Macros
#define	DIM_KER_X

#define	DIM_KER_Y

Functions
arm_status	arm_convolve_1x1_HWC_q7_fast_nonsquare (const q7_t Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t bufferA, q7_t bufferB)
	Fast Q7 version of 1x1 convolution (non-sqaure shape) More...

arm_status	arm_convolve_1x1_s8_fast (const q7_t input, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const uint16_t input_batches, const q7_t kernel, const uint16_t output_ch, const uint16_t pad_x, const uint16_t pad_y, const uint16_t stride_x, const uint16_t stride_y, const int32_t bias, q7_t output, const int32_t output_shift, const int32_t output_mult, const int32_t out_offset, const int32_t input_offset, const int32_t out_activation_min, const int32_t out_activation_max, const uint16_t output_x, const uint16_t output_y, q15_t *buffer_a)
	Fast s8 version for 1x1 convolution (non-square shape) More...

int32_t	arm_convolve_1x1_s8_fast_get_buffer_size (const uint16_t input_ch)
	Get the required buffer size for the fast 1x1 convolution (non-square shape) s8 convolution function. More...

arm_status	arm_convolve_HWC_q15_basic (const q15_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q15_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q15_t bias, const uint16_t bias_shift, const uint16_t out_shift, q15_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Basic Q15 convolution function. More...

arm_status	arm_convolve_HWC_q15_fast (const q15_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q15_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q15_t bias, const uint16_t bias_shift, const uint16_t out_shift, q15_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Fast Q15 convolution function. More...

arm_status	arm_convolve_HWC_q15_fast_nonsquare (const q15_t Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q15_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q15_t bias, const uint16_t bias_shift, const uint16_t out_shift, q15_t Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t bufferA, q7_t bufferB)
	Fast Q15 convolution function (non-sqaure shape) More...

arm_status	arm_convolve_HWC_q7_basic (const q7_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Basic Q7 convolution function. More...

arm_status	arm_convolve_HWC_q7_basic_nonsquare (const q7_t Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t bufferA, q7_t bufferB)
	Basic Q7 convolution function (non-sqaure shape) More...

arm_status	arm_convolve_HWC_q7_fast (const q7_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Fast Q7 convolution function. More...

arm_status	arm_convolve_HWC_q7_fast_nonsquare (const q7_t Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t bufferA, q7_t bufferB)
	Fast Q7 convolution function (non-sqaure shape) More...

arm_status	arm_convolve_HWC_q7_RGB (const q7_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Q7 convolution function for RGB image. More...

arm_status	arm_convolve_s8 (const q7_t input, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const uint16_t input_batches, const q7_t kernel, const uint16_t output_ch, const uint16_t kernel_x, const uint16_t kernel_y, const uint16_t pad_x, const uint16_t pad_y, const uint16_t stride_x, const uint16_t stride_y, const int32_t bias, q7_t output, const int32_t output_shift, const int32_t output_mult, const int32_t out_offset, const int32_t input_offset, const int32_t out_activation_min, const int32_t out_activation_max, const uint16_t output_x, const uint16_t output_y, q15_t *buffer_a)
	Basic s8 convolution function. More...

int32_t	arm_convolve_s8_get_buffer_size (const uint16_t input_ch, const uint16_t kernel_x, const uint16_t kernel_y)
	Get the required buffer size for s8 convolution function. More...

arm_status	arm_depthwise_conv_s8 (const q7_t input, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const q7_t kernel, const uint16_t output_ch, const uint16_t ch_mult, const uint16_t kernel_x, const uint16_t kernel_y, const uint16_t pad_x, const uint16_t pad_y, const uint16_t stride_x, const uint16_t stride_y, const int32_t bias, q7_t output, const int32_t output_shift, const int32_t output_mult, const uint16_t output_x, const uint16_t output_y, const int32_t output_offset, const int32_t input_offset, const int32_t output_activation_min, const int32_t output_activation_max, const uint16_t dilation_x, const uint16_t dilation_y, q15_t *buffer_a)
	Basic s8 depthwise convolution function. More...

