Internal Helpers

Collaboration diagram for Internal Helpers:

Files
file	noodle_internal.h
	Private declarations shared by Noodle implementation files.

Functions
float *	noodle_temp1_require (size_t required_floats)
	Ensure temp buffer 1 can hold a number of floats.
float *	noodle_temp2_require (size_t required_floats)
	Ensure temp buffer 2 can hold a number of floats.
void	noodle_temp_buffers_free (void)
	Free Noodle-owned scratch buffers and detach external scratch buffers.
float *	noodle_slice (float *flat, size_t W, size_t z)
	Return a channel plane from a packed [Z][W][W] tensor.
size_t	noodle_read_raw (NDL_File &f, void *dst, size_t n)
	Read raw bytes from a backend file handle.
size_t	noodle_write_raw (NDL_File &f, const void *src, size_t n)
	Write raw bytes to a backend file handle.
size_t	noodle_read_float_block (NDL_File &f, float *dst, size_t n_floats)
	Read a block of floats using the configured scalar file format.
float	noodle_dot_float_block (const float *x, const float *w, uint16_t n)
	Compute a dot product with a small unrolled loop.
uint16_t	noodle_do_conv1d (float *input, float *kernel, uint16_t W, uint16_t K, float *output, uint16_t P, uint16_t S)
	Accumulate one 1D convolution into an output sequence.
uint16_t	noodle_do_pooling1d (float *input, uint16_t W, uint16_t K, uint16_t S, const char *fn)
	Apply valid 1D max pooling and write to a file.
uint16_t	noodle_do_pooling1d (float *input, uint16_t W, uint16_t K, uint16_t S, NDL_File &fo)
	Apply valid 1D max pooling and write to an open file.
uint16_t	noodle_do_pooling1d (const float *input, uint16_t W, uint16_t K, uint16_t S, float *output)
	Apply valid 1D pooling and write to memory.
float	noodle_get_padded_x (byte *grid, int16_t i, int16_t j, int16_t W, int16_t P0, int16_t P1)
	Read a byte grid sample with asymmetric zero padding.
float	noodle_get_padded_x (float *grid, int16_t i, int16_t j, int16_t W, int16_t P0, int16_t P1)
	Read a float grid sample with asymmetric zero padding.
uint16_t	noodle_do_bias (float *output, float bias, uint16_t n)
	Add bias to a square output map and apply ReLU.
uint16_t	noodle_do_pooling (const float *input, uint16_t W, uint16_t K, uint16_t S, const char *fn)
	Apply 2D pooling and write to a file.
uint16_t	noodle_do_pooling (const float *input, uint16_t W, uint16_t K, uint16_t S, NDL_File &fo)
	Apply 2D pooling and write to an open file.
uint16_t	noodle_do_pooling (const float *input, uint16_t W, uint16_t K, uint16_t S, float *output)
	Apply 2D pooling and write to memory.
uint16_t	noodle_do_conv (byte *grid, const float *kernel, uint16_t K, uint16_t W, float *output, uint16_t P, uint16_t S)
	Accumulate one byte-input 2D convolution plane.
uint16_t	noodle_do_conv (float *grid, const float *kernel, uint16_t K, uint16_t W, float *output, uint16_t P, uint16_t S)
	Accumulate one float-input 2D convolution plane.
void	noodle_reset_buffer (float *buffer, uint16_t n)
	Clear a float buffer.
uint16_t	noodle_do_bias_act (float *output, float bias, uint16_t n, Activation act)
	Add bias to a square output map and apply the requested activation.
uint16_t	noodle_do_conv_transpose (float *input, const float *kernel, uint16_t K, uint16_t W, float *output, uint16_t P, uint16_t S, uint16_t OP)
	Accumulate one 2D transpose-convolution plane.
void	noodle_find_max (float *input, uint16_t n, float &max_val, uint16_t &max_idx)
	Find the maximum value and its index in a vector.
void	noodle_unpack_bn_params (const float *bn_params, uint16_t N, const float **gamma, const float **beta, const float **mean, const float **var)
	Split packed batch-normalization parameters into four arrays.
uint16_t	noodle_bn1d (float *x, uint16_t N, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Apply 1D batch normalization in place to a raw vector.
uint16_t	noodle_bn1d (float *x, uint16_t N, const float *bn_params, float eps)
	Apply 1D batch normalization from packed parameters.
uint16_t	noodle_bn1d_relu (float *x, uint16_t N, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Apply 1D batch normalization followed by ReLU in place.
uint16_t	noodle_bn1d_relu (float *x, uint16_t N, const float *bn_params, float eps)
	Apply packed 1D batch normalization followed by ReLU in place.
uint16_t	noodle_bn2d (float *x, uint16_t C, uint16_t W, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Apply 2D channel-wise batch normalization in place.
uint16_t	noodle_bn2d (float *x, uint16_t C, uint16_t W, const float *bn_params, float eps)
	Apply 2D channel-wise batch normalization from packed parameters.
uint16_t	noodle_bn2d_relu (float *x, uint16_t C, uint16_t W, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Apply 2D channel-wise batch normalization followed by ReLU.
uint16_t	noodle_bn2d_relu (float *x, uint16_t C, uint16_t W, const float *bn_params, float eps)
	Apply packed 2D batch normalization followed by ReLU.
uint16_t	noodle_compute_V (uint16_t K, uint16_t W, uint16_t P, uint16_t S)
	Compute 2D convolution output width.
uint16_t	noodle_compute_V_and_P (uint16_t K, uint16_t W, uint16_t P, uint16_t S, uint16_t &P0, uint16_t &P1)
	Compute 2D convolution output width and effective asymmetric padding.
uint16_t	noodle_valid_max_pool (float *inplace, uint16_t W, uint16_t C, const Pool &pool)
	Apply valid max pooling to a packed channel-first tensor in place.
uint16_t	noodle_compute_Vt (uint16_t K, uint16_t W, uint16_t P, uint16_t S, uint16_t OP)
	Compute transpose-convolution output width.
uint16_t	noodle_compute_Vt_and_P (uint16_t K, uint16_t W, uint16_t P, uint16_t S, uint16_t OP, uint16_t &P0, uint16_t &P1)
	Compute transpose-convolution output width and effective padding.
uint16_t	noodle_conv1d (float *in, uint16_t n_inputs, float *out, uint16_t n_outputs, uint16_t W, const ConvMem &conv, CBFPtr progress_cb)
	Raw memory-to-memory 1D convolution without pooling.
uint16_t	noodle_conv1d (float *in, uint16_t n_inputs, float *out, uint16_t n_outputs, uint16_t W, const ConvMem &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory 1D convolution with pooling.
uint16_t	noodle_conv1d (float *in, uint16_t n_inputs, const char *out_fn, uint16_t n_outputs, uint16_t W, const ConvMem &conv, CBFPtr progress_cb)
	Raw memory-to-file 1D convolution without pooling.
uint16_t	noodle_conv1d (const char *in_fn, uint16_t n_inputs, float *out, uint16_t n_outputs, uint16_t W, const ConvMem &conv, CBFPtr progress_cb)
	Raw file-to-memory 1D convolution without pooling.
