Convolution Functions

• GetBufferSizeNNConv
  • riscv_convolve_s16_get_buffer_size()
  • riscv_convolve_wrapper_s16_get_buffer_size()
  • riscv_convolve_wrapper_s16_get_buffer_size_dsp()
  • riscv_convolve_s4_get_buffer_size()
  • riscv_convolve_1_x_n_s4_get_buffer_size()
  • riscv_convolve_1x1_s4_fast_get_buffer_size()
  • riscv_convolve_wrapper_s4_get_buffer_size()
  • riscv_convolve_wrapper_s4_get_buffer_size_dsp()
  • riscv_convolve_s8_get_buffer_size()
  • riscv_convolve_1_x_n_s8_get_buffer_size()
  • riscv_convolve_1x1_s8_fast_get_buffer_size()
  • riscv_convolve_wrapper_s8_get_buffer_size()
  • riscv_convolve_wrapper_s8_get_buffer_size_dsp()
  • riscv_depthwise_conv_fast_s16_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s16_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s16_get_buffer_size_dsp()
  • riscv_depthwise_conv_s4_opt_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s4_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s4_get_buffer_size_dsp()
  • riscv_depthwise_conv_s8_opt_get_buffer_size_dsp()
  • riscv_depthwise_conv_s8_opt_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s8_get_buffer_size()
  • riscv_depthwise_conv_wrapper_s8_get_buffer_size_dsp()
  • riscv_transpose_conv_s8_get_buffer_size()
  • riscv_transpose_conv_s8_get_reverse_conv_buffer_size()

riscv_nmsis_nn_status riscv_convolve_1_x_n_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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)

riscv_nmsis_nn_status riscv_convolve_1x1_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_status riscv_convolve_1x1_s4_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_status riscv_convolve_1x1_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_status riscv_convolve_even_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *packed_filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_status riscv_convolve_s16(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 int16_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const nmsis_nn_bias_data *bias_data, const nmsis_nn_dims *output_dims, int16_t *output_data)

riscv_nmsis_nn_status riscv_convolve_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *packed_filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *upscale_dims, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_status riscv_convolve_wrapper_s16(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 int16_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const nmsis_nn_bias_data *bias_data, const nmsis_nn_dims *output_dims, int16_t *output_data)

riscv_nmsis_nn_status riscv_convolve_wrapper_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_status riscv_depthwise_conv_fast_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

static void __attribute__ ((unused))

static void depthwise_conv_s16_generic_s16(const int16_t *input, const uint16_t input_batches, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const int8_t *kernel, 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 int64_t *bias, int16_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_activation_min, const int32_t output_activation_max, const uint16_t dilation_x, const uint16_t dilation_y)

riscv_nmsis_nn_status riscv_depthwise_conv_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

static void depthwise_conv_s4_generic(const int8_t *input, const int32_t input_batches, 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, const int32_t dilation_x, const int32_t dilation_y)

riscv_nmsis_nn_status riscv_depthwise_conv_s4(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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_status riscv_depthwise_conv_s4_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

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 int8_t *input, const uint16_t input_batches, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const int8_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, int8_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)

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_status riscv_depthwise_conv_wrapper_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

riscv_nmsis_nn_status riscv_depthwise_conv_wrapper_s4(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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

riscv_nmsis_nn_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)

riscv_nmsis_nn_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)

riscv_nmsis_nn_status riscv_transpose_conv_s8(const nmsis_nn_context *ctx, const nmsis_nn_context *output_ctx, const nmsis_nn_transpose_conv_params *transpose_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_status riscv_transpose_conv_wrapper_s8(const nmsis_nn_context *ctx, const nmsis_nn_context *reverse_conv_ctx, const nmsis_nn_transpose_conv_params *transpose_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

group Convolution Functions

Collection of convolution, depthwise convolution functions and their variants.

The convolution is implemented in 2 steps: im2col and General Matrix Multiplication(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 followed by GEMM.

Functions

riscv_nmsis_nn_status riscv_convolve_1_x_n_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

1xn convolution for s4 weights

• Supported framework : TensorFlow Lite Micro

• The following constrains on the arguments apply

  1. stride.w * input_dims->c is a multiple of 4

  2. 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.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_convolve_1_x_n_s4_get_buffer_size will return the buffer_size if required The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension

• filter_data – [in] Filter data pointer. Data type: int8 as packed int4

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

1xn convolution

• 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.

