singa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "ASF subversion and git services (JIRA)" <>
Subject [jira] [Commented] (SINGA-182) CLean math function APIs and implementations
Date Sun, 12 Jun 2016 07:28:21 GMT


ASF subversion and git services commented on SINGA-182:

Commit 6d69047addc46e5c9f381b7e1d4cebd20ce9b2e3 in incubator-singa's branch refs/heads/dev
from [~flytosky]
[;h=6d69047 ]

SINGA-182 Clean math function APIs and implementations

Merge branch 'cuda' from #jinyangturbo.
Clean the cuda related code (tensor_math_cuda.h, kernel_math.h and
by unify the function arugments (names and arg order).
Need to reorder the functions.
Add Nrm2 for L2 norm using cblas and cublas.

> CLean math function APIs and implementations
> --------------------------------------------
>                 Key: SINGA-182
>                 URL:
>             Project: Singa
>          Issue Type: Improvement
>            Reporter: wangwei
> Since we are supporting different types of hardware devices using corresponding programming
languages, e.g., cpp, cuda and opencl,
> we need different math function implementations.
> It is important to make all math functions consistent in terms of their function APIs
and variable names. Here are some guides to make them consistent,
> 1. All function names should be like XxxYyy or XY, i.e., capitablize the first letter.
> 2. Order functions based on function name in alphabetical order.            
> 3. Function arguments order is {code}[const basic type] [const Blob] [mutable Blob]{code}
> 4. Function argument names, use 'num' for total number of elements in   elementwise operations;
use 'in1' 'in2' for input blobs; use 'out' for    output blob or value. With exceptions for
some functions, e.g.,          
>     {code} void Scale(const float alpha, const Blob* in, Blob* out);                
>   For such cases, use v, alpha, etc for scalar types.                   
>    For blas functions, follow the BLAS conventions for argument names.   
> 5. In the implementation, for Blob argument xxx, name its raw pointer as xxxPtr.
> 6. Add an argument 'Context *ctx' for every cuda kernel function.
> 7. Name the kernel functions as KernelXxxx.
> 8. Use size_t for number of elements, number of rows or columns.
> 9. Use the same name for the Tensor and Blob level math functions.

This message was sent by Atlassian JIRA

View raw message