diff --git a/CMakeLists.txt b/CMakeLists.txt
index f9e7eb2fe8cc9e731d42db2407a04f300e583b1b..8438ad53b026df19beb9f49becee379a2e45a795 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,8 +1,9 @@
 cmake_minimum_required(VERSION 3.6)
+option(USE_OPENMP "openmp support" ON)
+
 project(paddle-mobile)
 
 option(DEBUGING "enable debug mode" OFF)
-option(USE_OPENMP "openmp support" OFF)
 option(USE_EXCEPTION "use std exception" OFF)
 option(LOG_PROFILE "log profile" OFF)
 # select the platform to build
@@ -149,7 +150,17 @@ if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
     list(REMOVE_DUPLICATES CMAKE_CXX_FLAGS)
     add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
 elseif(IS_IOS)
-    add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
+    if(USE_OPENMP)
+        add_library(paddle-mobile-stage0 STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
+        add_custom_target(paddle-mobile ALL
+            COMMAND libtool -static -o ${CMAKE_BINARY_DIR}/libpaddle-mobile.a ${CMAKE_CURRENT_LIST_DIR}/tools/libomp.a $<TARGET_FILE:paddle-mobile-stage0>
+            WORKING_DIRECTORY ${CMAKE_BINARY_DIR}
+            DEPENDS paddle-mobile
+        )
+        add_dependencies(paddle-mobile paddle-mobile-stage0)
+    else()
+        add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
+    endif()
 else ()
     add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
 endif ()
diff --git a/src/ios_io/PaddleMobile.mm b/src/ios_io/PaddleMobile.mm
index 9298e7907709bbbed77b3d4d76528689cae7bd93..5c7b801be0ea7967ea0c94813325d41071bb890b 100644
--- a/src/ios_io/PaddleMobile.mm
+++ b/src/ios_io/PaddleMobile.mm
@@ -58,6 +58,7 @@ static std::mutex shared_mutex;
 - (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath{
   std::string model_path_str = std::string([modelPath UTF8String]);
   std::string weights_path_str = std::string([weighsPath UTF8String]);
+  pam_->SetThreadNum(2);
   if (loaded_ = pam_->Load(model_path_str, weights_path_str, true)) {
     return YES;
   } else {
diff --git a/src/operators/math/gemm.cpp b/src/operators/math/gemm.cpp
index 04a76465b834bad7553870e0a1eb0e5d47cdd71f..e5dc220f4c50922c2918251720619de8b4df7b98 100644
--- a/src/operators/math/gemm.cpp
+++ b/src/operators/math/gemm.cpp
@@ -3002,14 +3002,21 @@ void Sgemm(int m, int n, int k, float alpha, const float *A, int lda,
   NC = L2 / (KC * sizeof(float));
 
   // make sure MC is multiple of MR, and NC is multiple of NR
-  int mblock_num = (m + MC - 1) / MC;
-  MC = (m + mblock_num - 1) / mblock_num;
-  MC = (MC + MR - 1) / MR * MR;
+  if (MC == 0) {
+    MC = MR;
+  } else {
+    int mblock_num = (m + MC - 1) / MC;
+    MC = (m + mblock_num - 1) / mblock_num;
+    MC = (MC + MR - 1) / MR * MR;
+  }
   //  DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
-
-  int nblock_num = (n + NC - 1) / NC;
-  NC = (n + nblock_num - 1) / nblock_num;
-  NC = (NC + NR - 1) / NR * NR;
+  if (NC == 0) {
+    NC = NR;
+  } else {
+    int nblock_num = (n + NC - 1) / NC;
+    NC = (n + nblock_num - 1) / nblock_num;
+    NC = (NC + NR - 1) / NR * NR;
+  }
   //  DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
 
   packedA = static_cast<float *>(
@@ -3067,14 +3074,21 @@ void SgemmWithBn(int m, int n, int k, float alpha, const float *A, int lda,
   NC = L2 / (KC * sizeof(float));
 
   // make sure MC is multiple of MR, and NC is multiple of NR
-  int mblock_num = (m + MC - 1) / MC;
-  MC = (m + mblock_num - 1) / mblock_num;
-  MC = (MC + MR - 1) / MR * MR;
+  if (MC == 0) {
+    MC = MR;
+  } else {
+    int mblock_num = (m + MC - 1) / MC;
+    MC = (m + mblock_num - 1) / mblock_num;
+    MC = (MC + MR - 1) / MR * MR;
+  }
   //  DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
-
-  int nblock_num = (n + NC - 1) / NC;
-  NC = (n + nblock_num - 1) / nblock_num;
-  NC = (NC + NR - 1) / NR * NR;
+  if (NC == 0) {
+    NC = NR;
+  } else {
+    int nblock_num = (n + NC - 1) / NC;
+    NC = (n + nblock_num - 1) / nblock_num;
+    NC = (NC + NR - 1) / NR * NR;
+  }
   //  DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
 
   packedA = static_cast<float *>(
@@ -3133,14 +3147,21 @@ void SgemmWithPRelu(int m, int n, int k, const float *A, int lda,
   NC = L2 / (KC * sizeof(float));
 
