Fix #6697: Wrong dtype when using np.asarray on DeviceNDArray

[ashutoshvarma]

Closes #6697

I was not sure about where tests for this should go, for now, the test is in numba/cuda/tests/cudadrv/test_array_attr.py.

[gmarkall]

Many thanks for tracking down this issue and making the PR! I think it is looking great in general:

• Tests pass for me locally with hardware and the simulator.
• The reproducer given in Wrong dtype when using np.asarray on DeviceNDArray #6697 behaves correctly with this PR (it prints out int16 three times instead of int16 twice then float64).
• The starter patch from CUDA: Add support for passing tuples to kernels via vectorize #6563 applied on top of this PR passes all tests, and also allows the example from CUDA: Add support for passing tuples to kernels via vectorize #6563 (comment) to run correctly.

So, this PR really moves things in the right direction for both the issue #6697 and the feature request #6563.

Regarding the location of the test - the majority of tests that test device arrays, transferring them between the host and the device, and checking their dtypes and other properties are in numba/cuda/tests/cudadrv/test_cuda_ndarray.py - would you be happy to move the test there?

Many thanks again!

[ashutoshvarma]

Thanks for the review @gmarkall
I moved the test to numba/cuda/tests/cudadrv/test_cuda_ndarray.py (TestCudaNDArray) as you requested.

[gmarkall]

Many thanks for the update - this is looking good to me!

@esc Could this have a CUDA buildfarm run please?

[esc]

@gmarkall sure numba_smoketest_cuda_yaml_10

[esc]

BF passed.

[gmarkall]

@stuartarchibald @sklam Do you want to look at this before it goes to RTM?

[stuartarchibald]

Looks good, thanks for the patch. Thanks for reviewing @gmarkall .

[gmarkall]

Many thanks for the contribution @ashutoshvarma !