Deformable conv2d is slow with cuda

I test the model only with one dcn layer, ref by there, and I test it on Nvidia Jetson NX board.

Related software version:

pytorch 1.6.0 aarch64 version, torchvision 0.7.0

tvm main branch, whl file gives the exact version (tvm-0.8.dev1687+gf6a404447-cp36)

there is the time for inference evaluation:

with target = ‘llvm -mtriple=aarch64-linux-gnu’–>inference time (std dev): 197.94 ms (3.08 ms)

with target =“nx”)–>inference time (std dev): 481.27 ms (10.03 ms)

Did i do something wrong?

Code snippet

import os
import numpy as np
import torch
import torchvision
from torchvision.ops import DeformConv2d as dcn
import tvm
from tvm import relay
from tvm.relay import testing
import tvm.contrib.graph_executor as graph_runtime

if __name__ == "__main__":
    dtype = "float32"
    #target = 'llvm -mtriple=aarch64-linux-gnu'
    target ="nx")
    target_host = 'llvm -mtriple=aarch64-linux-gnu'
    cur_data = torch.randn([1, 3, 224, 224]).cuda()
    cur_data_np = cur_data.cpu().detach().numpy()

    cur_offset = torch.randn([1, 18, 224, 224]).cuda()
    cur_offset_np = cur_offset.cpu().detach().numpy()
    cur_weight = torch.randn([64, 3, 3, 3]).cuda()
    cur_weight_np = cur_weight.cpu().detach().numpy()

    # pytorch torchvision
    pytorch_model = dcn(3, 64, 3, padding=1, bias=False).cuda().eval()
    pytorch_model.weight = torch.nn.Parameter(cur_weight)
    scripted_model = torch.jit.trace(
        pytorch_model, [cur_data, cur_offset]
    pytorch_res = pytorch_model(cur_data, cur_offset)

    # tvm from pytorch
    shape_list = [("input0", [1, 3, 224, 224]), ("input1", [1, 18, 224, 224])]
    mod, params = relay.frontend.from_pytorch(scripted_model, shape_list)
    ctx = tvm.device(str(target), 0)
    with tvm.transform.PassContext(opt_level=3):
        lib =
            mod, target=target, target_host=target_host, params=params
    m = graph_runtime.GraphModule(lib["default"](ctx))
    m.set_input(0, tvm.nd.array(cur_data_np, ctx))
    m.set_input(1, tvm.nd.array(cur_offset_np, ctx))
    tvm_output = m.get_output(0)

    # evaluate
    print("Evaluate inference time cost on host...")
    ftimer = m.module.time_evaluator("run", ctx, number=1, repeat=100)
    prof_res = np.array(ftimer().results) * 1000
    print("Mean inference time (std dev): %.2f ms (%.2f ms)" %
            (np.mean(prof_res), np.std(prof_res)))

you can try to ansor-tune this op, and hopefully it will have good performance, I tested several cases before, and based on my tests, the ansor-tune results have great performance

You can try tuning it with autotvm or auto scheduler. But deformable_conv2d itself is difficult to optimize due to its memory access pattern, so it is expected to be much slower than conv2d. @comaniac has tried some optimizations to it.

AutoScheduler is definitely more effective in this case, but it really depends on the the data layout and the image size.

Thanks for your kindly reply, i will try it.