   // make sure MC is multiple of MR, and NC is multiple of NR
-  int mblock_num = (m + MC - 1) / MC;
-  MC = (m + mblock_num - 1) / mblock_num;
-  MC = (MC + MR - 1) / MR * MR;
+  if (MC == 0) {
+    MC = MR;
+  } else {
+    int mblock_num = (m + MC - 1) / MC;
+    MC = (m + mblock_num - 1) / mblock_num;
+    MC = (MC + MR - 1) / MR * MR;
+  }
   //  DLOG << "mblock_num = " << mblock_num << ", MC = " << MC << "\n";
-
-  int nblock_num = (n + NC - 1) / NC;
-  NC = (n + nblock_num - 1) / nblock_num;
-  NC = (NC + NR - 1) / NR * NR;
+  if (NC == 0) {
+    NC = NR;
+  } else {
+    int nblock_num = (n + NC - 1) / NC;
+    NC = (n + nblock_num - 1) / nblock_num;
+    NC = (NC + NR - 1) / NR * NR;
+  }
   //  DLOG << "nblock_num = " << nblock_num << ", NC = " << NC << "\n";
 
   packedA = static_cast<float *>(
@@ -3203,9 +3224,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
   if (m > n) {
     // 对 A 分块
     MC = L1 / (KC * sizeof(float));
-    int mblock_num = (m + MC - 1) / MC;
-    MC = (m + mblock_num - 1) / mblock_num;
-    MC = (MC + MR - 1) / MR * MR;
+    if (MC == 0) {
+      MC = MR;
+    } else {
+      int mblock_num = (m + MC - 1) / MC;
+      MC = (m + mblock_num - 1) / mblock_num;
+      MC = (MC + MR - 1) / MR * MR;
+    }
     // 补齐 B
     NC = (n + NR - 1) / NR * NR;
 
@@ -3227,9 +3252,13 @@ void Sgemm_omp(int m, int n, int k, float alpha, const float *A, int lda,
   } else {
     // 对 B 分块
     NC = L1 / (KC * sizeof(float));
-    int nblock_num = (n + NC - 1) / NC;
-    NC = (n + nblock_num - 1) / nblock_num;
-    NC = (NC + NR - 1) / NR * NR;
+    if (NC == 0) {
+      NC = NR;
+    } else {
+      int nblock_num = (n + NC - 1) / NC;
+      NC = (n + nblock_num - 1) / nblock_num;
+      NC = (NC + NR - 1) / NR * NR;
+    }
     // 补齐 A
     MC = (m + MR - 1) / MR * MR;
 
@@ -3311,9 +3340,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
   if (m > n) {
     // 对 A 分块
     MC = L1 / (KC * sizeof(float));
-    int mblock_num = (m + MC - 1) / MC;
-    MC = (m + mblock_num - 1) / mblock_num;
-    MC = (MC + MR - 1) / MR * MR;
+    if (MC == 0) {
+      MC = MR;
+    } else {
+      int mblock_num = (m + MC - 1) / MC;
+      MC = (m + mblock_num - 1) / mblock_num;
+      MC = (MC + MR - 1) / MR * MR;
+    }
     // 补齐 B
     NC = (n + NR - 1) / NR * NR;
 
@@ -3335,9 +3368,13 @@ void SgemmWithBn_omp(int m, int n, int k, float alpha, const float *A, int lda,
   } else {
     // 对 B 分块
     NC = L1 / (KC * sizeof(float));
-    int nblock_num = (n + NC - 1) / NC;
-    NC = (n + nblock_num - 1) / nblock_num;
-    NC = (NC + NR - 1) / NR * NR;
+    if (NC == 0) {
+      NC == NR;
+    } else {
+      int nblock_num = (n + NC - 1) / NC;
+      NC = (n + nblock_num - 1) / nblock_num;
+      NC = (NC + NR - 1) / NR * NR;
+    }
     // 补齐 A
     MC = (m + MR - 1) / MR * MR;
 
