diff --git a/cmake/util.cmake b/cmake/util.cmake
index 43a56378df0094200d3c7c95a704c27222654708..38299a87e95a151fe0abc535d1cc1968914442b3 100644
--- a/cmake/util.cmake
+++ b/cmake/util.cmake
@@ -156,6 +156,7 @@ function(link_paddle_test TARGET_NAME)
     link_paddle_exe(${TARGET_NAME})
     target_link_libraries(${TARGET_NAME}
                           paddle_test_main
+                          paddle_test_util
                           ${GTEST_LIBRARIES})
 endfunction()
 
diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt
index 0b3126155d0c0872a70fc83260d4ea34161cb717..42a9bd470c16dbfab6b6e6a6b713e471160514bd 100644
--- a/paddle/function/CMakeLists.txt
+++ b/paddle/function/CMakeLists.txt
@@ -17,9 +17,7 @@ if(WITH_TESTING)
     # file(GLOB test_files . *OpTest.cpp)
     # add_executable(${test_bin} EXCLUDE_FROM_ALL ${test_files})
     add_simple_unittest(CrossMapNormalOpTest)
-    add_unittest(ContextProjectionOpTest
-        ContextProjectionOpTest.cpp
-        ../gserver/tests/TestUtil.cpp)
+    add_simple_unittest(ContextProjectionOpTest)
 endif()
 endif()
 
diff --git a/paddle/function/ContextProjectionOpTest.cpp b/paddle/function/ContextProjectionOpTest.cpp
index 359428fc03d698145cb880bd735c908838f96f56..6223d2fd23ac3bbb4fbcf51d37d22feaf3b1330b 100644
--- a/paddle/function/ContextProjectionOpTest.cpp
+++ b/paddle/function/ContextProjectionOpTest.cpp
@@ -14,8 +14,8 @@ limitations under the License. */
 
 #include <gtest/gtest.h>
 #include "FunctionTest.h"
-#include "paddle/gserver/tests/TestUtil.h"
 #include "paddle/math/Matrix.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 
diff --git a/paddle/gserver/tests/CMakeLists.txt b/paddle/gserver/tests/CMakeLists.txt
index c26a2a7f06bc16c113f1812868b5d2b8a5060635..6775563b2ba9a443069b3923b25c5682988babea 100644
--- a/paddle/gserver/tests/CMakeLists.txt
+++ b/paddle/gserver/tests/CMakeLists.txt
@@ -2,8 +2,7 @@
 
 ################### test_ProtoDataProvider ############
 add_unittest_without_exec(test_ProtoDataProvider
-    test_ProtoDataProvider.cpp
-    TestUtil.cpp)
+    test_ProtoDataProvider.cpp)
 
 # test_ProtoDataProvider will mkdir as same name,
 # so if WORKING_DIRECTORY is default directory, then
@@ -15,53 +14,46 @@ add_test(NAME test_ProtoDataProvider
 ################# test_LayerGrad #######################
 add_unittest_without_exec(test_LayerGrad
     test_LayerGrad.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
 add_test(NAME test_LayerGrad
     COMMAND test_LayerGrad)
 
 add_unittest_without_exec(test_ActivationGrad
     test_ActivationGrad.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
 add_test(NAME test_ActivationGrad
     COMMAND test_ActivationGrad)
 ################# test_ConvTrans #######################
 add_unittest_without_exec(test_ConvTrans
     test_ConvTrans.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
 
 add_test(NAME test_ConvTrans
     COMMAND test_ConvTrans)
 ################# test_PriorBox #######################
 add_unittest_without_exec(test_PriorBox
     test_PriorBox.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
 
 add_test(NAME test_PriorBox
     COMMAND test_PriorBox)
 ################# test_ConvUnify #######################
 add_unittest_without_exec(test_ConvUnify
     test_ConvUnify.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
     
 add_test(NAME test_ConvUnify
     COMMAND test_ConvUnify)
 ################# test_BatchNorm #######################
 add_unittest_without_exec(test_BatchNorm
     test_BatchNorm.cpp
-    LayerGradUtil.cpp
-    TestUtil.cpp)
+    LayerGradUtil.cpp)
 
 add_test(NAME test_BatchNorm
     COMMAND test_BatchNorm)
 ################## test_Evaluator #######################
 add_unittest(test_Evaluator
-    test_Evaluator.cpp
-    TestUtil.cpp)
+    test_Evaluator.cpp)
 
