File syn_nn.h

FileList > src > syntropic > util > syn_nn.h

Go to the source code of this file

Zero-heap TinyML Neural Network Engine with Attention & Quantization Scaling. More...

• #include "syn_qmath.h"
• #include "syntropic/common/syn_defs.h"
• #include "syntropic/pt/syn_pt.h"

Classes

Type	Name
struct	SYN_NN_Layer Layer Descriptor Struct.
struct	SYN_NN_Model Model Descriptor Struct.
struct	syn_nn_quant_t Affine Quantization Scaling Parameters (Scale & Zero-Point).

Public Types

Type	Name
enum	SYN_NN_Activation Neural Network Activation Functions.
enum	SYN_NN_LayerType Layer Type Enum for Declarative Models.

Public Functions

Type	Name
size_t	syn_nn_argmax_q7 (const q7_t * outputs, size_t num_outputs) Find the class index with the highest output value (ArgMax).
SYN_Status	syn_nn_attention_q7 (const q7_t * q, const q7_t * k, const q7_t * v, size_t seq_len, size_t d_k, size_t d_v, q7_t * out, uint8_t attn_shift) INT8 Quantized Scaled Dot-Product Self-Attention Layer.
SYN_Status	syn_nn_avgpool1d_q7 (const q7_t * inputs, size_t seq_len, size_t num_channels, q7_t * outputs, size_t pool_size, size_t stride) 1D Average Pooling Layer for INT8 Feature Maps.
SYN_PT_Status	syn_nn_conv1d_pt (SYN_PT * pt, const q7_t * inputs, size_t seq_len, size_t num_channels, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_filters, size_t kernel_size, size_t stride, SYN_NN_Activation act, uint8_t out_shift, size_t * current_step, size_t chunk_size) Evaluate a 1D Temporal Convolution layer cooperatively inside a protothread.
SYN_Status	syn_nn_conv1d_q7 (const q7_t * inputs, size_t seq_len, size_t num_channels, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_filters, size_t kernel_size, size_t stride, SYN_NN_Activation act, uint8_t out_shift) Evaluate a 1D Temporal Convolutional Neural Network layer in INT8 (q7_t).
SYN_Status	syn_nn_conv1d_quant_q7 (const q7_t * inputs, size_t seq_len, size_t num_channels, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_filters, size_t kernel_size, size_t stride, SYN_NN_Activation act, const syn_nn_quant_t * quant) Evaluate a 1D Temporal Convolution layer with affine quantization scaling.
SYN_PT_Status	syn_nn_dense_pt (SYN_PT * pt, const q7_t * inputs, size_t num_inputs, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_outputs, SYN_NN_Activation act, uint8_t out_shift, size_t * current_neuron, size_t chunk_size) Evaluate a Dense Neural Network layer cooperatively inside a protothread.
SYN_Status	syn_nn_dense_q7 (const q7_t * inputs, size_t num_inputs, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_outputs, SYN_NN_Activation act, uint8_t out_shift) Evaluate a Dense (Fully Connected) Neural Network layer using INT8 (q7_t) weights.
SYN_Status	syn_nn_dense_quant_q7 (const q7_t * inputs, size_t num_inputs, const q7_t * weights, const q16_t * biases, q7_t * outputs, size_t num_outputs, SYN_NN_Activation act, const syn_nn_quant_t * quant) Evaluate a Dense layer with affine quantization scaling.
SYN_Status	syn_nn_maxpool1d_q7 (const q7_t * inputs, size_t seq_len, size_t num_channels, q7_t * outputs, size_t pool_size, size_t stride) 1D Max Pooling Layer for INT8 Feature Maps.
SYN_Status	syn_nn_softmax_q7 (const q7_t * inputs, q7_t * outputs, size_t num_inputs) Compute normalized Softmax probability distribution over input logits in Q7.

Detailed Description

Implements quantized INT8 (q7_t) Dense, Attention, Softmax, and Activation functions with zero dynamic memory allocation.