arm_status	arm_depthwise_conv_s8_opt (const q7_t input, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const q7_t kernel, const uint16_t output_ch, const uint16_t kernel_x, const uint16_t kernel_y, const uint16_t pad_x, const uint16_t pad_y, const uint16_t stride_x, const uint16_t stride_y, const int32_t bias, q7_t output, const int32_t output_shift, const int32_t output_mult, const uint16_t output_x, const uint16_t output_y, const int32_t output_offset, const int32_t input_offset, const int32_t output_activation_min, const int32_t output_activation_max, const uint16_t dilation_x, const uint16_t dilation_y, q15_t *buffer_a)
	Optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. More...

int32_t	arm_depthwise_conv_s8_opt_get_buffer_size (const uint16_t input_ch, const uint16_t kernel_x, const uint16_t kernel_y)
	Get the required buffer size for optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. More...

arm_status	arm_depthwise_conv_u8_basic_ver1 (const uint8_t input, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const uint8_t kernel, const uint16_t kernel_x, const uint16_t kernel_y, const int16_t ch_mult, const int16_t pad_x, const int16_t pad_y, const int16_t stride_x, const int16_t stride_y, const int16_t dilation_x, const int16_t dilation_y, const int32_t bias, const int32_t input_offset, const int32_t filter_offset, const int32_t output_offset, uint8_t output, const uint16_t output_x, const uint16_t output_y, const int32_t output_activation_min, const int32_t output_activation_max, const int32_t out_shift, const int32_t out_mult)
	uint8 depthwise convolution function with asymmetric quantization for even number of channel multiplier and input channels. Unless specified otherwise, arguments are mandatory. Both square and non-square inputs are accepted. More...

arm_status	arm_depthwise_separable_conv_HWC_q7 (const q7_t Im_in, const uint16_t dim_im_in, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel, const uint16_t padding, const uint16_t stride, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out, q15_t bufferA, q7_t bufferB)
	Q7 depthwise separable convolution function. More...

arm_status	arm_depthwise_separable_conv_HWC_q7_nonsquare (const q7_t Im_in, const uint16_t dim_im_in_x, const uint16_t dim_im_in_y, const uint16_t ch_im_in, const q7_t wt, const uint16_t ch_im_out, const uint16_t dim_kernel_x, const uint16_t dim_kernel_y, const uint16_t padding_x, const uint16_t padding_y, const uint16_t stride_x, const uint16_t stride_y, const q7_t bias, const uint16_t bias_shift, const uint16_t out_shift, q7_t Im_out, const uint16_t dim_im_out_x, const uint16_t dim_im_out_y, q15_t bufferA, q7_t bufferB)
	Q7 depthwise separable convolution function (non-square shape) More...

Description

Perform convolution layer

The convolution is implemented in 2 steps: im2col and GEMM

im2col is a process of converting each patch of image data into a column. After im2col, the convolution is computed as matrix-matrix multiplication.

To reduce the memory footprint, the im2col is performed partially. Each iteration, only a few column (i.e., patches) are generated and computed with GEMM kernels similar to CMSIS-DSP arm_mat_mult functions.

Macro Definition Documentation

#define DIM_KER_X

Referenced by arm_convolve_1x1_s8_fast().

#define DIM_KER_Y

Referenced by arm_convolve_1x1_s8_fast().

