Convolution Functions

Collection of convolution, depthwise convolution functions and their variants. More...

Functions
riscv_status	riscv_convolve_1_x_n_s8 (const nmsis_nn_context *ctx, const nmsis_nn_conv_params *conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input_data, const nmsis_nn_dims *filter_dims, const q7_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, q7_t *output_data)
	1xn convolution More...
int32_t	riscv_convolve_1_x_n_s8_get_buffer_size (const nmsis_nn_dims *input_dims, const nmsis_nn_dims *filter_dims)
	Get the required additional buffer size for 1xn convolution. More...
riscv_status	riscv_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...
riscv_status	riscv_convolve_1x1_s8_fast (const nmsis_nn_context *ctx, const nmsis_nn_conv_params *conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input_data, const nmsis_nn_dims *filter_dims, const q7_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, q7_t *output_data)
	Fast s8 version for 1x1 convolution (non-square shape) More...
int32_t	riscv_convolve_1x1_s8_fast_get_buffer_size (const nmsis_nn_dims *input_dims)
	Get the required buffer size for riscv_convolve_1x1_s8_fast. More...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_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...
riscv_status	riscv_convolve_s8 (const nmsis_nn_context *ctx, const nmsis_nn_conv_params *conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input_data, const nmsis_nn_dims *filter_dims, const q7_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, q7_t *output_data)
	Basic s8 convolution function. More...
int32_t	riscv_convolve_s8_get_buffer_size (const nmsis_nn_dims *input_dims, const nmsis_nn_dims *filter_dims)
	Get the required buffer size for s8 convolution function. More...
riscv_status	riscv_convolve_wrapper_s8 (const nmsis_nn_context *ctx, const nmsis_nn_conv_params *conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input_data, const nmsis_nn_dims *filter_dims, const q7_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, q7_t *output_data)
	s8 convolution layer wrapper function with the main purpose to call the optimal kernel available in nmsis-nn to perform the convolution. More...
int32_t	riscv_convolve_wrapper_s8_get_buffer_size (const nmsis_nn_conv_params *conv_params, const nmsis_nn_dims *input_dims, const nmsis_nn_dims *filter_dims, const nmsis_nn_dims *output_dims)
	Get the required buffer size for riscv_convolve_wrapper_s8. More...
riscv_status	riscv_depthwise_conv_3x3_s8 (const nmsis_nn_context *ctx, const nmsis_nn_dw_conv_params *dw_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input, const nmsis_nn_dims *filter_dims, const q7_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, q7_t *output)
	Optimized s8 depthwise convolution function for 3x3 kernel size with some constraints on the input arguments(documented below). Refer riscv_depthwise_conv_s8() for function argument details. More...
static void	depthwise_conv_s8_mult_4 (const int8_t *input, const int32_t input_x, const int32_t input_y, const int32_t input_ch, const int8_t *kernel, const int32_t output_ch, const int32_t ch_mult, const int32_t kernel_x, const int32_t kernel_y, const int32_t pad_x, const int32_t pad_y, const int32_t stride_x, const int32_t stride_y, const int32_t *bias, int8_t *output, const int32_t *output_shift, const int32_t *output_mult, const int32_t output_x, const int32_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)
static void	depthwise_conv_s8_generic (const q7_t *input, const uint16_t input_batches, 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)
riscv_status	riscv_depthwise_conv_s8 (const nmsis_nn_context *ctx, const nmsis_nn_dw_conv_params *dw_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input, const nmsis_nn_dims *filter_dims, const q7_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, q7_t *output)
	Basic s8 depthwise convolution function that doesn't have any constraints on the input dimensions. More...
riscv_status	riscv_depthwise_conv_s8_opt (const nmsis_nn_context *ctx, const nmsis_nn_dw_conv_params *dw_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input, const nmsis_nn_dims *filter_dims, const q7_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, q7_t *output)
	Optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. Refer riscv_depthwise_conv_s8() for function argument details. More...
int32_t	riscv_depthwise_conv_s8_opt_get_buffer_size (const nmsis_nn_dims *input_dims, const nmsis_nn_dims *filter_dims)
	Get the required buffer size for optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. More...
static void	depthwise_conv_u8_mult_4 (const uint8_t *input, const int32_t input_x, const int32_t input_y, const int32_t input_ch, const uint8_t *kernel, const int32_t output_ch, const int32_t ch_mult, const int32_t kernel_x, const int32_t kernel_y, const int32_t pad_x, const int32_t pad_y, const int32_t stride_x, const int32_t stride_y, const int32_t *bias, uint8_t *output, const int32_t output_shift, const int32_t output_mult, const int32_t output_x, const int32_t output_y, const int32_t output_offset, const int32_t input_offset, const int32_t filter_offset, const int32_t output_activation_min, const int32_t output_activation_max)
static void	depthwise_conv_u8_generic (const uint8_t *input, const int32_t input_x, const int32_t input_y, const int32_t input_ch, const uint8_t *kernel, const int32_t output_ch, const int32_t ch_mult, const int32_t kernel_x, const int32_t kernel_y, const int32_t pad_x, const int32_t pad_y, const int32_t stride_x, const int32_t stride_y, const int32_t *bias, uint8_t *output, const int32_t output_shift, const int32_t output_mult, const int32_t output_x, const int32_t output_y, const int32_t output_offset, const int32_t input_offset, const int32_t filter_offset, const int32_t output_activation_min, const int32_t output_activation_max)
riscv_status	riscv_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 output_shift, const int32_t output_mult)
	uint8 depthwise convolution function with asymmetric quantization More...
riscv_status	riscv_depthwise_conv_wrapper_s8 (const nmsis_nn_context *ctx, const nmsis_nn_dw_conv_params *dw_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const q7_t *input, const nmsis_nn_dims *filter_dims, const q7_t *filter, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, q7_t *output)
	Wrapper function to pick the right optimized s8 depthwise convolution function. More...
int32_t	riscv_depthwise_conv_wrapper_s8_get_buffer_size (const nmsis_nn_dw_conv_params *dw_conv_params, const nmsis_nn_dims *input_dims, const nmsis_nn_dims *filter_dims, const nmsis_nn_dims *output_dims)
	Get size of additional buffer required by riscv_depthwise_conv_wrapper_s8() More...
riscv_status	riscv_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...
riscv_status	riscv_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...

