OPENCL build error when run onnx model at windows.

Development
#1

I had download TVM code at the lastest commit on master branch. Then convert my model file (*.pb) to dll file in windows, dll can run by python system successfully. but when i use c++ to load dll and run the model, TVM throw error as bellow.

"TVMError: OpenCL build error for device=000002431E0BE180ptxas error : Entry function ‘fused_nn_conv2d_18_kernel0’ uses too much shared data (0xc9c4 bytes, 0xc000 max) Stack trace: File “f:\imgblur_new_tvm\imgblur\src\lib\tvm\opencl\opencl_module.cc”, line 314 " I had check the code again and again, from input data ,out put data, tvm_runtime_pack and so on, i had even modify cpp_deploy demo to run my model. It always throw the same error. Does anyone know how to solve the issue? Thanks!

#2

It would be great to look into what is the difference between the pytjon and c++ version, the difference is certainly interesting. Can you make sure to run the same opencl experiment in python? It is weird that one works and another don’t as they are the same dll

#3

Thanks for reply. Yes, I am sure i had run the same dll in python and c++. By the way, i had use another dll convert from tensorflow and it can run successfully in c++. So i am wander that if i had miss some source file about onnx in TVM?