Function TEST

src/core/test/mem_reuse.cpp:64–92

MGB_DYN_TYPE_OBJ_FINAL_IMPL(SharedDeviceTensorDirect);

TEST(TestMemReuse, PureMLP0) {
    auto graph = ComputingGraph::make();
    HostTensorGenerator<> gen;
    CompNode cn = CompNode::load("cpu0");
    //! FIXME currently recursive chooser does not support workspace_limit in
    //! heuristic
    auto host_inp = gen({256, 1, 64, 64}, cn), host_kern0 = gen({32, 1, 1, 1}, cn),
         host_kern1 = gen({32, 32, 1, 1}, cn);
    auto inp = opr::SharedDeviceTensor::make(*graph, *host_inp, {"inp"}),
         kern0 = opr::SharedDeviceTensor::make(*graph, *host_kern0, {"kern0"}),
         kern1 = opr::SharedDeviceTensor::make(*graph, *host_kern1, {"kern1"});
    constexpr size_t NR_LAYER = 7;
    SymbolVar layers[NR_LAYER];
    layers[0] = make_conv(inp, kern0).rename("l0");
    for (size_t i = 1; i < NR_LAYER; i++)
        layers[i] = make_conv(layers[i - 1], kern1).rename(ssprintf("l%zu", i));
    size_t alloc_size = 0;
    auto hdl = graph->event().register_receiver<cg::event::StaticMemAlloc>(
            [&](const cg::event::StaticMemAlloc& s) {
                if (s.comp_node.valid()) {
                    alloc_size = s.alloc_size;
                }
            });

    graph->options().allocate_static_mem_after_graph_compile = true;
    graph->compile({{layers[NR_LAYER - 1], [](DeviceTensorND&) {}}});

    EXPECT_EQ(host_inp->layout().span().dist_byte() * 32 * 2, alloc_size);
}

TEST(TestMemReuse, PureMLP1) {
    auto graph = ComputingGraph::make();