Public Types Documentation

enum SYN_NN_Activation

Neural Network Activation Functions.

enum SYN_NN_Activation {
    SYN_NN_ACT_NONE = 0,
    SYN_NN_ACT_RELU,
    SYN_NN_ACT_LEAKY_RELU,
    SYN_NN_ACT_SIGMOID,
    SYN_NN_ACT_TANH
};

---

enum SYN_NN_LayerType

Layer Type Enum for Declarative Models.

enum SYN_NN_LayerType {
    SYN_NN_LAYER_DENSE = 0,
    SYN_NN_LAYER_ATTENTION
};

---

Public Functions Documentation

function syn_nn_argmax_q7

Find the class index with the highest output value (ArgMax).

size_t syn_nn_argmax_q7 (
    const q7_t * outputs,
    size_t num_outputs
) 

Parameters:

• outputs Pointer to output vector.
• num_outputs Number of output elements.

Returns:

Index of maximum value element (0 if num_outputs == 0 or NULL).

---

function syn_nn_attention_q7

INT8 Quantized Scaled Dot-Product Self-Attention Layer.

SYN_Status syn_nn_attention_q7 (
    const q7_t * q,
    const q7_t * k,
    const q7_t * v,
    size_t seq_len,
    size_t d_k,
    size_t d_v,
    q7_t * out,
    uint8_t attn_shift
) 

Computes Attention(Q, K, V) = Softmax( (Q * K^T) >> attn_shift ) * V

Parameters:

• q Query matrix [seq_len * d_k].
• k Key matrix [seq_len * d_k].
• v Value matrix [seq_len * d_v].
• seq_len Sequence length / token count.
• d_k Query/Key dimension per token.
• d_v Value dimension per token.
• out Output matrix [seq_len * d_v].
• attn_shift Right bit-shift for dot-product scaling.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_avgpool1d_q7

1D Average Pooling Layer for INT8 Feature Maps.

SYN_Status syn_nn_avgpool1d_q7 (
    const q7_t * inputs,
    size_t seq_len,
    size_t num_channels,
    q7_t * outputs,
    size_t pool_size,
    size_t stride
) 

Parameters:

• inputs Pointer to input feature matrix.
• seq_len Input sequence length.
• num_channels Number of feature channels.
• outputs Destination output matrix.
• pool_size Pooling window size.
• stride Stride step size.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_conv1d_pt

Evaluate a 1D Temporal Convolution layer cooperatively inside a protothread.

SYN_PT_Status syn_nn_conv1d_pt (
    SYN_PT * pt,
    const q7_t * inputs,
    size_t seq_len,
    size_t num_channels,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_filters,
    size_t kernel_size,
    size_t stride,
    SYN_NN_Activation act,
    uint8_t out_shift,
    size_t * current_step,
    size_t chunk_size
) 

Parameters:

• pt Pointer to protothread state machine.
• inputs Pointer to input matrix.
• seq_len Sequence length.
• num_channels Number of input channels.
• weights Kernel weights matrix.
• biases Bias vector or NULL.
• outputs Destination output matrix.
• num_filters Number of output filters.
• kernel_size Kernel window size.
• stride Stride step.
• act Activation function.
• out_shift Right bit-shift scaling factor.
• current_step State variable tracking step progress across yields.
• chunk_size Number of output steps to evaluate per tick.

Returns:

SYN_PT_YIELDING while evaluating, SYN_PT_ENDED on completion.

---

function syn_nn_conv1d_q7

Evaluate a 1D Temporal Convolutional Neural Network layer in INT8 (q7_t).

SYN_Status syn_nn_conv1d_q7 (
    const q7_t * inputs,
    size_t seq_len,
    size_t num_channels,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_filters,
    size_t kernel_size,
    size_t stride,
    SYN_NN_Activation act,
    uint8_t out_shift
) 

Scans a 1D kernel filter matrix [num_filters * kernel_size * num_channels] across a time series [seq_len * num_channels].

Parameters:

• inputs Pointer to input matrix [seq_len * num_channels].
• seq_len Input sequence length (time steps).
• num_channels Number of input channels/features per time step.
• weights Flat kernel weights matrix [num_filters * kernel_size * num_channels].
• biases Bias vector of length num_filters in Q16 (or NULL).
• outputs Destination output matrix [out_steps * num_filters].
• num_filters Number of output filters/channels.
• kernel_size Size of 1D sliding window kernel.
• stride Stride step across time steps.
• act Activation function to apply to output filters.
• out_shift Right bit-shift scaling factor (0 to 15) to prevent overflow.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_conv1d_quant_q7

Evaluate a 1D Temporal Convolution layer with affine quantization scaling.