 ################ test_LinearChainCRF ####################
 add_simple_unittest(test_LinearChainCRF)
@@ -72,8 +64,7 @@ add_simple_unittest(test_MultinomialSampler)
 ############## test_PyDataProvider ########################
 if(WITH_PYTHON)
     add_unittest_without_exec(test_PyDataProvider
-        test_PyDataProvider.cpp
-        TestUtil.cpp)
+        test_PyDataProvider.cpp)
 
     add_test(NAME test_PyDataProvider
         COMMAND .set_python_path.sh -d ./gserver/tests:${PROJ_ROOT}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider
@@ -81,15 +72,12 @@ if(WITH_PYTHON)
 endif()
 
 ############### test_RecurrentLayer #######################
-add_unittest(test_RecurrentLayer
-    test_RecurrentLayer.cpp
-    TestUtil.cpp)
+add_simple_unittest(test_RecurrentLayer)
 
 ############### test_WarpCTCLayer #######################
 if(NOT WITH_DOUBLE)
     add_unittest_without_exec(test_WarpCTCLayer
-        test_WarpCTCLayer.cpp
-        TestUtil.cpp)
+        test_WarpCTCLayer.cpp)
 
     add_test(NAME test_WarpCTCLayer
         COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${PROJ_ROOT}/warp-ctc/build
@@ -108,8 +96,7 @@ add_test(NAME test_RecurrentGradientMachine
     WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
 
 add_unittest_without_exec(test_NetworkCompare
-    test_NetworkCompare.cpp
-    TestUtil.cpp)
+    test_NetworkCompare.cpp)
 if(WITH_GPU)
     add_test(NAME test_NetworkCompare
         COMMAND .set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h
index 4e88ac0e81ef2596f14995be53f7c5c20ddba2d7..9f68eb64d0b4ad27306d3b20387d74a7e438d910 100644
--- a/paddle/gserver/tests/LayerGradUtil.h
+++ b/paddle/gserver/tests/LayerGradUtil.h
@@ -17,7 +17,7 @@ limitations under the License. */
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/trainer/Trainer.h"
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 using namespace std;  // NOLINT
 
