Fix handling of matrix layout in PTR-REF and LAPACK-CSD

[braised-babbage]

This fixes a bug where LAPACK-CSD gives the wrong values for row-major matrices.

In previous versions of magicl, matrices constructed with FROM-ARRAY had their layout determined by the layout of the input array, and so essentially all matrices had :column-major by default. More recent changes f761ef9 allow this to be controlled by a keyword argument of FROM-ARRAY, and the default was changed to :row-major. Unfortunately LAPACK-CSD did not handle this correctly for two reasons:

• :column-major input layout was hardcoded by default in LAPACK-CSD
• The actual code responsible for dereferencing pointers for the underlying lapack library, PTR-REF, was incomplete.

This issue was not exercised by magicl-tests, but does come up in cl-quil-tests::compiler-hook-tests. We fix these and add a new test for LAPACK-CSD in magicl-tests.

This also suggests a change to the general magicl PR workflow: in addition to ensuring that magicl tests pass, we should also be routinely running cl-quil tests against the updated versions for magicl.

cc: @notmgsk @stylewarning

[notmgsk]

cl-quil-tests run pretty long. I wonder if executing a subset of the tests would suffice.

[notmgsk]

What does the error in quilc look like? I don't think I've hit it, so I guess I haven't been using an up-to-date magicl.

[braised-babbage]

What does the error in quilc look like? I don't think I've hit it, so I guess I haven't been using an up-to-date magicl.

Probably you (and the CI infrastructure) are running off of magicl 0.7, the latest in quicklisp. The bug appears on the sohaib.quil compiler hook test, and is a classic: "Matrices do not lie in the same projective class".

[notmgsk]

Classic Sohaib

[braised-babbage]

cl-quil-tests run pretty long. I wonder if executing a subset of the tests would suffice.

Another option (probably healthier) is to create an issue to make sure that magicl test coverage is at least as comprehensive as the code paths exercised by quilc. I noticed with the LAPACK-CSD function that it was not well tested in magicl.

[notmgsk]

cl-quil-tests run pretty long. I wonder if executing a subset of the tests would suffice.

Another option (probably healthier) is to create an issue to make sure that magicl test coverage is at least as comprehensive as the code paths exercised by quilc. I noticed with the LAPACK-CSD function that it was not well tested in magicl.

#122 (comment)