Detailed Description

Collection of convolution, depthwise convolution functions and their variants.

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 NMSIS-DSP riscv_mat_mult functions.

Function Documentation

depthwise_conv_s8_generic()

static void depthwise_conv_s8_generic ( const q7_t *  input, const uint16_t  input_batches, 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  )

static

depthwise_conv_s8_mult_4()

static void depthwise_conv_s8_mult_4 ( const int8_t *  input, const int32_t  input_x, const int32_t  input_y, const int32_t  input_ch, const int8_t *  kernel, const int32_t  output_ch, const int32_t  ch_mult, const int32_t  kernel_x, const int32_t  kernel_y, const int32_t  pad_x, const int32_t  pad_y, const int32_t  stride_x, const int32_t  stride_y, const int32_t *  bias, int8_t *  output, const int32_t *  output_shift, const int32_t *  output_mult, const int32_t  output_x, const int32_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  )

static

depthwise_conv_u8_generic()

static void depthwise_conv_u8_generic ( const uint8_t *  input, const int32_t  input_x, const int32_t  input_y, const int32_t  input_ch, const uint8_t *  kernel, const int32_t  output_ch, const int32_t  ch_mult, const int32_t  kernel_x, const int32_t  kernel_y, const int32_t  pad_x, const int32_t  pad_y, const int32_t  stride_x, const int32_t  stride_y, const int32_t *  bias, uint8_t *  output, const int32_t  output_shift, const int32_t  output_mult, const int32_t  output_x, const int32_t  output_y, const int32_t  output_offset, const int32_t  input_offset, const int32_t  filter_offset, const int32_t  output_activation_min, const int32_t  output_activation_max  )

