Synchronize internal github for version 1.2.0 (#1518)

3rd_party/flatbuffers/CMakeLists.txt

@@ -154,7 +154,7 @@ if(EXISTS "${CMAKE_TOOLCHAIN_FILE}")
   # is being configured externally
 elseif(APPLE)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -stdlib=libc++")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -pedantic -Werror -Wextra -Wno-unused-parameter")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -pedantic -Wextra -Wno-unused-parameter")
   set(FLATBUFFERS_PRIVATE_CXX_FLAGS "-Wold-style-cast")
 elseif(CMAKE_COMPILER_IS_GNUCXX)
   if(CYGWIN)
@@ -165,15 +165,15 @@ elseif(CMAKE_COMPILER_IS_GNUCXX)
       "${CMAKE_CXX_FLAGS} -std=c++0x")
   endif(CYGWIN)
   set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -Wall -pedantic -Werror -Wextra -Werror=shadow")
+    "${CMAKE_CXX_FLAGS} -Wall -pedantic -Wextra")
   set(FLATBUFFERS_PRIVATE_CXX_FLAGS "-Wold-style-cast")
   if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.4)
     if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
       set(CMAKE_CXX_FLAGS
-        "${CMAKE_CXX_FLAGS} -faligned-new -Werror=implicit-fallthrough=2")
+        "${CMAKE_CXX_FLAGS} -faligned-new")
     endif()
     set(CMAKE_CXX_FLAGS
-      "${CMAKE_CXX_FLAGS} -Wunused-result -Werror=unused-result -Wunused-parameter -Werror=unused-parameter")
+      "${CMAKE_CXX_FLAGS} -Wunused-result -Wunused-parameter")
   endif()
 
   # Certain platforms such as ARM do not use signed chars by default

3rd_party/flatbuffers/src/idl_gen_cpp.cpp

@@ -1422,9 +1422,8 @@ class CppGenerator : public BaseGenerator {
   }
 
   std::string GenFieldOffsetName(const FieldDef &field) {
-    std::string uname = Name(field);
-    std::transform(uname.begin(), uname.end(), uname.begin(), ToUpper);
-    return "VT_" + uname;
+    std::string uname = NumToString(field.value.offset);
+    return uname;
   }
 
   void GenFullyQualifiedNameGetter(const StructDef &struct_def,
@@ -1750,32 +1749,8 @@ class CppGenerator : public BaseGenerator {
       code_ += "  }";
     }
 
-
     GenFullyQualifiedNameGetter(struct_def, Name(struct_def));
 
-    // Generate field id constants.
-    if (struct_def.fields.vec.size() > 0) {
-      // We need to add a trailing comma to all elements except the last one as
-      // older versions of gcc complain about this.
-      code_.SetValue("SEP", "");
-      code_ += "  enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE {";
-      for (auto it = struct_def.fields.vec.begin();
-           it != struct_def.fields.vec.end(); ++it) {
-        const auto &field = **it;
-        if (field.deprecated) {
-          // Deprecated fields won't be accessible.
-          continue;
-        }
-
-        code_.SetValue("OFFSET_NAME", GenFieldOffsetName(field));
-        code_.SetValue("OFFSET_VALUE", NumToString(field.value.offset));
-        code_ += "{{SEP}}    {{OFFSET_NAME}} = {{OFFSET_VALUE}}\\";
-        code_.SetValue("SEP", ",\n");
-      }
-      code_ += "";
-      code_ += "  };";
-    }
-
     // Generate the accessors.
     for (auto it = struct_def.fields.vec.begin();
          it != struct_def.fields.vec.end(); ++it) {
@@ -1798,7 +1773,7 @@ class CppGenerator : public BaseGenerator {
       } else {
         accessor = "GetPointer<";
       }
-      auto offset_str = GenFieldOffsetName(field);
+      auto offset_str = NumToString(field.value.offset);
       auto offset_type =
           GenTypeGet(field.value.type, "", "const ", " *", false);
 
