Method get_kimpl

dnn/src/fallback/convolution/algos.cpp:399–427

Return the implement kernel

//! Return the implement kernel
SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoDefault::get_kimpl(
        ConvBiasImpl::AlgoBase* algo, const NCBKernSizeParam& param) {
    MIDOUT_BEGIN(megdnn_fallback_conv, midout_iv(0)) {
        // construct the conv_bias kern param
        ::ConvBiasImpl::NCBKernParam conv_bias_param;
        static_cast<::ConvBiasImpl::NCBKernSizeParam&>(conv_bias_param) =
                init_conv_bias_param(param);
        auto&& conv_bias_kerns = algo->dispatch_kerns(conv_bias_param);
        SmallVector<ConvolutionImpl::NCBKern> convolution_kerns;

        for (size_t i = 0; i < conv_bias_kerns.size(); i++) {
            auto&& kernel = conv_bias_kerns[i];
            //! If the kerenl batch parallel
            auto run = [conv_bias_param, kernel](
                               const NCBKernParam& p, const NCBKernIndex& ncb_index) {
                auto param = conv_bias_param;
                param.src_ptr = p.src_ptr;
                param.filter_ptr = p.filter_ptr;
                param.dst_ptr = p.dst_ptr;
                param.workspace_ptr = p.workspace_ptr;
                param.workspace_size = p.workspace_size;
                kernel.kern(param, {ncb_index.thread_id, ncb_index.ndrange_id});
            };
            convolution_kerns.push_back({run, kernel.global_size});
        }
        return convolution_kerns;
    }
    MIDOUT_END();
}

/////////////////////////// ConvolutionBackwardData /////////////////////
/* ===================== naive algo ===================== */