Merge pull request opencv#18323 from liqi-c:tengine-lite-update

Tengine lite update

* update tengine

* Modify for arm32 build.

* format optimization

* add teng_ befor some tengine api

* update graph_t to teng_graph_t

* update graph_t to teng_graph_t

* Code structure optimization

* optimization

* optimization

* remove space

* update tengine url

Co-authored-by: liqi <[email protected]>

Author: liqi-c

3rdparty/libtengine/tengine.cmake

@@ -20,9 +20,8 @@
 # Author: [email protected] or https://github.com/BUG1989
 #         [email protected]
 #         [email protected]
-#
 
-SET(TENGINE_COMMIT_VERSION "8a4c58e0e05cd850f4bb0936a330edc86dc0e28c")
+SET(TENGINE_COMMIT_VERSION "e89cf8870de2ff0a80cfe626c0b52b2a16fb302e")
 SET(OCV_TENGINE_DIR "${OpenCV_BINARY_DIR}/3rdparty/libtengine")
 SET(OCV_TENGINE_SOURCE_PATH "${OCV_TENGINE_DIR}/Tengine-${TENGINE_COMMIT_VERSION}")
 
@@ -32,11 +31,10 @@ IF(EXISTS "${OCV_TENGINE_SOURCE_PATH}")
 	SET(Tengine_FOUND ON)
 	SET(BUILD_TENGINE ON)
 ELSE()
-	SET(OCV_TENGINE_FILENAME "${TENGINE_COMMIT_VERSION}.zip")#name2
-	SET(OCV_TENGINE_URL "https://github.com/OAID/Tengine/archive/") #url2
-	SET(tengine_md5sum f51ca8f3963faeeff3f019a6f6edc206) #md5sum2
+	SET(OCV_TENGINE_FILENAME "${TENGINE_COMMIT_VERSION}.zip")#name
+	SET(OCV_TENGINE_URL "https://github.com/OAID/Tengine/archive/") #url
+	SET(tengine_md5sum 23f61ebb1dd419f1207d8876496289c5) #md5sum
 
-	#MESSAGE(STATUS "**** TENGINE DOWNLOAD BEGIN ****")
 	ocv_download(FILENAME ${OCV_TENGINE_FILENAME}
 						HASH ${tengine_md5sum}
 						URL
@@ -62,24 +60,17 @@ ENDIF()
 if(BUILD_TENGINE)
 	SET(HAVE_TENGINE 1)
 
-	# android system
-	if(ANDROID)
-	   if(${ANDROID_ABI} STREQUAL "armeabi-v7a")
-			   SET(CONFIG_ARCH_ARM32 ON)
-	   elseif(${ANDROID_ABI} STREQUAL "arm64-v8a")
-			   SET(CONFIG_ARCH_ARM64 ON)
-	   endif()
-	else()
+	if(NOT ANDROID)
 		# linux system
 		if(CMAKE_SYSTEM_PROCESSOR STREQUAL arm)
-			   SET(CONFIG_ARCH_ARM32 ON)
+			   SET(TENGINE_TOOLCHAIN_FLAG "-march=armv7-a")
 		elseif(CMAKE_SYSTEM_PROCESSOR STREQUAL aarch64) ## AARCH64
-			   SET(CONFIG_ARCH_ARM64 ON)
+			   SET(TENGINE_TOOLCHAIN_FLAG "-march=armv8-a")
 		endif()
 	endif()
 
 	SET(BUILT_IN_OPENCV ON) ## set for tengine compile discern .
-	SET(Tengine_INCLUDE_DIR  "${OCV_TENGINE_SOURCE_PATH}/core/include" CACHE INTERNAL "")
+	SET(Tengine_INCLUDE_DIR  "${OCV_TENGINE_SOURCE_PATH}/include" CACHE INTERNAL "")
 	if(EXISTS "${OCV_TENGINE_SOURCE_PATH}/CMakeLists.txt")
 		add_subdirectory("${OCV_TENGINE_SOURCE_PATH}" "${OCV_TENGINE_DIR}/build")
 	else()

modules/dnn/CMakeLists.txt

@@ -128,18 +128,11 @@ else()
   set(sources_options ${sources_options} EXCLUDE_CUDA)
 endif()
 
