_params_f16

Fully connected layer parameters for float32 operators.

cmsis_nn_activation_f16 activation

Output activation clamp range.

arm_nn_weight_format_flt weight_format

Weight storage format.

struct cmsis_nn_bmm_params_f16

Batched matrix multiplication parameters for float32 operators.

const bool adj_x

True when the left-hand-side operand is stored transposed.

const bool adj_y

True when the right-hand-side operand is stored transposed.

cmsis_nn_activation_f16 activation

Output activation clamp range.

arm_nn_weight_format_flt rhs_format

Right-hand-side operand storage format. ARM_NN_WEIGHT_FORMAT_NT_N_PACKED is currently supported only when adj_x == false and adj_y == false.

struct cmsis_nn_ew_params_f16

Elementwise operator parameters for float32 operators.

cmsis_nn_activation_f16 activation

Output activation clamp range.

struct cmsis_nn_transpose_params_f16

Transpose parameters for float32 operators.

int32_t num_dims

Number of active dimensions in the permutation.

int32_t perm[4]

Permutation indices.

arm_nn_tensor_layout layout

Layout convention used to interpret tensor dimensions.

struct cmsis_nn_svdf_params_f16

Singular value decomposition filter parameters for float32 operators.

int32_t rank

SVDF rank.

cmsis_nn_activation_f16 input_activation

Clamp range applied after the input projection.

cmsis_nn_activation_f16 output_activation

Clamp range applied to the final output.

struct cmsis_nn_lstm_gate_f16

Read-only weights and bias metadata for one float32 LSTM gate.

const float16_t * input_weights

Input-to-gate weight matrix.

const float16_t * hidden_weights

Hidden-state-to-gate weight matrix.

const float16_t * bias

Optional gate bias vector.

arm_nn_activation_type_flt activation_type

Gate activation selector.

struct cmsis_nn_lstm_params_f16

Parameters for a unidirectional float32 LSTM layer.

int32_t time_major

Non-zero when input/output tensors are time-major.

int32_t batch_size

Batch size processed per invocation.

int32_t time_steps

Number of time steps processed per invocation.

int32_t input_size

Input feature size per time step.

int32_t hidden_size

Hidden-state size.

float16_t cell_clip

Optional cell-state clip value.

cmsis_nn_lstm_gate_f16 forget_gate

Forget gate weights and activation.

cmsis_nn_lstm_gate_f16 input_gate

Input gate weights and activation.

cmsis_nn_lstm_gate_f16 cell_gate

Cell-update gate weights and activation.

cmsis_nn_lstm_gate_f16 output_gate

Output gate weights and activation.

struct cmsis_nn_lstm_context_f16

Scratch and mutable state buffers for a float32 LSTM invocation.

float16_t * temp1

Unused by the current implementation and may be NULL. Sized by arm_lstm_unidirectional_f16_temp1_get_buffer_size(), which reports 0; size from the query rather than hard-coding NULL if the buffer is arena-allocated.

float16_t * temp2

Unused by the current implementation and may be NULL. Sized by arm_lstm_unidirectional_f16_temp2_get_buffer_size().

float16_t * cell_state

Mutable cell-state buffer (in/out when streaming).

float16_t * hidden_state

struct cmsis_nn_gru_gate_f16

Weights and biases for a single float16 GRU gate.

const float16_t * input_weights

Input-to-gate weight matrix [hidden_size, input_size].

const float16_t * hidden_weights

Hidden-to-gate weight matrix [hidden_size, hidden_size].

const float16_t * input_bias

Optional input-projection bias [hidden_size]. May be NULL.

const float16_t * hidden_bias

Optional recurrent-projection bias [hidden_size]. May be NULL.

struct cmsis_nn_gru_params_f16

Parameters for a float16 unidirectional GRU invocation.

