Method prepare

src/cpu/operators/CpuDepthwiseConv2d.cpp:227–287

}

void CpuDepthwiseConv2d::CpuDepthwiseConv2dOptimizedInternal::prepare(ITensorPack &tensors)
{
    // if weights are not constant then we need to repack so that weights
    // can be updated in-place
    if (!_are_weights_const)
    {
        auto weights        = tensors.get_const_tensor(TensorType::ACL_SRC_1);
        auto bias           = tensors.get_const_tensor(TensorType::ACL_SRC_2);
        auto packed_weights = tensors.get_tensor(TensorType::ACL_INT_4);

        ITensorPack pack_opt;
        pack_opt.add_tensor(TensorType::ACL_SRC_1, weights);
        pack_opt.add_tensor(TensorType::ACL_SRC_2, bias);
        pack_opt.add_tensor(TensorType::ACL_INT_1, packed_weights);

        // Prepare optimized function
        _dwc_optimized_func->prepare(pack_opt);

        return;
    }

    if (!_is_prepared)
    {
        auto weights        = tensors.get_const_tensor(TensorType::ACL_SRC_1);
        auto bias           = tensors.get_const_tensor(TensorType::ACL_SRC_2);
        auto packed_weights = tensors.get_tensor(TensorType::ACL_INT_4);

        // Permute weights
        if (_permute)
        {
            auto permuted_weights = tensors.get_tensor(TensorType::ACL_INT_1);

            ITensorPack pack;
            pack.add_tensor(TensorType::ACL_SRC, weights);
            pack.add_tensor(TensorType::ACL_DST, permuted_weights);
            _permute_weights->run(pack);

            weights->mark_as_unused();

            ITensorPack pack_opt;
            pack_opt.add_const_tensor(TensorType::ACL_SRC_1, permuted_weights);
            pack_opt.add_tensor(TensorType::ACL_SRC_2, bias);
            pack_opt.add_tensor(TensorType::ACL_INT_1, packed_weights);

            // Prepare optimized function
            _dwc_optimized_func->prepare(pack_opt);
        }
        else
        {
            ITensorPack pack_opt;
            pack_opt.add_tensor(TensorType::ACL_SRC_1, weights);
            pack_opt.add_tensor(TensorType::ACL_SRC_2, bias);
            pack_opt.add_tensor(TensorType::ACL_INT_1, packed_weights);

            // Prepare optimized function
            _dwc_optimized_func->prepare(pack_opt);
        }