Parameters:

• ctx – [inout] 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 The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, WK, C_IN] where WK is the horizontal spatial filter dimension

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_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)

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

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in_x – [in] input tensor dimention x

• dim_im_in_y – [in] input tensor dimention y

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel_x – [in] filter kernel size x

• dim_kernel_y – [in] filter kernel size y

• padding_x – [in] padding size x

• padding_y – [in] padding size y

• stride_x – [in] convolution stride x

• stride_y – [in] convolution stride y

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out_x – [in] output tensor dimension x

• dim_im_out_y – [in] output tensor dimension y

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_status riscv_convolve_1x1_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

s4 version for 1x1 convolution with support for non-unity stride values

• Supported framework : TensorFlow Lite Micro

• The following constrains on the arguments apply

  1. conv_params->padding.w = conv_params->padding.h = 0

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. None is required by this function.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]

• filter_data – [in] Filter data pointer. Data type: int8 packed with 2x int4

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_status riscv_convolve_1x1_s4_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

Fast s4 version for 1x1 convolution (non-square shape)

• Supported framework : TensorFlow Lite Micro

• The following constrains on the arguments apply

  1. conv_params->padding.w = conv_params->padding.h = 0

  2. conv_params->stride.w = conv_params->stride.h = 1

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_convolve_1x1_s4_fast_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer ,if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]

• filter_data – [in] Filter data pointer. Data type: int8 packed with 2x int4

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_status riscv_convolve_1x1_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

s8 version for 1x1 convolution with support for non-unity stride values

• Supported framework : TensorFlow Lite Micro

• The following constrains on the arguments apply

  1. conv_params->padding.w = conv_params->padding.h = 0

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. None is required by this function.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

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

• Supported framework : TensorFlow Lite Micro

• The following constrains on the arguments apply

  1. conv_params->padding.w = conv_params->padding.h = 0

  2. conv_params->stride.w = conv_params->stride.h = 1

• 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

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

• ctx – [inout] 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

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, 1, 1, C_IN]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_status riscv_convolve_even_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *packed_filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

riscv_nmsis_nn_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.

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.

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimention

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_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.

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

dim_im_out is a multiple of 2

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimention

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_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)

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

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in_x – [in] input tensor dimention x

• dim_im_in_y – [in] input tensor dimention y

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel_x – [in] filter kernel size x

• dim_kernel_y – [in] filter kernel size y

• padding_x – [in] padding size x

• padding_y – [in] padding size y

• stride_x – [in] convolution stride x

• stride_y – [in] convolution stride y

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out_x – [in] output tensor dimension x

• dim_im_out_y – [in] output tensor dimension y

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_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.

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.

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimention

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_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:

• Im_in – [in] pointer to input tensor

• dim_im_in_x – [in] input tensor dimention x

• dim_im_in_y – [in] input tensor dimention y

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel_x – [in] filter kernel size x

• dim_kernel_y – [in] filter kernel size y

• padding_x – [in] padding size x

• padding_y – [in] padding size y

• stride_x – [in] convolution stride x

• stride_y – [in] convolution stride y

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out_x – [in] output tensor dimension x

• dim_im_out_y – [in] output tensor dimension y

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_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.

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.

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimention

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_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)

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

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in_x – [in] input tensor dimention x

• dim_im_in_y – [in] input tensor dimention y

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel_x – [in] filter kernel size x

• dim_kernel_y – [in] filter kernel size y

• padding_x – [in] padding size x

• padding_y – [in] padding size y

• stride_x – [in] convolution stride x

• stride_y – [in] convolution stride y

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out_x – [in] output tensor dimension x

• dim_im_out_y – [in] output tensor dimension y

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_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.

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.

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimention

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_status riscv_convolve_s16(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 int16_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const nmsis_nn_bias_data *bias_data, const nmsis_nn_dims *output_dims, int16_t *output_data)

Basic s16 convolution function.

