Function HybridConv

tensorflow/lite/kernels/internal/optimized/optimized_ops.h:1149–1227

}

inline void HybridConv(const ConvParams& params, float* scaling_factors_ptr,
                       const RuntimeShape& input_shape,
                       const int8_t* input_data,
                       const RuntimeShape& filter_shape,
                       const int8_t* filter_data,
                       const RuntimeShape& bias_shape, const float* bias_data,
                       const RuntimeShape& output_shape, float* output_data,
                       const RuntimeShape& im2col_shape, int8_t* im2col_data) {
  const int stride_width = params.stride_width;
  const int stride_height = params.stride_height;
  const float output_activation_min = params.float_activation_min;
  const float output_activation_max = params.float_activation_max;
  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);

  const int batch_size = input_shape.Dims(0);
  const int filter_width = filter_shape.Dims(2);
  const int filter_height = filter_shape.Dims(1);

  const int8_t* gemm_input_data = nullptr;
  int num_input;
  const bool need_im2col = stride_width != 1 || stride_height != 1 ||
                           filter_width != 1 || filter_height != 1;

  if (need_im2col) {
    TFLITE_DCHECK(im2col_data);
    // symmetric quantization assumes zero point of 0.
    const int input_zero_point = 0;

    Im2col(params, filter_height, filter_width, input_zero_point, input_shape,
           input_data, im2col_shape, im2col_data);
    gemm_input_data = im2col_data;
    num_input = im2col_shape.FlatSize();
  } else {
    TFLITE_DCHECK(!im2col_data);
    gemm_input_data = input_data;
    num_input = input_shape.FlatSize();
  }

  // Flatten 4D matrices into 2D matrices for matrix multiplication.

  // Flatten so that each filter has its own row.
  const int filter_rows = filter_shape.Dims(0);
  const int filter_cols = FlatSizeSkipDim(filter_shape, 0);

  // In MatrixBatchVectorMultiplyAccumulate, each output value is the
  // dot product of one row of the first matrix with one row of the second
  // matrix. Therefore, the number of cols in each matrix are equivalent.
  //
  // After Im2Col, each input patch becomes a row.
  const int gemm_input_cols = filter_cols;
  const int gemm_input_rows = num_input / gemm_input_cols;

  const int output_cols = output_shape.Dims(3);
  const int output_rows = FlatSizeSkipDim(output_shape, 3);
  TFLITE_DCHECK_EQ(output_cols, filter_rows);
  TFLITE_DCHECK_EQ(output_rows, gemm_input_rows);