Open sukamenev opened 6 months ago
Is it 32 or 64 bit atchitecture? need to track down which kernel fails.
I also suggest to try AMD official drivers and not Mesa only.
I recall that for AMD 560 closed source drivers worked way better than Mesa ones. Also check of ROCm drivers still work on Fiji they are also better.
Is it 32 or 64 bit atchitecture? need to track down which kernel fails.
My CPU have 64 bit architecture. GCN 3 (Fiji) - I don't know how many bit architecture.
Quote from AMD docs:
Every instruction is described with either 32 bits or 64 bits of microcode. • Vector Memory instructions are 64 bits. • Exports are 64 bits. • LDS and GDS are 64 bits. • Scalar ALU instructions are 32 bits but can have an additional 32 bits of literal constant data. • Vector ALU instructions can be 32 bits or 64 bits. The 32-bit versions can have an additional 32 bits of literal constant data.
On AMD OpenCL from amdgpu-pro also error
python tests/validate_network.py --device privateuseone:3
Testing resnet18
Accessing device #3:Fiji on AMD Accelerated Parallel Processing
Traceback (most recent call last):
File "/home/inetstar/Kamenev/programming/ZenDnn/pytorch_dlprim/tests/validate_network.py", line 280, in <module>
main(r)
File "/home/inetstar/Kamenev/programming/ZenDnn/pytorch_dlprim/tests/validate_network.py", line 221, in main
train_on_images(m,batch,args.device,args.eval,iter_size = args.iter_size,opt_steps = args.opt,fwd=args.fwd)
File "/home/inetstar/Kamenev/programming/ZenDnn/pytorch_dlprim/tests/validate_network.py", line 105, in train_on_images
ref = step(model,data,labels,opt_steps,iter_size,fwd=fwd,test=test)
File "/home/inetstar/Kamenev/programming/ZenDnn/pytorch_dlprim/tests/validate_network.py", line 85, in step
loss.backward()
File "/home/inetstar/Kamenev/programming/ZenDnn/lib/python3.10/site-packages/torch/_tensor.py", line 488, in backward
torch.autograd.backward(
File "/home/inetstar/Kamenev/programming/ZenDnn/lib/python3.10/site-packages/torch/autograd/__init__.py", line 197, in backward
Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: could not create a primitive descriptor iterator
I also suggest to try AMD official drivers and not Mesa only.
I recall that for AMD 560 closed source drivers worked way better than Mesa ones. Also check of ROCm drivers still work on Fiji they are also better.
Thank you! I got 8-9% speed impovement on amdgpu-pro OpenCL drivers.
Tested on your original code