Merge pull request #32 from MegEngine/update-internal

Update internal

README.md

@@ -33,6 +33,7 @@ MegCC supports Arm64/ArmV7/X86/BareMatal backend. You may want to check [support
 * Download release compiler suit from [release page](https://github.com/MegEngine/MegCC/releases)
 * Compiler from source, please fellow the [compiler doc](compiler/README.md)
 * Build the release tar, please fellow the [release doc](doc/how-to-release.md)
+* Get benchmark of different model please reference [benchmark](benchmark/README.md)
 
 #### How to use MegCC
 

benchmark/.gitignore

@@ -0,0 +1,4 @@
+model/benchmark_*
+model/generated_models
+config
+output

benchmark/CMakeLists.txt

@@ -0,0 +1,97 @@
+cmake_minimum_required(VERSION 3.15.2)
+set(CMAKE_EXPORT_COMPILE_COMMANDS
+    ON
+    CACHE INTERNAL "")
+
+project(Benchmarker)
+
+option(ENABLE_MEGENGINE_FRAMEWORK "build benchmark for megengine" OFF)
+configure_file(src/build_config.h.in
+               ${CMAKE_CURRENT_BINARY_DIR}/genfiles/build_config.h)
+# set megcc lib 
+if(NOT DEFINED RUNTIME_KERNEL_DIR)
+  message(FATAL_ERROR "build MegCC runtime kernel dir RUNTIME_KERNEL_DIR is empty, use -DRUNTIME_KERNEL_DIR=your_model_kernel_dir to set")
+else()
+  message(STATUS "build MegCC runtime with kernel dir ${RUNTIME_KERNEL_DIR}")
+endif()
+
+add_library(TinyNN STATIC IMPORTED)
+set_target_properties(
+  TinyNN PROPERTIES IMPORTED_LOCATION
+                    "${RUNTIME_KERNEL_DIR}/runtime/install/lib/libTinyNN.a")
+if(ENABLE_MEGENGINE_FRAMEWORK)
+  message(STATUS "build benchmark with megengine ${ENABLE_MEGENGINE_FRAMEWORK}")
+  option(X86_BACKEND "Build bechmarker with X86 megengine lib" ON)
+  # set megengine lib
+  if(NOT DEFINED MEGENGINE_INSTALL_DIR)
+    message(FATAL_ERROR "MEGENGINE_INSTALL_DIR is empty use -DMEGENGINE_INSTALL_DIR=your_megengine_install_dir to set")
+  else()
+    message(STATUS "MEGENGINE_INSTALL_DIR is ${MEGENGINE_INSTALL_DIR}")
+  endif()
+  add_library(mgb_imported INTERFACE IMPORTED)
+
+  if(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lite/lib/aarch64/liblite_static_all_in_one.a)
+    target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS})
+  elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "armv7-a")
+    set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lite/lib/armv7/liblite_static_all_in_one.a)
+    target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS})
+  elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL "riscv64")
+    set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lite/lib/riscv64/liblite_static_all_in_one.a)
+    target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS})
+  else()
+    if(X86_BACKEND)
+      if(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "x86_64" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "AMD64")
+        set(MKL_LIBS
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/x86_64/lib/libmkl_core.a
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/x86_64/lib/libmkl_sequential.a
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/x86_64/lib/libmkl_intel_ilp64.a
+        )
+        set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lite/lib/x86_64/liblite_static_all_in_one.a)
+        target_compile_definitions(mgb_imported INTERFACE -DMKL_ILP64)
+      # WARNING: i386 is not test locally 
+      elseif(${CMAKE_SYSTEM_PROCESSOR} STREQUAL "i386" OR ${CMAKE_SYSTEM_PROCESSOR} STREQUAL "i686")
+        set(MKL_LIBS
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/i386/lib/libmkl_core.a
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/x86_64/lib/libmkl_sequential.a
+          ${PROJECT_SOURCE_DIR}/../third_party/MegEngine/third_party/mkl/x86_64/lib/libmkl_intel_32.a
+        )
+        set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lite/lib/i386/liblite_static_all_in_one.a)
+      endif()
+      set(MKL_DNN_LIBS
+        ${MEGENGINE_INSTALL_DIR}/lib/libdnnl.a
+        ${MEGENGINE_INSTALL_DIR}/lib/libmkldnn.a
+      )
+
+      if(UNIX AND NOT APPLE)
+        target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS} ${MKL_DNN_LIBS} -Wl,--start-group -ldl ${MKL_LIBS} -Wl,--end-group) 
+      else()
+        target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS} ${MKL_DNN_LIBS} ${MKL_LIBS}) 
+      endif()
+    else()
+      set(MGE_INSTALL_LIBS ${MEGENGINE_INSTALL_DIR}/lib/libmegengine.a ${MEGENGINE_INSTALL_DIR}/lib/libflatbuffers.a)
+      target_link_libraries(mgb_imported INTERFACE  ${MGE_INSTALL_LIBS})  
+    endif()
+
+  endif()
+
+  target_include_directories(mgb_imported INTERFACE ${MEGENGINE_INSTALL_DIR}/include)
+endif()
+# benchmarker config
+file(GLOB_RECURSE SOURCES main.cpp src/*.cpp src/*.h)
+add_executable(benchmarker ${SOURCES})
+target_include_directories(
+    benchmarker PUBLIC $<BUILD_INTERFACE:${RUNTIME_KERNEL_DIR}/runtime/install/include> $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/genfiles>)
+if(ENABLE_MEGENGINE_FRAMEWORK)
+  target_link_libraries(benchmarker -pthread TinyNN mgb_imported)
+else()
+target_link_libraries(benchmarker -pthread TinyNN)
+endif()
+message(STATUS "${CMAKE_TOOLCHAIN_FILE}")
+if(CMAKE_TOOLCHAIN_FILE)
+  if(ANDROID)
+    target_link_libraries(benchmarker log)
+  endif()
+endif()
+
+install(TARGETS benchmarker LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX})