Function Documentation

arm_status arm_convolve_1x1_HWC_q7_fast_nonsquare	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in_x,
		const uint16_t	dim_im_in_y,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel_x,
		const uint16_t	dim_kernel_y,
		const uint16_t	padding_x,
		const uint16_t	padding_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out_x,
		const uint16_t	dim_im_out_y,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimention x
[in]	dim_im_in_y	input tensor dimention y
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel_x	filter kernel size x
[in]	dim_kernel_y	filter kernel size y
[in]	padding_x	padding size x
[in]	padding_y	padding size y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out_x	output tensor dimension x
[in]	dim_im_out_y	output tensor dimension y
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

This function is optimized for convolution with 1x1 kernel size (i.e., dim_kernel_x=1 and dim_kernel_y=1). It can be used for the second half of MobileNets [1] after depthwise separable convolution.

This function is the version with full list of optimization tricks, but with some contraints: ch_im_in is multiple of 4 ch_im_out is multiple of 2

[1] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications https://arxiv.org/abs/1704.04861

References arm_nn_mat_mult_kernel_q7_q15_reordered(), arm_q7_to_q15_reordered_no_shift(), and NN_ROUND.

arm_status arm_convolve_1x1_s8_fast	(	const q7_t *	input,
		const uint16_t	input_x,
		const uint16_t	input_y,
		const uint16_t	input_ch,
		const uint16_t	input_batches,
		const q7_t *	kernel,
		const uint16_t	output_ch,
		const uint16_t	pad_x,
		const uint16_t	pad_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const int32_t *	bias,
		q7_t *	output,
		const int32_t *	output_shift,
		const int32_t *	output_mult,
		const int32_t	out_offset,
		const int32_t	input_offset,
		const int32_t	out_activation_min,
		const int32_t	out_activation_max,
		const uint16_t	output_x,
		const uint16_t	output_y,
		q15_t *	buffer_a
	)

Parameters

[in]	input	pointer to input tensor. Format: [N, H, W, in_ch]
[in]	input_x	input tensor dimension x
[in]	input_y	input tensor dimension y
[in]	input_ch	number of input tensor channels
[in]	input_batches	number of input batches
[in]	kernel	pointer to kernel weights. Format: [out_ch, H, W, in_ch]
[in]	output_ch	number of filters, i.e., output tensor channels
[in]	pad_x	padding size x
[in]	pad_y	padding size y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to per channel bias. Range : int32
[in,out]	output	pointer to output tensor. Format: [H, W, out_ch]
[in]	output_shift	pointer to per output channel requantization shift parameter.
[in]	output_mult	pointer to per output channel requantization multiplier parameter.
[in]	out_offset	output tensor offset. Range: int8
[in]	input_offset	input tensor offset. Range: int8
[in]	out_activation_min	Minimum value to clamp the output to. Range: int8
[in]	out_activation_max	Minimum value to clamp the output to. Range: int8
[in]	output_x	output tensor width
[in]	output_y	output tensor height
[in]	buffer_a	pointer to buffer space used for input optimization(partial im2col) and is necessary when ARM_MATH_LOOPUNROLL and ARM_MATH_DSP is defined. Required space: 2 * input_ch * sizeof(q15_t) bytes Use arm_convolve_1x1_s8_fast_get_buffer_size() to get the size

Returns: The function returns either ARM_MATH_SIZE_MISMATCH if argument constraints fail. or, ARM_MATH_SUCCESS on successful completion.

Supported framework : TensorFlow Lite Micro
The following constrains on the arguments apply
1. input_ch is a multiple of 4
2. output_ch is a multiple of 2
3. padding equals 0
4. Stride equals 1
5. kernel dimension is 1x1 (Not provided in the argument list)

References arm_convolve_s8(), arm_nn_mat_mult_kernel_s8_s16_reordered(), arm_nn_read_q15x2_ia(), arm_nn_requantize(), arm_q7_to_q15_reordered_with_offset(), DIM_KER_X, DIM_KER_Y, MAX, and MIN.

int32_t arm_convolve_1x1_s8_fast_get_buffer_size ( const uint16_t input_ch )

Parameters

[in] input_ch number of input tensor channels

Returns: The function returns required buffer size

arm_status arm_convolve_HWC_q15_basic	(	const q15_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q15_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q15_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q15_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns ARM_MATH_SUCCESS

Buffer size:

bufferA size: ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

This basic version is designed to work for any input tensor and weight dimension.