1. Supported framework: TensorFlow Lite micro

2. Additional memory is required for optimization. Refer to argument ‘ctx’ for details.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_convolve_s16_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). conv_params->input_offset : Not used conv_params->output_offset : Not used

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int16

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Struct with optional bias data pointer. Bias data type can be int64 or int32 depending flag in struct.

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int16

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS if successful or RISCV_NMSIS_NN_ARG_ERROR if incorrect arguments or RISCV_NMSIS_NN_NO_IMPL_ERROR

riscv_nmsis_nn_status riscv_convolve_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *packed_filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

Basic s4 convolution function.

1. Supported framework: TensorFlow Lite micro

2. Additional memory is required for optimization. Refer to argument ‘ctx’ for details.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_convolve_s4_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer ,if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Packed Filter data pointer. Data type: int8 packed with 2x int4

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *upscale_dims, const nmsis_nn_dims *output_dims, int8_t *output_data)

Basic s8 convolution function.

1. Supported framework: TensorFlow Lite micro

2. Additional memory is required for optimization. Refer to argument ‘ctx’ for details.

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, CK] where HK, WK and CK are the spatial filter dimensions. CK != C_IN is used for grouped convolution, in which case the required conditions are C_IN = N * CK and C_OUT = N * M for N groups of size M.

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• upscale_dims – [in] Inserts zeroes to upscale the input in h/w dimensions if set to 2. This is used for tranposed convolution.

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS if successful or RISCV_NMSIS_NN_ARG_ERROR if incorrect arguments or RISCV_NMSIS_NN_NO_IMPL_ERROR

riscv_nmsis_nn_status riscv_convolve_wrapper_s16(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 int16_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const nmsis_nn_bias_data *bias_data, const nmsis_nn_dims *output_dims, int16_t *output_data)

s16 convolution layer wrapper function with the main purpose to call the optimal kernel available in nmsis-nn to perform the convolution.

Parameters:

• ctx – [inout] 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 The caller is expected to clear the buffer, if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). conv_params->input_offset : Not used conv_params->output_offset : Not used

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int16

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Struct with optional bias data pointer. Bias data type can be int64 or int32 depending flag in struct.

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int16

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_status riscv_convolve_wrapper_s4(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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

s4 convolution layer wrapper function with the main purpose to call the optimal kernel available in nmsis-nn to perform the convolution.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_convolve_wrapper_s4_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer ,if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8 packed with 2x int4

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_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 int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_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:

• ctx – [inout] 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. The caller is expected to clear the buffer ,if applicable, for security reasons.

• conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of conv_params->input_offset : [-127, 128] Range of conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_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.

• 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.

Returns:

The function returns one of the following RISCV_NMSIS_NN_ARG_ERROR - Unsupported dimension of tensors

• Unsupported pad size along the x axis RISCV_NMSIS_NN_SUCCESS - Successful operation

riscv_nmsis_nn_status riscv_depthwise_conv_fast_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

Optimized s16 depthwise convolution function with constraint that in_channel equals out_channel. Refer riscv_depthwise_conv_s16() for function argument details.

RISCV_NMSIS_NN_SUCCESS - Successful operation

• 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

• Reccomended when number of channels is 4 or greater.

Returns:

The function returns one of the following RISCV_NMSIS_NN_ARG_ERROR - ctx-buff == NULL and riscv_depthwise_conv_fast_s16_get_buffer_size() > 0 or input channel != output channel or ch_mult != 1

static void __attribute__ ((unused))

static void depthwise_conv_s16_generic_s16(const int16_t *input, const uint16_t input_batches, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const int8_t *kernel, 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 int64_t *bias, int16_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_activation_min, const int32_t output_activation_max, const uint16_t dilation_x, const uint16_t dilation_y)

riscv_nmsis_nn_status riscv_depthwise_conv_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

Basic s16 depthwise convolution function that doesn’t have any constraints on the input dimensions.

• Supported framework: TensorFlow Lite

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• dw_conv_params – [in] Depthwise convolution parameters (e.g. strides, dilations, pads,…) conv_params->input_offset : Not used conv_params->output_offset : Not used

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN] Batch argument N is not used.

