Method run_op

src/gpu/cl/kernels/ClScatterKernel.cpp:212–249

}

void ClScatterKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
    const auto updates =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0));
    const auto indices =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1));
    auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));

    const ITensorInfo *dst_info  = dst->info();
    const ITensorInfo *upd_info  = updates->info();
    const int          num_dims  = dst_info->num_dimensions();
    const int          ind_dims  = indices->info()->num_dimensions();
    const int          index_len = indices->info()->dimension(0);

    bool unsupported_padding_config =
        num_dims == index_len && index_len > 1 && (dst_info->has_padding() || upd_info->has_padding());
    if (unsupported_padding_config)
    {
        ARM_COMPUTE_ERROR("Unsupported Configuration! Padding not supported with these shapes.");
    }

    // calculate m-dimensional data block strides in updates and destination tensors
    const int upt_block_stride =
        updates->info()->strides_in_bytes()[updates->info()->num_dimensions() - (ind_dims - 1)];

    const int out_block_stride = dst_info->strides_in_bytes()[num_dims - index_len];

    unsigned int idx = 0;

    add_2D_tensor_argument(idx, updates, window);
    add_2D_tensor_argument(idx, indices, window);
    add_2D_tensor_argument(idx, dst, window);

    _kernel.setArg<cl_int>(idx++, upt_block_stride);
    _kernel.setArg<cl_int>(idx++, out_block_stride);

    enqueue(queue, *this, window, lws_hint());
}

} // namespace kernels
} // namespace opencl