Method configure

src/gpu/cl/kernels/ClMatMulLowpNativeKernel.cpp:137–238

    return Status{};
}
void ClMatMulLowpNativeKernel::configure(const ClCompileContext    &compile_context,
                                         ITensorInfo               *lhs,
                                         ITensorInfo               *rhs,
                                         ITensorInfo               *bias,
                                         ITensorInfo               *dst,
                                         const MatMulKernelInfo    &matmul_kernel_info,
                                         const ActivationLayerInfo &act_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs, dst, &compile_context, &matmul_kernel_info);
    ARM_COMPUTE_LOG_PARAMS(lhs, rhs, bias, dst, matmul_kernel_info);
    ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, bias, dst, matmul_kernel_info));

    // dst tensor auto initialization if not yet initialized
    auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(
                                 lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)));

    const int  m       = dst->dimension(1);
    const int  n       = dst->dimension(0);
    const int  k       = matmul_kernel_info.adj_lhs ? lhs->tensor_shape().y() : lhs->tensor_shape().x();
    const bool adj_lhs = matmul_kernel_info.adj_lhs;

    int m0 = adj_lhs ? adjust_vec_size(matmul_kernel_info.m0, m) : std::min(matmul_kernel_info.m0, m);
    int n0 = adjust_vec_size(matmul_kernel_info.n0, n);

    // Configure kernel window
    Window win = calculate_max_window(*dst, Steps(n0, m0));
    win        = win.collapse(win, Window::DimZ);
    IClKernel::configure_internal(win);

    // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
    const unsigned int partial_store_m0 = m % m0;
    const unsigned int partial_store_n0 = n % n0;

    CLBuildOptions build_opts;
    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(lhs->data_type()));
    build_opts.add_option("-DM0=" + support::cpp11::to_string(m0));
    build_opts.add_option("-DN0=" + support::cpp11::to_string(n0));
    build_opts.add_option("-DK0=" + support::cpp11::to_string(matmul_kernel_info.k0));
    build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
    build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
    build_opts.add_option("-DK=" + support::cpp11::to_string(k));

    const UniformQuantizationInfo lqinfo = lhs->quantization_info().uniform();
    const UniformQuantizationInfo rqinfo = rhs->quantization_info().uniform();
    const UniformQuantizationInfo dqinfo = dst->quantization_info().uniform();

    float multiplier        = lqinfo.scale * rqinfo.scale / dqinfo.scale;
    int   output_multiplier = 0;
    int   output_shift      = 0;
    arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);

    build_opts.add_option("-DDST_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
    build_opts.add_option("-DDST_SHIFT=" + support::cpp11::to_string(output_shift));

    // Note : Offset is not negated, unlike gemmlowp kernels
    build_opts.add_option("-DLHS_OFFSET=" + support::cpp11::to_string(lqinfo.offset));
    build_opts.add_option("-DRHS_OFFSET=" + support::cpp11::to_string(rqinfo.offset));
    build_opts.add_option("-DDST_OFFSET=" + support::cpp11::to_string(dqinfo.offset));