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int64

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [inout] Output data pointer. Data type: int16

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

static void depthwise_conv_s4_generic(const int8_t *input, const int32_t input_batches, 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, const int32_t dilation_x, const int32_t dilation_y)

riscv_nmsis_nn_status riscv_depthwise_conv_s4(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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

Basic s4 depthwise convolution function that doesn’t have any constraints on the input dimensions.

• Supported framework: TensorFlow Lite

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer ,if applicable, for security reasons.

• dw_conv_params – [in] 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]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN] Batch argument N is not used.

• input – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• kernel – [in] Filter data pointer. Data type: int8_t packed 4-bit weights, e.g four sequential weights [0x1, 0x2, 0x3, 0x4] packed as [0x21, 0x43].

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output – [inout] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_status riscv_depthwise_conv_s4_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

Optimized s4 depthwise convolution function with constraint that in_channel equals out_channel. Refer riscv_depthwise_conv_s4() for function argument details.

• 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

• Reccomended when number of channels is 4 or greater.

Returns:

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

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 int8_t *input, const uint16_t input_batches, const uint16_t input_x, const uint16_t input_y, const uint16_t input_ch, const int8_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, int8_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)

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

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

• Supported framework: TensorFlow Lite

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• dw_conv_params – [in] 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]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN] Batch argument N is not used.

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [inout] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *kernel, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

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

• 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

• Reccomended when number of channels is 4 or greater.

Returns:

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

riscv_nmsis_nn_status riscv_depthwise_conv_wrapper_s16(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 int16_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int64_t *bias, const nmsis_nn_dims *output_dims, int16_t *output)

Wrapper function to pick the right optimized s16 depthwise convolution function.

• Supported framework: TensorFlow Lite

• Picks one of the the following functions

  1. riscv_depthwise_conv_s16()

  2. riscv_depthwise_conv_fast_s16() - RISC-V CPUs with DSP extension only

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• dw_conv_params – [in] Depthwise convolution parameters (e.g. strides, dilations, pads,…) dw_conv_params->dilation is not used. Range of dw_conv_params->input_offset : Not used Range of dw_conv_params->output_offset : Not used

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [H, W, C_IN] Batch argument N is not used and assumed to be 1.

• input_data – [in] Input (activation) data pointer. Data type: int16

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int64

• output_dims – [in] Output tensor dimensions. Format: [1, H, W, C_OUT]

• output_data – [inout] Output data pointer. Data type: int16

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS - Successful completion.

riscv_nmsis_nn_status riscv_depthwise_conv_wrapper_s4(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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

Wrapper function to pick the right optimized s4 depthwise convolution function.

• Supported framework: TensorFlow Lite

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer ,if applicable, for security reasons.

• dw_conv_params – [in] 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]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [H, W, C_IN] Batch argument N is not used and assumed to be 1.

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• filter_data – [in] Filter data pointer. Data type: int8_t packed 4-bit weights, e.g four sequential weights [0x1, 0x2, 0x3, 0x4] packed as [0x21, 0x43].

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [1, H, W, C_OUT]

• output_data – [inout] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS - Successful completion.

riscv_nmsis_nn_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 int8_t *input, const nmsis_nn_dims *filter_dims, const int8_t *filter, const nmsis_nn_dims *bias_dims, const int32_t *bias, const nmsis_nn_dims *output_dims, int8_t *output)

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

• 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()

• Check details of riscv_depthwise_conv_s8_opt() for potential data that can be accessed outside of the boundary.

Parameters:

• ctx – [inout] 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. The caller is expected to clear the buffer, if applicable, for security reasons.

• dw_conv_params – [in] 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]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each output channel

• input_dims – [in] Input (activation) tensor dimensions. Format: [H, W, C_IN] Batch argument N is not used and assumed to be 1.

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [1, H, W, C_OUT]

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [1, H, W, C_OUT]

• output_data – [inout] Output data pointer. Data type: int8

Returns:

The function returns RISCV_NMSIS_NN_SUCCESS - Successful completion.

riscv_nmsis_nn_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.