References NN_ROUND.

arm_status arm_convolve_HWC_q15_fast	(	const q15_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q15_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q15_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q15_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

Input dimension constraints:

ch_im_in is multiple of 2

ch_im_out is multipe of 2

References NN_ROUND.

arm_status arm_convolve_HWC_q15_fast_nonsquare	(	const q15_t *	Im_in,
		const uint16_t	dim_im_in_x,
		const uint16_t	dim_im_in_y,
		const uint16_t	ch_im_in,
		const q15_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel_x,
		const uint16_t	dim_kernel_y,
		const uint16_t	padding_x,
		const uint16_t	padding_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const q15_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q15_t *	Im_out,
		const uint16_t	dim_im_out_x,
		const uint16_t	dim_im_out_y,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimention x
[in]	dim_im_in_y	input tensor dimention y
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel_x	filter kernel size x
[in]	dim_kernel_y	filter kernel size y
[in]	padding_x	padding size x
[in]	padding_y	padding size y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out_x	output tensor dimension x
[in]	dim_im_out_y	output tensor dimension y
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

Input dimension constraints:

ch_im_in is multiple of 2

ch_im_out is multipe of 2

References NN_ROUND.

arm_status arm_convolve_HWC_q7_basic	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns ARM_MATH_SUCCESS

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

This basic version is designed to work for any input tensor and weight dimension.

References arm_nn_mat_mult_kernel_q7_q15(), arm_q7_to_q15_no_shift(), and NN_ROUND.

arm_status arm_convolve_HWC_q7_basic_nonsquare	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in_x,
		const uint16_t	dim_im_in_y,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel_x,
		const uint16_t	dim_kernel_y,
		const uint16_t	padding_x,
		const uint16_t	padding_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out_x,
		const uint16_t	dim_im_out_y,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Basic Q7 convolution function (non-square shape)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimention x
[in]	dim_im_in_y	input tensor dimention y
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel_x	filter kernel size x
[in]	dim_kernel_y	filter kernel size y
[in]	padding_x	padding size x
[in]	padding_y	padding size y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out_x	output tensor dimension x
[in]	dim_im_out_y	output tensor dimension y
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns ARM_MATH_SUCCESS

References arm_nn_mat_mult_kernel_q7_q15(), arm_q7_to_q15_no_shift(), and NN_ROUND.

arm_status arm_convolve_HWC_q7_fast	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

Input dimension constraints:

ch_im_in is multiple of 4 ( because of the SIMD32 read and swap )

ch_im_out is multipe of 2 ( bacause 2x2 mat_mult kernel )

The im2col converts the Q7 tensor input into Q15 column, which is stored in bufferA. There is reordering happenning during this im2col process with arm_q7_to_q15_reordered_no_shift. For every four elements, the second and third elements are swapped.

The computation kernel arm_nn_mat_mult_kernel_q7_q15_reordered does the GEMM computation with the reordered columns.

To speed-up the determination of the padding condition, we split the computation into 3x3 parts, i.e., {top, mid, bottom} X {left, mid, right}. This reduces the total number of boundary condition checks and improves the data copying performance.

References arm_nn_mat_mult_kernel_q7_q15_reordered(), arm_q7_to_q15_reordered_no_shift(), and NN_ROUND.

