大多数情况，mathematica慢跟使用不当有关

原始引用

wolfram 博客文章：高效mathematica编程的10个技巧

下面这个应该是同一作者对同一主题的升级版的介绍。不过，好用的内容比原来少了，不再是10个技巧了；不过那些以往提到的10个技巧仍旧好用。此外，还增加的有PackedArray, SparseArray之类的新技巧。

视频培训资料可以这里试试

相关的notebook下载

这个原版英文的最权威，不过已经有不少中文翻译版本。个人感觉：明白了导致慢的原因和纠正的办法之后，大部分问题，更喜欢用mathematica，而不是matlab或maple。

当然，mathematica在少数问题上有局限的情况下，另当别论。不过这类情况较少。

中文翻译

百度贴吧Mathematica吧是个不错的地方，少见的繁荣。
其中有一个转帖的中文翻译，推荐

这个翻译来自另一个博客文章。

英文复制

When people tell me that Mathematica isn’t fast enough, I usually ask to see the offending code and often find that the problem isn’t a lack in Mathematica‘s performance, but sub-optimal use of Mathematica. I thought I would share the list of things that I look for first when trying to optimize Mathematica code.

1. Use floating-point numbers if you can, and use them early.

如果只是数值计算，及早使用浮点数（机器精度）。
Of the most common issues that I see when I review slow code is that the programmer has inadvertently asked Mathematica to do things more carefully than needed. Unnecessary use of exact arithmetic is the most common case.

In most numerical software, there is no such thing as exact arithmetic. 1/3 is the same thing as 0.33333333333333. That difference can be pretty important when you hit nasty, numerically unstable problems, but in the majority of tasks, floating-point numbers are good enough and, importantly, much faster. In Mathematica any number with a decimal point and less than 16 digits of input is automatically treated as a machine float, so always use the decimal point if you want speed ahead of accuracy (e.g. enter a third as 1./3.). Here is a simple example where working with floating-point numbers is nearly 50.6 times faster than doing the computation exactly and then converting the result to a decimal afterward. And in this case it gets the same result.

N[Det[Table[1/(1 + Abs[i - j]), {i, 1, 150}, {j, 1, 150}]]] // AbsoluteTiming

{3.9469012, 9.30311*10^-21}

Det[Table[1/(1. + Abs[i - j]), {i, 1., 150.}, {j, 1., 150.}]] // AbsoluteTiming

{0.0780020, 9.30311x10^-21}

The same is true for symbolic computation. If you don’t care about the symbolic answer and are not worried about stability, then substitute numerical values as soon as you can. For example, solving this polynomial symbolically before substituting the values in causes Mathematica to produce a five-page-long intermediate symbolic expression.

Solve[a x^4 + b x^3 + c x + d = 0, x] /. {a → 2., b → 4., c → 7., d → 11.} // AbsoluteTiming

{0.1872048, {{x → -2.20693}, {x → -1.1843}, {x → 0.695616 - 1.27296 &#F74E;}, {x → 0.695616 + 1.27296 I}}}

But do the substitution first, and Solve will use fast numerical methods.

Solve[a x^4 + b x^3 + c x + d = 0 /. {a → 2., b → 4., c → 7., d → 11.}, x] // AbsoluteTiming

{0.0468012, {{x → -2.20693}, {x → -1.1843}, {x → 0.695616 - 1.27296 i}, {x → 0.695616 + 1.27296 i}}}

When working with lists of data, be consistent in your use of reals. It only takes one exact value to cause the whole dataset to have to be held in a more flexible but less efficient form.

data = RandomReal[1, {1000000}]; ByteCount[Append[data, 1.]]

8000176

ByteCount[Append[data, 1]]

32000072 

2. Learn about Compile…

学习对代码进行编译的Compile函数
The Compile function takes Mathematica code and allows you to pre-declare the types (real, complex, etc.) and structures (value, list, matrix, etc.) of input arguments. This takes away some of the flexibility of the Mathematica language, but freed from having to worry about “What if the argument was symbolic?” and the like, Mathematica can optimize the program and create a byte code to run on its own virtual machine. Not everything can be compiled, and very simple code might not benefit, but complex low-level numerical code can get a really big speedup.