-
 if(HAVE_TENGINE)
 	list(APPEND include_dirs ${TENGINE_INCLUDE_DIRS})
-	if(EXISTS ${TENGINE_LIBRARIES})
-		list(APPEND libs ${TENGINE_LIBRARIES})
-	else()
-		ocv_add_dependencies(opencv_dnn tengine)
-		list(APPEND libs ${TENGINE_LIBRARIES})
-	endif()
+	list(APPEND libs -Wl,--whole-archive ${TENGINE_LIBRARIES} -Wl,--no-whole-archive)
 endif()
 
-
 ocv_module_include_directories(${include_dirs})
 if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   ocv_append_source_files_cxx_compiler_options(fw_srcs "-Wno-suggest-override")  # GCC

modules/dnn/src/dnn.cpp

@@ -1585,7 +1585,9 @@ struct Net::Impl : public detail::NetImplBase
     {
         CV_TRACE_FUNCTION();
         if (preferableBackend == DNN_BACKEND_OPENCV)
+        {
             CV_Assert(preferableTarget == DNN_TARGET_CPU || IS_DNN_OPENCL_TARGET(preferableTarget));
+        }
         else if (preferableBackend == DNN_BACKEND_HALIDE)
             initHalideBackend();
         else if (preferableBackend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)

modules/dnn/src/layers/convolution_layer.cpp

@@ -248,6 +248,10 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
     float power;
 #endif
 
+#ifdef HAVE_TENGINE
+    teng_graph_t tengine_graph;
+#endif
+
 #ifdef HAVE_CUDA
     cuda4dnn::ConvolutionConfiguration::FusionMode cudaFusionMode;
     cuda4dnn::ConvolutionConfiguration::ActivationType cudaActType;
@@ -266,8 +270,20 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
 #ifdef HAVE_CUDA
         cudaFusionMode = cuda4dnn::ConvolutionConfiguration::FusionMode::NONE;
         cudaActType = cuda4dnn::ConvolutionConfiguration::ActivationType::IDENTITY;
+#endif
+#ifdef HAVE_TENGINE
+        tengine_graph=NULL;
 #endif
     }
+#ifdef HAVE_TENGINE
+    ~ConvolutionLayerImpl()
+    {
+        if(NULL != tengine_graph )
+        {
+            tengine_release(tengine_graph);
+        }
+    }
+#endif
 
     MatShape computeColRowShape(const MatShape &inpShape, const MatShape &outShape) const CV_OVERRIDE
     {
@@ -391,6 +407,13 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
             for(int i = 0; i < numOutput; i++ )
                 biasvec[i] = biasMat.at<float>(i);
         }
+#ifdef HAVE_TENGINE
+        if(NULL != tengine_graph )
+        {
+            tengine_release(tengine_graph);
+            tengine_graph = NULL ;
+        }
+#endif
 #ifdef HAVE_OPENCL
         convolutionOp.release();
 #endif
@@ -1765,26 +1788,50 @@ class ConvolutionLayerImpl CV_FINAL : public BaseConvolutionLayerImpl
         }
 
 #ifdef HAVE_TENGINE
-        int inch = inputs[0].size[1]; 		// inch
-        int in_h = inputs[0].size[2]; 		// in_h
-        int in_w = inputs[0].size[3]; 		// in_w
+        bool tengine_ret = false; ;
+
+        std::vector<Mat> teng_in, teng_out;
+        inputs_arr.getMatVector(teng_in);
+        outputs_arr.getMatVector(teng_out);
+
+        int inch = teng_in[0].size[1];    // inch
+        int in_h = teng_in[0].size[2];    // in_h
+        int in_w = teng_in[0].size[3];    // in_w
 
