[ISSUE-131] Moving GPU memory management to OperationManager #841
Conversation
…oGPU Issue 825 refactor copy to gpu
…ot take in any parameter
…ngNeuronsDeviceProperties, now they return by reference so we can use it to intialize the allVerticesDevice_ and allEdgesDevice_
…le files so that it does not take in a parameter. this is the preparation step for the register to operationmanager
Issue 828 allocate op
…s not take in any parameter. This is to prepare the GPUModel::copySynapseIndexMapHostToDevice to be compatible to register to the operationManager
…ake in no parameter to be compatible with the operationManager
… such that it does not take in any parameter and thus can be registered to te operationManager. Also registered the copySynapseIndexMapHostToDevice In GPUModel.cpp to the OperationManager
…yToGPU to the operationManager
Issue 833 copy from gpu
… child class, so that it takes in no parameter. Modified functions copyEdgeDeviceToHost for allEdges class for similar functionality. register the copyFromGPU function to the operationMager. Copying simulation results are no longer done by long chain of responsibility of the neurons and synapses but are now through the centralized dispartch: operation Manager
modified function copyFromDevice for allVertices class as well as the…
… child vertex classes to take no parameters.Refactor deleteEdgeDeviceStruct in AllEdges and all child edge classes to take no parameters.registered the deleteSynapseImap in GPUModel to operationManager
…eviceStruct in AllVertices, now the deallocation is done through the operationManager
…Memory Issue 838 deallocate gpu memory
…d of having simulator.getModel().getLayout().getVertices().registerHistoryVariables(); we can just use operationManager to invokeit
…ryVariable registered registerHistoryVariable to the operationManager now instea…
|
This PR isn't named properly, and it doesn't indicate which issue(s) it closes. |
There was a problem hiding this comment.
I just looked at a couple files and I have many questions. I think that this PR is just too big and complex. This should have been done in smaller chunks, and/or it should have been made earlier so that interim code reviews could have been done. So, I will stop with 2 out of 26 files reviewed and around 10 questions. Once we get through that, we can move on to the next set.
I just included the issues it closes, also for the issue131(implementing the remaining operations) I included the subissues for it |
find .clang-format and print the .clang-format out
confirmed that local and remote's .clang-format is the same
… register the copyCPUtoGPU to operationManager
…Graphitti into BenDevelopment this merge is just to make sure my code on vsc is consistent on the github
…, get rid of the copySynapeseIndMap and replaced with copyCPUToGPU, to make the structure more clear
stiber
changed the title
|
|
||
| #if defined(USE_GPU) | ||
| GPUModel &gpuModel = static_cast<GPUModel &>(simulator.getModel()); | ||
| gpuModel.copyEdgeIndexMapHostToDevice(simulator.getModel().getConnections().getEdgeIndexMap(), | ||
| simulator.getTotalVertices()); | ||
| gpuModel.copyCPUtoGPU(); |
There was a problem hiding this comment.
Need to ask @d-kamath : why is only the GPUModel data being copied to the GPU? Why aren't we copying everything (which can be accomplished via the OperationManager now)?
There was a problem hiding this comment.
I contacted Dyvia and will post her answer later she replied
Closes
#131
#825
#821
Description
Summary
This PR completes the integration of the remaining GPU operations — copyFromGPU and deallocateGPUMemory — into the OperationManager framework, following the previously refactored copyToGPU. All related methods have been standardized to take no parameters and are now registered as std::function<void()> callbacks. This centralizes GPU operation dispatch, eliminates hardcoded call chains, and improves modularity and maintainability.
Changes
Refactored copyFromGPU and deallocateGPUMemory to be parameterless
Moved required context into member variables
Registered operations during class construction
Added #ifdef USE_GPU guards where applicable
Replaced explicit calls in GPUModel with OperationManager::execute(...)
Maintained consistency with the existing copyToGPU refactor
Checklist (Mandatory for new features)
Testing (Mandatory for all changes)
test-medium-connected.xmlPassedtest-large-long.xmlPassed