chrishkchris opened a new pull request #754:
URL: https://github.com/apache/singa/pull/754


   1. When I was building conda package of singa-gpu, I endountered the 
following error
   
   
![building](https://user-images.githubusercontent.com/38325429/86128724-53546000-bb14-11ea-8c1c-42bf5bf79b5c.png)
   
   One available solution is to call `roundf` in CUDA MATH API instead of 
`std::round `
   
https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__SINGLE.html#group__CUDA__MATH__SINGLE_1ga1c1521079e51b4f54771b16a7f8aeea
   
   With this modification, the round test case can pass:
   ```
   root@e7f452d7c67a:~/dcsysh/singa/test/python# python3 test_api.py
   .................................
   ----------------------------------------------------------------------
   Ran 33 tests in 1.409s
   
   OK
   ```
   By the way, in the docker version of singa, it does not need this change, 
maybe it is the difference of GCC version in the two environments.
   
   2. This PR updates the numpy to 1.18.5
   
   3. Seems the `cudaDeviceReset()` in destructor returns "driver shutting 
down" code, we don't need the `CUDA_CHECK`.
   
   ```
   root@e7f452d7c67a:~/dcsysh/singa/test/python# python3 test_api.py
   .................................
   ----------------------------------------------------------------------
   Ran 33 tests in 1.383s
   
   OK
   WARNING: Logging before InitGoogleLogging() is written to STDERR
   F0630 12:40:39.892601  3077 cuda_gpu.cc:49] Check failed: error == 
cudaSuccess (29 vs. 0)  driver shutting down
   *** Check failure stack trace: ***
   Aborted (core dumped)
   ```
   After changed, it works well
   ```
   root@e7f452d7c67a:~/dcsysh/singa/test/python# python3 test_api.py
   .................................
   ----------------------------------------------------------------------
   Ran 33 tests in 1.564s
   
   OK
   ```
   


----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to