master_june24: CID_ACCESS::kernelAccessConst should be called only once (it is now called once per diagram)

[valassi]

Another issue introduced in #830 and being reviewed in #882.

This is a conceptual mistake (in the sense that it differs from the rest of the design of the code), and a (minor) performance penalty. It is also an unncessary increase of bloatware code.

CID_ACCESS::kernelAccessConst is now called once per diagram
https://github.com/valassi/madgraph4gpu/blob/8e312bc02d9d072615fcb1052b5db54754498517/epochX/cudacpp/gg_tt.mad/SubProcesses/P1_gg_ttx/CPPProcess.cc#L339

      // Amplitude(s) for diagram number 1
      FFV1_0<W_ACCESS, A_ACCESS, CD_ACCESS>( w_fp[3], w_fp[2], w_fp[4], COUPs[1], 1.0, &amp_fp[0] );
#ifdef MGONGPU_SUPPORTS_MULTICHANNEL
      if( channelIds != 0 )
      {
        channelids_sv = CID_ACCESS::kernelAccessConst( channelIds );
...

and https://github.com/valassi/madgraph4gpu/blob/8e312bc02d9d072615fcb1052b5db54754498517/epochX/cudacpp/gg_tt.mad/SubProcesses/P1_gg_ttx/CPPProcess.cc#L365

      // Amplitude(s) for diagram number 2
      FFV1_0<W_ACCESS, A_ACCESS, CD_ACCESS>( w_fp[3], w_fp[4], w_fp[1], COUPs[1], 1.0, &amp_fp[0] );
#ifdef MGONGPU_SUPPORTS_MULTICHANNEL
      if( channelIds != 0 )
      {
        channelids_sv = CID_ACCESS::kernelAccessConst( channelIds );

and so on

The channelid array is the same for all diagrams (which can be thousands!!!), so it can and should be retrieved only once. This is what is done for numerators and denominators
https://github.com/valassi/madgraph4gpu/blob/8e312bc02d9d072615fcb1052b5db54754498517/epochX/cudacpp/gg_tt.mad/SubProcesses/P1_gg_ttx/CPPProcess.cc#L320

#ifdef MGONGPU_SUPPORTS_MULTICHANNEL
      // Numerators and denominators for the current event (CUDA) or SIMD event page (C++)
      fptype_sv& numerators_sv = NUM_ACCESS::kernelAccess( numerators );
      fptype_sv& denominators_sv = DEN_ACCESS::kernelAccess( denominators );
#endif

Note in addition that the if checks on channelID should be changed to use nullptr, as described in #892

[roiser]

I agree that a single access would probably slightly enhance the performance. I vaguely remember that I had done this but ran into a problem, but not 100 % sure anymore, pls give it a try as I understand that you want to make the code changes.

[valassi]

I fixed this in afea373 and 89c1a81.

(In any case calculate_wavefunction was later changed even more, to go back essentially to what it is in upstream/master, all other changes being unnecessary).

Closing, fixed in #882