Merge pull request #529 from uraj/sgx

Fix SGX build failure

Author: Zhang Shuai

cmake/compiler_options.cmake

@@ -57,27 +57,34 @@ anakin_add_compile_option(-Wno-missing-field-initializers)
 anakin_add_compile_option(-Wno-extra)
 
 if(ENABLE_NOISY_WARNINGS)
-	anakin_add_compile_option(-Wcast-align)
-	anakin_add_compile_option(-Wstrict-aliasing=2)
-	anakin_add_compile_option(-Wundef)
-	anakin_add_compile_option(-Wsign-compare)
+    anakin_add_compile_option(-Wcast-align)
+    anakin_add_compile_option(-Wstrict-aliasing=2)
+    anakin_add_compile_option(-Wundef)
+    anakin_add_compile_option(-Wsign-compare)
 else()
-	anakin_add_compile_option(-Wno-undef)
-	anakin_add_compile_option(-Wno-narrowing)
-	anakin_add_compile_option(-Wno-unknown-pragmas)
-	anakin_add_compile_option(-Wno-delete-non-virtual-dtor)
-	anakin_add_compile_option(-Wno-comment)
-	anakin_add_compile_option(-Wno-sign-compare)
+    anakin_add_compile_option(-Wno-undef)
+    anakin_add_compile_option(-Wno-narrowing)
+    anakin_add_compile_option(-Wno-unknown-pragmas)
+    anakin_add_compile_option(-Wno-delete-non-virtual-dtor)
+    anakin_add_compile_option(-Wno-comment)
+    anakin_add_compile_option(-Wno-sign-compare)
     anakin_add_compile_option(-Wno-write-strings)
     anakin_add_compile_option(-Wno-ignored-qualifiers)
     anakin_add_compile_option(-Wno-enum-compare)
     anakin_add_compile_option(-Wno-missing-field-initializers)
 endif()
 
+if(USE_SGX)
+    # SGX build uses MKL instead of MKLMKL, possibly a higer version
+    # Some APIs may be deprecated by later MKL. We want to ignore
+    # these warnings
+    anakin_add_compile_option(-Wno-deprecated-declarations)
+endif()
+
 if(CMAKE_BUILD_TYPE MATCHES Debug)
     anakin_add_compile_option(-O0)
-	anakin_add_compile_option(-g)
-	anakin_add_compile_option(-gdwarf-2) # for old version gcc and gdb. see: http://stackoverflow.com/a/15051109/673852
+    anakin_add_compile_option(-g)
+    anakin_add_compile_option(-gdwarf-2) # for old version gcc and gdb. see: http://stackoverflow.com/a/15051109/673852
 else()
     if(USE_SGX)
       anakin_add_compile_option(-Os)
@@ -95,16 +102,16 @@ else()
 endif()
 
 if(TARGET_ANDROID)
-	anakin_add_compile_option(-pie)
+    anakin_add_compile_option(-pie)
     add_compile_options(-ldl)
-	anakin_add_compile_option(-lc)
+    anakin_add_compile_option(-lc)
     set(ANAKIN_EXTRA_CXX_FLAGS "${ANAKIN_EXTRA_CXX_FLAGS} ${ANDROID_CXX_FLAGS}")
     set(CMAKE_SHARED_LINKER_FLAGS "${CMAKE_SHARED_LINKER_FLAGS} -Wl,--gc-sections")
     set(MAKE_STATIC_LINKER_FLAGS "${MAKE_STATIC_LINKER_FLAGS} -Wl,--gc-sections")
 endif()
 
 if(TARGET_IOS)
-	# none temp
+    # none temp
 endif()
 
 if(BUILD_STATIC OR X86_COMPILE_482)
@@ -141,18 +148,11 @@ endif()
 
 # The -Wno-long-long is required in 64bit systems when including sytem headers.
 if(X86_64)
-	anakin_add_compile_option(-Wno-long-long)
+    anakin_add_compile_option(-Wno-long-long)
 endif()
 
 set(CMAKE_CXX_FLAGS  "${CMAKE_CXX_FLAGS} ${ANAKIN_EXTRA_CXX_FLAGS}")
 
