Method run_op

src/gpu/cl/kernels/ClSoftmaxKernel.cpp:186–256

}

void ClSoftmaxKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
{
    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);

    const auto src =
        utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
    auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
    ICLTensor *tmp = (_tmp_info.total_size() > 0)
                         ? utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_INT_0))
                         : nullptr;

    if (!_prepared)
    {
        _prepared = true;

        const auto *src_info    = src->info();
        const auto *dst_info    = dst->info();
        auto        src_strides = src_info->strides_in_bytes();
        auto        dst_strides = dst_info->strides_in_bytes();

        const auto src_stride_axis = src_strides[_axis];
        const auto dst_stride_axis = dst_strides[_axis];

        // This axis has been removed from execution window, hence we remove it from the list of strides
        // provided to the kernel.
        // In case axis > 0, src/dst_stride_axis will be provided in dedicated argument independent from global ID.
        src_strides.remove(_axis);
        dst_strides.remove(_axis);

        // Argument 0: src_ptr.
        _kernel.setArg<cl_uint>(1, src_strides[0]);
        _kernel.setArg<cl_uint>(2, src_strides[1]);
        _kernel.setArg<cl_uint>(3, src_strides[2]);
        _kernel.setArg<cl_uint>(4, src_info->offset_first_element_in_bytes());

        // Argument 5: dst_ptr.
        _kernel.setArg<cl_uint>(6, dst_strides[0]);
        _kernel.setArg<cl_uint>(7, dst_strides[1]);
        _kernel.setArg<cl_uint>(8, dst_strides[2]);
        _kernel.setArg<cl_uint>(9, dst_info->offset_first_element_in_bytes());

        if (tmp != nullptr)
        {
            const auto *tmp_info    = tmp->info();
            const auto &tmp_strides = tmp_info->strides_in_bytes();

            // Argument 10: tmp_ptr.
            _kernel.setArg<cl_uint>(11, tmp_strides[1]);
            _kernel.setArg<cl_uint>(12, tmp_strides[2]);
            _kernel.setArg<cl_uint>(13, tmp_strides[3]);
            _kernel.setArg<cl_uint>(14, 0);
        }

        if (_axis > 0)
        {
            _kernel.setArg<cl_uint>(15, src_stride_axis);
            _kernel.setArg<cl_uint>(16, dst_stride_axis);