static

depthwise_conv_u8_mult_4()

static void depthwise_conv_u8_mult_4 ( const uint8_t *  input, const int32_t  input_x, const int32_t  input_y, const int32_t  input_ch, const uint8_t *  kernel, const int32_t  output_ch, const int32_t  ch_mult, const int32_t  kernel_x, const int32_t  kernel_y, const int32_t  pad_x, const int32_t  pad_y, const int32_t  stride_x, const int32_t  stride_y, const int32_t *  bias, uint8_t *  output, const int32_t  output_shift, const int32_t  output_mult, const int32_t  output_x, const int32_t  output_y, const int32_t  output_offset, const int32_t  input_offset, const int32_t  filter_offset, const int32_t  output_activation_min, const int32_t  output_activation_max  )

static

riscv_convolve_1_x_n_s8()

riscv_status riscv_convolve_1_x_n_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_conv_params *	conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

1xn convolution

Parameters

[in,out]	ctx	Function context that contains the additional buffer if required by the function. riscv_convolve_1_x_n_s8_get_buffer_size will return the buffer_size if required
[in]	conv_params	Convolution parameters (e.g. strides, dilations, pads,...). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Optional bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [N, H, W, C_OUT]
[out]	output_data	Output data pointer. Data type: int8

Returns

The function returns either RISCV_MATH_SIZE_MISMATCH if argument constraints fail. or, RISCV_MATH_SUCCESS on successful completion.

• Supported framework : TensorFlow Lite Micro
• The following constrains on the arguments apply
  1. input_dims->n equals 1
  2. ouput_dims->w is a multiple of 4
  3. Explicit constraints(since it is for 1xN convolution) -## input_dims->h equals 1 -## output_dims->h equals 1 -## filter_dims->h equals 1

     Todo:

     Remove constraint on output_dims->w to make the function generic.

riscv_convolve_1_x_n_s8_get_buffer_size()

int32_t riscv_convolve_1_x_n_s8_get_buffer_size	(	const nmsis_nn_dims *	input_dims,
		const nmsis_nn_dims *	filter_dims
	)

Get the required additional buffer size for 1xn convolution.

Parameters

[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension

Returns

The function returns required buffer size(bytes)

riscv_convolve_1x1_HWC_q7_fast_nonsquare()

riscv_status riscv_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)

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 RISCV_MATH_SIZE_MISMATCH or RISCV_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 constraints: 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

riscv_convolve_1x1_s8_fast()

riscv_status riscv_convolve_1x1_s8_fast	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_conv_params *	conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Fast s8 version for 1x1 convolution (non-square shape)

Parameters

[in,out]	ctx	Function context that contains the additional buffer if required by the function. riscv_convolve_1x1_s8_fast_get_buffer_size will return the buffer_size if required
[in]	conv_params	Convolution parameters (e.g. strides, dilations, pads,...). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Optional bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [N, H, W, C_OUT]
[out]	output_data	Output data pointer. Data type: int8

Returns

The function returns either RISCV_MATH_SIZE_MISMATCH if argument constraints fail. or, RISCV_MATH_SUCCESS on successful completion.

• Supported framework : TensorFlow Lite Micro
• The following constrains on the arguments apply
  1. input_dims->c is a multiple of 4
  2. conv_params->padding.w = conv_params->padding.h = 0
  3. conv_params->stride.w = conv_params->stride.h = 1

riscv_convolve_1x1_s8_fast_get_buffer_size()

int32_t riscv_convolve_1x1_s8_fast_get_buffer_size

(

const nmsis_nn_dims *

input_dims

)

Get the required buffer size for riscv_convolve_1x1_s8_fast.