 namespace paddle {
diff --git a/paddle/gserver/tests/TestUtil.cpp b/paddle/gserver/tests/TestUtil.cpp
deleted file mode 100644
index c691fe26255914811c8861cff80495c821990179..0000000000000000000000000000000000000000
--- a/paddle/gserver/tests/TestUtil.cpp
+++ /dev/null
@@ -1,219 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "TestUtil.h"
-#include <gflags/gflags.h>
-#include "paddle/math/SparseMatrix.h"
-
-DEFINE_int32(fixed_seq_length, 0, "Produce some sequence of fixed length");
-
-namespace paddle {
-
-std::string randStr(const int len) {
-  std::string str =
-      "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-  std::string s = "";
-  for (int i = 0; i < len; ++i) s += str[(rand() % 62)];  // NOLINT
-  return s;
-}
-
-MatrixPtr makeRandomSparseMatrix(size_t height,
-                                 size_t width,
-                                 bool withValue,
-                                 bool useGpu,
-                                 bool equalNnzPerSample) {
-  std::vector<int64_t> ids(height);
-  std::vector<int64_t> indices(height + 1);
-  indices[0] = 0;
-
-  std::function<size_t()> randomer = [] { return uniformRandom(10); };
-  if (equalNnzPerSample) {
-    size_t n = 0;
-    do {
-      n = uniformRandom(10);
-    } while (!n);
-    randomer = [=] { return n; };
-  }
-  for (size_t i = 0; i < height; ++i) {
-    indices[i + 1] = indices[i] + std::min(randomer(), width);
-    ids[i] = i;
-  }
-
-  if (!withValue) {
-    std::vector<sparse_non_value_t> data;
-    data.resize(indices[height] - indices[0]);
-    for (size_t i = 0; i < data.size(); ++i) {
-      data[i].col = uniformRandom(width);
-    }
-    auto mat = Matrix::createSparseMatrix(
-        height, width, data.size(), NO_VALUE, SPARSE_CSR, false, useGpu);
-    if (useGpu) {
-      std::dynamic_pointer_cast<GpuSparseMatrix>(mat)->copyFrom(
-          ids.data(), indices.data(), data.data(), HPPL_STREAM_DEFAULT);
-    } else {
-      std::dynamic_pointer_cast<CpuSparseMatrix>(mat)->copyFrom(
-          ids.data(), indices.data(), data.data());
-    }
-    return mat;
-  } else {
-    std::vector<sparse_float_value_t> data;
-    data.resize(indices[height] - indices[0]);
-    for (size_t i = 0; i < data.size(); ++i) {
-      data[i].col = uniformRandom(width);
-      data[i].value = rand() / static_cast<float>(RAND_MAX);  // NOLINT
-    }
-    auto mat = Matrix::createSparseMatrix(
-        height, width, data.size(), FLOAT_VALUE, SPARSE_CSR, false, useGpu);
-    if (useGpu) {
-      std::dynamic_pointer_cast<GpuSparseMatrix>(mat)->copyFrom(
-          ids.data(), indices.data(), data.data(), HPPL_STREAM_DEFAULT);
-    } else {
-      std::dynamic_pointer_cast<CpuSparseMatrix>(mat)->copyFrom(
-          ids.data(), indices.data(), data.data());
-    }
-    return mat;
-  }
-}
-
-void generateSequenceStartPositions(size_t batchSize,
-                                    IVectorPtr& sequenceStartPositions) {
-  ICpuGpuVectorPtr gpuCpuVec;
-  generateSequenceStartPositions(batchSize, gpuCpuVec);
-  sequenceStartPositions = gpuCpuVec->getMutableVector(false);
-}
-
-void generateSequenceStartPositions(size_t batchSize,
-                                    ICpuGpuVectorPtr& sequenceStartPositions) {
-  int numSeqs;
-  if (FLAGS_fixed_seq_length != 0) {
-    numSeqs = std::ceil((float)batchSize / (float)FLAGS_fixed_seq_length);
-  } else {
-    numSeqs = batchSize / 10 + 1;
-  }
-  sequenceStartPositions =
-      ICpuGpuVector::create(numSeqs + 1, /* useGpu= */ false);
-  int* buf = sequenceStartPositions->getMutableData(false);
-  int64_t pos = 0;
-  int len = FLAGS_fixed_seq_length;
-  int maxLen = 2 * batchSize / numSeqs;
-  for (int i = 0; i < numSeqs; ++i) {
-    if (FLAGS_fixed_seq_length == 0) {
-      len = uniformRandom(
-                std::min<int64_t>(maxLen, batchSize - pos - numSeqs + i)) +
-            1;
-    }
-    buf[i] = pos;
-    pos += len;
-    VLOG(1) << " len=" << len;