benchmark/README.md

@@ -0,0 +1,80 @@
+# How to use megcc benchmark
+
+## introduction
+megcc benchmark is a easy tool to get the benchmark result of different model in megcc
+the file struction  is shown bellow:
+```
+├── clean.sh
+├── CMakeLists.txt
+├── main.cpp
+├── model
+│   ├── model_arm.json
+│   ├── model_riscv.json
+│   ├── model_x86.json
+│   └── request.txt
+├── python
+│   ├── example.py
+│   ├── format.sh
+│   └── src
+│       ├── benchmark.py
+│       └── models.py
+├── README.md
+├── src
+│   ├── benchmark.h
+│   ├── build_config.h.in
+│   ├── CCbenchmark.cpp
+│   ├── CCbenchmark.h
+│   ├── MGEbenchmark.cpp
+│   └── MGEbenchmark.h
+└── tools
+    ├── cc_analysis.py
+    └── inference_visual.py
+```
+
+in src, it is a c++ application to run benchmark result on different platform.
+in python, the model convertion, other related preparing work and the benchmarker example is given
+the tools contains some usable scripts to analysis benchmark results
+## supported model
+mobilenetv2,  resnet18, efficientnetb0 shufflenetv2 vgg16
+## request
+```bash
+mgeconvert > v.1.0.2 
+onnx==1.11.0
+torch==1.10.0
+# or 
+git clone https://github.com/MegEngine/mgeconvert.git
+cd mgeconvert 
+git checkout master
+python3 -m pip install . --user --install-option="--targets=onnx"
+
+```
+the mgeconvert can be install by following command:
+```bash
+git clone https://github.com/MegEngine/mgeconvert.git
+cd mgeconvert 
+git checkout master
+python3 -m pip install . --user --install-option="--targets=onnx"
+
+```
+## get model and run benchmark example 
+``` bash
+cd megcc/benchmark
+export MEGCC_MGB_TO_TINYNN_PATH=<your_mgb_to_tinynn_path>
+python3  python/example.py
+```
+if you want to run in other platform, please reference the example add your new run_platform_xxx function in BenchmarkRunner, 
+the example given a ssh remote device test template
+
+## analysis megcc log
+
+the `output` directory is generated by `example.py`
+
+### visualize the inference result of different model
+```bash
+python3 benchmark/tools/inference_visual.py benchmark/output -o figure_dir
+```
+
+### visualize the profile result of different kernel in different model
+```bash
+python3 benchmark/tools/cc_analysis.py benchmark/output -o figure_dir
+```

benchmark/clean.sh

@@ -0,0 +1,3 @@
+# /bin/bash -e
+set -x
+rm -rf ./build* ./output ./config ./model/benchmark* ./model/generate*

