Method dispatch_kerns

dnn/src/fallback/conv_bias/algos.cpp:202–231

}

SmallVector<ConvBiasImpl::NCBKern> ConvBiasImpl::AlgoNaive::dispatch_kerns(
        const NCBKernSizeParam& p) const {
    size_t workspace_size = get_workspace(p);
    //! When group>1 or n>1, this algo will parallel by group and n
    size_t nr_threads = p.nr_threads;
    size_t GROUP = p.filter_meta.group;
    size_t N = p.n;
    size_t workspace_per_thread = workspace_size / nr_threads;
    auto kern = [workspace_per_thread](
                        const NCBKernParam& param, const NCBKernIndex& ncb_index) {
        MIDOUT_BEGIN(megdnn_fallback_naive, 2) {
            size_t group_id = ncb_index.ndrange_id[0];
            size_t batch_id = ncb_index.ndrange_id[1];
            size_t thread_id = ncb_index.thread_id;
            auto thread_param = param;
            thread_param.workspace_ptr = reinterpret_cast<void*>(
                    reinterpret_cast<ptrdiff_t>(param.workspace_ptr) +
                    thread_id * workspace_per_thread);

            thread_param.filter_ptr += param.filter_offset(group_id);
            thread_param.dst_ptr += param.dst_offset(batch_id, group_id);
            thread_param.src_ptr += param.src_offset(batch_id, group_id);
            thread_param.bias_ptr += param.bias_offset(batch_id, group_id);

            kern_default(thread_param);
        }
        MIDOUT_END();
    };
    return {{kern, {GROUP, N, 1_z}}};
}

MIDOUT_DECL(megdnn_fallback_winograd)
/* ======================= AlgoWinogradF32 ======================== */