uint16_t	noodle_conv_float (const char *in_fn, uint16_t n_inputs, uint16_t n_outputs, float *output, uint16_t W, const Conv &conv, const Pool &pool, CBFPtr progress_cb)
	File-to-memory 2D convolution with file-backed parameters.
uint16_t	noodle_conv_float (float *input, uint16_t n_inputs, uint16_t n_outputs, const char *out_fn, uint16_t W, const Conv &conv, const Pool &pool, CBFPtr progress_cb)
	Memory-to-file 2D convolution with file-backed parameters.
uint16_t	noodle_conv_float (float *input, uint16_t n_inputs, uint16_t n_outputs, const char *out_fn, uint16_t W, const ConvMem &conv, const Pool &pool, CBFPtr progress_cb)
	Memory-to-file 2D convolution with memory-backed parameters.
uint16_t	noodle_conv_float (float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, uint16_t W, const Conv &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory 2D convolution with file-backed parameters.
uint16_t	noodle_conv_float (float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, uint16_t W, const ConvMem &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory 2D convolution with memory-backed parameters.
uint16_t	noodle_conv_float (float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, uint16_t W, const ConvProgmem &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory 2D convolution with near-PROGMEM parameters.
uint16_t	noodle_conv_transpose_float (float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, uint16_t W, const ConvMem &conv, CBFPtr progress_cb)
	Raw memory-to-memory 2D transpose convolution.
uint16_t	noodle_dwconv_float (float *input, uint16_t n_channels, float *output, uint16_t W, const Conv &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory depthwise convolution with file-backed parameters.
uint16_t	noodle_dwconv_float (float *input, uint16_t n_channels, float *output, uint16_t W, const ConvMem &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory depthwise convolution with memory-backed parameters.
uint16_t	noodle_dwconv_float (float *input, uint16_t n_channels, float *output, uint16_t W, const ConvProgmem &conv, const Pool &pool, CBFPtr progress_cb)
	Raw memory-to-memory depthwise convolution with near-PROGMEM parameters.
uint16_t	noodle_fcn (const byte *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNFile &fcn, CBFPtr progress_cb)
	Byte-input fully connected layer with file-backed parameters.
uint16_t	noodle_fcn (const int8_t *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNFile &fcn, CBFPtr progress_cb)
	Int8-input fully connected layer with file-backed parameters.
uint16_t	noodle_fcn (const float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNFile &fcn, CBFPtr progress_cb)
	Float-input fully connected layer with file-backed parameters.
uint16_t	noodle_fcn (const float *input, uint16_t n_inputs, uint16_t n_outputs, const char *out_fn, const FCNFile &fcn, CBFPtr progress_cb)
	Float-input fully connected layer that writes output to a file.
uint16_t	noodle_fcn (const char *in_fn, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNFile &fcn, CBFPtr progress_cb)
	File-input fully connected layer that writes output to memory.
uint16_t	noodle_fcn (const float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNMem &fcn, CBFPtr progress_cb)
	Float-input fully connected layer with memory-backed parameters.
uint16_t	noodle_fcn (const float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const FCNProgmem &fcn, CBFPtr progress_cb)
	Float-input fully connected layer with far-PROGMEM parameters.
uint16_t	noodle_fcn_progmem (const float *input, uint16_t n_inputs, uint16_t n_outputs, float *output, const float *weight, const float *bias, Activation act, CBFPtr progress_cb)
	Float-input fully connected layer with near-PROGMEM parameters.
void	noodle_array_to_file (float *array, NDL_File &fo, uint16_t n)
	Write a float array to an already-open file.
void	noodle_grid_to_file (byte *grid, NDL_File &fo, uint16_t n)
	Write a byte grid to an already-open file.
void	noodle_grid_to_file (float *grid, NDL_File &fo, uint16_t n)
	Write a float grid to an already-open file.
void	noodle_array_from_file (NDL_File &fi, float *buffer, uint16_t K)
	Read a float array from an already-open file.
void	noodle_grid_from_file (NDL_File &fi, byte *buffer, uint16_t K)
	Read a byte grid from an already-open file.
void	noodle_grid_from_file (NDL_File &fi, int8_t *buffer, uint16_t K)
	Read an int8 grid from an already-open file.
void	noodle_grid_from_file (NDL_File &fi, float *buffer, uint16_t K)
	Read a float grid from an already-open file.
void	noodle_copy_kernel_progmem (const float *w, uint32_t base, uint16_t K, float *kernel)
	Copy one square kernel from near-PROGMEM into RAM.
uint16_t	noodle_flat (const char *in_fn, float *output, uint16_t V, uint16_t n_filters)
	Flatten a packed file tensor into an HWC-like raw vector.
uint16_t	noodle_flat (float *input, float *output, uint16_t V, uint16_t n_filters)
	Flatten a packed memory tensor into an HWC-like raw vector.
uint16_t	noodle_reshape (const float *src_hwc, float *dst_chw, uint16_t W, uint16_t C)
	Convert HWC-like raw data to packed channel-first raw data.
uint16_t	noodle_gap (float *inout, uint16_t C, uint16_t W)
	Apply global average pooling in place to packed channel-first maps.
uint16_t	noodle_gmp (float *inout, uint16_t C, uint16_t W)
	Apply global max pooling in place to packed channel-first data.
uint16_t	noodle_soft_max (float *input_output, uint16_t n)
	Apply numerically stabilized softmax in place.
uint16_t	noodle_sigmoid (float *input_output, uint16_t n)
	Apply sigmoid in place.
float	noodle_sigmoidf (float x)
	Compute sigmoid for one scalar.
uint16_t	noodle_logit (float *input_output, uint16_t n)
	Apply logistic sigmoid in place.
uint16_t	noodle_relu (float *input_output, uint16_t n)
	Apply ReLU in place.
uint16_t	noodle_bn (float *x, uint16_t C, uint16_t W, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Backward-compatible raw alias for noodle_bn2d().
uint16_t	noodle_bn (float *x, uint16_t C, uint16_t W, const float *bn_params, float eps)
	Backward-compatible raw alias for packed-parameter noodle_bn2d().
uint16_t	noodle_bn_relu (float *x, uint16_t C, uint16_t W, const float *gamma, const float *beta, const float *mean, const float *var, float eps)
	Backward-compatible raw alias for noodle_bn2d_relu().
uint16_t	noodle_bn_relu (float *x, uint16_t C, uint16_t W, const float *bn_params, float eps)
	Backward-compatible raw alias for packed-parameter noodle_bn2d_relu().

