Run on stampede

[rmojgani]

Hi @wadaniel , I tried compiling korali on stampede, seems I could not get OneDNN to work, and I get the following

[Korali] Warning: Neural Network's engine set to OneDNN, but Korali was installed without support for OneDNN. Using Korali's default NN Engine

The training is "very very" slow, do you think it's the reason?

[wadaniel]

Hi Rambod,

yes, if oneDNN cannot be installed, it uses a “hand made” NN built uppon the eigen library. oneDNN is the industry standard NN, also used by tensorflow an pytorch.

any ideas why oneDNN can’t be installed? Daniel

On 25 Apr 2023, at 04:21, rmojgani @.***> wrote:

Hi @wadaniel https://github.com/wadaniel , I tried compiling korali on stampede, seems I could not get OneDNN to work, and I get the following

[Korali] Warning: Neural Network's engine set to OneDNN, but Korali was installed without support for OneDNN. Using Korali's default NN Engine

The training is "very very" slow, do you think it's the reason?

— Reply to this email directly, view it on GitHub https://github.com/rmojgani/RLonKorali/issues/9, or unsubscribe https://github.com/notifications/unsubscribe-auth/AK6AZECZ3SMGEWI45QUDMGLXC4YKZANCNFSM6AAAAAAXKLM2YY. You are receiving this because you were mentioned.

[rmojgani]

@wadaniel

so I have installed Korali with

meson setup build --buildtype=release --prefix=/jet/home/rmojgani/.local

with installs the hand-made one, when I install with

meson setup build --buildtype=release --prefix=/jet/home/rmojgani/.local -Donednn=true

I get the following

ModuleNotFoundError: No module named 'libkorali'

[rmojgani]

@wadaniel These are some of the compiler failing messages

../source/modules/neuralNetwork/layer/activation/activation.cpp:58:46: error: ‘desc’ is not a member of ‘dnnl::eltwise_forward’ auto activationDesc = eltwise_forward::desc( ^~~~ ../source/modules/neuralNetwork/layer/activation/activation.cpp:77:46: error: ‘desc’ is not a member of ‘dnnl::softmax_forward’ auto activationDesc = softmax_forward::desc(_propKind, _prevLayer->_outputMem[0].get_desc(), axis); ^~~~ ../source/modules/neuralNetwork/layer/activation/activation.cpp: In member function ‘virtual void korali::neuralNetwork::layer::Activation::createBackwardPipeline()’: ../source/modules/neuralNetwork/layer/activation/activation.cpp:122:47: error: ‘desc’ is not a member of ‘dnnl::eltwise_backward’ auto activationDesc = eltwise_backward::desc(_activationAlgorithm, _prevLayer->_outputMem[0].get_desc(), _outputMem[0].get_desc(), _alpha, _beta); ^~~~ ../source/modules/neuralNetwork/layer/activation/activation.cpp:136:47: error: ‘desc’ is not a member of ‘dnnl::softmax_backward’ auto activationDesc = softmax_backward::desc(_prevLayer->_outputMem[0].get_desc(), _outputMem[0].get_desc(), axis); ^~~~ [41/103] Compiling C++ object source/libkorali.cp...neuralNetwork_layer_convolution_convolution.cpp.o FAILED: source/libkorali.cpython-39-x86_64-linux-gnu.so.p/modules_neuralNetwork_layer_convolution_convolution.cpp.o