Here is an example:

arg = Range[ -50., 50, 0.25];

fn = Function[{x}, Block[{sum = 1.0, inc = 1.0}, Do[inc = inc*x/i; sum = sum + inc, {i, 10000}]; sum]];

Map[fn, arg]; // AbsoluteTiming

{21.5597528, Null}

Using Compile instead of Function makes the execution over 80 times faster.

cfn = Compile[{x}, Block[{sum = 1.0, inc = 1.0}, Do[inc = inc*x/i; sum = sum + inc, {i, 10000}]; sum]];

Map[cfn, arg]; // AbsoluteTiming

{0.2652068, Null}

But we can go further by giving Compile some hints about the parallelizable nature of the code, getting an even better result.

cfn2 = Compile[{x}, Block[{sum = 1.0, inc = 1.0}, Do[inc = inc*x/i; sum = sum + inc, {i, 10000}]; sum], RuntimeAttributes → {Listable}, Parallelization → True];

cfn2[arg]; // AbsoluteTiming

{0.1404036, Null}

On my dual-core machine I get a result 150 times faster than the original; the benefit would be even greater with more cores.

Be aware though that many Mathematica functions like Table, Plot, NIntegrate, and so on automatically compile their arguments, so you won’t see any improvement when passing them compiled versions of your code.

2.5. …and use Compile to generate C code.

编译成C代码、执行效率更高
Furthermore, if your code is compilable, then you can also use the option CompilationTarget->“C” to generate C code, call your C compiler to compile it to a DLL, and link the DLL back into Mathematica, all automatically. There is more overhead in the compilation stage, but the DLL runs directly on your CPU, not on the Mathematica virtual machine, so the results can be even faster.

cfn2C = Compile[{x}, Block[{sum = 1.0, inc = 1.0}, Do[inc = inc*x/i; sum = sum + inc, {i, 10000}]; sum], RuntimeAttributes → {Listable}, Parallelization → True, CompilationTarget → “C”];

cfn2C[arg]; // AbsoluteTiming

{0.0470015, Null} 

3. Use built-in functions.

尽可能优先考虑内置函数，不要想当然自己写
Mathematica has a lot of functions. More than the average person would care to sit down and learn in one go. So it is not surprising that I often see code where someone has implemented some operation without having realized that Mathematica already knows how to do it. Not only is it a waste of time re-implementing work that is already done, but our guys are paid to worry about what the best algorithms are for different kinds of input and how to implement them efficiently, so most built-in functions are really fast.

If you find something close-but-not-quite-right, then check the options and optional arguments; often they generalize functions to cover many specialized uses or abstracted applications.

Here is such an example. If I have a list of a million 2×2 matrices that I want to turn into a list of a million flat lists of 4 elements, the conceptually easiest way might be to Map the basic Flatten operation down the list of them.

data = RandomReal[1, {1000000, 2, 2}];

Map[Flatten, data]; // AbsoluteTiming

{0.2652068, Null}

But Flatten knows how to do this whole task on its own when you specify that levels 2 and 3 of the data structure should be merged and level 1 be left alone. Specifying such details might be comparatively fiddly, but staying within Flatten to do the whole flattening job turns out to be nearly 4 times faster than re-implementing that sub-feature yourself.

Flatten[data, {{1}, {2, 3}}]; // AbsoluteTiming

{0.0780020, Null}

So remember—do a search in the Help menu before you implement anything.

4. Use Wolfram Workbench.

用专门的集成开发环境（编辑器）调试和优化代码
Mathematica can be quite forgiving of some kinds of programming mistakes—it will proceed happily in symbolic mode if you forget to initialize a variable at the right point and doesn’t care about recursion or unexpected data types. That’s great when you just need to get a quick answer, but it will also let you get away with less than optimal solutions without realizing it.

Workbench helps in several ways. First it lets you debug and organize large code projects better, and having clean, organized code should make it easier to write good code. But the key feature in this context is the profiler that lets you see which lines of code used up the time, and how many times they were called.

