Function getKernel

src/backend/cuda/jit.cpp:320–347

}

static CUfunction getKernel(const vector<Node*>& output_nodes,
                            const vector<int>& output_ids,
                            const vector<Node*>& full_nodes,
                            const vector<Node_ids>& full_ids,
                            const bool is_linear, const bool loop0,
                            const bool loop1, const bool loop2,
                            const bool loop3) {
    const string funcName{getFuncName(output_nodes, output_ids, full_nodes,
                                      full_ids, is_linear, loop0, loop1, loop2,
                                      loop3)};
    // A forward lookup in module cache helps avoid recompiling
    // the JIT source generated from identical JIT-trees.
    const auto entry{
        findModule(getActiveDeviceId(), deterministicHash(funcName))};

    if (!entry) {
        const string jitKer{getKernelString(funcName, full_nodes, full_ids,
                                            output_ids, is_linear, loop0, loop1,
                                            loop2, loop3)};
        saveKernel(funcName, jitKer, ".cu");

        const common::Source jit_src{jitKer.c_str(), jitKer.size(),
                                     deterministicHash(jitKer)};

        return common::getKernel(funcName, {{jit_src}}, {}, {}, true).get();
    }
    return common::getKernel(entry, funcName, true).get();
}

template<typename T>
void evalNodes(vector<Param<T>>& outputs, const vector<Node*>& output_nodes) {