Ep2 w4

[oliviermattelaer]

I open a PR.
Such that we have a place attached to the code to report finding on this idea.

So this branch implements the idea to replace the W[6] of the wavefuntions by a W[4]
and retrieving from global memory (hoping from L1 cache) the removed information:

Current point of concern with the current implementation:
[ ] need to investigate if the convention for the 4 momenta is general enough
[ ] is this a smart idea

[oliviermattelaer]

Here is some comparison for the e+ e- > mu+ mu- process:

blue is the "W4" branch and orange is "W6"
So we can see that this is going on the wrong direction (decrease compute and increase memory)

we can see the increase in global memory call in this graph.

and the absence of register gain in this plot.

So as expected this does not seems a good idea for ee_mumu....

[oliviermattelaer]

Here is the same comparison for g g > t t~ g g
(blue is W4 andd red is W6

No big difference here, and seems to go in the correct direction.

Here the L1 hit rate is much higher, (as well as the amount of transfer)

and the number of register is still to maximum possible...

So not clear if this is worthed or not ....

[oliviermattelaer]

So one nice comparison that this impact significantly the number of register is the number of register used during the kernel execution: (yellow means maximum number of register)so here is the plot for W6

and here is the new plot for w4.

So we have a couple of bottelneck that we have to look at but this might be worth to look at

[valassi]

Hi Olivier, thanks a lot, very interesting and very good idea to open an issue to document it!

Not sure, maybe keep it in progress for the time being, and we do more tests trying to understand this better? It seems that it is not a game changer. But maybe, if we do find other game changers (cuda streams/graph...?), maybe we shoudl reevaluate this and we will see larger effects. Just brainstorming.

How do you produce the plots of registers actualy used by the way? Interesting plots.

Thanks Andrea