Detailed Description

Maintainer-facing helpers used by Noodle implementation files.

Function Documentation

noodle_array_from_file()

void noodle_array_from_file	(	NDL_File &	fi,
		float *	buffer,
		uint16_t	K )

Read a float array from an already-open file.

noodle_array_to_file()

void noodle_array_to_file	(	float *	array,
		NDL_File &	fo,
		uint16_t	n )

Write a float array to an already-open file.

noodle_bn() [1/2]

uint16_t noodle_bn	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	bn_params,
		float	eps )

Backward-compatible raw alias for packed-parameter noodle_bn2d().

Treats x as packed [C][W][W]. bn_params is packed as [gamma[C]][beta[C]][mean[C]][var[C]].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn() [2/2]

uint16_t noodle_bn	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Backward-compatible raw alias for noodle_bn2d().

Treats x as packed [C][W][W].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn1d() [1/2]

uint16_t noodle_bn1d	(	float *	x,
		uint16_t	N,
		const float *	bn_params,
		float	eps )

Apply 1D batch normalization from packed parameters.

bn_params is packed as [gamma[N]][beta[N]][mean[N]][var[N]].

Parameters

x	Vector updated in place.
N	Number of vector elements.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

N.

noodle_bn1d() [2/2]

uint16_t noodle_bn1d	(	float *	x,
		uint16_t	N,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Apply 1D batch normalization in place to a raw vector.

gamma, beta, mean, and var each contain N values.

Parameters

x	Vector updated in place.
N	Number of vector elements.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

N.

noodle_bn1d_relu() [1/2]

uint16_t noodle_bn1d_relu	(	float *	x,
		uint16_t	N,
		const float *	bn_params,
		float	eps )

Apply packed 1D batch normalization followed by ReLU in place.

bn_params is packed as [gamma[N]][beta[N]][mean[N]][var[N]].

