Time breakdown of GNN inferencing

Questions
#1

Hi everyone! Is the copy time (t4-t3) for CPU-GPU data transfer correctly measured in this inference function? It would also be very helpful if someone could confirm if other timings are correctly measured and if there is a similar way to profile the CUDA memory allocation for each invocation of inference(node_id).

# Inference Function
def inference(node_id):
    # get computation graph
    t1 = time.time()
    _, _, computation_graphs = sampler.sample_blocks(g, node_id)
    t2 = time.time()

    # copy to device
    t3 = time.time()
    computation_graphs = [cg.to(device) for cg in computation_graphs]
    x = computation_graphs[0].srcdata['feat']
    t4 = time.time()

    # forward
    t5 = time.time()
    with torch.no_grad():
        y_pred = model(computation_graphs, x)
        pred = y_pred.argmax(1)
    t6 = time.time()

    # deallocate memory
    t7 = time.time()
    del computation_graphs
    torch.cuda.empty_cache()
    t8 = time.time()

    return pred, (t2 - t1) * 1000, (t4 - t3) * 1000, (t6 - t5) * 1000, (t8 - t7) * 1000, (t8 - t1) * 1000
#2

CUDA operations are in general asynchronous so you might need extra attention when timing those operators. A typical practice is to add cuda.device.synchronize before each time.time call.