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)

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in – [in] input tensor dimension

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel – [in] filter kernel size

• padding – [in] padding sizes

• stride – [in] convolution stride

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out – [in] output tensor dimension

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_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)

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

Parameters:

• Im_in – [in] pointer to input tensor

• dim_im_in_x – [in] input tensor dimension x

• dim_im_in_y – [in] input tensor dimension y

• ch_im_in – [in] number of input tensor channels

• wt – [in] pointer to kernel weights

• ch_im_out – [in] number of filters, i.e., output tensor channels

• dim_kernel_x – [in] filter kernel size x

• dim_kernel_y – [in] filter kernel size y

• padding_x – [in] padding sizes x

• padding_y – [in] padding sizes y

• stride_x – [in] convolution stride x

• stride_y – [in] convolution stride y

• bias – [in] pointer to bias

• bias_shift – [in] amount of left-shift for bias

• out_shift – [in] amount of right-shift for output

• Im_out – [inout] pointer to output tensor

• dim_im_out_x – [in] output tensor dimension x

• dim_im_out_y – [in] output tensor dimension y

• bufferA – [inout] pointer to buffer space for input

• bufferB – [inout] pointer to buffer space for output

Returns:

The function returns either RISCV_NMSIS_NN_SIZE_MISMATCH or RISCV_NMSIS_NN_SUCCESS based on the outcome of size checking.

riscv_nmsis_nn_status riscv_transpose_conv_s8(const nmsis_nn_context *ctx, const nmsis_nn_context *output_ctx, const nmsis_nn_transpose_conv_params *transpose_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

Basic s8 transpose convolution function.

1. Supported framework: TensorFlow Lite micro

2. Additional memory is required for optimization. Refer to arguments ‘ctx’ and ‘output_ctx’ for details.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_transpose_conv_s8_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer, if applicable, for security reasons.

• output_ctx – [inout] Temporary scratch buffer. The size required size is: output width * output height * output channel * 4 The caller is expected to clear the buffer, if applicable, for security reasons.

• transpose_conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of transpose_conv_params->input_offset : [-127, 128] Range of transpose_conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each out channel.

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.

riscv_nmsis_nn_status riscv_transpose_conv_wrapper_s8(const nmsis_nn_context *ctx, const nmsis_nn_context *reverse_conv_ctx, const nmsis_nn_transpose_conv_params *transpose_conv_params, const nmsis_nn_per_channel_quant_params *quant_params, const nmsis_nn_dims *input_dims, const int8_t *input_data, const nmsis_nn_dims *filter_dims, const int8_t *filter_data, const nmsis_nn_dims *bias_dims, const int32_t *bias_data, const nmsis_nn_dims *output_dims, int8_t *output_data)

Wrapper to select optimal transposed convolution algorithm depending on parameters.

1. Supported framework: TensorFlow Lite micro

2. Additional memory is required for optimization. Refer to arguments ‘ctx’ and ‘output_ctx’ for details.

Parameters:

• ctx – [inout] Function context that contains the additional buffer if required by the function. riscv_transpose_conv_s8_get_buffer_size will return the buffer_size if required. The caller is expected to clear the buffer, if applicable, for security reasons.

• output_ctx – [inout] Temporary scratch buffer. The size required size is: output width * output height * output channel * 4 The caller is expected to clear the buffer, if applicable, for security reasons.

• transpose_conv_params – [in] Convolution parameters (e.g. strides, dilations, pads,…). Range of transpose_conv_params->input_offset : [-127, 128] Range of transpose_conv_params->output_offset : [-128, 127]

• quant_params – [in] Per-channel quantization info. It contains the multiplier and shift values to be applied to each out channel.

• input_dims – [in] Input (activation) tensor dimensions. Format: [N, H, W, C_IN]

• input_data – [in] Input (activation) data pointer. Data type: int8

• filter_dims – [in] Filter tensor dimensions. Format: [C_OUT, HK, WK, C_IN] where HK and WK are the spatial filter dimensions

• filter_data – [in] Filter data pointer. Data type: int8

• bias_dims – [in] Bias tensor dimensions. Format: [C_OUT]

• bias_data – [in] Optional bias data pointer. Data type: int32

• output_dims – [in] Output tensor dimensions. Format: [N, H, W, C_OUT]

• output_data – [out] Output data pointer. Data type: int8

Returns:

The function returns either RISCV_NMSIS_NN_ARG_ERROR if argument constraints fail. or, RISCV_NMSIS_NN_SUCCESS on successful completion.