IntelPython / dpnp / 39462497481

83%

push

github

Avoid copying uninitialized float64 buffers to the device in tests (#3110)

Four tests built an output array with `numpy.empty` in float64 and
copied it to the device with `dpnp.array`:

- `TestDot::test_out_scalar` in `test_product.py`
(`numpy.result_type(int, float32_array)` is float64)
- `TestMatmul::test_syrk_out_order` in `test_product.py`
(`generate_random_numpy_array` defaults to float64)
- `TestVecdot::test_out_0D` in `test_product.py`
- `TestNanProdSum::test_out` in `test_nanfunctions.py`

On a device without fp64 support the copy casts the buffer to float32.
If the uninitialized memory happens to hold a value beyond the float32
range, NumPy raises `RuntimeWarning: overflow encountered in cast`, so
the warning appears at random in test runs on such devices (it was seen
for `TestDot::test_out_scalar[float32]` on a Windows GPU).

The buffers are now created with `numpy.zeros`. In
`TestNanProdSum::test_out`, where only a wrongly shaped `out` array is
needed, it is allocated on the device with `dpnp.empty` instead. Only
test files are changed.

1604 of 2994 branches covered (53.57%)

Branch coverage included in aggregate %.

29394 of 34528 relevant lines covered (85.13%)

19614.12 hits per line