Method exec

dnn/src/cuda/convolution3d/forward/group_conv.cpp:91–124

}

void Convolution3DForwardImpl::AlgoGroupConvGeneral::exec(const ExecArgs& args) const {
    auto bundle = get_workspace_bundle(args.workspace.raw_ptr, args);
    {
        auto config = prepare_sub_opr(args);
        TensorND tsrc{args.src_tensor->raw_ptr(), config.first[0]};
        TensorND tfilter{args.filter_tensor->raw_ptr(), config.first[1]};
        TensorND tdst{args.dst_tensor->raw_ptr(), config.first[2]};

        size_t c_pos;
        if (args.filter_meta.format == Param::Format::NCDHW) {
            c_pos = 1;
        } else {
            megdnn_assert(
                    args.filter_meta.format == Param::Format::NDHWC,
                    "invalid conv format");
            c_pos = 4;
        }

        auto grp = args.filter_meta.group;

        auto&& fm = args.filter_meta;
        auto strd_src = tsrc.layout.stride[c_pos] * fm.icpg * tsrc.layout.dtype.size(),
             strd_dst = tdst.layout.stride[c_pos] * fm.ocpg * tdst.layout.dtype.size(),
             strd_flt = fm.icpg * fm.ocpg * fm.spatial[0] * fm.spatial[1] *
                        fm.spatial[2] * tfilter.layout.dtype.size();

        for (uint32_t g = 0; g < grp; ++g) {
            config.second->exec(tsrc, tfilter, tdst, bundle.get_workspace(0));
            incr_refp(tsrc.get_ref_ptr(), strd_src);
            incr_refp(tdst.get_ref_ptr(), strd_dst);
            incr_refp(tfilter.get_ref_ptr(), strd_flt);
        }
    }
}

// vim: syntax=cpp.doxygen