-  }
-  buf[numSeqs] = batchSize;
-}
-
-void generateSubSequenceStartPositions(
-    const ICpuGpuVectorPtr& sequenceStartPositions,
-    ICpuGpuVectorPtr& subSequenceStartPositions) {
-  int numSeqs = sequenceStartPositions->getSize() - 1;
-  const int* buf = sequenceStartPositions->getData(false);
-  int numOnes = 0;
-  for (int i = 0; i < numSeqs; ++i) {
-    if (buf[i + 1] - buf[i] == 1) {
-      ++numOnes;
-    }
-  }
-  // each seq has two sub-seq except length 1
-  int numSubSeqs = numSeqs * 2 - numOnes;
-  subSequenceStartPositions =
-      ICpuGpuVector::create(numSubSeqs + 1, /* useGpu= */ false);
-  int* subBuf = subSequenceStartPositions->getMutableData(false);
-  int j = 0;
-  for (int i = 0; i < numSeqs; ++i) {
-    if (buf[i + 1] - buf[i] == 1) {
-      subBuf[j++] = buf[i];
-    } else {
-      int len = uniformRandom(buf[i + 1] - buf[i] - 1) + 1;
-      subBuf[j++] = buf[i];
-      subBuf[j++] = buf[i] + len;
-    }
-  }
-  subBuf[j] = buf[numSeqs];
-}
-
-void generateMDimSequenceData(const IVectorPtr& sequenceStartPositions,
-                              IVectorPtr& cpuSequenceDims) {
-  /* generate sequences with 2 dims */
-  int numSeqs = sequenceStartPositions->getSize() - 1;
-  int numDims = 2;
-
-  cpuSequenceDims = IVector::create(numSeqs * numDims, /* useGpu= */ false);
-  int* bufStarts = sequenceStartPositions->getData();
-  int* bufDims = cpuSequenceDims->getData();
-
-  for (int i = 0; i < numSeqs; i++) {
-    int len = bufStarts[i + 1] - bufStarts[i];
-    /* get width and height randomly */
-    std::vector<int> dimVec;
-    for (int j = 0; j < len; j++) {
-      if (len % (j + 1) == 0) {
-        dimVec.push_back(1);
-      }
-    }
-    int idx = rand() % dimVec.size();  // NOLINT use rand_r
-    bufDims[i * numDims] = dimVec[idx];
-    bufDims[i * numDims + 1] = len / dimVec[idx];
-  }
-}
-
-void generateMDimSequenceData(const ICpuGpuVectorPtr& sequenceStartPositions,
-                              IVectorPtr& cpuSequenceDims) {
-  /* generate sequences with 2 dims */
-  int numSeqs = sequenceStartPositions->getSize() - 1;
-  int numDims = 2;
-
-  cpuSequenceDims = IVector::create(numSeqs * numDims, /* useGpu= */ false);
-  const int* bufStarts = sequenceStartPositions->getData(false);
-  int* bufDims = cpuSequenceDims->getData();
-
-  for (int i = 0; i < numSeqs; i++) {
-    int len = bufStarts[i + 1] - bufStarts[i];
-    /* get width and height randomly */
-    std::vector<int> dimVec;
-    for (int j = 0; j < len; j++) {
-      if (len % (j + 1) == 0) {
-        dimVec.push_back(1);
-      }
-    }
-    int idx = rand() % dimVec.size();  // NOLINT use rand_r
-    bufDims[i * numDims] = dimVec[idx];
-    bufDims[i * numDims + 1] = len / dimVec[idx];
-  }
-}
-
-void checkMatrixEqual(const MatrixPtr& a, const MatrixPtr& b) {
-  EXPECT_EQ(a->getWidth(), b->getWidth());
-  EXPECT_EQ(a->getHeight(), b->getHeight());
-  EXPECT_EQ(a->isTransposed(), b->isTransposed());
-  for (size_t r = 0; r < a->getHeight(); ++r) {
-    for (size_t c = 0; c < a->getWidth(); ++c) {
-      EXPECT_FLOAT_EQ(a->getElement(r, c), b->getElement(r, c));
-    }
-  }
-}
-
-void checkVectorEqual(const IVectorPtr& a, const IVectorPtr& b) {
-  EXPECT_EQ(a->getSize(), b->getSize());
-  for (size_t r = 0; r < a->getSize(); ++r) {
-    EXPECT_FLOAT_EQ(a->get(r), b->get(r));
-  }
-}
-}  // namespace paddle
diff --git a/paddle/gserver/tests/TestUtil.h b/paddle/gserver/tests/TestUtil.h
deleted file mode 100644
index ec86469aebbafbf5406a21e6825eda6c105a6b9d..0000000000000000000000000000000000000000
--- a/paddle/gserver/tests/TestUtil.h
+++ /dev/null
@@ -1,78 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#pragma once
-
-#include <gtest/gtest.h>
-#include "paddle/math/Matrix.h"
-
-namespace paddle {
-
-std::string randStr(const int len);
-
-inline int uniformRandom(int n) { return n == 0 ? 0 : rand() % n; }
-
-inline bool approximatelyEqual(float a, float b, float epsilon) {
