feat(lite): load_and_run support optimize for inference

GitOrigin-RevId: d9abb8de9e

lite/load_and_run/src/models/model.cpp

@@ -36,16 +36,19 @@ std::shared_ptr<ModelBase> ModelBase::create_model(std::string model_path) {
 
     auto model_type = get_model_type(model_path);
 
-    if (ModelType::LITE_MODEL == model_type) {
+    if (FLAGS_lite) {
+        mgb_log("run model force lite mode\n");
+        return std::make_shared<ModelLite>(model_path);
+    } else if (FLAGS_mdl) {
+        mgb_log("run model force mdl mode\n");
+        return std::make_shared<ModelMdl>(model_path);
+    } else if (ModelType::LITE_MODEL == model_type) {
         return std::make_shared<ModelLite>(model_path);
-    } else if (ModelType::MEGDL_MODEL == model_type) {
-        if (FLAGS_lite)
-            return std::make_shared<ModelLite>(model_path);
-        else
-            return std::make_shared<ModelMdl>(model_path);
     } else {
-        return nullptr;
+        mgb_assert(ModelType::MEGDL_MODEL == model_type);
+        return std::make_shared<ModelMdl>(model_path);
     }
 }
 DEFINE_bool(lite, false, "use megengine lite interface to run model");
+DEFINE_bool(mdl, false, "use megengine mdl interface to run model");
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

lite/load_and_run/src/models/model.h

@@ -4,6 +4,7 @@
 #include "helpers/common.h"
 #include "megbrain/utils/json.h"
 DECLARE_bool(lite);
+DECLARE_bool(mdl);
 
 namespace lar {
 /*!

lite/load_and_run/src/models/model_mdl.h

@@ -42,6 +42,8 @@ public:
         return m_load_result;
     }
 
+    void update_mdl_load_result(const mgb::SymbolVarArray& output_var_array);
+
     //! get load config for megDL model
     mgb::serialization::GraphLoadConfig& get_mdl_config() { return m_load_config; }
 

lite/load_and_run/src/options/fastrun_options.cpp

@@ -31,6 +31,13 @@ void FastRunOption::config_model_internel<ModelLite>(
             LITE_LOG("enable fast-run strategy for algo profile");
             strategy = static_cast<uint32_t>(Strategy::LITE_ALGO_PROFILE) |
                        static_cast<uint32_t>(Strategy::LITE_ALGO_OPTIMIZED) | strategy;
+        } else if ((!m_fast_run_cache.empty() &&
+                    !access(m_fast_run_cache.c_str(), F_OK))) {
+            LITE_LOG(
+                    "detect fast-run cache usable set LITE_ALGO_PROFILE for algo "
+                    "profile");
+            strategy = static_cast<uint32_t>(Strategy::LITE_ALGO_PROFILE) |
+                       static_cast<uint32_t>(Strategy::LITE_ALGO_HEURISTIC) | strategy;
         } else {
             strategy = static_cast<uint32_t>(Strategy::LITE_ALGO_HEURISTIC) | strategy;
         }

lite/load_and_run/src/options/optimize_options.cpp

@@ -299,6 +299,75 @@ void FuseConvBiasElemwiseAddOption::config_model(
     CONFIG_MODEL_FUN;
 }
 