Take this example, a truly horrible way (computationally speaking) to implement Fibonacci numbers. If you didn’t think about the consequences of the double recursion, you might be surprised by the 22 seconds it takes to evaluate fib[35] (about the same time it takes the built-in function to calculate all 208,987,639 digits of Fibonacci[1000000000] [see tip 3]).

fib[n_] := fib[n - 1] + fib[n - 2]; fib[1] = 1; fib[2] = 1; fib[35]; // AbsoluteTiming

{22.3709736, Null}

Running the code in the profiler reveals the reason. The main rule is invoked 9,227,464 times, and the fib[1] value is requested 18,454,929 times.

Being told what your code really does, rather than what you thought it would do, can be a real eye-opener.

5. Remember values that you will need in the future.

保留那些还会在代码用到的值或赋值
This is good programming advice in any language. The Mathematica construct that you will want to know is this:

f[x_] := f[x] =(What the function does)

It saves the result of calling f on any value, so that if it is called again on the same value, Mathematica will not need to work it out again. You are trading speed for memory here, so it isn’t appropriate if your function is likely to be called for a huge number of values, but rarely the same ones twice. But if the possible input set is constrained, this can really help. Here it is rescuing the program that I used to illustrate tip 3. Change the first rule to this:

fib[n_] := fib[n] = fib[n - 1] + fib[n - 2];

And it becomes immeasurably fast, since fib[35] now only requires the main rule to be evaluated 33 times. Looking up previous results prevents the need to repeatedly recurse down to fib[1].

6. Parallelize.

尝试并行化
An increasing number of Mathematica operations will automatically parallelize over local cores (most linear algebra, image processing, and statistics), and, as we have seen, so does Compile if manually requested. But for other operations, or if you want to parallelize over remote hardware, you can use the built-in parallel programming constructs.

There is a collection of tools for this, but for very independent tasks, you can get quite a long way with just ParallelTable, ParallelMap, and ParallelTry. Each of these automatically takes care of communication, worker management, and collection of results. There is some overhead for sending the task and retrieving the result, so there is a trade-off of time gained versus time lost. Your Mathematica comes with four compute kernels, and you can scale up with gridMathematica if you have access to additional CPU power. Here, ParallelTable gives me double the performance, since it is running on my dual-core machine. More CPUs would give a better speedup.

Table[PrimeQ[x], {x, 10^1000, 10^1000 + 5000}]; // AbsoluteTiming

{8.8298264, Null}

ParallelTable[PrimeQ[x], {x, 10^1000, 10^1000 + 5000}]; // AbsoluteTiming

{4.9921280, Null}

Anything that Mathematica can do, it can also do in parallel. For example, you could send a set of parallel tasks to remote hardware, each of which compiles and runs in C or on a GPU.

6.5. Think about CUDALink and OpenCLLink.

如果显卡支持，Mathematica也提供了这类并行计算的可能
If you have GPU hardware, there are some really fast things you can do with massive parallelization. Unless one of the built-in CUDA-optimized functions happens to do what you want, you will need to do a little work, but the CUDALink and OpenCLLink tools automate a lot of the messy details for you.

7. Use Sow and Reap to accumulate large amounts of data (not AppendTo).

这个我从来没用过，也不懂；不过似乎跟Mathematica特殊的数据结构有关的。看百度贴吧转帖的翻译，也昏昏昭昭。应该是如果在循环中需要改变list的大小，和matlab类似，都会导致低效。参考其代码例子的使用方法即可。

Because of the flexibility of Mathematica data structures, AppendTo can’t assume that you will be appending a number, because you might equally append a document or a sound or an image. As a result, AppendTo must create a fresh copy of all of the data, restructured to accommodate the appended information. This makes it progressively slower as the data accumulates. (And the construct data=Append[data,value] is the same as AppendTo.)