-  return fabs(a - b) <= ((fabs(a) < fabs(b) ? fabs(b) : fabs(a)) * epsilon);
-}
-
-MatrixPtr makeRandomSparseMatrix(size_t height,
-                                 size_t width,
-                                 bool withValue,
-                                 bool useGpu,
-                                 bool equalNnzPerSample = false);
-
-/**
- * @brief generate sequenceStartPositions for INPUT_SEQUENCE_DATA,
- *        INPUT_HASSUB_SEQUENCE_DATA and INPUT_SEQUENCE_LABEL
- *
- * @param batchSize                      batchSize
- *        sequenceStartPositions[out] generation output
- */
-void generateSequenceStartPositions(size_t batchSize,
-                                    IVectorPtr& sequenceStartPositions);
-
-void generateSequenceStartPositions(size_t batchSize,
-                                    ICpuGpuVectorPtr& sequenceStartPositions);
-
-/**
- * @brief generate subSequenceStartPositions for INPUT_HASSUB_SEQUENCE_DATA
- *        according to sequenceStartPositions
- *
- * @param sequenceStartPositions[in]     input
- *        subSequenceStartPositions[out] generation output
- */
-void generateSubSequenceStartPositions(const IVectorPtr& sequenceStartPositions,
-                                       IVectorPtr& subSequenceStartPositions);
-
-void generateSubSequenceStartPositions(
-    const ICpuGpuVectorPtr& sequenceStartPositions,
-    ICpuGpuVectorPtr& subSequenceStartPositions);
-
-/**
- * @brief generate cpuSequenceDims for INPUT_SEQUENCE_MDIM_DATA according to
- *        sequenceStartPositions
- *
- * @param sequenceStartPositions[in]     input
- *        cpuSequenceDims[out]              generation output
- */
-void generateMDimSequenceData(const IVectorPtr& sequenceStartPositions,
-                              IVectorPtr& cpuSequenceDims);
-void generateMDimSequenceData(const ICpuGpuVectorPtr& sequenceStartPositions,
-                              IVectorPtr& cpuSequenceDims);
-
-void checkMatrixEqual(const MatrixPtr& a, const MatrixPtr& b);
-
-void checkVectorEqual(const IVectorPtr& a, const IVectorPtr& b);
-}  // namespace paddle
diff --git a/paddle/gserver/tests/test_ActivationGrad.cpp b/paddle/gserver/tests/test_ActivationGrad.cpp
index 7d7e68da5c5a9dbcba024002a988f26f7613b724..b201ba8a5a4146ab28cd96454f434f889d72a968 100644
--- a/paddle/gserver/tests/test_ActivationGrad.cpp
+++ b/paddle/gserver/tests/test_ActivationGrad.cpp
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_BatchNorm.cpp b/paddle/gserver/tests/test_BatchNorm.cpp
index 7f5fcb670b70aed9f0a04180d344556a0390122f..822db5a3c40a80bfed4d06f8f28d21253ae44f87 100644
--- a/paddle/gserver/tests/test_BatchNorm.cpp
+++ b/paddle/gserver/tests/test_BatchNorm.cpp
@@ -22,7 +22,7 @@ limitations under the License. */
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_ConvTrans.cpp b/paddle/gserver/tests/test_ConvTrans.cpp
index dd3378304b433c135881310eb89273b6bf492af2..40bb1e2d73c81280a9b12114c13de851285c276b 100644
--- a/paddle/gserver/tests/test_ConvTrans.cpp
+++ b/paddle/gserver/tests/test_ConvTrans.cpp
@@ -23,7 +23,7 @@ limitations under the License. */
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_ConvUnify.cpp b/paddle/gserver/tests/test_ConvUnify.cpp
index ad99b50245cf56eb7db227fa582f6e3f41b47a7a..207fc0566fcf4a0d2e971f3c169a14a64146155b 100644
--- a/paddle/gserver/tests/test_ConvUnify.cpp
+++ b/paddle/gserver/tests/test_ConvUnify.cpp
@@ -23,7 +23,7 @@ limitations under the License. */
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_Evaluator.cpp b/paddle/gserver/tests/test_Evaluator.cpp
index e07066dad84aa6326c2447fc5ee80fa496735fbf..8165eb8269336193858962edac4f9637c2fc1c2f 100644
--- a/paddle/gserver/tests/test_Evaluator.cpp
+++ b/paddle/gserver/tests/test_Evaluator.cpp
@@ -15,7 +15,7 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include <vector>
 #include "ModelConfig.pb.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 #include "paddle/trainer/Trainer.h"
 
