SMOP
是小型Matlab和八度到Python编译器。SMOP
将matlab翻译成python。尽管matlab和数字python之间有明显的相似之处,但在现实生活中有足够的差异使手工翻译不可行。 SMOP
生成人类可读的蟒蛇,这似乎也比八度快。速度有多快?表1显示了“移动家具”的计时结果。似乎对于该程序,转换为python导致加速大约两倍,并且 使用cython 将SMOP
运行时库编译runtime.py
为C 实现了额外的两倍加速。这个伪基准测量标量性能,而我的解释是标量计算对八度组不太感兴趣。源代码
Working example
$ cd smop/smop $ python main.py solver.m $ python solver.py
We will translate solver.m
to present a sample of smop features. The program was borrowed from the matlab programming competition in 2004 (Moving Furniture).To the left is solver.m
. To the right is a.py
--- its translation to python. Though only 30 lines long, this example shows many of the complexities of converting matlab code to python.
01 function mv = solver(ai,af,w) 01 def solver_(ai,af,w,nargout=1):
02 nBlocks = max(ai(:)); 02 nBlocks=max_(ai[:])
03 [m,n] = size(ai); 03 m,n=size_(ai,nargout=2)
02 | Matlab uses round brackets both for array indexing and for function calls. To figure out which is which, SMOP computes local use-def information, and then applies the following rule: undefined names are functions, while defined are arrays. |
03 | Matlab function size returns variable number of return values, which corresponds to returning a tuple in python. Since python functions are unaware of the expected number of return values, their number must be explicitly passed in nargout . |
04 I = [0 1 0 -1]; 04 I=matlabarray([0,1,0,- 1])
05 J = [1 0 -1 0]; 05 J=matlabarray([1,0,- 1,0])
06 a = ai; 06 a=copy_(ai)
07 mv = []; 07 mv=matlabarray([])
04 | Matlab array indexing starts with one; python indexing starts with zero. New class matlabarray derives fromndarray , but exposes matlab array behaviour. For example, matlabarray instances always have at least two dimensions -- the shape of I and J is [1 4]. |
06 | Matlab array assignment implies copying; python assignment implies data sharing. We use explicit copy here. |
07 | Empty matlabarray object is created, and then extended at line 28. Extending arrays by out-of-bounds assignment is deprecated in matlab, but is widely used never the less. Python ndarray can't be resized except in some special cases. Instances of matlabarray can be resized except where it is too expensive. |
08 while ~isequal(af,a) 08 while not isequal_(af,a):
09 bid = ceil(rand*nBlocks); 09 bid=ceil_(rand_() * nBlocks)
10 [i,j] = find(a==bid); 10 i,j=find_(a == bid,nargout=2)
11 r = ceil(rand*4); 11 r=ceil_(rand_() * 4)
12 ni = i + I(r); 12 ni=i + I[r]
13 nj = j + J(r); 13 nj=j + J[r]
09 | Matlab functions of zero arguments, such as rand , can be used without parentheses. In python, parentheses are required. To detect such cases, used but undefined variables are assumed to be functions. |
10 | The expected number of return values from the matlab function find is explicitly passed in nargout . |
12 | Variables I and J contain instances of the new class matlabarray , which among other features uses one based array indexing. |
14 if (ni<1) || (ni>m) || 14 if (ni < 1) or (ni > m) or
(nj<1) || (nj>n) (nj < 1) or (nj > n):
15 continue 15 continue
16 end 16
17 if a(ni,nj)>0 17 if a[ni,nj] > 0:
18 continue 18 continue
19 end 19
20 [ti,tj] = find(af==bid); 20 ti,tj=find_(af == bid,nargout=2)
21 d = (ti-i)^2 + (tj-j)^2; 21 d=(ti - i) ** 2 + (tj - j) ** 2
22 dn = (ti-ni)^2 + (tj-nj)^2; 22 dn=(ti - ni) ** 2 + (tj - nj) ** 2
23 if (d<dn) && (rand>0.05) 23 if (d < dn) and (rand_() > 0.05):
24 continue 24 continue
25 end 25
26 a(ni,nj) = bid; 26 a[ni,nj]=bid
27 a(i,j) = 0; 27 a[i,j]=0
28 mv(end+1,[1 2]) = [bid r]; 28 mv[mv.shape[0] + 1,[1,2]]=[bid,r]
29 end 29
30 30 return mv
Implementation status
Random remarks
- With less than five thousands lines of python code
-
SMOP
does not pretend to compete with such polished products as matlab or octave. Yet, it is not a toy. There is an attempt to follow the original matlab semantics as close as possible. Matlab language definition (never published afaik) is full of dark corners, andSMOP
tries to follow matlab as precisely as possible. - There is a price, too.
- The generated sources are matlabic, rather than pythonic, which means that library maintainers must be fluent in both languages, and the old development environment must be kept around.
- Should the generated program be pythonic or matlabic?
-
For example should array indexing start with zero (pythonic) or with one (matlabic)?
I beleive now that some matlabic accent is unavoidable in the generated python sources. Imagine matlab program is using regular expressions, matlab style. We are not going to translate them to python style, and that code will remain forever as a reminder of the program's matlab origin.
Another example. Matlab code opens a file; fopen returns -1 on error. Pythonic code would raise exception, but we are not going to do that. Instead, we will live with the accent, and smop takes this to the extreme --- the matlab program remains mostly unchanged.
It turns out that generating matlabic` allows for moving much of the project complexity out of the compiler (which is already complicated enough) and into the runtime library, where there is almost no interaction between the library parts.
- Which one is faster --- python or octave? I don't know.
-
Doing reliable performance measurements is notoriously hard, and is of low priority for me now. Instead, I wrote a simple driver
go.m
andgo.py
and rewrote rand so that python and octave versions run the same code. Then I ran the above example on my laptop. The results are twice as fast for the python version. What does it mean? Probably nothing. YMMV.
ai = zeros(10,10);
af = ai;
ai(1,1)=2;
ai(2,2)=3;
ai(3,3)=4;
ai(4,4)=5;
ai(5,5)=1;
af(9,9)=1;
af(8,8)=2;
af(7,7)=3;
af(6,6)=4;
af(10,10)=5;
tic;
mv = solver(ai,af,0);
toc
Running the test suite:
$ cd smop $ make check $ make test
Command-line options
lei@dilbert ~/smop-github/smop $ python main.py -h
SMOP compiler version 0.25.1
Usage: smop [options] file-list
Options:
-V --version
-X --exclude=FILES Ignore files listed in comma-separated list FILES
-d --dot=REGEX For functions whose names match REGEX, save debugging
information in "dot" format (see www.graphviz.org).
You need an installation of graphviz to use --dot
option. Use "dot" utility to create a pdf file.
For example:
$ python main.py fastsolver.m -d "solver|cbest"
$ dot -Tpdf -o resolve_solver.pdf resolve_solver.dot
-h --help
-o --output=FILENAME By default create file named a.py
-o- --output=- Use standard output
-s --strict Stop on the first error
-v --verbose