Referenced by main().

arm_status arm_convolve_HWC_q7_fast_nonsquare	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in_x,
		const uint16_t	dim_im_in_y,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel_x,
		const uint16_t	dim_kernel_y,
		const uint16_t	padding_x,
		const uint16_t	padding_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out_x,
		const uint16_t	dim_im_out_y,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimention x
[in]	dim_im_in_y	input tensor dimention y
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel_x	filter kernel size x
[in]	dim_kernel_y	filter kernel size y
[in]	padding_x	padding size x
[in]	padding_y	padding size y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out_x	output tensor dimension x
[in]	dim_im_out_y	output tensor dimension y
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

This function is the version with full list of optimization tricks, but with some contraints: ch_im_in is multiple of 4 ch_im_out is multiple of 2

References arm_nn_mat_mult_kernel_q7_q15_reordered(), arm_q7_to_q15_reordered_no_shift(), and NN_ROUND.

arm_status arm_convolve_HWC_q7_RGB	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Q7 version of convolution for RGB image.

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

Input dimension constraints:

ch_im_in equals 3

This kernel is written exclusively for convolution with ch_im_in equals 3. This applies on the first layer of CNNs which has input image with RGB format.

References arm_nn_mat_mult_kernel_q7_q15(), arm_nnword::half_words, NN_ROUND, and arm_nnword::word.

Referenced by main().

arm_status arm_convolve_s8	(	const q7_t *	input,
		const uint16_t	input_x,
		const uint16_t	input_y,
		const uint16_t	input_ch,
		const uint16_t	input_batches,
		const q7_t *	kernel,
		const uint16_t	output_ch,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y,
		const uint16_t	pad_x,
		const uint16_t	pad_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const int32_t *	bias,
		q7_t *	output,
		const int32_t *	output_shift,
		const int32_t *	output_mult,
		const int32_t	out_offset,
		const int32_t	input_offset,
		const int32_t	output_activation_min,
		const int32_t	output_activation_max,
		const uint16_t	output_x,
		const uint16_t	output_y,
		q15_t *	buffer_a
	)

Parameters

[in]	input	pointer to input tensor. Range: int8, format: [N,H,W,in_ch]
[in]	input_x	input tensor width
[in]	input_y	input tensor height
[in]	input_ch	number of input tensor channels
[in]	input_batches	number of input batches
[in]	kernel	pointer to kernel weights. Range: int8, format: [out_ch, H, W, in_ch]
[in]	output_ch	number of filters, i.e., output tensor channels
[in]	kernel_x	filter/kernel width
[in]	kernel_y	filter/kernel height
[in]	pad_x	padding along width
[in]	pad_y	padding along height
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to per output channel bias. Range: int32
[in,out]	output	pointer to output tensor. format: [H, W, out_ch]
[in]	output_shift	pointer to per output channel requantization shift parameter.
[in]	output_mult	pointer to per output channel requantization multiplier parameter.
[in]	out_offset	output tensor offset. Range: int8
[in]	input_offset	input tensor offset. Range: int8
[in]	output_activation_min	Minimum value to clamp the output to. Range: int8
[in]	output_activation_max	Minimum value to clamp the output to. Range: int8
[in]	output_x	output tensor width
[in]	output_y	output tensor height
[in]	buffer_a	pointer to buffer space used for input optimization(partial im2col) and is necessary when both ARM_MATH_LOOPUNROLL and ARM_MATH_DSP are defined. Required space: (2 * input_ch * kernel_x * kernel_y) * sizeof(q15_t) bytes Use arm_convolve_s8_get_buffer_size() to get the size.

Returns: The function returns ARM_MATH_SUCCESS

Supported framework: TensorFlow Lite micro
q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs.
Additional memory is required for optimization. Refer to argument 'buffer_a' for details.

References arm_nn_mat_mult_kernel_s8_s16(), arm_nn_read_q15x2_ia(), arm_nn_requantize(), arm_q7_to_q15_with_offset(), MAX, and MIN.