Parameters

[in]

input_dims

Input (activation) dimensions

Returns

The function returns the required buffer size in bytes

riscv_convolve_HWC_q15_basic()

riscv_status riscv_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.

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 RISCV_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.

riscv_convolve_HWC_q15_fast()

riscv_status riscv_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.

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 RISCV_MATH_SIZE_MISMATCH or RISCV_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 multiple of 2

riscv_convolve_HWC_q15_fast_nonsquare()

riscv_status riscv_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)

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 RISCV_MATH_SIZE_MISMATCH or RISCV_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 multiple of 2

riscv_convolve_HWC_q7_basic()

riscv_status riscv_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.

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 RISCV_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.

riscv_convolve_HWC_q7_basic_nonsquare()

riscv_status riscv_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)

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 RISCV_MATH_SUCCESS

riscv_convolve_HWC_q7_fast()

riscv_status riscv_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.

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 RISCV_MATH_SIZE_MISMATCH or RISCV_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 multiple 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 riscv_q7_to_q15_reordered_no_shift. For every four elements, the second and third elements are swapped.

The computation kernel riscv_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.

riscv_convolve_HWC_q7_fast_nonsquare()

riscv_status riscv_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)

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 RISCV_MATH_SIZE_MISMATCH or RISCV_MATH_SUCCESS based on the outcome of size checking.

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

riscv_convolve_HWC_q7_RGB()

riscv_status riscv_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.

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 RISCV_MATH_SIZE_MISMATCH or RISCV_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.

riscv_convolve_s8()

riscv_status riscv_convolve_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_conv_params *	conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Basic s8 convolution function.

Parameters

[in,out]	ctx	Function context that contains the additional buffer if required by the function. riscv_convolve_s8_get_buffer_size will return the buffer_size if required
[in]	conv_params	Convolution parameters (e.g. strides, dilations, pads,...). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Optional bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [N, H, W, C_OUT]
[out]	output_data	Output data pointer. Data type: int8

Returns

The function returns RISCV_MATH_SUCCESS

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

riscv_convolve_s8_get_buffer_size()

int32_t riscv_convolve_s8_get_buffer_size	(	const nmsis_nn_dims *	input_dims,
		const nmsis_nn_dims *	filter_dims
	)

Get the required buffer size for s8 convolution function.

Parameters

[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

Returns

The function returns required buffer size(bytes)

riscv_convolve_wrapper_s8()

riscv_status riscv_convolve_wrapper_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_conv_params *	conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

s8 convolution layer wrapper function with the main purpose to call the optimal kernel available in nmsis-nn to perform the convolution.

Parameters

[in,out]	ctx	Function context that contains the additional buffer if required by the function. riscv_convolve_wrapper_s8_get_buffer_size will return the buffer_size if required
[in]	conv_params	Convolution parameters (e.g. strides, dilations, pads,...). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [N, H, W, C_IN]
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [N, H, W, C_OUT]
[out]	output_data	Output data pointer. Data type: int8

Returns

The function returns either RISCV_MATH_SIZE_MISMATCH if argument constraints fail. or, RISCV_MATH_SUCCESS on successful completion.

riscv_convolve_wrapper_s8_get_buffer_size()

int32_t riscv_convolve_wrapper_s8_get_buffer_size	(	const nmsis_nn_conv_params *	conv_params,
		const nmsis_nn_dims *	input_dims,
		const nmsis_nn_dims *	filter_dims,
		const nmsis_nn_dims *	output_dims
	)

Get the required buffer size for riscv_convolve_wrapper_s8.

Parameters

[in]	conv_params	Convolution parameters (e.g. strides, dilations, pads,...). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]
[in]	input_dims	Input (activation) dimensions. Format: [N, H, W, C_IN]
[in]	filter_dims	Filter dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions
[in]	output_dims	Output tensor dimensions. Format: [N, H, W, C_OUT]

Returns

The function returns required buffer size(bytes)

riscv_depthwise_conv_3x3_s8()

riscv_status riscv_depthwise_conv_3x3_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_dw_conv_params *	dw_conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Optimized s8 depthwise convolution function for 3x3 kernel size with some constraints on the input arguments(documented below). Refer riscv_depthwise_conv_s8() for function argument details.

