Method configure_conv_fc

src/gpu/cl/operators/ClFullyConnected.cpp:287–314

}

void ClFullyConnected::configure_conv_fc(const CLCompileContext        &compile_context,
                                         ITensorInfo                   *src,
                                         ITensorInfo                   *weights,
                                         ITensorInfo                   *bias,
                                         ITensorInfo                   *dst,
                                         const FullyConnectedLayerInfo &fc_info)
{
    // MatMul fuses transpose operation, so we use the first dimension for comparison where appropriate.
    ARM_COMPUTE_ERROR_ON((weights->dimension((_use_matmul && _transpose_weights) ? 0 : 1) !=
                          (src->dimension(0) * src->dimension(1) * src->dimension(2))));

    // If the fully connected layer is called after a convolution layer, the input tensor must be linearized

    // Initialize output tensor for flatten
    _flattened_src = src->clone()
                         ->set_is_resizable(true)
                         .reset_padding()
                         .set_tensor_shape(compute_flatten_shape(src))
                         .set_data_layout(DataLayout::NCHW);

    // Configure flatten kernel
    _flatten = std::make_unique<ClFlatten>();
    _flatten->configure(compile_context, src, &_flattened_src);

    // Note: if flatten has > 1 dimensions after, these dimensions are batch
    // Configure matrix multiply kernel
    configure_mm(compile_context, &_flattened_src, weights, bias, dst, fc_info);
}

void ClFullyConnected::configure_fc_fc(const CLCompileContext        &compile_context,
                                       ITensorInfo                   *src,