Referenced by arm_convolve_1x1_s8_fast().

int32_t arm_convolve_s8_get_buffer_size	(	const uint16_t	input_ch,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y
	)

Parameters

[in]	input_ch	number of input tensor channels
[in]	kernel_x	filter/kernel width
[in]	kernel_y	filter/kernel height

Returns: The function returns required buffer size

arm_status arm_depthwise_conv_s8	(	const q7_t *	input,
		const uint16_t	input_x,
		const uint16_t	input_y,
		const uint16_t	input_ch,
		const q7_t *	kernel,
		const uint16_t	output_ch,
		const uint16_t	ch_mult,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y,
		const uint16_t	pad_x,
		const uint16_t	pad_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const int32_t *	bias,
		q7_t *	output,
		const int32_t *	output_shift,
		const int32_t *	output_mult,
		const uint16_t	output_x,
		const uint16_t	output_y,
		const int32_t	output_offset,
		const int32_t	input_offset,
		const int32_t	output_activation_min,
		const int32_t	output_activation_max,
		const uint16_t	dilation_x,
		const uint16_t	dilation_y,
		q15_t *	buffer_a
	)

Parameters

[in]	input	pointer to input tensor. Range: int8, format: [H,W,in_ch]
[in]	input_x	input tensor width
[in]	input_y	input tensor height
[in]	input_ch	number of input tensor channels
[in]	kernel	pointer to kernel weights. Range: int8, format: [in_ch, H, W, out_ch]
[in]	output_ch	Number of output channels. output_ch = ch_mult * input_ch
[in]	ch_mult	channel multiplier.
[in]	kernel_x	filter/kernel width
[in]	kernel_y	filter/kernel height
[in]	pad_x	padding along width
[in]	pad_y	padding along height
[in]	stride_x	convolution stride along width
[in]	stride_y	convolution stride along height
[in]	bias	pointer to per output channel bias. Range: int32
[in,out]	output	pointer to output tensor. Format: [H, W, out_ch]
[in]	output_shift	pointer to per output channel requantization shift parameter.
[in]	output_mult	pointer to per output channel requantization multiplier parameter.
[in]	output_x	output tensor width
[in]	output_y	output tensor height
[in]	output_offset	offset to elements of output tensor
[in]	input_offset	offset to elements of input tensor
[in]	output_activation_min	Minimum value to clamp the output to. Range: int8
[in]	output_activation_max	Minimum value to clamp the output to. Range: int8
[in]	dilation_x	dilation along x. Not used. Dilation factor of 1 is used.
[in]	dilation_y	dilation along y. Not used. Dilation factor of 1 is used.
[in]	buffer_a	Not used.

Returns: The function returns ARM_MATH_SUCCESS

Supported framework: TensorFlow Lite
q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs.
Optimization using DSP extension is not available for the generic case where channel multiplier is > 1.

References arm_nn_requantize(), MAX, and MIN.

Referenced by arm_depthwise_conv_s8_opt().

arm_status arm_depthwise_conv_s8_opt	(	const q7_t *	input,
		const uint16_t	input_x,
		const uint16_t	input_y,
		const uint16_t	input_ch,
		const q7_t *	kernel,
		const uint16_t	output_ch,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y,
		const uint16_t	pad_x,
		const uint16_t	pad_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const int32_t *	bias,
		q7_t *	output,
		const int32_t *	output_shift,
		const int32_t *	output_mult,
		const uint16_t	output_x,
		const uint16_t	output_y,
		const int32_t	output_offset,
		const int32_t	input_offset,
		const int32_t	output_activation_min,
		const int32_t	output_activation_max,
		const uint16_t	dilation_x,
		const uint16_t	dilation_y,
		q15_t *	buffer_a
	)