Parameters

x	Vector updated in place.
N	Number of vector elements.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

N.

noodle_bn1d_relu() [2/2]

uint16_t noodle_bn1d_relu	(	float *	x,
		uint16_t	N,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Apply 1D batch normalization followed by ReLU in place.

gamma, beta, mean, and var each contain N values.

Parameters

x	Vector updated in place.
N	Number of vector elements.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

N.

noodle_bn2d() [1/2]

uint16_t noodle_bn2d	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	bn_params,
		float	eps )

Apply 2D channel-wise batch normalization from packed parameters.

Treats x as packed [C][W][W]. bn_params is packed as [gamma[C]][beta[C]][mean[C]][var[C]].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn2d() [2/2]

uint16_t noodle_bn2d	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Apply 2D channel-wise batch normalization in place.

Treats x as packed [C][W][W]. The parameter arrays each contain C values, one per channel.

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn2d_relu() [1/2]

uint16_t noodle_bn2d_relu	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	bn_params,
		float	eps )

Apply packed 2D batch normalization followed by ReLU.

Treats x as packed [C][W][W]. bn_params is packed as [gamma[C]][beta[C]][mean[C]][var[C]].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn2d_relu() [2/2]

uint16_t noodle_bn2d_relu	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Apply 2D channel-wise batch normalization followed by ReLU.

Treats x as packed [C][W][W]. The parameter arrays each contain C values, one per channel.

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn_relu() [1/2]

uint16_t noodle_bn_relu	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	bn_params,
		float	eps )

Backward-compatible raw alias for packed-parameter noodle_bn2d_relu().

Treats x as packed [C][W][W]. bn_params is packed as [gamma[C]][beta[C]][mean[C]][var[C]].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
bn_params	Packed batch-normalization parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_bn_relu() [2/2]

uint16_t noodle_bn_relu	(	float *	x,
		uint16_t	C,
		uint16_t	W,
		const float *	gamma,
		const float *	beta,
		const float *	mean,
		const float *	var,
		float	eps )

Backward-compatible raw alias for noodle_bn2d_relu().

Treats x as packed [C][W][W].

Parameters

x	Tensor updated in place.
C	Number of channels.
W	Plane width and height.
gamma	Scale parameters.
beta	Offset parameters.
mean	Moving-mean parameters.
var	Moving-variance parameters.
eps	Small value added to variance before inversion.

Returns

W.

noodle_compute_V()

uint16_t noodle_compute_V	(	uint16_t	K,
		uint16_t	W,
		uint16_t	P,
		uint16_t	S )

Compute 2D convolution output width.

Parameters

K	Kernel width.
W	Input width.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.

Returns

Output width.

noodle_compute_V_and_P()

uint16_t noodle_compute_V_and_P	(	uint16_t	K,
		uint16_t	W,
		uint16_t	P,
		uint16_t	S,
		uint16_t &	P0,
		uint16_t &	P1 )

Compute 2D convolution output width and effective asymmetric padding.

Parameters

K	Kernel width.
W	Input width.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.
P0	Receives top/left padding.
P1	Receives bottom/right padding.

Returns

Output width.

noodle_compute_Vt()

uint16_t noodle_compute_Vt	(	uint16_t	K,
		uint16_t	W,
		uint16_t	P,
		uint16_t	S,
		uint16_t	OP )

Compute transpose-convolution output width.

For explicit padding, this computes V = (W - 1) * S - 2 * P + K + OP; callers choose OP to match the desired output width. With P == 65535, SAME-style output uses V = W * S.

Parameters

K	Kernel width.
W	Input width.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.
OP	User-computed output padding for explicit padding.

Returns

Output width.

noodle_compute_Vt_and_P()

uint16_t noodle_compute_Vt_and_P	(	uint16_t	K,
		uint16_t	W,
		uint16_t	P,
		uint16_t	S,
		uint16_t	OP,
		uint16_t &	P0,
		uint16_t &	P1 )

Compute transpose-convolution output width and effective padding.

For explicit padding, this computes V = (W - 1) * S - 2 * P + K + OP; callers choose OP to match the desired output width. With P == 65535, SAME-style output uses V = W * S and derives asymmetric crop values.

Parameters

K	Kernel width.
W	Input width.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.
OP	User-computed output padding for explicit padding.
P0	Receives top/left padding.
P1	Receives bottom/right padding.

Returns

Output width.

noodle_conv1d() [1/4]

uint16_t noodle_conv1d	(	const char *	in_fn,
		uint16_t	n_inputs,
		float *	out,
		uint16_t	n_outputs,
		uint16_t	W,
		const ConvMem &	conv,
		CBFPtr	progress_cb )

Raw file-to-memory 1D convolution without pooling.

noodle_conv1d() [2/4]

uint16_t noodle_conv1d	(	float *	in,
		uint16_t	n_inputs,
		const char *	out_fn,
		uint16_t	n_outputs,
		uint16_t	W,
		const ConvMem &	conv,
		CBFPtr	progress_cb )

Raw memory-to-file 1D convolution without pooling.

noodle_conv1d() [3/4]

uint16_t noodle_conv1d	(	float *	in,
		uint16_t	n_inputs,
		float *	out,
		uint16_t	n_outputs,
		uint16_t	W,
		const ConvMem &	conv,
		CBFPtr	progress_cb )