@@ -2021,7 +1996,7 @@ class CppGenerator : public BaseGenerator {
         // }
         code_.SetValue("FIELD_NAME", Name(field));
         code_.SetValue("FIELD_TYPE", GenTypeWire(field.value.type, " ", true));
-        code_.SetValue("ADD_OFFSET", Name(struct_def) + "::" + offset);
+        code_.SetValue("ADD_OFFSET", offset);
         code_.SetValue("ADD_NAME", name);
         code_.SetValue("ADD_VALUE", value);
         if (is_scalar) {
@@ -2068,7 +2043,7 @@ class CppGenerator : public BaseGenerator {
       if (!field.deprecated && field.required) {
         code_.SetValue("FIELD_NAME", Name(field));
         code_.SetValue("OFFSET_NAME", GenFieldOffsetName(field));
-        code_ += "    fbb_.Required(o, {{STRUCT_NAME}}::{{OFFSET_NAME}});";
+        code_ += "    fbb_.Required(o, {{OFFSET_NAME}});";
       }
     }
     code_ += "    return o;";

CMakeLists.txt

@@ -74,6 +74,14 @@ option(MNN_USE_SSE "Use SSE optimization for x86 if possiable" ON)
 option(MNN_BUILD_CODEGEN "Build with codegen" OFF)
 option(MNN_ENABLE_COVERAGE "Build with coverage enable" OFF)
 
+IF (NOT DEFINED MNN_USE_SPARSE_COMPUTE)
+  IF (TARGET_OS_IPHONE OR TARGET_OS_SIMULATOR)
+    set(MNN_USE_SPARSE_COMPUTE OFF)
+  ELSE()
+    set(MNN_USE_SPARSE_COMPUTE ON)
+  ENDIF()
+ENDIF()
+
 IF(NOT MNN_BUILD_SHARED_LIBS AND MNN_SEP_BUILD)
   message(WARNING "Close MNN_SEP_BUILD for static library")
   SET(MNN_SEP_BUILD OFF CACHE BOOL "<docstring>" FORCE)
@@ -117,7 +125,6 @@ IF(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND NOT MNN_BUILD_SHARED_LIBS AND NOT (M
   ENDIF()
 ENDIF()
 
-
 if(MNN_FORBID_MULTI_THREAD)
     add_definitions(-DMNN_FORBIT_MULTI_THREADS)
 endif()
@@ -146,6 +153,7 @@ option(MNN_ONEDNN "Enable oneDNN" OFF)
 option(MNN_AVX512 "Enable AVX512" OFF)
 option(MNN_CUDA "Enable CUDA" OFF)
 option(MNN_TENSORRT "Enable TensorRT" OFF)
+option(MNN_COREML "Enable CoreML" OFF)
 
 if (MNN_USE_THREAD_POOL)
     message(STATUS "Use Threadpool, forbid openmp")
@@ -182,6 +190,7 @@ message(STATUS "\tVulkan: ${MNN_VULKAN}")
 message(STATUS "\tARM82: ${MNN_ARM82}")
 message(STATUS "\toneDNN: ${MNN_ONEDNN}")
 message(STATUS "\tTensorRT: ${MNN_TENSORRT}")
+message(STATUS "\tCoreML: ${MNN_COREML}")
 message(STATUS "\tCUDA: ${MNN_CUDA}")
 message(STATUS "\tOpenMP: ${MNN_OPENMP}")
 message(STATUS "\tBF16: ${MNN_SUPPORT_BF16}")
@@ -319,8 +328,14 @@ FILE(GLOB MNN_CV_SRC ${CMAKE_CURRENT_LIST_DIR}/source/cv/*)
 add_library(MNNCV OBJECT ${MNN_CV_SRC})
 list(APPEND MNN_OBJECTS_TO_LINK $<TARGET_OBJECTS:MNNCV>)
 list(APPEND MNN_TARGETS MNNCV)
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64)|(X86_64)|(x64)|(X64)|(amd64)|(AMD64)|(i686)")
+    if (APPLE)
+        add_definitions(-fno-stack-check) # Workaround a Xcode 11.X bug
+    endif()
+endif()
 if (MNN_USE_SSE)
     if(CMAKE_SYSTEM_PROCESSOR MATCHES "(x86_64)|(X86_64)|(x64)|(X64)|(amd64)|(AMD64)|(i686)")
+        target_compile_options(MNNCV PRIVATE -DMNN_USE_SSE)
         if (NOT MSVC)
             target_compile_options(MNNCV PRIVATE -msse4.1)
         endif()
@@ -401,8 +416,6 @@ if (NOT APPLE)
     endif()
 endif()
 
-set(CMAKE_CXX_FLAGS_ORIGIN ${CMAKE_CXX_FLAGS})
-set(CMAKE_C_FLAGS_ORIGIN ${CMAKE_C_FLAGS})
 if ((NOT MSVC) AND MNN_HIDDEN)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fvisibility-inlines-hidden -fvisibility=hidden")
     set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fvisibility=hidden")
@@ -435,6 +448,21 @@ if(MNN_METAL AND APPLE)
     list(APPEND MNN_OBJECTS_TO_LINK $<TARGET_OBJECTS:MNNMetal>)
 endif()
 