Parameters

[in]	input	pointer to input tensor. Range: int8, format: [H,W,in_ch]
[in]	input_x	input tensor width
[in]	input_y	input tensor height
[in]	input_ch	number of input tensor channels
[in]	kernel	pointer to kernel weights. Range: int8, Format: [in_ch, H, W, out_ch]
[in]	output_ch	Number of output channels.
[in]	kernel_x	filter/kernel width
[in]	kernel_y	filter/kernel height
[in]	pad_x	padding along width
[in]	pad_y	padding along height
[in]	stride_x	convolution stride along width
[in]	stride_y	convolution stride along height
[in]	bias	pointer to per output channel bias. Range: int8
[in,out]	output	pointer to output tensor. Format: [H, W, out_ch]
[in]	output_shift	pointer to per output channel requantization shift parameter.
[in]	output_mult	pointer to per output channel requantization multiplier parameter.
[in]	output_x	output tensor width
[in]	output_y	output tensor height
[in]	output_offset	offset to elements of output tensor
[in]	input_offset	offset to elements of input tensor
[in]	output_activation_min	Minimum value to clamp the output to. Range: int8
[in]	output_activation_max	Minimum value to clamp the output to. Range: int8
[in]	dilation_x	dilation along x. Not used. Dilation factor of 1 is used.
[in]	dilation_y	dilation along y. Not used. Dilation factor of 1 is used.
[in]	buffer_a	Buffer for partial im2col optimization. This is mandatory when ARM_MATH_LOOPUNROLL and ARM_MATH_DSP are defined. Required space: (2 * input_ch * kernel_x * kernel_y) * sizeof(q15_t) bytes Use arm_depthwise_conv_s8_opt_get_buffer_size() to get the size.

Returns: The function returns one of the following ARM_MATH_SIZE_MISMATCH - Unsupported dimension of tensors ARM_MATH_SUCCESS - Successful operation

Supported framework: TensorFlow Lite
q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs.
Reccomended when number of channels is 4 or greater.

References arm_depthwise_conv_s8(), arm_nn_read_q15x2(), arm_nn_read_q7x4(), arm_nn_requantize(), arm_q7_to_q15_with_offset(), col_buffer, MAX, and MIN.

int32_t arm_depthwise_conv_s8_opt_get_buffer_size	(	const uint16_t	input_ch,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y
	)

Parameters

[in]	input_ch	number of input tensor channels
[in]	kernel_x	filter/kernel width
[in]	kernel_y	filter/kernel height

Returns: The function returns required buffer size

arm_status arm_depthwise_conv_u8_basic_ver1	(	const uint8_t *	input,
		const uint16_t	input_x,
		const uint16_t	input_y,
		const uint16_t	input_ch,
		const uint8_t *	kernel,
		const uint16_t	kernel_x,
		const uint16_t	kernel_y,
		const int16_t	ch_mult,
		const int16_t	pad_x,
		const int16_t	pad_y,
		const int16_t	stride_x,
		const int16_t	stride_y,
		const int16_t	dilation_x,
		const int16_t	dilation_y,
		const int32_t *	bias,
		const int32_t	input_offset,
		const int32_t	filter_offset,
		const int32_t	output_offset,
		uint8_t *	output,
		const uint16_t	output_x,
		const uint16_t	output_y,
		const int32_t	output_activation_min,
		const int32_t	output_activation_max,
		const int32_t	out_shift,
		const int32_t	out_mult
	)

uint8 depthwise convolution function with asymmetric quantization for even number of channel multiplier and input channels. Unless specified otherwise, arguments are mandatory.

Parameters

[in]	input	Pointer to input tensor
[in]	input_x	Width of input tensor
[in]	input_y	Height of input tensor
[in]	input_ch	Channels in input tensor
[in]	kernel	Pointer to kernel weights
[in]	kernel_x	Width of kernel
[in]	kernel_y	Height of kernel
[in]	ch_mult	Number of channel multiplier
[in]	pad_x	Padding sizes x
[in]	pad_y	Padding sizes y
[in]	stride_x	Convolution stride along the width
[in]	stride_y	Convolution stride along the height
[in]	dilation_x	Dilation along width. Not used and intended for future enhancement.
[in]	dilation_y	Dilation along height. Not used and intended for future enhancement.
[in]	bias	Pointer to optional bias values. If no bias is availble, NULL is expected
[in]	input_offset	Input tensor zero offset
[in]	filter_offset	Kernel tensor zero offset
[in]	output_offset	Output tensor zero offset
[in,out]	output	Pointer to output tensor
[in]	output_x	Width of output tensor
[in]	output_y	Height of output tensor
[in]	output_activation_min	Minimum value to clamp the output to. Range : {0, 255}
[in]	output_activation_max	Minimum value to clamp the output to. Range : {0, 255}
[in]	out_shift	Amount of right-shift for output
[in]	out_mult	Output multiplier for requantization