Raw memory-to-memory 1D convolution without pooling.

noodle_conv1d() [4/4]

uint16_t noodle_conv1d	(	float *	in,
		uint16_t	n_inputs,
		float *	out,
		uint16_t	n_outputs,
		uint16_t	W,
		const ConvMem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory 1D convolution with pooling.

noodle_conv_float() [1/6]

uint16_t noodle_conv_float	(	const char *	in_fn,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		uint16_t	W,
		const Conv &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

File-to-memory 2D convolution with file-backed parameters.

noodle_conv_float() [2/6]

uint16_t noodle_conv_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		const char *	out_fn,
		uint16_t	W,
		const Conv &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Memory-to-file 2D convolution with file-backed parameters.

noodle_conv_float() [3/6]

uint16_t noodle_conv_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		const char *	out_fn,
		uint16_t	W,
		const ConvMem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Memory-to-file 2D convolution with memory-backed parameters.

noodle_conv_float() [4/6]

uint16_t noodle_conv_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		uint16_t	W,
		const Conv &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory 2D convolution with file-backed parameters.

noodle_conv_float() [5/6]

uint16_t noodle_conv_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		uint16_t	W,
		const ConvMem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory 2D convolution with memory-backed parameters.

noodle_conv_float() [6/6]

uint16_t noodle_conv_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		uint16_t	W,
		const ConvProgmem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory 2D convolution with near-PROGMEM parameters.

noodle_conv_transpose_float()

uint16_t noodle_conv_transpose_float	(	float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		uint16_t	W,
		const ConvMem &	conv,
		CBFPtr	progress_cb )

Raw memory-to-memory 2D transpose convolution.

Input and output use packed channel-first layout. conv.weight is packed as [O][I][K][K]. For explicit padding, callers set conv.OP so V = (W - 1) * conv.S - 2 * conv.P + conv.K + conv.OP matches the desired output width.

Parameters

input	Packed [I][W][W] input tensor.
n_inputs	Number of input channels.
n_outputs	Number of output channels.
output	Packed [O][V][V] output tensor.
W	Input width and height.
conv	Memory-backed transpose convolution parameters.
progress_cb	Optional progress callback.

Returns

Output width, or 0 on null input/output/weights or invalid shape.

noodle_copy_kernel_progmem()

void noodle_copy_kernel_progmem	(	const float *	w,
		uint32_t	base,
		uint16_t	K,
		float *	kernel )

Copy one square kernel from near-PROGMEM into RAM.

Parameters

w	Base pointer to packed PROGMEM weights.
base	Element offset of the first kernel value.
K	Kernel width.
kernel	Destination buffer with room for K * K floats.

noodle_do_bias()

uint16_t noodle_do_bias	(	float *	output,
		float	bias,
		uint16_t	n )

Add bias to a square output map and apply ReLU.

Parameters

output	Output map with n * n values.
bias	Bias scalar.
n	Map width and height.

Returns

n.

noodle_do_bias_act()

uint16_t noodle_do_bias_act	(	float *	output,
		float	bias,
		uint16_t	n,
		Activation	act )

Add bias to a square output map and apply the requested activation.

ACT_RELU clamps negative values to zero. ACT_NONE leaves biased values unchanged. Other activation values are ignored here.

Parameters

output	Output map with n * n values.
bias	Bias scalar.
n	Map width and height.
act	Activation to apply.

Returns

n.

noodle_do_conv() [1/2]

uint16_t noodle_do_conv	(	byte *	grid,
		const float *	kernel,
		uint16_t	K,
		uint16_t	W,
		float *	output,
		uint16_t	P,
		uint16_t	S )

Accumulate one byte-input 2D convolution plane.

The input plane is [W][W]; the kernel is [K][K]; output is accumulated in [V][V] order instead of cleared.

Parameters

grid	Input plane.
kernel	Kernel values.
K	Kernel width.
W	Input width and height.
output	Output accumulator.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.

Returns

Output width before pooling.

noodle_do_conv() [2/2]

uint16_t noodle_do_conv	(	float *	grid,
		const float *	kernel,
		uint16_t	K,
		uint16_t	W,
		float *	output,
		uint16_t	P,
		uint16_t	S )

Accumulate one float-input 2D convolution plane.

The input plane is [W][W]; the kernel is [K][K]; output is accumulated in [V][V] order instead of cleared.

Parameters

grid	Input plane.
kernel	Kernel values.
K	Kernel width.
W	Input width and height.
output	Output accumulator.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.

Returns

Output width before pooling.

noodle_do_conv1d()

uint16_t noodle_do_conv1d	(	float *	input,
		float *	kernel,
		uint16_t	W,
		uint16_t	K,
		float *	output,
		uint16_t	P,
		uint16_t	S )

Accumulate one 1D convolution into an output sequence.

Values outside the input sequence are treated as zero.

Parameters

input	Input sequence with W values.
kernel	Kernel with K values.
W	Input sequence length.
K	Kernel length.
output	Accumulator receiving V values.
P	Zero padding per side.
S	Stride.