-#if(WIN32)
-#    if(MSVC)
-#    	message(STATUS "Using msvc compiler")
-#        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /D_SCL_SECURE_NO_WARNINGS")
-#    endif()
-#endif()
-
 if(USE_CUDA)
     if(CMAKE_BUILD_TYPE MATCHES Debug)
         anakin_add_compile_option("-Xcompiler -fPIC" NVCC)

cmake/external/xbyak.cmake

@@ -28,7 +28,7 @@ message(STATUS "Scanning external modules ${Green}xbyak${ColourReset} ...")
 include_directories(${XBYAK_INC_DIR})
 
 if(USE_SGX)
-    set(SGX_PATCH_CMD "cd ${ANAKIN_TEMP_THIRD_PARTY_PATH} && patch -p0 <${ANAKIN_THIRD_PARTY_PATH}/xbyak.patch")
+    set(SGX_PATCH_CMD "cd ${XBYAK_INSTALL_ROOT} && patch -p1 <${ANAKIN_THIRD_PARTY_PATH}/xbyak.patch")
 else()
     # use a whitespace as nop so that sh won't complain about missing argument
     set(SGX_PATCH_CMD " ")

framework/core/functor.h

@@ -35,7 +35,7 @@ class Functor {
 	Functor(){}
     Functor(FuncType& func):_func(func){};
 
-    Functor<RetType, ParamTypes...>& operator=(FuncType& func) { _func=func; return ;}
+    Functor<RetType, ParamTypes...>& operator=(FuncType& func) { _func=func; return *this;}
 
 	  /// must be overwritten
     virtual RetType operator()(ParamTypes ...parameters) = 0;

framework/core/net/net.cpp

@@ -917,7 +917,9 @@ void Net<Ttype, Ptype, RunType>::fusion_init(graph::Graph<Ttype, Ptype>& graph,
     // shallow copy
     _graph_p->CopyFrom(graph);
     auto node_names_in_exec_order = graph.get_nodes_in_order();
+#ifndef USE_SGX
     load_calibrator_config(graph,!_has_loaded_layout_from_file,auto_config_layout);
+#endif
 #ifndef USE_BM_PLACE // anbl add 
     // infer basic shape and parsing parameter from graph
     for (auto& node_name : node_names_in_exec_order) {

framework/graph/llvm/scheduler.cpp

@@ -126,7 +126,6 @@ void Scheduler::Run() {
 
             // debug info
             for (const auto& op : this->_wait_que) {
-                std::ostringstream oss;
                 LOG(INFO) << "op.name=" << op.name;
                 int i = 0;
                 for (const auto& in : op_inputs[op.name]) {

saber/CMakeLists.txt

@@ -71,11 +71,11 @@ endif()
 
 if(USE_SGX)
     set(SGX_INCOMPATIBLE_SRC
-        ${ANAKIN_SABER}/funcs/impl/x86/mkl_gemm.cpp
-        ${ANAKIN_SABER}/funcs/impl/x86/mkldnn_helper.cpp
-        ${ANAKIN_SABER}/funcs/impl/x86/saber_lstmp.cpp
-        ${ANAKIN_SABER}/funcs/impl/x86/saber_lstm.cpp
-        ${ANAKIN_SABER}/funcs/impl/x86/mkl_packed_int8_gemm.cpp
+	    ${ANAKIN_SABER}/funcs/impl/x86/mkl_gemm.cpp
+	    ${ANAKIN_SABER}/funcs/impl/x86/mkldnn_helper.cpp
+	${ANAKIN_SABER}/funcs/impl/x86/saber_lstmp.cpp
+	#${ANAKIN_SABER}/funcs/impl/x86/saber_lstm.cpp
+	${ANAKIN_SABER}/funcs/impl/x86/mkl_packed_int8_gemm.cpp
     )
     foreach(toremove ${SGX_INCOMPATIBLE_SRC})
         list(REMOVE_ITEM ANAKIN_SABER_BASE_SRC ${toremove})

saber/funcs/impl/x86/vender_fc.cpp

@@ -58,7 +58,6 @@ ::create(const std::vector<Tensor<X86> *>& inputs,
     for (int i = 0; i < inputs.size(); i++) {
         cblas_int IC = inputs[i]->count_valid(param.axis, inputs[i]->dims());
         packed_weights.push_back(cblas_sgemm_alloc(CblasAMatrix, OC, MB, IC));
-        // LOG(INFO) << "anakin input[" << i << "] alloc passed";
         cblas_sgemm_pack(CblasColMajor,
                          CblasAMatrix,
                          param.is_transpose_weights ? CblasNoTrans : CblasTrans,
@@ -67,7 +66,6 @@ ::create(const std::vector<Tensor<X86> *>& inputs,
                          weights + total_IC * OC, IC,
                          packed_weights[i]);
         total_IC += IC;
-        // LOG(INFO) << "anakin input[" << i << "] pack passed";
     }
 