Returns

The function returns one of the following RISCV_MATH_SIZE_MISMATCH - Unsupported dimension of tensors RISCV_MATH_ARGUMENT_ERROR - Unsupported pad size along the x axis RISCV_MATH_SUCCESS - Successful operation

• Supported framework : TensorFlow Lite Micro
• The following constrains on the arguments apply
  1. Number of input channel equals number of output channels
  2. Filter height and width equals 3
  3. Padding along x is either 0 or 1.

riscv_depthwise_conv_s8()

riscv_status riscv_depthwise_conv_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_dw_conv_params *	dw_conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Basic s8 depthwise convolution function that doesn't have any constraints on the input dimensions.

Parameters

[in,out]	ctx	Function context (e.g. temporary buffer). Check the function definition file to see if an additional buffer is required. Optional function {API}_get_buffer_size() provides the buffer size if an additional buffer is required. exists if additional memory is.
[in]	dw_conv_params	Depthwise convolution parameters (e.g. strides, dilations, pads,...) dw_conv_params->dilation is not used. Range of dw_conv_params->input_offset : [-127, 128] Range of dw_conv_params->input_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [1, H, W, C_IN] Batch argument N is not used.
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [1, H, W, C_OUT]
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [1, H, W, C_OUT]
[in,out]	output_data	Output data pointer. Data type: int8

Returns

The function returns RISCV_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.

riscv_depthwise_conv_s8_opt()

riscv_status riscv_depthwise_conv_s8_opt	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_dw_conv_params *	dw_conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Optimized s8 depthwise convolution function with constraint that in_channel equals out_channel. Refer riscv_depthwise_conv_s8() for function argument details.

Returns

The function returns one of the following RISCV_MATH_SIZE_MISMATCH - input channel != output channel or ch_mult != 1 RISCV_MATH_SUCCESS - Successful operation

Note

If number of channels is not a multiple of 4, upto 3 elements outside the boundary will be read out for the following if MVE optimizations(Arm Helium Technology) are used.

• Output shift
• Output multiplier
• Output bias
• kernel

• Supported framework: TensorFlow Lite
• The following constrains on the arguments apply
  1. Number of input channel equals number of output channels or ch_mult equals 1
• 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.

riscv_depthwise_conv_s8_opt_get_buffer_size()

int32_t riscv_depthwise_conv_s8_opt_get_buffer_size	(	const nmsis_nn_dims *	input_dims,
		const nmsis_nn_dims *	filter_dims
	)

Get the required buffer size for optimized s8 depthwise convolution function with constraint that in_channel equals out_channel.

Parameters

[in]	input_dims	Input (activation) tensor dimensions. Format: [1, H, W, C_IN] Batch argument N is not used.
[in]	filter_dims	Filter tensor dimensions. Format: [1, H, W, C_OUT]

Returns

The function returns required buffer size in bytes

riscv_depthwise_conv_u8_basic_ver1()

riscv_status riscv_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	output_shift,
		const int32_t	output_mult
	)

uint8 depthwise convolution function with asymmetric quantization

uint8 depthwise convolution function with asymmetric quantization 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 available, 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]	output_shift	Amount of right-shift for output
[in]	output_mult	Output multiplier for requantization

Returns

The function returns one of the following RISCV_MATH_SIZE_MISMATCH - Not supported dimension of tensors RISCV_MATH_SUCCESS - Successful operation RISCV_MATH_ARGUMENT_ERROR - Implementation not available

riscv_depthwise_conv_wrapper_s8()

riscv_status riscv_depthwise_conv_wrapper_s8	(	const nmsis_nn_context *	ctx,
		const nmsis_nn_dw_conv_params *	dw_conv_params,
		const nmsis_nn_per_channel_quant_params *	quant_params,
		const nmsis_nn_dims *	input_dims,
		const q7_t *	input_data,
		const nmsis_nn_dims *	filter_dims,
		const q7_t *	filter_data,
		const nmsis_nn_dims *	bias_dims,
		const int32_t *	bias_data,
		const nmsis_nn_dims *	output_dims,
		q7_t *	output_data
	)

Wrapper function to pick the right optimized s8 depthwise convolution function.

