Method configure

src/gpu/cl/kernels/ClSoftmaxKernel.cpp:92–184

}

void ClSoftmaxKernel::configure(const CLCompileContext  &compile_context,
                                const ITensorInfo       &src,
                                ITensorInfo             &dst,
                                const SoftmaxKernelInfo &info)
{
    ARM_COMPUTE_UNUSED(compile_context, src, dst, info);

    const auto &dst_shape = dst.tensor_shape();

    const auto data_type    = src.data_type();
    const auto element_size = src.element_size();

    const auto is_quantized = data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED;
    const auto src_qinfo    = src.quantization_info().uniform();
    const auto dst_qinfo    = dst.quantization_info().uniform();

    const auto axis   = wrap_around(info.axis, static_cast<int32_t>(src.num_dimensions()));
    const auto length = dst_shape[axis];

    const auto tmp_data_type = is_quantized ? DataType::F32 : data_type;

    const auto vec_size          = adjust_vec_size(16 / element_size, dst_shape[0]);
    const auto vec_size_leftover = dst_shape[0] % vec_size;

    std::string    kernel_name("softmax");
    CLBuildOptions build_opts;

    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
    build_opts.add_option("-DTMP_DATA_TYPE=" + get_cl_type_from_data_type(tmp_data_type));
    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size));
    build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_leftover));
    build_opts.add_option("-DLENGTH=" + support::cpp11::to_string(length));
    build_opts.add_option_if(info.is_log, "-DIS_LOG");
    build_opts.add_option("-DBETA=" + float_to_string_with_full_precision(info.beta));

    build_opts.add_option_if(is_quantized, "-DIS_QUANTIZED");
    build_opts.add_option_if(is_quantized, "-DSRC_OFFSET=" + float_to_string_with_full_precision(src_qinfo.offset));
    build_opts.add_option_if(is_quantized, "-DSRC_SCALE=" + float_to_string_with_full_precision(src_qinfo.scale));
    build_opts.add_option_if(is_quantized, "-DDST_OFFSET=" + float_to_string_with_full_precision(dst_qinfo.offset));
    build_opts.add_option_if(is_quantized, "-DDST_SCALE=" + float_to_string_with_full_precision(dst_qinfo.scale));

    if (axis == 0)
    {
        kernel_name += "_x";
        build_opts.add_option("-DSOFTMAX_X");

        if (is_quantized)
        {
            _tmp_info = TensorInfo(dst_shape, 1, tmp_data_type);
        }
    }
    else
    {
        kernel_name += "_non_x";
        build_opts.add_option("-DSOFTMAX_NON_X");

        TensorShape tmp_shape;