https://github.com/dubzzz/cuda-compile-4-cuda-and-noncuda
Design a CUDA-program able to run on a non-CUDA computer. When dealing with a non-CUDA computer it should use CPU version of the code to do the same operation.
https://github.com/dubzzz/cuda-compile-4-cuda-and-noncuda
Last synced: 7 months ago
JSON representation
Design a CUDA-program able to run on a non-CUDA computer. When dealing with a non-CUDA computer it should use CPU version of the code to do the same operation.
- Host: GitHub
- URL: https://github.com/dubzzz/cuda-compile-4-cuda-and-noncuda
- Owner: dubzzz
- License: mit
- Created: 2014-06-06T17:13:35.000Z (over 11 years ago)
- Default Branch: master
- Last Pushed: 2014-06-07T17:20:15.000Z (over 11 years ago)
- Last Synced: 2025-01-25T22:39:05.747Z (9 months ago)
- Language: Python
- Size: 188 KB
- Stars: 0
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
#Compile 4 CUDA & Non-CUDA
Design a CUDA-program able to run on a non-CUDA computer. When dealing with a non-CUDA computer it should use CPU version of the code to do the same operation.
##How does it work?
This code is designed to build a Python extension. However if you want to create a standalone program the same technics apply:
1. tweak the Makefile to check for CUDA
2. if CUDA: use nvcc
3. if not CUDA: use g++ or gcc with the option '-x c++ -c' in front of each *.cu file (for c++ code) / '-x c -c' (for c code)
Please refer to `src/Makefile` for further details.
Please note that you also need to adapt your C/C++ sourcecode in order to be able to cope with cases where CUDA does not exist. In order to do so I used the `__CUDACC__` variable which is defined if and only if using nvcc.
When compiling with another compiler (different from nvcc), every reference to CUDA code need to be removed. Please refer to `src/cudanoncuda-src.cu`.
##Special cases
Even if nvcc exists in the system, it is not sure that it can be used for GPU computations.
Compiling with `nvcc` or having a copy of `libcudart.so` does not prove that the program will be able to launch a kernel on your system. Moreover launching a CUDA program and having `libcudart.so` on the computer will not raise any exceptions when launching kernels, except cuda errors which need to be handle manually. For that reason, the C/C++ code needs to check GPU cards at runtime.
##Test it!
Here is a sample code to test the behaviour of the generated library:
```Python
import numpy as np
import cudanoncuda as cnc
# Create vectors
va = np.random.random(5)
vb = np.random.random(5)
# Use the library
# Depending on your environment it should run CUDA or CPU code
vc = cnc.add(va, vb)
```
Remark: a code compiled on a computer with `nvcc` on it and executed on another computer without CUDA (or more precisely: without `libcudart.so`) will fail to run.