fail to find a function wehn build unit-test

[lzhao4ever]

In #1337, I got error messages from Travis CI as follows. This error doesn't appear in my local build. Could you please help me?

[ 76%] Built target paddle_gserver

Scanning dependencies of target paddle_network

[ 76%] Building CXX object paddle/pserver/CMakeFiles/paddle_network.dir/LightNetwork.cpp.o

[ 77%] Building CXX object paddle/pserver/CMakeFiles/paddle_network.dir/SocketChannel.cpp.o

[ 77%] Building CXX object paddle/pserver/CMakeFiles/paddle_network.dir/ProtoServer.cpp.o

Linking CXX static library libpaddle_network.a

Done processing LightNetwork.cpp

Total errors found: 0

Done processing ProtoServer.cpp

Total errors found: 0

Done processing SocketChannel.cpp

Total errors found: 0

Done processing LightNetwork.h

Total errors found: 0

Done processing ProtoServer.h

Total errors found: 0

Done processing SocketChannel.h

Total errors found: 0

[ 77%] Built target paddle_network

Scanning dependencies of target paddle_pserver

[ 77%] Building CXX object paddle/pserver/CMakeFiles/paddle_pserver.dir/BaseClient.cpp.o

[ 78%] Building CXX object paddle/pserver/CMakeFiles/paddle_pserver.dir/ParameterClient2.cpp.o

[ 78%] Building CXX object paddle/pserver/CMakeFiles/paddle_pserver.dir/ParameterServer2.cpp.o

[ 78%] Building CXX object paddle/pserver/CMakeFiles/paddle_pserver.dir/SparseParameterDistribution.cpp.o

[ 79%] Building CXX object paddle/pserver/CMakeFiles/paddle_pserver.dir/ParameterServerController.cpp.o

Linking CXX static library libpaddle_pserver.a

Done processing BaseClient.cpp

Total errors found: 0

Done processing ParameterClient2.cpp

Total errors found: 0

Done processing ParameterServer2.cpp

Total errors found: 0

Done processing ParameterServerController.cpp

Total errors found: 0

Done processing SparseParameterDistribution.cpp

Total errors found: 0

Done processing BaseClient.h

Total errors found: 0

Done processing ParameterClient2.h

Total errors found: 0

Done processing ParameterServer2.h

Total errors found: 0

Done processing ParameterServerController.h

Total errors found: 0

Done processing SparseParameterDistribution.h

Total errors found: 0

[ 79%] Built target paddle_pserver

Scanning dependencies of target test_CustomStackTrace

[ 79%] Building CXX object paddle/utils/tests/CMakeFiles/test_CustomStackTrace.dir/test_CustomStackTrace.cpp.o

Linking CXX executable test_CustomStackTrace

Done processing test_CustomStackTrace.cpp

Total errors found: 0

../../math/libpaddle_math.a(Matrix.cpp.o): In function `paddle::GpuMatrix::classificationError(paddle::Matrix&, paddle::VectorT<int>&, unsigned long)':

/home/travis/build/PaddlePaddle/Paddle/paddle/math/Matrix.cpp:821: undefined reference to `hl_matrix_classification_error(float*, int, int*, float*, int, int, int, int, int*, float*)'

collect2: error: ld returned 1 exit status

make[2]: *** [paddle/utils/tests/test_CustomStackTrace] Error 1

make[1]: *** [paddle/utils/tests/CMakeFiles/test_CustomStackTrace.dir/all] Error 2

make: *** [all] Error 2

[gangliao]

这个是因为你要考虑cpu stub啊。stub/hl_matrix_stub.h
添加：

inline int hl_matrix_classification_error(Elem_t* topVal, int ldv,
                                          int* topIds,
                                          Elem_t* src, int lds,
                                          int dim,
                                          int topkSize,
                                          int numSamples,
                                          int* label,
                                          Elem_t* recResult) {
  return HPPL_RET_OK;
}

[gangliao]

可以考虑加一些单测：

test_matrixCompare.cpp

void testMatrixClassificationError(int samples, int dim, int beamSize) {
  MatrixPtr cpuOut = std::make_shared<CpuMatrix>(samples, 1);
  MatrixPtr gpuOut = std::make_shared<GpuMatrix>(samples, 1);
  MatrixPtr cpuSrc = std::make_shared<CpuMatrix>(samples, dim);
  MatrixPtr gpuSrc = std::make_shared<GpuMatrix>(samples, dim);
  IVectorPtr cpuLbl = std::make_shared<CpuIVector>(samples);
  IVectorPtr gpuLbl = std::make_shared<GpuIVector>(samples);

  cpuSrc->randomizeUniform();
  gpuSrc->copyFrom(*cpuSrc);

  int* labels = cpuLbl->getData();
  for (int i = 0; i < samples; ++i) {
    labels[i] = (int)rand() % dim;
    CHECK_LT(labels[i], dim);
  }
  gpuLbl->copyFrom(*cpuLbl);

  cpuOut->classificationError(cpuSrc, cpuLbl, beamSize);
  gpuOut->classificationError(gpuSrc, gpuLbl, beamSize);

  MatrixPtr gpuRes = std::make_shared<CpuMatrix>(samples, 1);
  gpuRes->copyFrom(*gpuOut);

  MatrixCheckEqual(*gpuRes, *cpuOut);
}

TEST(Matrix, classificationError) {
  for (auto samples : {1, 5, 31, 90, 150, 300}) {
    for (auto dim : {1, 5, 8, 10, 15, 64, 80, 120, 256, 300,
                    1280, 5120, 50000}) {
      for (auto beamSize : {1, 5, 10, 20, 40, (int)rand() % dim + 1}) {
        if (beamSize > dim ) continue;
        VLOG(3) << " sample= " << samples
                << " beamSize= " << beamSize
                << " dim= " << dim;
        testMatrixClassificationError(samples, dim, beamSize);
      }
    }
  }
}

[hedaoyuan]

#892 (comment) 第三条。
修改hl_matrix_stub.h里面的hl_matrix_classification_error定义。

[lzhao4ever]

Got it. Thanks.