FYI.

A paper describing the OpenCL implementation and evaluation of AutoDock-GPU was uploaded as a preprint in ChemRxiv

https://chemrxiv.org/articles/Accelerating_Au...ased_Local_Search/9702389

Sadly, we had no updates this week regarding GPU's.

Thank you very much, robertmiles, for your post.
To me, this post, this thread and especially the paper here:
https://chemrxiv.org/articles/Accelerating_Au...ased_Local_Search/9702389

seem to be really important.

Knowing little about coding it is hard to give some good advice, but let me nonetheless
say the following:

It seems to me, that it would be good to carry out the process of porting autodock to
GPU usability very thoroughly.

This case of porting a BOINC sub program (i.e. autodock) to GPU could serve as a model case. I whish that we would take the time to write a good handbook with the title:

"How to port Programms distributed computing programms to GPU"

Let us not forget, that there are prgramms for distributed other than outodock.

If we had a good handbook on porting programms to GPU available, it will be a lot easieer to port other and future distributed computing programms from CPU to GPU.

In my view, all projects on BOINC should be able to use GPU power.

And with a good handbook on how to adapt them all from GPU this process
of porting them would much much easier.

If there was such a handbook, less talented programmers could take part.

I ask for comments, please.

Links to already existing handbooks on programming code for GPU are very welcome.

[snip]


Sections needed to such a handbook:
How to program in a Linux environment if you have never used Linux before
How to program for BOINC
How to program for CUDA
How to attach a CUDA program to a BOINC application
How to program for OpenCL
How to attach an OpenCL program to a BOINC application
How to set up a Linux server for BOINC if you have never used Linux before

In my view, all projects on BOINC should be able to use GPU power.

If you don't care how fast. For programs where you can't find enough groups of steps that can be done simultaneously because they don't write to any location used by a different step, the GPU version can actually be SLOWER than the CPU version, since the GPU's clock is typically about a quarter of the speed of the CPU's clock. Add a section for identifying groups of this kind.

I've already taken a class in CUDA, and can identify SOME of the groups of items that can be done in parallel, so I can do PART of the work. Only for Nvidia GPUs, so far.

I'm still looking for an online class in OpenCL at a reasonable price.

Let me know if you get the other sections needed.

Which has been done. The AutoDock OpenCL/CUDA they created is AutoDock 4 and they ported reactive docking and whatever other protocols they needed to that application. They also updated the AD4 code on the GPU side with new algorithms that make it faster and more accurate for larger molecules which I guess could some day be applied to the standard CPU app.

We just need an update from the WCG techs to know how the implementation process is looking like.