Returns: The function returns one of the following ARM_MATH_SIZE_MISMATCH - Not supported dimension of tensors ARM_MATH_SUCCESS - Successful operation ARM_MATH_ARGUMENT_ERROR - Implementation not available

Input constraints ch_mult is multiple of 2 kernel_x is multiple of 2

References arm_nn_divide_by_power_of_two(), arm_nn_sat_doubling_high_mult(), DILATION_X, DILATION_Y, LEFT_SHIFT, and RIGHT_SHIFT.

arm_status arm_depthwise_separable_conv_HWC_q7	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel,
		const uint16_t	padding,
		const uint16_t	stride,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimention
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel	filter kernel size
[in]	padding	padding sizes
[in]	stride	convolution stride
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out	output tensor dimension
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

Buffer size:

bufferA size: 2*ch_im_in*dim_kernel*dim_kernel

bufferB size: 0

Input dimension constraints:

ch_im_in equals ch_im_out

Implementation: There are 3 nested loop here: Inner loop: calculate each output value with MAC instruction over an accumulator Mid loop: loop over different output channel Outer loop: loop over different output (x, y)

References arm_nnword::bytes, NN_ROUND, and arm_nnword::word.

arm_status arm_depthwise_separable_conv_HWC_q7_nonsquare	(	const q7_t *	Im_in,
		const uint16_t	dim_im_in_x,
		const uint16_t	dim_im_in_y,
		const uint16_t	ch_im_in,
		const q7_t *	wt,
		const uint16_t	ch_im_out,
		const uint16_t	dim_kernel_x,
		const uint16_t	dim_kernel_y,
		const uint16_t	padding_x,
		const uint16_t	padding_y,
		const uint16_t	stride_x,
		const uint16_t	stride_y,
		const q7_t *	bias,
		const uint16_t	bias_shift,
		const uint16_t	out_shift,
		q7_t *	Im_out,
		const uint16_t	dim_im_out_x,
		const uint16_t	dim_im_out_y,
		q15_t *	bufferA,
		q7_t *	bufferB
	)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimention x
[in]	dim_im_in_y	input tensor dimention y
[in]	ch_im_in	number of input tensor channels
[in]	wt	pointer to kernel weights
[in]	ch_im_out	number of filters, i.e., output tensor channels
[in]	dim_kernel_x	filter kernel size x
[in]	dim_kernel_y	filter kernel size y
[in]	padding_x	padding sizes x
[in]	padding_y	padding sizes y
[in]	stride_x	convolution stride x
[in]	stride_y	convolution stride y
[in]	bias	pointer to bias
[in]	bias_shift	amount of left-shift for bias
[in]	out_shift	amount of right-shift for output
[in,out]	Im_out	pointer to output tensor
[in]	dim_im_out_x	output tensor dimension x
[in]	dim_im_out_y	output tensor dimension y
[in,out]	bufferA	pointer to buffer space for input
[in,out]	bufferB	pointer to buffer space for output

Returns: The function returns either ARM_MATH_SIZE_MISMATCH or ARM_MATH_SUCCESS based on the outcome of size checking.

This function is the version with full list of optimization tricks, but with some contraints: ch_im_in is equal to ch_im_out

References arm_nnword::bytes, NN_ROUND, and arm_nnword::word.

Macros

Functions

Description

Macro Definition Documentation

Function Documentation