Instead use Sow and Reap. Sow throws out the values that you want to accumulate, and Reap collects them and builds a data object once at the end. The following are equivalent:

低效的代码：

data = {}; Do[AppendTo[data, RandomReal[x]], {x, 0, 40000}]; // AbsoluteTiming

{5.8813508, Null}

使用Reap和Sow这一对函数组合提高效率之后的代码：

data = Reap[Do[Sow[RandomReal[x]], {x, 0, 40000}]][[2]]; // AbsoluteTiming

效率提高50多倍：

{0.1092028, Null} 

8. Use Block or With rather than Module.

对于局部代码段，使用Block或With而不是Module（为啥不去掉它？）,
Block, With, and Module are all localization constructs with slightly different properties. In my experience, Block and Module are interchangeable in at least 95% of code that I write, but Block is usually faster, and in some cases With (effectively Block with the variables in a read-only state) is faster still.

Do[Module[{x = 2.}, 1/x], {1000000}]; // AbsoluteTiming

{4.1497064, Null}

Do[Block[{x = 2.}, 1/x], {1000000}]; // AbsoluteTiming

效率提高3倍多：

{1.4664376, Null} 

9. Go easy on pattern matching.

模式匹配的代码低效，尽可能少用。

Pattern matching is great. It can make complicated tasks easy to program. But it isn’t always fast, especially the fuzzier patterns like BlankNullSequence (usually written as “_”), which can search long and hard through your data for patterns that you, as a programmer, might already know will never be there. If execution speed matters, use tighter patterns, or none at all.

As an example, here is a rather neat way to implement a bubble sort in a single line of code using patterns:

data = RandomReal[1, {200}]; data //. {a___, b_, c_, d___} /; b > c → {a, c, b, d}; // AbsoluteTiming

{2.5272648, Null}

Conceptually neat, but slow compared to this procedural approach that I was taught when I first learned programming:

(flag = True; While[TrueQ[flag], flag = False; Do[If[data[[i]] > data[[i + 1]], temp = data[[i]]; data[[i]] = data[[i + 1]]; data[[i + 1]] = temp; flag = True], {i, 1, Length[data] - 1}]]; data); // AbsoluteTiming

{0.1716044, Null}

Of course in this case you should use the built-in function (see tip 3), which will use better sorting algorithms than bubble sort.

Sort[RandomReal[1, {200}]]; // AbsoluteTiming

{0., Null} 

10. Try doing things differently.

这个软件的语法诡异，需要多尝试Mathematica独特的编程风格，而不是常规高级语言的解决问题的思路。
http://mathprogramming-intro.org/
这个教程Mathematica® Programming - an advanced introduction非常好，推荐。

One of Mathematica‘s great strengths is that it can tackle the same problem in different ways. It allows you to program the way you think, as opposed to reconceptualizing the problem for the style of the programming language. However, conceptual simplicity is not always the same as computational efficiency. Sometimes the easy-to-understand idea does more work than is necessary.

But another issue is that because special optimizations and smart algorithms are applied automatically in Mathematica, it is often hard to predict when something clever is going to happen. For example, here are two ways of calculating factorial, but the second is over 10 times faster.

temp = 1; Do[temp = temp i, {i, 2^16}]; // AbsoluteTiming

{0.8892228, Null}

效率提升：

Apply[Times, Range[2^16]]; // AbsoluteTiming

{0.0624016, Null}

Why? You might guess that the Do loop is slow, or all those assignments to temp take time, or that there is something else “wrong” with the first implementation, but the real reason is probably quite unexpected. Times knows a clever binary splitting trick that can be used when you have a large number of integer arguments. It is faster to recursively split the arguments into two smaller products, (1*2*…32767)(32768*…*65536), rather than working through the arguments from first to last. It still has to do the same number of multiplications, but fewer of them involve very big integers, and so, on average, are quicker to do. There are lots of such pieces of hidden magic in Mathematica, and more get added with each release.

Of course the best way here is to use the built-in function (tip 3 again):

AbsoluteTiming[65536!;]

{0.0156004, Null}

Mathematica is capable of superb computational performance, and also superb robustness and accuracy, but not always both at the same time. I hope that these tips will help you to balance the sometimes conflicting needs for rapid programming, rapid execution, and accurate results.

Download this post as a Computable Document Format (CDF) file.

All timings use a Windows 7 64-bit PC with 2.66 GHz Intel Core 2 Duo and 6 GB RAM.