Function arm_fully_connected_per_channel_s8

Function Documentation

arm_cmsis_nn_status arm_fully_connected_per_channel_s8(const cmsis_nn_context *ctx, const cmsis_nn_fc_params *fc_params, const cmsis_nn_per_channel_quant_params *quant_params, const cmsis_nn_dims *input_dims, const int8_t *input_data, const cmsis_nn_dims *filter_dims, const int8_t *filter_data, const cmsis_nn_dims *bias_dims, const int32_t *bias_data, const cmsis_nn_dims *output_dims, int8_t *output_data)

Basic s8 Fully Connected function using per channel quantization.

  • Supported framework: TensorFlow Lite

Parameters:

  • ctx – [in] Per-output-channel kernel sums, supplied by the caller - not scratch memory that this function fills in. The library never populates ctx->buf for this function, and on builds with the MVE extension clearing it makes the output silently wrong, because the sums also carry the bias term. Fill it with arm_vector_sum_s8(), passing filter_dims->n as vector_cols, output_dims->c as vector_rows, filter_data as vector_data, fc_params->input_offset as lhs_offset, fc_params->filter_offset as rhs_offset and the same bias_data given here, so that entry j holds bias_data[j] plus input_offset times the sum of the weights of output channel j, where each of those filter_dims->n weights first has filter_offset added to it. That helper is available on every build and returns ARM_CMSIS_NN_SUCCESS. This function only reads the buffer and never writes it, so it is filled once and may then be reused for as long as filter_data, bias_data, fc_params->input_offset and fc_params->filter_offset are unchanged. Pass a valid context on every build; ctx itself is dereferenced unconditionally. Currently the contents are read only on builds with the MVE extension (ARM_MATH_MVEI), where the bias_data argument is ignored because the sums already carry it and a NULL buf is reported as ARM_CMSIS_NN_ARG_ERROR; an unfilled or cleared buffer there yields wrong output while still returning ARM_CMSIS_NN_SUCCESS. On other builds this function adds bias_data directly and does not read ctx->buf. None of this is a guarantee about future versions. There is no per-channel sizer; arm_fully_connected_s8_get_buffer_size() returns the same quantity this function needs, filter_dims->c * sizeof(int32_t) where the sums are used and 0 otherwise. Do NOT clear this buffer between calls: zeroing it is indistinguishable from leaving it unfilled, and produces the silently wrong output described above. If it must be cleared for security reasons, clear it after the last call that uses it, and refill it before any further call.

  • fc_params – [in] Fully Connected layer parameters. Range of fc_params->input_offset : [-127, 128] fc_params->filter_offset : 0 Range of fc_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 dimension is taken as Nx(H * W * C_IN)

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

  • filter_dims – [in] Two dimensional filter dimensions. Format: [N, C] N : accumulation depth and equals (H * W * C_IN) from input_dims C : output depth and equals C_OUT in output_dims H & W : Not used

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

  • bias_dims – [in] Bias tensor dimensions. Format: [C_OUT] N, H, W : Not used

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

  • output_dims – [in] Output tensor dimensions. Format: [N, C_OUT] N : Batches C_OUT : Output depth H & W : Not used.

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

Returns:

The function returns either ARM_CMSIS_NN_ARG_ERROR if argument constraints fail. or, ARM_CMSIS_NN_SUCCESS on successful completion.