Function Conv

tensorflow/lite/kernels/internal/optimized/optimized_ops.h:1050–1147

}

inline void Conv(const ConvParams& params, const RuntimeShape& input_shape,
                 const float* input_data, const RuntimeShape& filter_shape,
                 const float* filter_data, const RuntimeShape& bias_shape,
                 const float* bias_data, const RuntimeShape& output_shape,
                 float* output_data, const RuntimeShape& im2col_shape,
                 float* im2col_data, CpuBackendContext* cpu_backend_context) {
  const int stride_width = params.stride_width;
  const int stride_height = params.stride_height;
  const int dilation_width_factor = params.dilation_width_factor;
  const int dilation_height_factor = params.dilation_height_factor;
  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);

  (void)im2col_data;
  (void)im2col_shape;
  gemmlowp::ScopedProfilingLabel label("Conv");

  // NB: the float 0.0f value is represented by all zero bytes.
  const uint8 float_zero_byte = 0x00;
  const float* gemm_input_data = nullptr;
  const RuntimeShape* gemm_input_shape = nullptr;
  const int filter_width = filter_shape.Dims(2);
  const int filter_height = filter_shape.Dims(1);
  const bool need_dilated_im2col =
      dilation_width_factor != 1 || dilation_height_factor != 1;
  const bool need_im2col = stride_width != 1 || stride_height != 1 ||
                           filter_width != 1 || filter_height != 1;
  if (need_dilated_im2col) {
    DilatedIm2col(params, float_zero_byte, input_shape, input_data,
                  filter_shape, output_shape, im2col_data);
    gemm_input_data = im2col_data;
    gemm_input_shape = &im2col_shape;
  } else if (need_im2col) {
    TFLITE_DCHECK(im2col_data);
    Im2col(params, filter_height, filter_width, float_zero_byte, input_shape,
           input_data, im2col_shape, im2col_data);
    gemm_input_data = im2col_data;
    gemm_input_shape = &im2col_shape;
  } else {
    // TODO(aselle): We need to make sure to not send im2col if it is not
    // needed.
    TFLITE_DCHECK(!im2col_data);
    gemm_input_data = input_data;
    gemm_input_shape = &input_shape;
  }

  const int gemm_input_dims = gemm_input_shape->DimensionsCount();
  int m = FlatSizeSkipDim(*gemm_input_shape, gemm_input_dims - 1);
  int n = output_shape.Dims(3);
  int k = gemm_input_shape->Dims(gemm_input_dims - 1);

#if defined(TF_LITE_USE_CBLAS) && defined(__APPLE__)
  // The following code computes matrix multiplication c = a * transponse(b)
  // with CBLAS, where:
  // * `a` is a matrix with dimensions (m, k).