Feature-request: TensorRT 7.1.3 supporting

[mochechan]

Current tkDNN supports TensorRT 6.0.1. However, newer TensorRT was released.
Is it possible that tkDNN supports TensorRT 7.1.3?

[mive93]

Hi @mochechan ,

Have you tried it? Because we actually support 7.1.0 and I don't think it should be that much different.

[mochechan]

In my evaluation, to install the following TensorRT 6.0.1 libraries can compile the tkDNN without error. The compiled result can be executed as well.
apt-get install -y libnvinfer6=6.0.1-1+cuda10.2 libnvinfer-dev=6.0.1-1+cuda10.2 libnvinfer-plugin6=6.0.1-1+cuda10.2 libeigen3-dev libopencv-dev

However, to install the following TensorRT 7.1.3 libraries occurs errors when compiling the tkDNN.
apt-get install -y libnvinfer7=7.1.3-1+cuda10.2 libnvinfer-dev=7.1.3-1+cuda10.2 libnvinfer-plugin7=7.1.3-1+cuda10.2 libeigen3-dev libopencv-dev

The error messages are described as the follows.
In file included from /root/tkDNN/include/tkDNN/Layer.h:6:0,
from /root/tkDNN/src/Conv2d.cpp:3:
/root/tkDNN/src/Conv2d.cpp: In member function 'void tk::dnn::Conv2d::initCUDNN(bool)':
/root/tkDNN/src/Conv2d.cpp:68:57: error: 'CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST' was not declared in this scope
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &bwAlgo) );
^
/root/tkDNN/src/Conv2d.cpp:66:9: note: in expansion of macro 'checkCUDNN'
checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle,
^~~~~~~~~~
/root/tkDNN/src/Conv2d.cpp:68:57: note: suggested alternative: 'CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT'
CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST, 0, &bwAlgo) );
^
/root/tkDNN/src/Conv2d.cpp:66:9: note: in expansion of macro 'checkCUDNN'
checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle,
^~~~~~~~~~
/root/tkDNN/src/Conv2d.cpp:66:21: error: 'cudnnGetConvolutionBackwardDataAlgorithm' was not declared in this scope
checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle,
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
/root/tkDNN/src/Conv2d.cpp:66:21: note: suggested alternative: 'cudnnGetConvolutionBackwardDataAlgorithm_v7'
checkCUDNN( cudnnGetConvolutionBackwardDataAlgorithm(net->cudnnHandle,
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
/root/tkDNN/src/Conv2d.cpp:79:57: error: 'CUDNN_CONVOLUTION_FWD_PREFER_FASTEST' was not declared in this scope
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo) );
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
/root/tkDNN/src/Conv2d.cpp:79:57: note: suggested alternative: 'CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3'
CUDNN_CONVOLUTION_FWD_PREFER_FASTEST, 0, &algo) );
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
/root/tkDNN/src/Conv2d.cpp:77:21: error: 'cudnnGetConvolutionForwardAlgorithm' was not declared in this scope
checkCUDNN( cudnnGetConvolutionForwardAlgorithm(net->cudnnHandle,
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
/root/tkDNN/src/Conv2d.cpp:77:21: note: suggested alternative: 'cudnnGetConvolutionForwardAlgorithm_v7'
checkCUDNN( cudnnGetConvolutionForwardAlgorithm(net->cudnnHandle,
^
/root/tkDNN/include/tkDNN/utils.h:68:9: note: in definition of macro 'checkCUDNN'
if (status != CUDNN_STATUS_SUCCESS) {
^~~~~~
CMakeFiles/tkDNN.dir/build.make:120: recipe for target 'CMakeFiles/tkDNN.dir/src/Conv2d.cpp.o' failed
make[2]: *** [CMakeFiles/tkDNN.dir/src/Conv2d.cpp.o] Error 1
CMakeFiles/Makefile2:436: recipe for target 'CMakeFiles/tkDNN.dir/all' failed
make[1]: *** [CMakeFiles/tkDNN.dir/all] Error 2
Makefile:148: recipe for target 'all' failed
make: *** [all] Error 2

I haven't test TensorRT 7.1.0 yet.

[minhtcai]

Hi, do you have any updates on that? I got the same problem.

[ceccocats]

This is a cudnn error, not tensorrt.
Can you check what version of cuDNN?
we support until 8.0.0, I bet you have 8.0.1

[minhtcai]

Yeah that's the 8.0.1. I then rolled back to TensorRT 6.0.1 and everything works perfectly!

[alexanderfrey]

For Xavier NX you can not roll back since everything is part of JetPack 4.4. Any updates on this ?

[mive93]

We do support the Jetpack 4.4. We made several test on the Jeston NANO using that.

[masip85]

This happens to me with cuDNN: ver 8.0.0. In fact 8.0.0.1 and tensorRT 7.1.3.
#267