+# CoreML
+IF(MNN_COREML)
+    add_definitions(-DMNN_COREML_ENABLED=1)
+    add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/source/backend/coreml/)
+    list(APPEND MNN_DEPS MNNCoreML)
+    list(APPEND MNN_OBJECTS_TO_LINK $<TARGET_OBJECTS:MNNCoreML>)
+
+    find_library(COREML CoreML)
+    find_library(FOUNDATION Foundation)
+    find_library(METAL Metal)
+    list(APPEND MNN_EXTRA_DEPENDS ${COREML})
+    list(APPEND MNN_EXTRA_DEPENDS ${FOUNDATION})
+    list(APPEND MNN_EXTRA_DEPENDS ${METAL})
+ENDIF()
+
 # Vulkan
 IF(MNN_VULKAN)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/source/backend/vulkan/)

benchmark/benchmark.cpp

@@ -94,9 +94,7 @@ std::vector<Model> findModelFiles(const char* dir) {
 
 void setInputData(MNN::Tensor* tensor) {
     float* data = tensor->host<float>();
-    for (int i = 0; i < tensor->elementSize(); i++) {
-        data[i] = Revert::getRandValue();
-    }
+    Revert::fillRandValue(data, tensor->elementSize());
 }
 
 static inline uint64_t getTimeInUs() {
@@ -117,9 +115,9 @@ static inline uint64_t getTimeInUs() {
 }
 
 std::vector<float> doBench(Model& model, int loop, int warmup = 10, int forward = MNN_FORWARD_CPU, bool only_inference = true,
-                           int numberThread = 4, int precision = 2) {
+                           int numberThread = 4, int precision = 2, float sparsity = 0.0f, int sparseBlockOC = 1) {
     auto revertor = std::unique_ptr<Revert>(new Revert(model.model_file.c_str()));
-    revertor->initialize();
+    revertor->initialize(sparsity, sparseBlockOC);
     auto modelBuffer      = revertor->getBuffer();
     const auto bufferSize = revertor->getBufferSize();
     auto net = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromBuffer(modelBuffer, bufferSize));
@@ -348,15 +346,36 @@ void set_cpu_affinity()
 #endif
 }
 
+#if TARGET_OS_IPHONE
+void iosBenchAll(const char* modelPath) {
+    std::cout << "MNN benchmark" << std::endl;
+    int loop               = 20;
+    int warmup             = 10;
+    MNNForwardType forward = MNN_FORWARD_CPU;
+    forward = MNN_FORWARD_NN;
+    int numberThread       = 4;
+    int precision = 2;
+    std::cout << "Forward type: **" << forwardType(forward) << "** thread=" << numberThread << "** precision=" <<precision << std::endl;
+    std::vector<Model> models = findModelFiles(modelPath);
+    std::cout << "--------> Benchmarking... loop = " << loop << ", warmup = " << warmup << std::endl;
 
+    for (auto& m : models) {
+        std::vector<float> costs = doBench(m, loop, warmup, forward, false, numberThread, precision);
+        displayStats(m.name, costs);
+    }
+}
+#else
 int main(int argc, const char* argv[]) {
     std::cout << "MNN benchmark" << std::endl;
     int loop               = 10;
     int warmup             = 10;
     MNNForwardType forward = MNN_FORWARD_CPU;
     int numberThread       = 4;
+    int precision = 2;
+    float sparsity = 0.0f;
+    int sparseBlockOC = 1;
     if (argc <= 2) {
-        std::cout << "Usage: " << argv[0] << " models_folder [loop_count] [warmup] [forwardtype] [numberThread] [precision]" << std::endl;
+        std::cout << "Usage: " << argv[0] << " models_folder [loop_count] [warmup] [forwardtype] [numberThread] [precision] [weightSparsity]" << std::endl;
         return 1;
     }
     if (argc >= 3) {
@@ -371,11 +390,20 @@ int main(int argc, const char* argv[]) {
     if (argc >= 6) {
         numberThread = atoi(argv[5]);
     }
-    int precision = 2;
+
     if (argc >= 7) {
         precision = atoi(argv[6]);
     }
-    std::cout << "Forward type: **" << forwardType(forward) << "** thread=" << numberThread << "** precision=" <<precision << std::endl;
+
+    if(argc >= 8) {
+        sparsity = atof(argv[7]);
+    }
+
+    if(argc >= 9) {
+        sparseBlockOC = atoi(argv[8]);
+    }
+
+    std::cout << "Forward type: **" << forwardType(forward) << "** thread=" << numberThread << "** precision=" <<precision << "** sparsity=" <<sparsity << "** sparseBlockOC=" << sparseBlockOC << std::endl;
     std::vector<Model> models = findModelFiles(argv[1]);
 