+///////////////////////// optimize for inference options ///////////////
+bool OptimizeForInferenceOption::m_valid;
+namespace lar {
+template <>
+void OptimizeForInferenceOption::config_model_internel<ModelLite>(
+        RuntimeParam& runtime_param, std::shared_ptr<ModelLite> model) {
+    LITE_MARK_USED_VAR(model);
+    if (runtime_param.stage == RunStage::BEFORE_MODEL_LOAD) {
+        auto optimize_for_infer =
+                std::static_pointer_cast<lar::Bool>(m_option["optimize_for_inference"])
+                        ->get_value();
+        if (optimize_for_infer) {
+            LITE_THROW(
+                    "optimize for inference not supported in lite "
+                    "model");
+        }
+    }
+}
+
+template <>
+void OptimizeForInferenceOption::config_model_internel<ModelMdl>(
+        RuntimeParam& runtime_param, std::shared_ptr<ModelMdl> model) {
+    if (runtime_param.stage == RunStage::AFTER_MODEL_LOAD) {
+        auto optimize_for_infer =
+                std::static_pointer_cast<lar::Bool>(m_option["optimize_for_inference"])
+                        ->get_value();
+        if (optimize_for_infer) {
+            mgb_log("enable optimize for inference optimization");
+            auto&& load_result = model->get_mdl_load_result();
+            mgb::cg::GraphCommonOptimizeOptions opt =
+                    model->get_mdl_load_result().graph->options().graph_opt;
+            auto inference_opt2 = mgb::gopt::OptimizeForInferenceOptions(opt);
+            auto output_var_list = mgb::gopt::optimize_for_inference(
+                    load_result.output_var_list, inference_opt2);
+            model->get_mdl_load_result().update_output_var_list(output_var_list);
+            model->get_mdl_load_result().graph->options().graph_opt.clear();
+        }
+    }
+}
+}  // namespace lar
+
+void OptimizeForInferenceOption::update() {
+    m_option_name = "optimize_for_inference";
+    m_option = {{"optimize_for_inference", lar::Bool::make(false)}};
+    std::static_pointer_cast<lar::Bool>(m_option["optimize_for_inference"])
+            ->set_value(FLAGS_optimize_for_inference);
+}
+
+bool OptimizeForInferenceOption::is_valid() {
+    bool ret = FLAGS_optimize_for_inference;
+    return ret || m_valid;
+}
+
+std::shared_ptr<OptionBase> OptimizeForInferenceOption::create_option() {
+    static std::shared_ptr<OptimizeForInferenceOption> option(
+            new OptimizeForInferenceOption);
+    if (OptimizeForInferenceOption::is_valid()) {
+        option->update();
+        return option;
+    } else {
+        return nullptr;
+    }
+}
+
+void OptimizeForInferenceOption::config_model(
+        RuntimeParam& runtime_param, std::shared_ptr<ModelBase> model) {
+    CONFIG_MODEL_FUN;
+}
+
 ///////////////////////// graph retrict options /////////////////////////
 bool GraphRecordOption::m_valid;
 namespace lar {
@@ -646,6 +715,9 @@ DEFINE_bool(
         enable_fuse_conv_bias_with_z, false,
         "fuse conv, bias (elemwise add), z(elemwise add) into one opr "
         "(only support on GPU)");
+DEFINE_bool(
+        optimize_for_inference, false,
+        "whether to optimize_for_inference, fuse bn and many base optimize");
 
 ///////////////////////// graph retrict options /////////////////////////
 DEFINE_bool(
@@ -700,6 +772,11 @@ REGIST_OPTION_VALIDATER(
         fuse_conv_bias_nonlinearity, lar::FuseConvBiasNonlinearOption::set_valid);
 
 REGIST_OPTION_CREATOR(
+        optimize_for_inference, lar::OptimizeForInferenceOption::create_option);
+REGIST_OPTION_VALIDATER(
+        optimize_for_inference, lar::OptimizeForInferenceOption::set_valid);
+
+REGIST_OPTION_CREATOR(
         fuse_conv_bias_with_z, lar::FuseConvBiasElemwiseAddOption::create_option);
 REGIST_OPTION_VALIDATER(
         fuse_conv_bias_with_z, lar::FuseConvBiasElemwiseAddOption::set_valid);

lite/load_and_run/src/options/optimize_options.h

@@ -5,6 +5,7 @@
 #include "option_base.h"
 
 DECLARE_bool(enable_fuse_preprocess);
+DECLARE_bool(optimize_for_inference);
 DECLARE_bool(fuse_grain);
 DECLARE_bool(weight_preprocess);
 DECLARE_bool(enable_fuse_conv_bias_nonlinearity);
@@ -216,6 +217,34 @@ private:
     uint64_t workspace_limit;
 };
 