 using namespace paddle;  // NOLINT
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
index 2cc25f6b211e367fc82c07c30082c3e12c04e53d..66a70ecd41091b9590038dab3194dd2a0c59dd03 100644
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -21,7 +21,7 @@ limitations under the License. */
 #include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_NetworkCompare.cpp b/paddle/gserver/tests/test_NetworkCompare.cpp
index 0d261059555c971cd509e64802d6c70abc9d2fef..4db30f37a5bc92d4348caed0aebdd8a589b55712 100644
--- a/paddle/gserver/tests/test_NetworkCompare.cpp
+++ b/paddle/gserver/tests/test_NetworkCompare.cpp
@@ -18,7 +18,7 @@ limitations under the License. */
 #include <algorithm>
 #include <cstdlib>
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 #include "paddle/trainer/Trainer.h"
 #include "paddle/utils/Stat.h"
 
diff --git a/paddle/gserver/tests/test_PriorBox.cpp b/paddle/gserver/tests/test_PriorBox.cpp
index a6d6a242696633e66a05bf9fc9eee81a468ed056..ae0e3bc3d24c54eb84c7b5f5053e629607ef4310 100644
--- a/paddle/gserver/tests/test_PriorBox.cpp
+++ b/paddle/gserver/tests/test_PriorBox.cpp
@@ -17,7 +17,7 @@ limitations under the License. */
 #include <vector>
 