Returns

Output length before pooling.

noodle_do_conv_transpose()

uint16_t noodle_do_conv_transpose	(	float *	input,
		const float *	kernel,
		uint16_t	K,
		uint16_t	W,
		float *	output,
		uint16_t	P,
		uint16_t	S,
		uint16_t	OP )

Accumulate one 2D transpose-convolution plane.

The input plane is [W][W]; the kernel is [K][K]; output is accumulated in [Vt][Vt] order instead of cleared.

For explicit padding, callers choose OP so Vt = (W - 1) * S - 2 * P + K + OP matches the desired output width. With P == 65535, SAME-style transpose output uses Vt = W * S.

Parameters

input	Input plane.
kernel	Kernel values.
K	Kernel width.
W	Input width and height.
output	Output accumulator.
P	Padding per side, or 65535 for SAME-style padding.
S	Stride.
OP	User-computed output padding for explicit padding.

Returns

Output width.

noodle_do_pooling() [1/3]

uint16_t noodle_do_pooling	(	const float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		const char *	fn )

Apply 2D pooling and write to a file.

When NOODLE_POOL_MODE is NOODLE_POOL_NONE, this writes the input map unchanged.

Parameters

input	Input map with W * W values.
W	Input width and height.
K	Pool window size.
S	Pool stride.
fn	Output file.

Returns

Output width, or 0 for invalid pooling parameters.

noodle_do_pooling() [2/3]

uint16_t noodle_do_pooling	(	const float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		float *	output )

Apply 2D pooling and write to memory.

Parameters

input	Input map with W * W values.
W	Input width and height.
K	Pool window size.
S	Pool stride.
output	Destination map.

Returns

Output width, or W for identity/no pooling.

noodle_do_pooling() [3/3]

uint16_t noodle_do_pooling	(	const float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		NDL_File &	fo )

Apply 2D pooling and write to an open file.

Parameters

input	Input map with W * W values.
W	Input width and height.
K	Pool window size.
S	Pool stride.
fo	Open output file handle.

Returns

Output width, or 0 for invalid pooling parameters.

noodle_do_pooling1d() [1/3]

uint16_t noodle_do_pooling1d	(	const float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		float *	output )

Apply valid 1D pooling and write to memory.

K <= 1 copies the input unchanged. S == 0 defaults the stride to K. The helper computes mean pooling only when NOODLE_POOL_MODE selects mean pooling; otherwise it computes max pooling.

Parameters

input	Input sequence with W values.
W	Input sequence length.
K	Pool window size.
S	Pool stride, or 0 to use K.
output	Destination sequence.

Returns

Output sequence length.

noodle_do_pooling1d() [2/3]

uint16_t noodle_do_pooling1d	(	float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		const char *	fn )

Apply valid 1D max pooling and write to a file.

Parameters

input	Input sequence with W values.
W	Input sequence length.
K	Pool window size.
S	Pool stride.
fn	Output file.

Returns

Output sequence length.

noodle_do_pooling1d() [3/3]

uint16_t noodle_do_pooling1d	(	float *	input,
		uint16_t	W,
		uint16_t	K,
		uint16_t	S,
		NDL_File &	fo )

Apply valid 1D max pooling and write to an open file.

Parameters

input	Input sequence with W values.
W	Input sequence length.
K	Pool window size.
S	Pool stride.
fo	Open output file handle.

Returns

Output sequence length.

noodle_dot_float_block()

float noodle_dot_float_block	(	const float *	x,
		const float *	w,
		uint16_t	n )

Compute a dot product with a small unrolled loop.

Parameters

x	Input vector.
w	Weight vector.
n	Number of elements.

Returns

Sum of element-wise products.

noodle_dwconv_float() [1/3]

uint16_t noodle_dwconv_float	(	float *	input,
		uint16_t	n_channels,
		float *	output,
		uint16_t	W,
		const Conv &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory depthwise convolution with file-backed parameters.

noodle_dwconv_float() [2/3]

uint16_t noodle_dwconv_float	(	float *	input,
		uint16_t	n_channels,
		float *	output,
		uint16_t	W,
		const ConvMem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory depthwise convolution with memory-backed parameters.

noodle_dwconv_float() [3/3]

uint16_t noodle_dwconv_float	(	float *	input,
		uint16_t	n_channels,
		float *	output,
		uint16_t	W,
		const ConvProgmem &	conv,
		const Pool &	pool,
		CBFPtr	progress_cb )

Raw memory-to-memory depthwise convolution with near-PROGMEM parameters.

noodle_fcn() [1/7]

uint16_t noodle_fcn	(	const byte *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNFile &	fcn,
		CBFPtr	progress_cb )

Byte-input fully connected layer with file-backed parameters.

noodle_fcn() [2/7]

uint16_t noodle_fcn	(	const char *	in_fn,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNFile &	fcn,
		CBFPtr	progress_cb )

File-input fully connected layer that writes output to memory.

noodle_fcn() [3/7]

uint16_t noodle_fcn	(	const float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		const char *	out_fn,
		const FCNFile &	fcn,
		CBFPtr	progress_cb )

Float-input fully connected layer that writes output to a file.

noodle_fcn() [4/7]

uint16_t noodle_fcn	(	const float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNFile &	fcn,
		CBFPtr	progress_cb )

Float-input fully connected layer with file-backed parameters.

