update roundeven backend algorithm

[dcslin]

No description provided.

[dcslin]

HI @joddiy , this is updated according to the round even requirement, could you please help to review?

[chrishkchris]

error message at math_kernel.cu line 143:

[ 14%] Building NVCC (Device) object src/CMakeFiles/cuda_compile_1.dir/core/tensor/cuda_compile_1_generated_math_kernel.cu.o
/root/miniconda/conda-bld/singa_1593705642277/work/src/core/tensor/math_kernel.cu(143): error: explicit type is missing ("int" assu                                                               med)

1 error detected in the compilation of "/tmp/tmpxft_00000574_00000000-6_math_kernel.cpp1.ii".
CMake Error at cuda_compile_1_generated_math_kernel.cu.o.cmake:280 (message):
  Error generating file
  /root/miniconda/conda-bld/singa_1593705642277/work/build/src/CMakeFiles/cuda_compile_1.dir/core/tensor/./cuda_compile_1_generated                                                               _math_kernel.cu.o

it needs explicit type instead of auto?

Maybe I try later version of GCC first

[chrishkchris]

error message at math_kernel.cu line 143:

[ 14%] Building NVCC (Device) object src/CMakeFiles/cuda_compile_1.dir/core/tensor/cuda_compile_1_generated_math_kernel.cu.o
/root/miniconda/conda-bld/singa_1593705642277/work/src/core/tensor/math_kernel.cu(143): error: explicit type is missing ("int" assu                                                               med)

1 error detected in the compilation of "/tmp/tmpxft_00000574_00000000-6_math_kernel.cpp1.ii".
CMake Error at cuda_compile_1_generated_math_kernel.cu.o.cmake:280 (message):
  Error generating file
  /root/miniconda/conda-bld/singa_1593705642277/work/build/src/CMakeFiles/cuda_compile_1.dir/core/tensor/./cuda_compile_1_generated                                                               _math_kernel.cu.o

it needs explicit type instead of auto?

Maybe I try later version of GCC first

Resolved after upgrading C and CXX complier versions in #759
-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0

[chrishkchris]

This auto causes problem in CXX complier version 5.4, upgraded to 7.5.0 will then be okay. Resolved in #759

[dcslin]

Thank you Chris for the prompt response! I could also replace the auto in this PR. Yes I think updating to newer gcc is a good idea, although the "new"(c++11) one is already widely used and not that new, it would be helpful to support running c++11 code.

[chrishkchris]

I guess there is no need to change the auto, auto seems to be a good way to prevent implicit type casting (sometime developer us could be careless). The only problem of auto is just the code readability for the developer to know the exact datatype.

In any case, I think need to keep the same version of complier between the docker version and conda version (as in #759), because a few day ago we had the similar complier error because of the version different.

[chrishkchris]

@dcslin The test case looks good to me
https://blogs.sas.com/content/iml/2019/11/11/round-to-even.html
As the example given in the link above, rounde rounds 0.5 to 0, and 2.5 to 2
The test case passed so it should be correct.

root@71ac539cda77:~/dcsysh/singa/test/python# python3 test_api.py
.................................
----------------------------------------------------------------------
Ran 33 tests in 0.758s

OK