Hi @oliviermattelaer @roiser @jtchilders I am just filing this as an issue after our discussion on Monday (also copying @joshmcfayden).
Personally I had not understood that for NLO calculations Madgraph calls (one? two?) external libraries rather than doing this in-house through the ALOHA code.
Another thing we are discussing via email is that some of these calculations (eg OpenLoops) require quadruple precision, ie FP128. Not at all clear to me how we would handle this on GPUs.
Various things we should probably do, incomplete list:
write down which libraries are needed (is there a choice? are versions of these available for GPU?)
contact the authors of these libraries and understand if they are working on a GPU port
profile some piece of NLO code to understand how much of the CPU time goes into the loop calculation
understand if this code could run on GPUs at all or not (with which precision, FP128? is this C++ code, can this be easily ported to GPUs?)
Hi @oliviermattelaer @roiser @jtchilders I am just filing this as an issue after our discussion on Monday (also copying @joshmcfayden).
Personally I had not understood that for NLO calculations Madgraph calls (one? two?) external libraries rather than doing this in-house through the ALOHA code.
Another thing we are discussing via email is that some of these calculations (eg OpenLoops) require quadruple precision, ie FP128. Not at all clear to me how we would handle this on GPUs.
Various things we should probably do, incomplete list: