Function TEST

src/gopt/test/profiler.cpp:63–120

#if CUDA_VERSION >= 10020
TEST(TestProfiler, Conv) {
    REQUIRE_GPU(1);
    auto cn = CompNode::load("gpu0");
    cn.activate();
    REQUIRE_CUDA_COMPUTE_CAPABILITY_EQ(7, 5);
    auto ctx = make_ctx();

    HostTensorGenerator<dtype::Int8> gen;
    auto graph = ComputingGraph::make();
    graph->options().graph_opt_level = 0;
    auto mkvar = [&](const char* name, const TensorShape& shp, const DType& dtype) {
        return opr::TypeCvt::make(
                opr::Host2DeviceCopy::make(*graph, gen(shp, cn)).rename(name), dtype);
    };
    auto mkcvar = [&](const char* name, const TensorShape& shp, const DType& dtype) {
        return opr::TypeCvt::make(
                opr::SharedDeviceTensor::make(*graph, *gen(shp, cn)).rename(name),
                dtype);
    };
    auto x =
            mkvar("x", {64, 48, 14, 14},
                  dtype::Quantized4Asymm(2.5f, static_cast<uint8_t>(4)));
    auto w1 = mkcvar("w1", {48, 48, 3, 3}, dtype::QuantizedS4(2.5f));
    auto b1 = mkcvar("b1", {1, 48, 1, 1}, dtype::QuantizedS32(6.25f));
    opr::ConvBias::Param param;
    param.format = opr::ConvBias::Param::Format::NCHW;
    param.nonlineMode = opr::ConvBias::Param::NonlineMode::IDENTITY;
    param.stride_h = param.stride_w = 1;
    param.pad_h = param.pad_w = 1;
    auto c1 = opr::ConvBias::make(
            x, w1, b1, param, {},
            OperatorNodeConfig(
                    dtype::Quantized4Asymm(12.345f, static_cast<uint8_t>(5))));
    x = opr::TypeCvt::make(c1, dtype::QuantizedS8(12.345f));
    auto w2 = mkcvar("w2", {48, 48, 3, 3}, dtype::QuantizedS8(2.5f));
    auto b2 = mkcvar("b2", {1, 48, 1, 1}, dtype::QuantizedS32(12.345f * 2.5f));
    auto c2 = opr::ConvBias::make(
            x, w2, b2, param, {}, OperatorNodeConfig(dtype::QuantizedS8(2.5f)));

    using S = opr::mixin::AlgoChooserHelper::ExecutionPolicy::Strategy;
    S strategy = S::PROFILE;
    gopt::modify_opr_algo_strategy_inplace({c2}, strategy);
    SubGraphExtractor extractor(ctx->opr_list());
    auto partitions = extractor.extract({c2});
    ASSERT_EQ(partitions.size(), 1u);
    Problem problem(partitions[0], *ctx);
    auto profiler = ProfilerBase::make_profiler();
    auto rst = profiler->profile(problem);
    const auto& opr_rst = rst.opr_record;
    const auto& var_rst = rst.var_record;
    EXPECT_TRUE(opr_rst.count(c1.node()->owner_opr()) > 0);
    EXPECT_TRUE(opr_rst.count(c2.node()->owner_opr()) > 0);
    EXPECT_TRUE(opr_rst.count(x.node()->owner_opr()) > 0);
    EXPECT_TRUE(var_rst.count(w1.node()) == 0);
    EXPECT_TRUE(var_rst.count(b1.node()) == 0);
    EXPECT_TRUE(var_rst.count(w2.node()) == 0);
    EXPECT_TRUE(var_rst.count(b2.node()) == 0);
}