Method dispatch_kern

dnn/src/fallback/convolution/algos.cpp:195–223

}

SmallVector<ConvolutionImpl::NCBKern> ConvolutionImpl::AlgoFallback::dispatch_kern(
        const NCBKernSizeParam& param) const {
    MIDOUT_BEGIN(megdnn_fallback_conv, midout_iv("AlgoFallback::dispatch_kern"_hash)) {
        size_t group = param.filter_meta.group;
        size_t N = param.n;
        size_t nr_threads = param.nr_threads;
        size_t workspace_per_thread = get_workspace(param) / nr_threads;
        auto kern_fallback = [workspace_per_thread](
                                     const NCBKernParam& p,
                                     const NCBKernIndex& ncb_index) {
            UNPACK_CONV_F32_NCB_KERN_SIZES(p);
            size_t batch_id = ncb_index.ndrange_id[1];
            size_t group_id = ncb_index.ndrange_id[0];
            MEGDNN_MARK_USED_VAR(N);
            auto src = p.src<float>(batch_id, group_id),
                 filter = p.filter<float>(group_id);
            auto dst = p.dst<float>(batch_id, group_id);
            size_t thread_id = ncb_index.thread_id;
            void* workspace_ptr = reinterpret_cast<void*>(
                    reinterpret_cast<ptrdiff_t>(p.workspace_ptr) +
                    workspace_per_thread * thread_id);
            convolution::run_conv(
                    src, filter, dst, workspace_ptr, IH, IW, IC, FH, FW, OH, OW, OC, PH,
                    PW, SH, SW, !p.filter_meta.should_flip);
        };
        return {{kern_fallback, {group, N, 1_z}}};
    }
    MIDOUT_END();
}

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