noodle_fcn() [5/7]

uint16_t noodle_fcn	(	const float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNMem &	fcn,
		CBFPtr	progress_cb )

Float-input fully connected layer with memory-backed parameters.

noodle_fcn() [6/7]

uint16_t noodle_fcn	(	const float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNProgmem &	fcn,
		CBFPtr	progress_cb )

Float-input fully connected layer with far-PROGMEM parameters.

noodle_fcn() [7/7]

uint16_t noodle_fcn	(	const int8_t *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const FCNFile &	fcn,
		CBFPtr	progress_cb )

Int8-input fully connected layer with file-backed parameters.

noodle_fcn_progmem()

uint16_t noodle_fcn_progmem	(	const float *	input,
		uint16_t	n_inputs,
		uint16_t	n_outputs,
		float *	output,
		const float *	weight,
		const float *	bias,
		Activation	act,
		CBFPtr	progress_cb )

Float-input fully connected layer with near-PROGMEM parameters.

noodle_find_max()

void noodle_find_max	(	float *	input,
		uint16_t	n,
		float &	max_val,
		uint16_t &	max_idx )

Find the maximum value and its index in a vector.

Parameters

input	Input vector.
n	Number of values.
max_val	Receives the maximum value.
max_idx	Receives the index of the maximum value.

noodle_flat() [1/2]

uint16_t noodle_flat	(	const char *	in_fn,
		float *	output,
		uint16_t	V,
		uint16_t	n_filters )

Flatten a packed file tensor into an HWC-like raw vector.

Reads packed [C][V][V] input and writes output[pixel * n_filters + channel].

Parameters

in_fn	Input file containing packed channel-first planes.
output	Destination buffer with room for V * V * n_filters floats.
V	Input plane width and height.
n_filters	Number of channel planes.

Returns

Number of floats written.

noodle_flat() [2/2]

uint16_t noodle_flat	(	float *	input,
		float *	output,
		uint16_t	V,
		uint16_t	n_filters )

Flatten a packed memory tensor into an HWC-like raw vector.

Reads input as packed [C][V][V] and writes output[pixel * n_filters + channel].

Parameters

input	Source buffer with packed channel-first planes.
output	Destination buffer with room for V * V * n_filters floats.
V	Input plane width and height.
n_filters	Number of channel planes.

Returns

Number of floats written.

noodle_gap()

uint16_t noodle_gap	(	float *	inout,
		uint16_t	C,
		uint16_t	W )

Apply global average pooling in place to packed channel-first maps.

Reduces [C][W][W] to [C] by writing each channel mean into the first C positions of inout.

Parameters

inout	Buffer containing packed [C][W][W] data.
C	Number of channels.
W	Plane width and height.

Returns

C.

noodle_get_padded_x() [1/2]

float noodle_get_padded_x	(	byte *	grid,
		int16_t	i,
		int16_t	j,
		int16_t	W,
		int16_t	P0,
		int16_t	P1 )

Read a byte grid sample with asymmetric zero padding.

Parameters

grid	Input plane with W * W byte values.
i	Padded-row coordinate.
j	Padded-column coordinate.
W	Input width and height.
P0	Top/left padding.
P1	Bottom/right padding.

Returns

Input value, or 0.0 outside the unpadded grid.

noodle_get_padded_x() [2/2]

float noodle_get_padded_x	(	float *	grid,
		int16_t	i,
		int16_t	j,
		int16_t	W,
		int16_t	P0,
		int16_t	P1 )

Read a float grid sample with asymmetric zero padding.

Parameters

grid	Input plane with W * W float values.
i	Padded-row coordinate.
j	Padded-column coordinate.
W	Input width and height.
P0	Top/left padding.
P1	Bottom/right padding.

Returns

Input value, or 0.0 outside the unpadded grid.

noodle_gmp()

uint16_t noodle_gmp	(	float *	inout,
		uint16_t	C,
		uint16_t	W )

Apply global max pooling in place to packed channel-first data.

Reduces each channel to one maximum value in the first C positions of inout. The current helper scans W values per channel.

Parameters

inout	Buffer containing packed channel data.
C	Number of channels.
W	Number of values scanned per channel.

Returns

C.

noodle_grid_from_file() [1/3]

void noodle_grid_from_file	(	NDL_File &	fi,
		byte *	buffer,
		uint16_t	K )

Read a byte grid from an already-open file.

noodle_grid_from_file() [2/3]

void noodle_grid_from_file	(	NDL_File &	fi,
		float *	buffer,
		uint16_t	K )

Read a float grid from an already-open file.

noodle_grid_from_file() [3/3]

void noodle_grid_from_file	(	NDL_File &	fi,
		int8_t *	buffer,
		uint16_t	K )

Read an int8 grid from an already-open file.

noodle_grid_to_file() [1/2]

void noodle_grid_to_file	(	byte *	grid,
		NDL_File &	fo,
		uint16_t	n )

Write a byte grid to an already-open file.

noodle_grid_to_file() [2/2]

void noodle_grid_to_file	(	float *	grid,
		NDL_File &	fo,
		uint16_t	n )

Write a float grid to an already-open file.

noodle_logit()

uint16_t noodle_logit	(	float *	input_output,
		uint16_t	n )

Apply logistic sigmoid in place.

