Issue #261: adding missing dense descriptor and delete unnecessary se…

[aristeidis-mastoras]

Closes issue #261 .

Closes issue #305 .

• [x] Adding dense descriptor
• [x] Spreading the descriptor template parameter in the eWiseApply and eWiseLambda functions
• [x] work around invert_mask

[anyzelman]

Note: I also added the todo of removing the invert_mask in the PR algorithm. If no objection I could do that quickly around now?

[anyzelman]

Note: I also added the todo of removing the invert_mask in the PR algorithm. If no objection I could do that quickly around now?

This is done. I did run into a new bug in nonblocking, I opened issue #305 for it.

[anyzelman]

@aristeidis-mastoras @byjtew -- this one is now conflicted but was the next on my list. I'll try resolve the conflicts now

[anyzelman]

Rebase complete, smoke tests with LPF OK. Unit tests with LPF and GitHub CI running overnight

[anyzelman]

(If anyone of @aristeidis-mastoras @byjtew wants to review how I handled the conflicts lemme know before tomorrow -- if no message by e.g. noon, I'll proceed with review/merge:)

[anyzelman]

Crossref #305 which this MR should also fix (TBC)

[anyzelman]

Review OK but found one issue in the bsp1d backend. Re-running smoke & unit tests with LPF.

[anyzelman]

All tests OK, will merge

[anyzelman]

Concept release notes:

Prior to this MR, the PageRank implementation does not pass the dense descriptor when it could, while it includes the use of the expensive invert_mask in its performance-critical loop. Mask inversion is expensive since it requires iterating over the full vector range, while, in this case, the inverted mask is constant. Hence, this MR introduces the following changes:

• the inverted mask is precomputed once, while in the performance-critical loop the precomputed inversion is reused as a structural mask; and
• the PR algorithm is reviewed to pass the dense descriptor whenever possible.

This improvement does come at the cost of one additional work-space vector that is now required for the PageRank implementation.

In implementing these fixes, the following feature has been added:

• the grb::eWiseLambda now takes a descriptor template argument. This was previously accepted in the nonblocking backend, but was missing from the specification as well as missing in the implementations of other backends.

The following bugs were encountered and fixed with this MR:

• mixed-domain vector-to-scalar reductions were bugged for the nonblocking backend: when the second (underlying) monoid domain type did not match its third (output) domain type, a compiler error would be returned. This MR fixes the behaviour for both left- and right-handed reductions;
• the descriptors for multiple-vector eWiseLambdas were not correctly propagated for the nonblocking backend.

[anyzelman]

Found one more minor issue in nonblocking, now fixed -- re-running unit tests excluding the LPF ones.