Function TEST

test/rotate.cpp:162–201

///////////////////////////// CPP //////////////////////////////////////

////////////////////////////////// CPP //////////////////////////////////////
//
TEST(Rotate, CPP) {
    const unsigned resultIdx = 0;
    const float angle        = 180;
    const bool crop          = false;

    vector<dim4> numDims;
    vector<vector<float>> in;
    vector<vector<float>> tests;
    readTests<float, float, float>(string(TEST_DIR "/rotate/rotate1.test"),
                                   numDims, in, tests);

    dim4 dims   = numDims[0];
    float theta = angle * PI / 180.0f;

    array input(dims, &(in[0].front()));
    array output = rotate(input, theta, crop, AF_INTERP_NEAREST);

    // Get result
    float* outData = new float[tests[resultIdx].size()];
    output.host((void*)outData);

    // Compare result
    size_t nElems = tests[resultIdx].size();

    // This is a temporary solution. The reason we need this is because of
    // floating point error in the index computations on CPU/GPU, some
    // elements of GPU(CUDA/OpenCL) versions are different from the CPU version.
    // That is, the input index of CPU/GPU may differ by 1 (rounding error) on
    // x or y, hence a different value is copied.
    // We expect 99.99% values to be same between the CPU/GPU versions and
    // ASSERT_EQ (in comments below) to pass for CUDA & OpenCL backends
    size_t fail_count = 0;
    for (size_t i = 0; i < nElems; i++) {
        if (fabs(tests[resultIdx][i] - outData[i]) > 0.0001) fail_count++;
    }
    ASSERT_EQ(true, ((fail_count / (float)nElems) < 0.01));

    // Delete
    delete[] outData;
}