Parameters

input_output	Vector updated in place.
n	Number of vector elements.

Returns

n.

noodle_read_float_block()

size_t noodle_read_float_block	(	NDL_File &	f,
		float *	dst,
		size_t	n_floats )

Read a block of floats using the configured scalar file format.

In binary mode this reads raw float32 bytes. In text mode it calls noodle_read_float() once per value.

Parameters

f	Open input file.
dst	Destination float buffer.
n_floats	Number of floats requested.

Returns

Number of floats read.

noodle_read_raw()

size_t noodle_read_raw	(	NDL_File &	f,
		void *	dst,
		size_t	n )

Read raw bytes from a backend file handle.

Parameters

f	Open file handle.
dst	Destination buffer.
n	Number of bytes to read.

Returns

Number of bytes read, or 0 for the no-filesystem backend.

noodle_relu()

uint16_t noodle_relu	(	float *	input_output,
		uint16_t	n )

Apply ReLU in place.

Parameters

input_output	Vector updated in place.
n	Number of vector elements.

Returns

n.

noodle_reset_buffer()

void noodle_reset_buffer	(	float *	buffer,
		uint16_t	n )

Clear a float buffer.

Parameters

buffer	Buffer to fill with zero.
n	Number of float elements.

noodle_reshape()

uint16_t noodle_reshape	(	const float *	src_hwc,
		float *	dst_chw,
		uint16_t	W,
		uint16_t	C )

Convert HWC-like raw data to packed channel-first raw data.

Reads src_hwc[pixel * C + channel] and writes packed [C][W][W] data into dst_chw.

Parameters

src_hwc	Source buffer in HWC-like order.
dst_chw	Destination buffer with room for W * W * C floats.
W	Output plane width and height.
C	Number of channel planes.

Returns

Number of floats written.

noodle_sigmoid()

uint16_t noodle_sigmoid	(	float *	input_output,
		uint16_t	n )

Apply sigmoid in place.

Parameters

input_output	Vector updated in place.
n	Number of vector elements.

Returns

n.

noodle_sigmoidf()

float noodle_sigmoidf

(

float

x

)

Compute sigmoid for one scalar.

Parameters

x	Scalar input.

Returns

Logistic sigmoid of x.

noodle_slice()

float * noodle_slice	(	float *	flat,
		size_t	W,
		size_t	z )

Return a channel plane from a packed [Z][W][W] tensor.

Parameters

flat	Base pointer to the packed tensor.
W	Width and height of each plane.
z	Plane/channel index.

Returns

Pointer to the first element of plane z.

noodle_soft_max()

uint16_t noodle_soft_max	(	float *	input_output,
		uint16_t	n )

Apply numerically stabilized softmax in place.

Parameters

input_output	Vector updated in place.
n	Number of vector elements.

Returns

n.

noodle_temp1_require()

float * noodle_temp1_require

(

size_t

required_floats

)

Ensure temp buffer 1 can hold a number of floats.

Automatically allocated buffers grow when needed. A buffer installed with noodle_setup_temp_buffers() has unknown capacity and is returned as-is.

Parameters

required_floats

Required capacity in float elements.

Returns

Usable float pointer, or NULL on allocation failure/zero request.

noodle_temp2_require()

float * noodle_temp2_require

(

size_t

required_floats

)

Ensure temp buffer 2 can hold a number of floats.

Automatically allocated buffers grow when needed. A buffer installed with noodle_setup_temp_buffers() has unknown capacity and is returned as-is.

Parameters

required_floats

Required capacity in float elements.

Returns

Usable float pointer, or NULL on allocation failure/zero request.

noodle_temp_buffers_free()

void noodle_temp_buffers_free

(

void

)

Free Noodle-owned scratch buffers and detach external scratch buffers.

Free Noodle-owned scratch buffers and detach external scratch buffers.

External buffers installed with noodle_setup_temp_buffers() are detached but not freed.

noodle_unpack_bn_params()

void noodle_unpack_bn_params	(	const float *	bn_params,
		uint16_t	N,
		const float **	gamma,
		const float **	beta,
		const float **	mean,
		const float **	var )

Split packed batch-normalization parameters into four arrays.

Parameters

bn_params	Packed [gamma[N]][beta[N]][mean[N]][var[N]] values.
N	Number of elements per parameter array.
gamma	Receives pointer to gamma values.
beta	Receives pointer to beta values.
mean	Receives pointer to mean values.
var	Receives pointer to variance values.

noodle_valid_max_pool()

uint16_t noodle_valid_max_pool	(	float *	inplace,
		uint16_t	W,
		uint16_t	C,
		const Pool &	pool )

Apply valid max pooling to a packed channel-first tensor in place.

Parameters

inplace	Tensor in packed [C][W][W] order; compacted in place.
W	Input width and height.
C	Number of channels.
pool	Pooling parameters.

Returns

Output width, or 0 for invalid parameters.

noodle_write_raw()

size_t noodle_write_raw	(	NDL_File &	f,
		const void *	src,
		size_t	n )

Write raw bytes to a backend file handle.

Parameters

f	Open file handle.
src	Source buffer.
n	Number of bytes to write.

Returns

Number of bytes written, or 0 for the no-filesystem backend.