Implementing L2 cache persistence into python

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello, I’m recently studying Device Memory L2 Access Management from nvidia document.
I’m into implementing examples of nvidia document on python now.
However, declaring cuda graph node “cudaGraphNode_t node” is becoming an obstacle now.
To be more precise, I want to convert code below into python.

cudaGraph_t graph;
cudaGraphExec_t graphExec;
cudaStream_t stream;

size_t numNodes = 0;
cudaGraphNode_t* nodes = nullptr;
cudaGraphNode_t node;

cudaGraphGetNodes(graph, nullptr, &numNodes);

nodes = new cudaGraphNode_t[numNodes];
CUDA_CHECK(cudaGraphGetNodes(graph, nodes, &numNodes));
node = nodes[0];

cudaKernelNodeAttrValue node_attribute;
node_attribute.accessPolicyWindow.base_ptr = reinterpret_cast<void*>(d_data);
node_attribute.accessPolicyWindow.num_bytes = total_size;
node_attribute.accessPolicyWindow.hitRatio = hitRatio;
node_attribute.accessPolicyWindow.hitProp = cudaAccessPropertyPersisting;
node_attribute.accessPolicyWindow.missProp = cudaAccessPropertyStreaming;
CUDA_CHECK(cudaGraphKernelNodeSetAttribute(node, cudaKernelNodeAttributeAccessPolicyWindow, &node_attribute));

cudaGraphInstantiate(&graphExec, graph, nullptr, nullptr, 0);

can anyone help me with converting code above into python?
Thank you in advance.

2

I believe that cuda-python should have everything you need, since it has basically a 1:1 correspondence with CUDA C++ that you have written, and you haven’t expressed a toolchain preference (e.g. numba, cupy, etc.) Sorry I don’t have a code example for you. The following may be of interest: 1 (note the runtime API directory on the right) 2 (basic tutorial) 3 (github)