GRU has three gates (update, reset, candidate) and, unlike LSTM, no cell state. The hidden state is the layer output.

int32_t time_major

Non-zero when input/output tensors are time-major.

int32_t batch_size

Batch size processed per invocation.

int32_t time_steps

Number of time steps processed per invocation.

int32_t input_size

Input feature size per time step.

int32_t hidden_size

Hidden-state size.

int32_t reset_after

Non-zero: reset gate applied after the recurrent matmul (Keras/TFLite default).

cmsis_nn_gru_gate_f16 update_gate

Update gate (z), sigmoid activation.

cmsis_nn_gru_gate_f16 reset_gate

Reset gate (r), sigmoid activation.

cmsis_nn_gru_gate_f16 candidate_gate

Candidate/new gate (n), tanh activation.

struct cmsis_nn_gru_context_f16

Scratch buffers for a float16 GRU invocation.

float16_t * temp1

Scratch required when reset_after == 0; sized by arm_gru_unidirectional_f16_temp1_get_buffer_size().

float16_t * hidden_state

Optional in/out persistent hidden state [hidden_size] for streaming (batch_size == 1).

enum arm_nn_tensor_layout

Tensor layout selector for floating-point APIs.

Float public APIs currently accept NHWC layout only.

ARM_NN_LAYOUT_NHWC = 0

Tensor dimensions are ordered as [N, H, W, C].

enum arm_nn_activation_type

Enum for specifying activation function types

ARM_SIGMOID = 0

Sigmoid activation function

ARM_TANH = 1

Tanh activation function

enum arm_nn_activation_type_flt

Activation selector for floating-point operator APIs.

Numeric values intentionally live in a dedicated floating-point range to avoid overlap with the legacy integer public activation enum.

ARM_NN_FLT_ACT_NONE = 32

Identity activation function.

ARM_NN_FLT_ACT_SIGMOID = 33

Sigmoid activation function.

ARM_NN_FLT_ACT_TANH = 34

Hyperbolic tangent activation function.

ARM_NN_FLT_ACT_RELU = 35

ReLU activation function.

ARM_NN_FLT_ACT_RELU6 = 36

ReLU6 activation function.

ARM_NN_FLT_ACT_HARDSWISH = 37

Hard-swish activation function.

ARM_NN_FLT_ACT_LEAKY_RELU = 38

Leaky ReLU activation function.

enum arm_nn_compare_operation

Enum for specifying comparison operator

ARM_COMPARE_EQUAL = 0

Returns 1 if lhs == rhs else 0

ARM_COMPARE_NOT_EQUAL = 1

Returns 1 if lhs != rhs else 0

ARM_COMPARE_GREATER = 2

Returns 1 if lhs > rhs else 0

ARM_COMPARE_GREATER_EQUAL = 3

Returns 1 if lhs >= rhs else 0

ARM_COMPARE_LESS = 4

Returns 1 if lhs < rhs else 0

ARM_COMPARE_LESS_EQUAL = 5

Returns 1 if lhs <= rhs else 0

enum arm_cmsis_nn_status

Function return codes

ARM_CMSIS_NN_SUCCESS = 0

No error

ARM_CMSIS_NN_ARG_ERROR = -1

One or more arguments are incorrect

ARM_CMSIS_NN_NO_IMPL_ERROR = -2

No implementation available

ARM_CMSIS_NN_FAILURE = -3

Logical error

enum arm_nn_dw_kernel_layout_f32

Depthwise kernel storage layout selector for floating-point kernels.

Public float depthwise entry points currently use KC storage ([k][c]).

ARM_NN_DW_KERNEL_KC = 0

Depthwise kernel stored as [kernel][channel].

ARM_NN_DW_KERNEL_CK = 1

Depthwise kernel stored as [channel][kernel].

enum arm_nn_weight_format_flt

Weight storage format selector for floating-point operators.

This enum allows frameworks to describe whether weights are provided in the standard public operator layout or in a backend-specific packed layout.

ARM_NN_WEIGHT_FORMAT_NT_N_PACKED matches the packed RHS layout consumed by arm_nn_mat_mult_nt_n_packed_f16/f32.

The packed NTxN layout exists because MVE kernels typically perform best when output-channel blocks can be loaded contiguously. With the standard NT x T formulation, vectorizing over output channels tends to require gather-load accesses to the RHS, which is less efficient than a packed non-transposed RHS layout.

For operators that support ARM_NN_WEIGHT_FORMAT_NT_N_PACKED, supplying offline-repacked constant weights in this layout is therefore generally the preferred way to achieve the best MVE performance.

ARM_NN_WEIGHT_FORMAT_STANDARD = 0

Standard public operator layout.

ARM_NN_WEIGHT_FORMAT_NT_N_PACKED = 1

Packed [K][N-block] layout for NTxN matmul helpers.

typedef arm_nn_dw_kernel_layout_f32 arm_nn_dw_kernel_layout_f16