How to measure the RK3399 running time


I follow this tutorial: I use the measure time code like this:
print (finish-start)
when I set “”, it have the warning “src/codegen/ OpenCL runtime not enabled, return a source module…”,the running time is 20 second.
when set "target=“llvm” the running time is 3 second.
Is it too long? How can I get the real time.


When building tvm are you compiling with OpenCL support? (specified in


I building tvm in server without OpenCL support and in the device with the OpenCL support.


My understanding is that you need to compile with OpenCL support on the server side, even if the kernel itself is built from source at runtime.


Should I let theLLVM_CONFIG = /home/daming/software/clang+llvm-6.0.0-aarch64-linux-gnu/bin/llvm-config?
I let LLVM-CONFIG = /home/daming/software/clang+llvm-5.0.1-x86_64-linux-gnu-ubuntu-16.04/bin now.


Either 6.0.0 or 5.0.1 should work; I would just check the path is correct e.g., it points to a binary eventually.


I mean that I shoud let LLVM-CONFIG = arrch64 or = x86_64 in the server?


You should compile with the target host for the RPC device. If the RPC device is the RK3999 board it should be
llvm -target=aarch64-linux-gnu -mattr=+neon. See @merrymercy’s repo for more details:


@daming5432’s question is how to measure the time cost.

TVM provides a time_evaluator function to measure time cost which excludes the cost of compilation, rpc and first dry run.

You can see my benchmark script for examples

When doing benchmark, we should always do some warmup runs. For opencl, the first run will compile kernel code on the board, which takes a lot of time.


Thank you. I have the right measure time.
But When I run your examle of "merrymercy/tvm-mali/blob/3a116f4ec374dd541298a28822a8961ea4fc02f9/ "
The command :python --target-host ‘llvm -target=aarch64-linux-gnu -mattr=+neon’ --host --port 9091 --model all

and error:

[14:50:35] src/codegen/ OpenCL runtime not enabled, return a source module…
Traceback (most recent call last):
File “”, line 107, in
run_case(model, dtype)
File “”, line 47, in run_case
rparams = {k: tvm.nd.array(v, ctx) for k, v in params.items()}
File “”, line 47, in
rparams = {k: tvm.nd.array(v, ctx) for k, v in params.items()}
File “/home/daming/.local/lib/python2.7/site-packages/tvm-0.2.0-py2.7-linux-x86_64.egg/tvm/”, line 199, in array
return empty(arr.shape, arr.dtype, ctx).copyfrom(arr)
File “/home/daming/.local/lib/python2.7/site-packages/tvm-0.2.0-py2.7-linux-x86_64.egg/tvm/_ffi/”, line 165, in copyfrom
File “/home/daming/.local/lib/python2.7/site-packages/tvm-0.2.0-py2.7-linux-x86_64.egg/tvm/_ffi/”, line 232, in copyto
self.handle, target.handle, None))
File “/home/daming/.local/lib/python2.7/site-packages/tvm-0.2.0-py2.7-linux-x86_64.egg/tvm/_ffi/”, line 66, in check_call
raise TVMError(py_str(_LIB.TVMGetLastError()))
tvm._ffi.base.TVMError: [14:50:50] src/runtime/rpc/…/…/common/socket.h:259: Socket SockChannel::Send Error:Connection reset by peer


There are some breaking changes in the RPC protocol recently. You should update the tvm version on your board and host to the latest one.

Or do you have more than 2.5 GB memory on your board? vgg16 needs at least 2.5 GB memory to run on my board. If you don’t, you can skip vgg16 by replace --model all with --model mobilenet or --model resnet18


Thank you very much! I test resnet18 mobilenet. It is ok.
I see your “”.you load model like this:
elif model == ‘resnet18’:
net, params = nnvm.testing.resnet.get_workload(num_layers=18,
batch_size=1, image_shape=image_shape, dtype=dtype)
Do you use mxnet model?
I don’t have mxnet in my RK3399 device.How can I load mxnet model?


nnvm.testing.resnet.get_workload uses random weights, it is only used for testing

If you want to use mxnet model with pretrained weights, you can follow this tutorial
NNVM supports download pretrained model from gluon model zoo


Excuse me ,what’s the cost when you measure the RK3399 running time ?