Tom Neep (58613cb7) at 28 Mar 17:01
I think we only want to transfer the internal class information once? When running multiple events I get a memory leak which causes a crash if I don't do this. @mslater is this right?
Setting the cuda device once might not be the best thing in case we want to be able to run on multiple devices but the way that we set the code up at the moment I'm not sure this is a problem for now?
Tom Neep (130b4f20) at 28 Mar 17:01
Merge branch 'single-internal-class-transfer-gpu' into 'gpu-integra...
... and 1 more commit
Tom Neep (f890dc20) at 28 Mar 17:00
I was testing running on lxplus-gpu
which gives me a Tesla T4
GPU.
Setting up LCG_105/x86_64-el9-gcc12-opt
I found that I needed to include thrust headers to make the code compile.
After doing so the GEM example runs fine.
Tom Neep (34dcf7a3) at 28 Mar 17:00
Merge branch 'thrust-lxplus-gpu' into 'gpu-integration'
... and 1 more commit
Tom Neep (ce71fc3a) at 05 Mar 12:42
Merge branch 'avalanche-size-limit' into 'master'
... and 1 more commit
I think we only want to transfer the internal class information once? When running multiple events I get a memory leak which causes a crash if I don't do this. @mslater is this right?
Setting the cuda device once might not be the best thing in case we want to be able to run on multiple devices but the way that we set the code up at the moment I'm not sure this is a problem for now?
I was testing running on lxplus-gpu
which gives me a Tesla T4
GPU.
Setting up LCG_105/x86_64-el9-gcc12-opt
I found that I needed to include thrust headers to make the code compile.
After doing so the GEM example runs fine.
Tom Neep (f6ff704f) at 09 Feb 18:39
Fix some missing GPUFLOATs in ComponentGPU
Tom Neep (de0159cd) at 09 Feb 18:16
GPU GEM exmaple, set electron energy regardless of run mode
... and 5 more commits
Tom Neep (023f9876) at 05 Feb 14:36
Fix bug in MediumGPU
Tom Neep (256711b4) at 05 Feb 12:20
Bug fix to ensure something actually happens in CPU avalanche
I don't really know why, but in 0f7a7b2e the TransportElectron
method of AvalancheMicroscopic
was split into three cases for regular, with bfield, and semi-conductor versions. Since we haven't implemented the GPU code for the bfield and semi-conductor versions anyway it works out well because we can safely remove commented out code from our version.
Does this seem ok to you @mslater? My thinking is the difference with other commented code in that method is that we might want to implement some of the other functionality at some point
Tom Neep (54fe90df) at 01 Feb 12:59
Remove bfield and band structure code from GPU avalanche (single pa...
... and 300 more commits
Tom Neep (a7d6a4c7) at 26 Jan 11:56
Using triple pointer rather than stuct hack
... and 1 more commit
Tom Neep (98705053) at 26 Jan 11:15
Tom Neep (98705053) at 26 Jan 11:14
Add CUDA capability to AvalancheMicroscopic