 #include "LayerGradUtil.h"
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_ProtoDataProvider.cpp b/paddle/gserver/tests/test_ProtoDataProvider.cpp
index 8fc0aaab69548ae60100696db04d5611570df110..e11bf402c27898b8fdbd3fceeb8aeff8906352db 100644
--- a/paddle/gserver/tests/test_ProtoDataProvider.cpp
+++ b/paddle/gserver/tests/test_ProtoDataProvider.cpp
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/gserver/dataproviders/ProtoDataProvider.h"
 #include "paddle/utils/Util.h"
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace std;  // NOLINT
 
diff --git a/paddle/gserver/tests/test_PyDataProvider.cpp b/paddle/gserver/tests/test_PyDataProvider.cpp
index 0f264ecf91837f6681f0577b93be7e35be268c04..db883543c306c1938eb9da188ce20ed768018efb 100644
--- a/paddle/gserver/tests/test_PyDataProvider.cpp
+++ b/paddle/gserver/tests/test_PyDataProvider.cpp
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/gserver/dataproviders/PyDataProvider.h"
 #include "paddle/utils/Util.h"
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace std;     // NOLINT
 using namespace paddle;  // NOLINT
diff --git a/paddle/gserver/tests/test_RecurrentLayer.cpp b/paddle/gserver/tests/test_RecurrentLayer.cpp
index f91c788863b6963df92b735dbfef2bacee1fff45..16ab0e6aecb6a895b20389992a44dc542eb3b00a 100644
--- a/paddle/gserver/tests/test_RecurrentLayer.cpp
+++ b/paddle/gserver/tests/test_RecurrentLayer.cpp
@@ -19,7 +19,7 @@ limitations under the License. */
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/gserver/layers/Layer.h"
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/gserver/tests/test_WarpCTCLayer.cpp b/paddle/gserver/tests/test_WarpCTCLayer.cpp
index dab6366588b7894a6700c00a5331d436ca2a410c..23ae95852e84216c9065f1b123d35ce868fbb90f 100644
--- a/paddle/gserver/tests/test_WarpCTCLayer.cpp
+++ b/paddle/gserver/tests/test_WarpCTCLayer.cpp
@@ -20,7 +20,7 @@ limitations under the License. */
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/gserver/layers/WarpCTCLayer.h"
 
-#include "TestUtil.h"
+#include "paddle/testing/TestUtil.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT
diff --git a/paddle/math/tests/CMakeLists.txt b/paddle/math/tests/CMakeLists.txt
index a3ea078509704f305672d0b02d272de0f6c97f51..06fc10bae7232fb1278e89e8d9cbdf477fc27b60 100644
--- a/paddle/math/tests/CMakeLists.txt
+++ b/paddle/math/tests/CMakeLists.txt
@@ -7,8 +7,7 @@ add_simple_unittest(test_SparseMatrix)
 
 # TODO(yuyang18): Refactor TestUtil.cpp. Remove this cross module reference.
 add_unittest(test_matrixCompare
-    test_matrixCompare.cpp
-    ../../gserver/tests/TestUtil.cpp)
+    test_matrixCompare.cpp)
 
 add_simple_unittest(test_sparseMatrixCompare)
 add_simple_unittest(test_perturbation)
diff --git a/paddle/math/tests/test_GpuProfiler.cpp b/paddle/math/tests/test_GpuProfiler.cpp
index d490078d909e7940e83a6f461f9386eeda02f53c..e6b5dba446b5a0022ade76b188895c4e0e2a22b4 100644
--- a/paddle/math/tests/test_GpuProfiler.cpp
+++ b/paddle/math/tests/test_GpuProfiler.cpp
@@ -15,9 +15,9 @@ limitations under the License. */
 #ifndef PADDLE_ONLY_CPU
 
 #include <gtest/gtest.h>
-#include "paddle/gserver/tests/TestUtil.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/math/SparseMatrix.h"
+#include "paddle/testing/TestUtil.h"
 #include "paddle/utils/Stat.h"
 #include "paddle/utils/Util.h"
 
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index 98d63438a57b48340bc3b05ac7ac3d6c5cd90fb0..3a780d26c050ac5870824f2ef35c87edc61900a2 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -18,9 +18,9 @@ limitations under the License. */
 
 #include <gtest/gtest.h>
 #include "TensorCheck.h"
-#include "paddle/gserver/tests/TestUtil.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/math/SparseMatrix.h"
+#include "paddle/testing/TestUtil.h"
 #include "paddle/utils/Stat.h"
 #include "paddle/utils/Util.h"
 
diff --git a/paddle/testing/CMakeLists.txt b/paddle/testing/CMakeLists.txt
index 584498c8602ee5faad3e21a8588af7bb802d7377..c47add04b081cbdf78b5a5d3bca3a71025b3d9ac 100644
--- a/paddle/testing/CMakeLists.txt
+++ b/paddle/testing/CMakeLists.txt
@@ -3,4 +3,6 @@
 if(WITH_TESTING)
   add_library(paddle_test_main STATIC TestMain.cpp)
   add_dependencies(paddle_test_main gen_proto_cpp)
+  add_library(paddle_test_util STATIC TestUtil.cpp)
+  add_dependencies(paddle_test_util gen_proto_cpp)
 endif()