Function MatMulCheck

test/blas.cpp:51–111

template<typename T, bool isBVector>
void MatMulCheck(string TestFile) {
    SUPPORTED_TYPE_CHECK(T);

    vector<dim4> numDims;

    vector<vector<T>> hData;
    vector<vector<T>> tests;
    readTests<T, T, int>(TestFile, numDims, hData, tests);

    af_array a, aT, b, bT;
    ASSERT_SUCCESS(af_create_array(&a, &hData[0].front(), numDims[0].ndims(),
                                   numDims[0].get(),
                                   (af_dtype)dtype_traits<T>::af_type));
    dim4 atdims = numDims[0];
    {
        dim_t f   = atdims[0];
        atdims[0] = atdims[1];
        atdims[1] = f;
    }
    ASSERT_SUCCESS(af_moddims(&aT, a, atdims.ndims(), atdims.get()));
    ASSERT_SUCCESS(af_create_array(&b, &hData[1].front(), numDims[1].ndims(),
                                   numDims[1].get(),
                                   (af_dtype)dtype_traits<T>::af_type));
    dim4 btdims = numDims[1];
    {
        dim_t f   = btdims[0];
        btdims[0] = btdims[1];
        btdims[1] = f;
    }
    ASSERT_SUCCESS(af_moddims(&bT, b, btdims.ndims(), btdims.get()));

    vector<af_array> out(tests.size(), 0);
    if (isBVector) {
        ASSERT_SUCCESS(af_matmul(&out[0], aT, b, AF_MAT_NONE, AF_MAT_NONE));
        ASSERT_SUCCESS(af_matmul(&out[1], bT, a, AF_MAT_NONE, AF_MAT_NONE));
        ASSERT_SUCCESS(af_matmul(&out[2], b, a, AF_MAT_TRANS, AF_MAT_NONE));
        ASSERT_SUCCESS(af_matmul(&out[3], bT, aT, AF_MAT_NONE, AF_MAT_TRANS));
        ASSERT_SUCCESS(af_matmul(&out[4], b, aT, AF_MAT_TRANS, AF_MAT_TRANS));
    } else {
        ASSERT_SUCCESS(af_matmul(&out[0], a, b, AF_MAT_NONE, AF_MAT_NONE));
        ASSERT_SUCCESS(af_matmul(&out[1], a, bT, AF_MAT_NONE, AF_MAT_TRANS));
        ASSERT_SUCCESS(af_matmul(&out[2], a, bT, AF_MAT_TRANS, AF_MAT_NONE));
        ASSERT_SUCCESS(af_matmul(&out[3], aT, bT, AF_MAT_TRANS, AF_MAT_TRANS));
    }

    for (size_t i = 0; i < tests.size(); i++) {
        dim4 dd;
        dim_t *d = dd.get();
        af_get_dims(&d[0], &d[1], &d[2], &d[3], out[i]);
        ASSERT_VEC_ARRAY_NEAR(tests[i], dd, out[i], 1e-3);
    }

    ASSERT_SUCCESS(af_release_array(a));
    ASSERT_SUCCESS(af_release_array(aT));
    ASSERT_SUCCESS(af_release_array(b));
    ASSERT_SUCCESS(af_release_array(bT));

    for (size_t i = 0; i < out.size(); i++) {