     CHECK_EQ(inputs.size(), 1);
@@ -182,7 +180,6 @@ ::dispatch(const std::vector<Tensor<X86> *>& inputs,
         cblas_int IC = inputs[i]->count_valid(param.axis, inputs[i]->dims());
 
         if (i == 0) {
-            // C := alpha * op(A) * op(B) + beta * C
             cblas_sgemm_compute(CblasColMajor,                                     // Layout
                                 CblasPacked,                                       // a
                                 CblasNoTrans,                                      // b是否转置
@@ -201,14 +198,6 @@ ::dispatch(const std::vector<Tensor<X86> *>& inputs,
                                 1.0,                                               // beta
                                 dst, OC);                                          // c, ldc
         }
-
-        //LOG(INFO) << "anakin compute[" << i << "] passed";
-
-        // LOG(INFO) << "inputs[]:dims: " << inputs[0]->dims();
-        // LOG(INFO) << "inputs:size: " << inputs.size();
-        // LOG(INFO) << "inputs:capacity: " << inputs.capacity();
-        // LOG(INFO) << "output:size: " << outputs.size();
-        // LOG(INFO) << "OC, MB, IC: " << OC << " "<< MB << " " << IC;
     }
 
     if (bias) {
@@ -246,7 +235,6 @@ SaberStatus VenderFc<X86, AK_INT8>::create(const std::vector<Tensor<X86> *>& inp
         ws_ = nullptr;
     }
 
-    //    LOG(INFO)<<"batch size = "<<_batch_size<<","<<_output_channel;
     ws_ = zmalloc(_batch_size * _output_channel * sizeof(int), 256);
 
     if (ws_ == nullptr) {
@@ -291,7 +279,6 @@ SaberStatus VenderFc<X86, AK_INT8>::init(const std::vector<Tensor<X86> *>& input
         _need_weights_trans = true;
         _weights_trans.re_alloc(param.weights->valid_shape(), AK_INT8);
         utils::ScaleUtils::scale_fc_weights_to_nchw_host(_weights_trans, *param.weights);
-        //        LOG(INFO)<<"input shape "<<inputs[0]->valid_shape()<<" , weights shape "<<param.weights->valid_shape();
     }
 
     if (_need_weights_trans) {
@@ -369,18 +356,9 @@ SaberStatus VenderFc<X86, AK_INT8>::dispatch(const std::vector<Tensor<X86> *>& i
         auto weight = static_cast<const int8_t*>(param.weights->data()) + total_ic * _output_channel;
 
         if (_need_weights_trans) {
-            //            LOG(INFO)<<"weights trans";
             weight = static_cast<const int8_t*>(_weights_trans.data()) + total_ic * _output_channel;
-            //            print_tensor(_weights_trans);
         }
 
-        //        for(auto a:_scale){
-        //            LOG(INFO)<<"scale = "<<a;
-        //        }
-        //        LOG(INFO)<<"m,n,k = "<<_output_channel<<","<<_batch_size<<","<<IC;
-        //        print_tensor(_bias_scale);
-        /* c = scale * { op(A) + a_offset_scale * a_offset } *
-               { op(B) + b_offset_scale * b_offset } + beta * C + c_offset */
         if (i == 0) {
             cblas_gemm_s8u8s32(CblasColMajor,                       // Layout
                                _is_transpose_weights,                // a need to transpose or not

saber/funcs/impl/x86/vender_fc.h

@@ -68,9 +68,9 @@ class VenderFc<X86, OpDtype> : public ImplBase<X86, OpDtype, FcParam<X86> > {
     Tensor<X86> _input_scale;
     Tensor<X86> _bias_scale;
 
-#ifndef USE_SGX
+//#ifndef USE_SGX
     PackedMKLInt8Gemm _packed_int8_gemm;
-#endif
+//#endif
 };
 
