pytorch-gpgpu-sim "PYTORCH_BIN" about gpgpu-sim_distribution HOT 4 CLOSED

I also wonder what PYTORCH_BIN exactly means.

from gpgpu-sim_distribution.

Hello,
As I know, libcuda in gpgpu-sim uses the PYTORCH_BIN variable to extract ptx code.
Therefore, you can set the variable as existing libcudnn.so file (downloaded from Nvidia web site and installed).

from gpgpu-sim_distribution.

Not yet. I'm facing another problem to run the pytorch-gpgpu-sim. (Issue #111)
Your problem may be the next step of my situation...

from gpgpu-sim_distribution.

Hello, do you have any progress in this issue?
After I solve Issue #111, I've met similar issue with you.

I set several environment values as following.

PYTORCH_BIN=/usr/local/cuda/lib64/libcudnn.so
LD_LIBRARY_PATH=/home/ubuntu/gpgpu-sim_distribution/lib/gcc-5.4.0/cuda-8000/release/:/usr/local/cuda/lib64

And I removed my original libcudart.so from /usr/local/cuda/lib64 directory and replace it with my libcudart.so (the result library file of gpgpu-sim_distribution).

However, when I run my pytorch model code(written in python, including only one convolution layer), I met following message.

GPGPU-Sim: detected inactive GPU simulation thread
GPGPU-Sim: synchronize waiting for inactive GPU simulation
GPGPU-Sim API: Stream Manager State
GPGPU-Sim: detected inactive GPU simulation thread
Done my python code
GPGPU-Sim: *** exit detected ***

I could not met "no NVIDIA deviced found" nor "CuDNN error: CUDNN_STATUS_NOT_INITIALIZED", but any cudaLaunch() callback in GPGPU-Sim is not executed.

I guess that dynamically-linked libcudnn.so does not call libcudart.so with GPGPU-Sim, but it calls original libcudart.so...

Do you have any idea?

from gpgpu-sim_distribution.