-        int out_b = outputs[0].size[0];     // out batch size
-        int outch = outputs[0].size[1]; 	// outch
-        int out_h = outputs[0].size[2]; 	// out_h
-        int out_w = outputs[0].size[3]; 	// out_w
+        int out_b = teng_out[0].size[0];  // out batch size
+        int outch = teng_out[0].size[1];  // outch
+        int out_h = teng_out[0].size[2];  // out_h
+        int out_w = teng_out[0].size[3];  // out_w
 
-        float *input_  = inputs[0].ptr<float>();
-        float *output_ = outputs[0].ptr<float>();
+        float *input_  = teng_in[0].ptr<float>();
+        float *output_ = teng_out[0].ptr<float>();
         float *kernel_ = weightsMat.ptr<float>();
         float *teg_bias = &biasvec[0];
 
-        bool tengine_ret = tengine_forward(input_, inch, ngroups, in_h, in_w,
-                                    output_, out_b, outch, out_h, out_w,
-                                    kernel_, kernel_size.size(), kernel.height, kernel.width,
-                                    teg_bias, stride.height, stride.width,
-                                    pad.height,  pad.width, dilation.height, dilation.width,
-                                    weightsMat.step1(), padMode);
+        int nstripes = std::max(getNumThreads(), 1);
+
+        /* tengine_init will run when first time. */
+        if(NULL == tengine_graph)
+        {
+            tengine_graph = tengine_init(name.c_str(), input_, inch, ngroups, in_h, in_w,
+                                         output_, out_b, outch, out_h, out_w,
+                                         kernel_, kernel_size.size(), kernel.height, kernel.width,
+                                         teg_bias, stride.height, stride.width,
+                                         pad.height,  pad.width, dilation.height, dilation.width,
+                                         weightsMat.step1(), padMode, tengine_graph, nstripes);
+            /*printf("Init(%s):  input=%p(%d %d %d %d ),output=%p(%d %d %d %d ),kernel=%p(%ld %d %d ), bias=%p ,"
+                   "stride(%d %d), pad(%d %d), dilation(%d %d) ,weightsMat=%ld, padMode=%s ,tengine_graph = %p \n",
+                   name.c_str(),input_, inch, ngroups, in_h, in_w,
+                   output_, out_b, outch, out_h, out_w,
+                   kernel_, kernel_size.size(), kernel.height, kernel.width,
+                   teg_bias, stride.height, stride.width,
+                   pad.height,  pad.width, dilation.height, dilation.width,
+                   weightsMat.step1(), padMode.c_str() ,tengine_graph);*/
+        }
+        if(NULL != tengine_graph)
+        {
+            tengine_ret = tengine_forward(tengine_graph);
+        }
         /* activation */
         if((true == tengine_ret) && activ )
         {

modules/dnn/src/tengine4dnn/include/tengine_graph_convolution.hpp

@@ -26,17 +26,24 @@
 #define TENGINE_GRAPH_CONVOLUTION_HPP
 
 #define FLOAT_TO_REALSIZE (4)
+#ifdef HAVE_TENGINE
+
+#include "tengine_c_api.h"
 
 namespace cv
 {
 namespace dnn
 {
-bool tengine_forward(float *input_, int inch, int group, int in_h, int in_w,
+teng_graph_t  tengine_init(const char* name , float* input_, int inch, int group, int in_h, int in_w,
                         float *output_, int out_b, int outch, int out_h, int out_w,
                         float *kernel_,int kernel_s , int kernel_h, int kernel_w,
                         float *teg_bias, int stride_h,int stride_w,
                         int pad_h, int pad_w,  int dilation_h, int dilation_w,
-                        size_t wstep, const std::string padMode) ;
+                        size_t wstep, const std::string padMode , teng_graph_t& graph, int nstripes) ;
+
+bool tengine_forward(teng_graph_t& graph) ;
+bool tengine_release(teng_graph_t& graph) ;
 }
 }
-#endif /* TENGINE_GRAPH_CONVOLUTION_HPP */
+#endif
+#endif /* TENGINE_GRAPH_CONVOLUTION_HPP */