benchmark/main.cpp

@@ -0,0 +1,64 @@
+/**
+ * \file benchmark/main.cpp
+ *
+ * This file is part of MegCC, a deep learning compiler developed by Megvii.
+ *
+ * \copyright Copyright (c) 2021-2022 Megvii Inc. All rights reserved.
+ */
+#include <cmath>
+#include <cstdio>
+#include <memory>
+#include "src/CCbenchmark.h"
+#include "src/MGEbenchmark.h"
+
+using namespace megcc;
+using namespace Benchmark;
+int main(int argc, char** argv) {
+    if (argc < 2 && argc > 4) {
+        fprintf(stderr, "cmdline error, please run with:\n");
+        fprintf(stderr, "benchmarker <input_model> [options] ... \n");
+        fprintf(stderr,
+                "tips:\n\t you can use --profile and --mge to profile model "
+                "and enable megengine framework (\"megcc\" is default)\n");
+        return -1;
+    }
+    int log_level = 3;
+    std::string framework = "megcc";
+    std::string model_path = argv[1];
+    int idx = 2;
+    while (idx < argc) {
+        std::string args = argv[idx];
+        if (args == "--profile") {
+            log_level = 0;
+        } else if (args == "--mge") {
+            framework = "mge";
+        } else {
+            fprintf(stderr, "invalid option: %s\n", argv[idx]);
+        }
+        ++idx;
+    }
+    std::vector<std::shared_ptr<Benchmarker>> benchmarkers;
+    if (framework == "megcc") {
+        benchmarkers.push_back(
+                std::make_shared<CCBenchmarker>(model_path, log_level));
+    }
+#if ENABLE_MEGENGINE_FRAMEWORK
+    else if (framework == "mge") {
+        benchmarkers.push_back(
+                std::make_shared<MGEBenchmarker>(model_path, log_level));
+    }
+#endif
+    else {
+        fprintf(stderr,
+                "unsupport framework: %s, megcc, mge(export "
+                "ENABLE_MEGENGINE_FRAMEWORK=ON) is supported\n",
+                framework.c_str());
+    }
+
+    for (size_t i = 0; i < benchmarkers.size(); ++i) {
+        benchmarkers[i]->load_model();
+        benchmarkers[i]->profile();
+    }
+
+    return 0;
+}

benchmark/model/model_arm.json

@@ -0,0 +1,47 @@
+{
+    "dump_dir": "./benchmark_kernel_arm/",
+    "models": [
+        {
+            "model_name": "mobilenetv2",
+            "model_path": "./generated_models/mobilenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "resnet18",
+            "model_path": "./generated_models/resnet18.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "resnet50",
+            "model_path": "./generated_models/resnet50.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "efficientnetb0",
+            "model_path": "./generated_models/efficientnetb0.mge",
+            "input_shape_str": "data=(1,3,256,256)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "shufflenetv2",
+            "model_path": "./generated_models/shufflenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "vgg11",
+            "model_path": "./generated_models/vgg11.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        },
+        {
+            "model_name": "vgg16",
+            "model_path": "./generated_models/vgg16.mge",
+            "input_shape_str": "data=(1,3,224,224)",
+            "enable_nchw44": true
+        }
+    ]
+}

benchmark/model/model_riscv.json

@@ -0,0 +1,40 @@
+{
+    "dump_dir": "./benchmark_kernel_riscv/",
+    "models": [
+        {
+            "model_name": "mobilenetv2",
+            "model_path": "./generated_models/mobilenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "resnet18",
+            "model_path": "./generated_models/resnet18.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "resnet50",
+            "model_path": "./generated_models/resnet50.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "efficientnetb0",
+            "model_path": "./generated_models/efficientnetb0.mge",
+            "input_shape_str": "data=(1,3,256,256)"
+        },
+        {
+            "model_name": "shufflenetv2",
+            "model_path": "./generated_models/shufflenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "vgg11",
+            "model_path": "./generated_models/vgg11.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "vgg16",
+            "model_path": "./generated_models/vgg16.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        }
+    ]
+}

benchmark/model/model_x86.json

@@ -0,0 +1,40 @@
+{
+    "dump_dir": "./benchmark_kernel_x86/",
+    "models": [
+        {
+            "model_name": "mobilenetv2",
+            "model_path": "./generated_models/mobilenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "resnet18",
+            "model_path": "./generated_models/resnet18.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "resnet50",
+            "model_path": "./generated_models/resnet50.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "efficientnetb0",
+            "model_path": "./generated_models/efficientnetb0.mge",
+            "input_shape_str": "data=(1,3,256,256)"
+        },
+        {
+            "model_name": "shufflenetv2",
+            "model_path": "./generated_models/shufflenetv2.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "vgg11",
+            "model_path": "./generated_models/vgg11.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        },
+        {
+            "model_name": "vgg16",
+            "model_path": "./generated_models/vgg16.mge",
+            "input_shape_str": "data=(1,3,224,224)"
+        }
+    ]
+}