CUDA Graph Node Count

Problem

Count the nodes of a CUDA graph as kernels + memcpys + events, and the launch overhead saved versus eager execution as (nodes - 1) * launch_overhead_us, clamped at 0.

The signature is cuda_graph_nodes(num_kernels, num_memcpys, num_events, launch_overhead_us). Return [nodes, saved_us].

Hint

Graph capture replays many launches as a single submission, cutting CPU overhead.

Test cases

case 1

in: [10,2,0,5]

expected: [12,55]

case 2

in: [1,0,0,5]

expected: [1,0]

case 3

in: [50,10,4,2.5]

expected: [64,157.5]

case 4

in: [0,0,0,5]

expected: [0,0]

Starter codePython

def cuda_graph_nodes(num_kernels, num_memcpys, num_events, launch_overhead_us):
    # Your code here
    pass

4 test cases · run your code against them in the editor