+///////////////////////// optimize for inference options /////////////////////////
+class OptimizeForInferenceOption final : public OptionBase {
+public:
+    static bool is_valid();
+
+    static std::shared_ptr<OptionBase> create_option();
+
+    void config_model(
+            RuntimeParam& runtime_param, std::shared_ptr<ModelBase> model) override;
+
+    static void set_valid(bool val) { m_valid = val; }
+
+    std::string option_name() const override { return m_option_name; };
+
+    OptionValMap* get_option() override { return &m_option; }
+
+    void update() override;
+
+private:
+    OptimizeForInferenceOption() = default;
+    template <typename ModelImpl>
+    void config_model_internel(RuntimeParam&, std::shared_ptr<ModelImpl>){};
+
+    std::string m_option_name;
+    static bool m_valid;
+    OptionValMap m_option;
+};
+
 ///////////////////////// other options for optimization /////////////////
 class JITOption final : public OptionBase {
 public:

lite/src/mge/network_impl.cpp

@@ -366,19 +366,7 @@ void NetworkImplDft::layout_transform_optimization() {
         mgb::ThinHashMap<mgb::SymbolVar, mgb::SymbolVar> out_var_map;
         auto output_var_array = mgb::gopt::layout_transform(
                 m_load_result.output_var_list, m_layout_transform_target);
-        // replace symvar in output_var_list
-        for (size_t idx = 0; idx < output_var_array.size(); ++idx) {
-            out_var_map[m_load_result.output_var_list[idx]] = output_var_array[idx];
-            m_load_result.output_var_list[idx] = output_var_array[idx];
-        }
-        // replace symvar in output_var_map_id
-        for (auto&& item : m_load_result.output_var_map_id) {
-            item.second = out_var_map[item.second];
-        }
-        // replace symvar in output_var_map
-        for (auto&& item : m_load_result.output_var_map) {
-            item.second = out_var_map[item.second];
-        }
+        m_load_result.update_output_var_list(output_var_array);
     } else if (m_user_config->auto_optimize_inference) {
         //! set model weight preprocess
         m_load_config.comp_graph->options().graph_opt.weight_preprocess = true;

lite/test/test_common.h

@@ -8,6 +8,7 @@
 #include "../src/misc.h"
 #include "lite/network.h"
 #include "lite/tensor.h"
+#include "megbrain/comp_node.h"
 #include "megbrain/graph/bases.h"
 #include "megbrain/plugin/opr_io_dump.h"
 #include "megbrain/plugin/profiler.h"
@@ -167,4 +168,18 @@ __attribute__((unused)) static std::shared_ptr<Tensor> mgb_lar(
 
 #endif
 
+static inline bool check_gpu_available(size_t num) {
+    if (mgb::CompNode::get_device_count(mgb::CompNode::DeviceType::CUDA) < num) {
+        mgb_log_warn("skip test case that requires %zu GPU(s)", num);
+        return false;
+    }
+    return true;
+}
+#define REQUIRE_CUDA()                 \
+    {                                  \
+        if (!check_gpu_available(1)) { \
+            return;                    \
+        }                              \
+    }                                  \
+    while (0)
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

lite/test/test_lar_options.cpp

@@ -0,0 +1,51 @@
+#include <gtest/gtest.h>
+#include <string.h>
+#include <memory>
+#include "test_common.h"
+#include "test_options.h"
+
+using namespace lar;
+DECLARE_bool(lite);
+DECLARE_bool(cpu);
+DECLARE_bool(optimize_for_inference);
+#if LITE_WITH_CUDA
+DECLARE_bool(cuda);
+#endif
+
+namespace {
+
+BOOL_OPTION_WRAP(optimize_for_inference);
+
+BOOL_OPTION_WRAP(lite);
+BOOL_OPTION_WRAP(cpu);
+#if LITE_WITH_CUDA
+BOOL_OPTION_WRAP(cuda);
+#endif
+}  // anonymous namespace
+
+TEST(TestLarOption, OPTIMIZE_FOR_INFERENCE) {
+    DEFINE_WRAP(cpu);
+    std::string model_path = "./shufflenet.mge";
+
+    TEST_BOOL_OPTION(optimize_for_inference);
+}
+
+#if LITE_WITH_OPENCL
+TEST(TestLarOption, OPTIMIZE_FOR_INFERENCE_OPENCL) {
+    REQUIRE_OPENCL();
+    DEFINE_WRAP(opencl);
+    std::string model_path = "./shufflenet.mge";
+
+    TEST_BOOL_OPTION(optimize_for_inference);
+}
+#endif
+
+#if LITE_WITH_CUDA
+TEST(TestLarOption, OPTIMIZE_FOR_INFERENCE_CUDA) {
+    REQUIRE_CUDA();
+    DEFINE_WRAP(cuda);
+    std::string model_path = "./shufflenet.mge";
+
+    TEST_BOOL_OPTION(optimize_for_inference);
+}
+#endif