@@ -3430,9 +3467,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
   if (m > n) {
     // 对 A 分块
     MC = L1 / (KC * sizeof(float));
-    int mblock_num = (m + MC - 1) / MC;
-    MC = (m + mblock_num - 1) / mblock_num;
-    MC = (MC + MR - 1) / MR * MR;
+    if (MC == 0) {
+      MC = MR;
+    } else {
+      int mblock_num = (m + MC - 1) / MC;
+      MC = (m + mblock_num - 1) / mblock_num;
+      MC = (MC + MR - 1) / MR * MR;
+    }
     // 补齐 B
     NC = (n + NR - 1) / NR * NR;
 
@@ -3454,9 +3495,13 @@ void SgemmWithPRelu_omp(int m, int n, int k, const float *A, int lda,
   } else {
     // 对 B 分块
     NC = L1 / (KC * sizeof(float));
-    int nblock_num = (n + NC - 1) / NC;
-    NC = (n + nblock_num - 1) / nblock_num;
-    NC = (NC + NR - 1) / NR * NR;
+    if (NC == 0) {
+      NC = NR;
+    } else {
+      int nblock_num = (n + NC - 1) / NC;
+      NC = (n + nblock_num - 1) / nblock_num;
+      NC = (NC + NR - 1) / NR * NR;
+    }
     // 补齐 A
     MC = (m + MR - 1) / MR * MR;
 
diff --git a/src/operators/math/math_function.cpp b/src/operators/math/math_function.cpp
index 1ef06372292cd2e8311dfd25ae84b22be03676cd..6ef9fb2a8252e82014ebebc22f82066eeb324c0d 100644
--- a/src/operators/math/math_function.cpp
+++ b/src/operators/math/math_function.cpp
@@ -28,16 +28,9 @@ void matmul<float>(const framework::Tensor &matrix_a, bool trans_a,
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
-  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
-  //  dim_out.size() ==
-  //  2,
-  //                 "The input and output of matmul be matrix");
-  //
-  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
-  //                     platform::is_cpu_place(matrix_b.place())
-  //                     &&
-  //                     platform::is_cpu_place(matrix_out->place()),
-  //                 "Matrix must all be in CPUPlace");
+  PADDLE_MOBILE_ENFORCE(
+      dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+      "The input and output of matmul be matrix");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -61,16 +54,9 @@ void matmulWithBn<float>(const framework::Tensor &matrix_a, bool trans_a,
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
-  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
-  //  dim_out.size() ==
-  //  2,
-  //                 "The input and output of matmul be matrix");
-  //
-  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
-  //                     platform::is_cpu_place(matrix_b.place())
-  //                     &&
-  //                     platform::is_cpu_place(matrix_out->place()),
-  //                 "Matrix must all be in CPUPlace");
+  PADDLE_MOBILE_ENFORCE(
+      dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+      "The input and output of matmul be matrix");
 
   int M = dim_out[0];
   int N = dim_out[1];
@@ -95,16 +81,9 @@ void matmulWithPRelu(const framework::Tensor &matrix_a, bool trans_a,
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
-  //  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 &&
-  //  dim_out.size() ==
-  //  2,
-  //                 "The input and output of matmul be matrix");
-  //
-  //  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
-  //                     platform::is_cpu_place(matrix_b.place())
-  //                     &&
-  //                     platform::is_cpu_place(matrix_out->place()),
-  //                 "Matrix must all be in CPUPlace");
+  PADDLE_MOBILE_ENFORCE(
+      dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+      "The input and output of matmul be matrix");
 
   int M = dim_out[0];
   int N = dim_out[1];
diff --git a/tools/ios-cmake/ios.toolchain.cmake b/tools/ios-cmake/ios.toolchain.cmake
index 4db079d01de8db35fca8fbe63b59e58fd5a3463e..6000f7a8e5dffcd8693b56539f4519840ddd8be8 100644
--- a/tools/ios-cmake/ios.toolchain.cmake
+++ b/tools/ios-cmake/ios.toolchain.cmake
@@ -50,8 +50,13 @@ endif (CMAKE_UNAME)
 #include (CMakeForceCompiler)
 #CMAKE_C_COMPILER (/usr/bin/gcc)
 #CMAKE_CXX_COMPILER (/usr/bin/g++)
-set(CMAKE_C_COMPILER /usr/bin/gcc)
-set(CMAKE_CXX_COMPILER /usr/bin/g++)
+if(USE_OPENMP)
+    set(CMAKE_C_COMPILER /usr/local/opt/llvm/bin/clang)
+    set(CMAKE_CXX_COMPILER /usr/local/opt/llvm/bin/clang++)
+else()
+    set(CMAKE_C_COMPILER /usr/bin/gcc)
+    set(CMAKE_CXX_COMPILER /usr/bin/g++)
+endif()
 set(CMAKE_AR ar CACHE FILEPATH "" FORCE)
 
 # Skip the platform compiler checks for cross compiling