Parameters

[in,out]	ctx	Function context (e.g. temporary buffer). Check the function definition file to see if an additional buffer is required. Optional function {API}_get_buffer_size() provides the buffer size if required.
[in]	dw_conv_params	Depthwise convolution parameters (e.g. strides, dilations, pads,...) dw_conv_params->dilation is not used. Range of dw_conv_params->input_offset : [-127, 128] Range of dw_conv_params->output_offset : [-128, 127]
[in]	quant_params	Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel
[in]	input_dims	Input (activation) tensor dimensions. Format: [H, W, C_IN] Batch argument N is not used and assumed to be 1.
[in]	input_data	Input (activation) data pointer. Data type: int8
[in]	filter_dims	Filter tensor dimensions. Format: [1, H, W, C_OUT]
[in]	filter_data	Filter data pointer. Data type: int8
[in]	bias_dims	Bias tensor dimensions. Format: [C_OUT]
[in]	bias_data	Bias data pointer. Data type: int32
[in]	output_dims	Output tensor dimensions. Format: [1, H, W, C_OUT]
[in,out]	output_data	Output data pointer. Data type: int8

Returns

The function returns RISCV_MATH_SUCCESS - Successful completion.

• Supported framework: TensorFlow Lite
• Picks one of the the following functions
  1. riscv_depthwise_conv_s8()
  2. riscv_depthwise_conv_3x3_s8() - RISC-V CPUs with DSP extension only
  3. riscv_depthwise_conv_s8_opt()
• q7 is used as data type eventhough it is s8 data. It is done so to be consistent with existing APIs.
• Check details of riscv_depthwise_conv_s8_opt() for potential data that can be accessed outside of the boundary.

riscv_depthwise_conv_wrapper_s8_get_buffer_size()

int32_t riscv_depthwise_conv_wrapper_s8_get_buffer_size	(	const nmsis_nn_dw_conv_params *	dw_conv_params,
		const nmsis_nn_dims *	input_dims,
		const nmsis_nn_dims *	filter_dims,
		const nmsis_nn_dims *	output_dims
	)

Get size of additional buffer required by riscv_depthwise_conv_wrapper_s8()

Parameters

[in]	dw_conv_params	Depthwise convolution parameters (e.g. strides, dilations, pads,...) dw_conv_params->dilation is not used. Range of dw_conv_params->input_offset : [-127, 128] Range of dw_conv_params->input_offset : [-128, 127]
[in]	input_dims	Input (activation) tensor dimensions. Format: [H, W, C_IN] Batch argument N is not used and assumed to be 1.
[in]	filter_dims	Filter tensor dimensions. Format: [1, H, W, C_OUT]
[in]	output_dims	Output tensor dimensions. Format: [1, H, W, C_OUT]

Returns

Size of additional memory required for optimizations in bytes.

riscv_depthwise_separable_conv_HWC_q7()

riscv_status riscv_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.

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in	input tensor dimension
[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 RISCV_MATH_SIZE_MISMATCH or RISCV_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)

riscv_depthwise_separable_conv_HWC_q7_nonsquare()

riscv_status riscv_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)

Parameters

[in]	Im_in	pointer to input tensor
[in]	dim_im_in_x	input tensor dimension x
[in]	dim_im_in_y	input tensor dimension 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 RISCV_MATH_SIZE_MISMATCH or RISCV_MATH_SUCCESS based on the outcome of size checking.

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