lite/test/test_layout_options.cpp

@@ -1,6 +1,7 @@
 #include <gtest/gtest.h>
 #include <string.h>
 #include <memory>
+#include "test_common.h"
 #include "test_options.h"
 
 using namespace lar;

src/core/include/megbrain/graph/cg.h

@@ -109,6 +109,16 @@ struct GraphCommonOptimizeOptions {
                      ///< support on Nvidia GPU
     };
     LayoutTransform layout_transform = LayoutTransform::DEFAULT;
+    void clear() {
+        f16_io_f32_comp = false;
+        f16_io_comp = false;
+        fuse_conv_bias_nonlinearity = false;
+        fuse_conv_bias_with_z = false;
+        weight_preprocess = false;
+        fuse_preprocess = false;
+        fuse_grain = false;
+        layout_transform = LayoutTransform::DEFAULT;
+    }
 
 #define SET(n)                                  \
     GraphCommonOptimizeOptions& enable_##n() {  \

src/gopt/include/megbrain/gopt/inference.h

@@ -312,6 +312,9 @@ public:
 };
 
 struct OptimizeForInferenceOptions : cg::GraphCommonOptimizeOptions {
+    OptimizeForInferenceOptions() = default;
+    OptimizeForInferenceOptions(const cg::GraphCommonOptimizeOptions& opt)
+            : cg::GraphCommonOptimizeOptions(opt){};
     uint64_t serialize() {
         uint64_t ret = 0;
         ret |= (uint64_t)layout_transform << 32;

src/serialization/impl/serializer.cpp

@@ -17,6 +17,25 @@ std::unique_ptr<cg::AsyncExecutable> GraphLoader::LoadResult::graph_compile(
     return ret;
 }
 
+void GraphLoader::LoadResult::update_output_var_list(
+        const SymbolVarArray& output_var_array) {
+    mgb::ThinHashMap<mgb::SymbolVar, mgb::SymbolVar> out_var_map;
+    mgb_assert(output_var_array.size() == output_var_list.size());
+    // replace symvar in output_var_list
+    for (size_t idx = 0; idx < output_var_array.size(); ++idx) {
+        out_var_map[output_var_list[idx]] = output_var_array[idx];
+        output_var_list[idx] = output_var_array[idx];
+    }
+    // replace symvar in output_var_map_id
+    for (auto&& item : output_var_map_id) {
+        item.second = out_var_map[item.second];
+    }
+    // replace symvar in output_var_map
+    for (auto&& item : output_var_map) {
+        item.second = out_var_map[item.second].rename(item.first);
+    }
+}
+
 void GraphLoader::LoadResult::graph_compile_ahead() {
     //! when force_output_use_user_specified_memory is set, the output var may
     //! be changed by gopt, then the var in LoadResult can not exist, so here

src/serialization/include/megbrain/serialization/serializer.h

@@ -45,6 +45,13 @@ public:
         //! GraphDumper::dump
         SymbolVarArray output_var_list;
 
+        /**
+         * \brief update  output_var_list with output_var_map, output_var_map_id
+         *
+         */
+        MGE_WIN_DECLSPEC_FUC void update_output_var_list(
+                const SymbolVarArray& output_var_array);
+
         /*!
          * \brief call graph->compile() but also checks for comp seq rec
          *