SYN_Status syn_nn_conv1d_quant_q7 (
    const q7_t * inputs,
    size_t seq_len,
    size_t num_channels,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_filters,
    size_t kernel_size,
    size_t stride,
    SYN_NN_Activation act,
    const syn_nn_quant_t * quant
) 

Parameters:

• inputs Pointer to input matrix.
• seq_len Sequence length.
• num_channels Number of input channels.
• weights Kernel weights matrix.
• biases Bias vector or NULL.
• outputs Destination output matrix.
• num_filters Number of output filters.
• kernel_size Kernel window size.
• stride Stride step.
• act Activation function.
• quant Pointer to affine quantization scaling parameters.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_dense_pt

Evaluate a Dense Neural Network layer cooperatively inside a protothread.

SYN_PT_Status syn_nn_dense_pt (
    SYN_PT * pt,
    const q7_t * inputs,
    size_t num_inputs,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_outputs,
    SYN_NN_Activation act,
    uint8_t out_shift,
    size_t * current_neuron,
    size_t chunk_size
) 

Parameters:

• pt Pointer to protothread state machine.
• inputs Pointer to input vector.
• num_inputs Number of input features.
• weights Flat weight matrix.
• biases Bias vector or NULL.
• outputs Destination output vector.
• num_outputs Number of output neurons.
• act Activation function to apply.
• out_shift Right bit-shift scaling factor.
• current_neuron State variable tracking progress across yields.
• chunk_size Number of neurons to evaluate per protothread tick.

Returns:

SYN_PT_YIELDING while evaluating, SYN_PT_ENDED on completion.

---

function syn_nn_dense_q7

Evaluate a Dense (Fully Connected) Neural Network layer using INT8 (q7_t) weights.

SYN_Status syn_nn_dense_q7 (
    const q7_t * inputs,
    size_t num_inputs,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_outputs,
    SYN_NN_Activation act,
    uint8_t out_shift
) 

Computes: Output[i] = Activation( ((Sum(Input[j] * Weight[i][j]) + Bias[i]) >> out_shift) )

Parameters:

• inputs Pointer to input vector (length = num_inputs).
• num_inputs Number of input features.
• weights Flat weight matrix [num_outputs * num_inputs].
• biases Bias vector of length num_outputs in Q16.16 (or NULL).
• outputs Destination buffer for outputs (length = num_outputs).
• num_outputs Number of output neurons in layer.
• act Activation function to apply.
• out_shift Right bit-shift scaling factor (0 to 15) to prevent overflow.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_dense_quant_q7

Evaluate a Dense layer with affine quantization scaling.

SYN_Status syn_nn_dense_quant_q7 (
    const q7_t * inputs,
    size_t num_inputs,
    const q7_t * weights,
    const q16_t * biases,
    q7_t * outputs,
    size_t num_outputs,
    SYN_NN_Activation act,
    const syn_nn_quant_t * quant
) 

Parameters:

• inputs Pointer to input vector.
• num_inputs Number of input features.
• weights Flat weight matrix.
• biases Bias vector or NULL.
• outputs Destination output vector.
• num_outputs Number of output neurons.
• act Activation function.
• quant Pointer to affine quantization scaling parameters.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_maxpool1d_q7

1D Max Pooling Layer for INT8 Feature Maps.

SYN_Status syn_nn_maxpool1d_q7 (
    const q7_t * inputs,
    size_t seq_len,
    size_t num_channels,
    q7_t * outputs,
    size_t pool_size,
    size_t stride
) 

Parameters:

• inputs Pointer to input feature matrix.
• seq_len Input sequence length.
• num_channels Number of feature channels.
• outputs Destination output matrix.
• pool_size Pooling window size.
• stride Stride step size.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

function syn_nn_softmax_q7

Compute normalized Softmax probability distribution over input logits in Q7.

SYN_Status syn_nn_softmax_q7 (
    const q7_t * inputs,
    q7_t * outputs,
    size_t num_inputs
) 

Output values sum to 127 (+1.0 in Q7).

Parameters:

• inputs Pointer to logit input vector.
• outputs Pointer to destination probability vector.
• num_inputs Number of features / classes.

Returns:

SYN_OK on success, SYN_INVALID_PARAM on failure.

---

---

The documentation for this class was generated from the following file src/syntropic/util/syn_nn.h