First two entries of the w are momenta. Change 6 dimensional wave functions to 4 dimensions (possible?)

[roiser]

No description provided.

[valassi]

Note for instance https://github.com/roiser/madgraph4gpu/blob/812f7438462c01e1052c4be48cafcdaef104195c/examples/gpu/eemumu_AV/SubProcesses/P1_Sigma_sm_epem_mupmum/CPPProcess.cc#L326 : in function FFV1_0, the first two dimensions our of 6 are not used at all.

In https://github.com/roiser/madgraph4gpu/blob/812f7438462c01e1052c4be48cafcdaef104195c/examples/gpu/eemumu_AV/SubProcesses/P1_Sigma_sm_epem_mupmum/CPPProcess.cc#L433 instead, in FFV1P0_3 those two dimensions are simply used to be added (internal particle has momenta that are the sums of incomding particles)

[oliviermattelaer]

Yes this is an idea that I had in mind, you can in principle replace those entry by a couple of binary number such that the information can be used if/when needed. (if you pass a binary number like 000101 for each Wavefunction, you can recompute that information from the allmomenta (global memory) when needed. -- the 0/1 indicates if that particle momenta need to be summed or not--

This is certainly not useful for e+ e- > mu+ mu- (since we are compute bound) but for the other processes this can make a lot of sense since we are memory bound

[oliviermattelaer]

Hi,

This is now implemented in the branch ep2_w4.
@roiser could you run the profiler on it?

Would be nice to see if this has any impact.
By just looking at the performance for g g > t t~ g g this does not seem to improve anything at all.
But having the plot and the nsight-compute will allow to give us a better idea of the situation.

[valassi]

Keeping this open as I am not sure we made progress on this.

Personally my impression is that there is a fine balance between passing more stuff around and having to recompute it all the time. In terms of registers, I guess that those get occupied even if you recalculate some temporary data, rather than read it from global memory. I mean, unless we significantly rewrite the basic formulas for spinors, which we probably do not want to do, those 6 components are quite handy. One example is here

madgraph4gpu/epochX/cudacpp/gg_ttgg.auto/src/HelAmps_sm.cc

Line 889 in e41c142

const fptype_sv P2[4] = { +cxreal( V2[0] ), +cxreal( V2[1] ), +cximag( V2[1] ), +cximag( V2[0] ) };

En passant, I was wondering if we could save something by having a more clever real/imag functions which pass references, rather than values (but all of std complex, thrust and cucomplex return by value there). It would imply rewriting a custom cxtype class like I did for cxtype_v. I started this now for fun, but it gets wuite complex, and there are other priorities. Not sure it would help at all.

Keep this open...