     std::cout << "--------> Benchmarking... loop = " << argv[2] << ", warmup = " << warmup << std::endl;
@@ -384,7 +412,8 @@ int main(int argc, const char* argv[]) {
     // set_cpu_affinity();
 
     for (auto& m : models) {
-        std::vector<float> costs = doBench(m, loop, warmup, forward, false, numberThread, precision);
+        std::vector<float> costs = doBench(m, loop, warmup, forward, false, numberThread, precision, sparsity, sparseBlockOC);
         displayStats(m.name, costs);
     }
 }
+#endif

benchmark/benchmark.h

@@ -0,0 +1,9 @@
+//
+//  benchmark.h
+//  MNN
+//
+//  Created by MNN on 2021/04/13.
+//  Copyright © 2018, Alibaba Group Holding Limited
+//
+
+void iosBenchAll(const char* modelPath);

demo/exec/CMakeLists.txt

@@ -1,20 +1,40 @@
 message(STATUS "MNN Depends:" ${MNN_DEPS})
+set(MNN_EXEC_DEMOS "")
+
 add_executable(pictureRecognition.out ${CMAKE_CURRENT_LIST_DIR}/pictureRecognition.cpp)
-target_link_libraries(pictureRecognition.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS pictureRecognition.out)
+add_executable(pictureRecognition_module.out ${CMAKE_CURRENT_LIST_DIR}/pictureRecognition_module.cpp)
+list(APPEND MNN_EXEC_DEMOS pictureRecognition_module.out)
+
 add_executable(pictureRotate.out ${CMAKE_CURRENT_LIST_DIR}/pictureRotate.cpp)
-target_link_libraries(pictureRotate.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS pictureRotate.out)
 
 add_executable(multiPose.out ${CMAKE_CURRENT_LIST_DIR}/multiPose.cpp)
-target_link_libraries(multiPose.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS multiPose.out)
 
 add_executable(segment.out ${CMAKE_CURRENT_LIST_DIR}/segment.cpp)
-target_link_libraries(segment.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS segment.out)
 
 add_executable(expressDemo.out ${CMAKE_CURRENT_LIST_DIR}/expressDemo.cpp)
-target_link_libraries(expressDemo.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS expressDemo.out)
 
 add_executable(transformerDemo.out ${CMAKE_CURRENT_LIST_DIR}/transformerDemo.cpp)
-target_link_libraries(transformerDemo.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS transformerDemo.out)
 
 add_executable(rasterDemo.out ${CMAKE_CURRENT_LIST_DIR}/rasterDemo.cpp)
-target_link_libraries(rasterDemo.out ${MNN_DEPS})
+list(APPEND MNN_EXEC_DEMOS rasterDemo.out)
+
+add_executable(nluDemo.out ${CMAKE_CURRENT_LIST_DIR}/nluDemo.cpp)
+list(APPEND MNN_EXEC_DEMOS nluDemo.out)
+
+foreach(TARGET ${MNN_EXEC_DEMOS})
+    target_link_libraries(${TARGET} ${MNN_DEPS})
+    if (MSVC)
+        target_compile_definitions(${TARGET} PRIVATE "_CRT_SECURE_NO_WARNINGS")
+        if (NOT MNN_BUILD_SHARED_LIBS)
+            foreach (DEPEND ${MNN_DEPS})
+                target_link_options(${TARGET} PRIVATE /WHOLEARCHIVE:$<TARGET_FILE:${DEPEND}>)
+            endforeach ()
+        endif()
+    endif()
+endforeach()

demo/exec/expressDemo.cpp

@@ -1,3 +1,11 @@
+//
+//  expressDemo.cpp
+//  MNN
+//
+//  Created by MNN on b'2019/08/19'.
+//  Copyright © 2018, Alibaba Group Holding Limited
+//
+
 #include <MNN/expr/Expr.hpp>
 #include <MNN/expr/ExprCreator.hpp>
 #include <MNN/expr/Executor.hpp>