 

sgx/CMakeLists.txt

@@ -165,8 +165,10 @@ target_link_libraries(anakin_app
 set(CLI11_SRC ${UNTRUSTED_DIR}/CLI11.hpp)
 add_custom_command(
   OUTPUT ${CLI11_SRC}
-  WORKING_DIRECTORY ${UNTRUSTED_DIR}
-  COMMAND wget ARGS -q https://github.com/CLIUtils/CLI11/releases/download/v1.7.1/CLI11.hpp
+  COMMAND ${CMAKE_COMMAND} -E make_directory ${UNTRUSTED_DIR}
+  COMMAND wget ARGS
+     -q https://github.com/CLIUtils/CLI11/releases/download/v1.7.1/CLI11.hpp
+     -O ${UNTRUSTED_DIR}/CLI11.hpp
   COMMENT "Downloading CLI11.hpp..."
 )
 add_custom_target(cli11 ALL

third-party/xbyak.patch

@@ -0,0 +1,77 @@
+--- xbyak/include/xbyak/xbyak.h.orig    2019-01-17 11:58:36.391601030 -0800
++++ xbyak/include/xbyak/xbyak.h    2019-01-17 11:59:09.367746049 -0800
+@@ -27,9 +27,7 @@
+ // #define XBYAK_DISABLE_AVX512
+ 
+ //#define XBYAK_USE_MMAP_ALLOCATOR
+-#if !defined(__GNUC__) || defined(__MINGW32__)
+-	#undef XBYAK_USE_MMAP_ALLOCATOR
+-#endif
++#undef XBYAK_USE_MMAP_ALLOCATOR
+ 
+ #ifdef __GNUC__
+ 	#define XBYAK_GNUC_PREREQ(major, minor) ((__GNUC__) * 100 + (__GNUC_MINOR__) >= (major) * 100 + (minor))
+@@ -69,7 +67,6 @@
+ 	#include <malloc.h>
+ #elif defined(__GNUC__)
+ 	#include <unistd.h>
+-	#include <sys/mman.h>
+ 	#include <stdlib.h>
+ #endif
+ #if !defined(_MSC_VER) || (_MSC_VER >= 1600)
+@@ -253,15 +250,8 @@
+ 
+ inline void *AlignedMalloc(size_t size, size_t alignment)
+ {
+-#ifdef __MINGW32__
+-	return __mingw_aligned_malloc(size, alignment);
+-#elif defined(_WIN32)
+-	return _aligned_malloc(size, alignment);
+-#else
+-	void *p;
+-	int ret = posix_memalign(&p, alignment, size);
+-	return (ret == 0) ? p : 0;
+-#endif
++	abort();
++	return nullptr;
+ }
+ 
+ inline void AlignedFree(void *p)
+@@ -970,35 +960,7 @@
+ 	*/
+ 	static inline bool protect(const void *addr, size_t size, int protectMode)
+ 	{
+-#if defined(_WIN32)
+-		const DWORD c_rw = PAGE_READWRITE;
+-		const DWORD c_rwe = PAGE_EXECUTE_READWRITE;
+-		const DWORD c_re = PAGE_EXECUTE_READ;
+-		DWORD mode;
+-#else
+-		const int c_rw = PROT_READ | PROT_WRITE;
+-		const int c_rwe = PROT_READ | PROT_WRITE | PROT_EXEC;
+-		const int c_re = PROT_READ | PROT_EXEC;
+-		int mode;
+-#endif
+-		switch (protectMode) {
+-		case PROTECT_RW: mode = c_rw; break;
+-		case PROTECT_RWE: mode = c_rwe; break;
+-		case PROTECT_RE: mode = c_re; break;
+-		default:
+-			return false;
+-		}
+-#if defined(_WIN32)
+-		DWORD oldProtect;
+-		return VirtualProtect(const_cast<void*>(addr), size, mode, &oldProtect) != 0;
+-#elif defined(__GNUC__)
+-		size_t pageSize = sysconf(_SC_PAGESIZE);
+-		size_t iaddr = reinterpret_cast<size_t>(addr);
+-		size_t roundAddr = iaddr & ~(pageSize - static_cast<size_t>(1));
+-		return mprotect(reinterpret_cast<void*>(roundAddr), size + (iaddr - roundAddr), mode) == 0;
+-#else
+-		return true;
+-#endif
++		return false;
+ 	}
+ 	/**
+ 		get aligned memory pointer
+