As it currently stands, this question is not a good fit for our Q&A format. We expect answers to be supported by facts, references, or expertise, but this question will likely solicit debate, arguments, polling, or extended discussion. If you feel that this question can be improved and possibly reopened, visit the help center for guidance.
Closed 10 years ago.
Some years ago I was on a panel that was interviewing candidates for a relatively senior embedded C programmer position.
One of the standard questions that I asked was about optimisation techniques. I was quite surprised that some of the candidates didn't have answers.
So, in the interests of putting together a list for posterity - what techniques and constructs do you normally use when optimising C programs?
Answers to optimisation for speed and size both accepted.
First things first - don't optimise too early. It's not uncommon to spend time carefully optimising a chunk of code only to find that it wasn't the bottleneck that you thought it was going to be. Or, to put it another way "Before you make it fast, make it work"
Investigate whether there's any option for optimising the algorithm before optimising the code. It'll be easier to find an improvement in performance by optimising a poor algorithm than it is to optimise the code, only then to throw it away when you change the algorithm anyway.
And work out why you need to optimise in the first place. What are you trying to achieve? If you're trying, say, to improve the response time to some event work out if there is an opportunity to change the order of execution to minimise the time critical areas. For example when trying to improve the response to some external interrupt can you do any preparation in the dead time between events?
Once you've decided that you need to optimise the code, which bit do you optimise? Use a profiler. Focus your attention (first) on the areas that are used most often.
So what can you do about those areas?
minimise condition checking. Checking conditions (eg. terminating conditions for loops) is time that isn't being spent on actual processing. Condition checking can be minimised with techniques like loop-unrolling.
In some circumstances condition checking can also be eliminated by using function pointers. For example if you are implementing a state machine you may find that implementing the handlers for individual states as small functions (with a uniform prototype) and storing the "next state" by storing the function pointer of the next handler is more efficient than using a large switch statement with the handler code implemented in the individual case statements. YMMV.
minimise function calls. Function calls usually carry a burden of context saving (eg. writing local variables contained in registers to the stack, saving the stack pointer), so if you don't have to make a call this is time saved. One option (if you're optimising for speed and not space) is to make use of inline functions.
If function calls are unavoidable minimise the data that is being passed to the functions. For example passing pointers is likely to be more efficient than passing structures.
When optimising for speed choose datatypes that are the native size for your platform. For example on a 32bit processor it is likely to be more efficient to manipulate 32bit values than 8 or 16 bit values. (side note - it is worth checking that the compiler is doing what you think it is. I've had situations where I've discovered that my compiler insisted on doing 16 bit arithmetic on 8 bit values with all of the to and from conversions to go with them)
Find data that can be precalculated, and either calculate during initialisation or (better yet) at compile time. For example when implementing a CRC you can either calculate your CRC values on the fly (using the polynomial directly) which is great for size (but dreadful for performance), or you can generate a table of all of the interim values - which is a much faster implementation, to the detriment of the size.
Localise your data. If you're manipulating a blob of data often your processor may be able to speed things up by storing it all in cache. And your compiler may be able to use shorter instructions that are suited to more localised data (eg. instructions that use 8 bit offsets instead of 32 bit)
In the same vein, localise your functions. For the same reasons.
Work out the assumptions that you can make about the operations that you're performing and find ways of exploiting them. For example, on an 8 bit platform if the only operation that at you're doing on a 32 bit value is an increment you may find that you can do better than the compiler by inlining (or creating a macro) specifically for this purpose, rather than using a normal arithmetic operation.
Avoid expensive instructions - division is a prime example.
The "register" keyword can be your friend (although hopefully your compiler has a pretty good idea about your register usage). If you're going to use "register" it's likely that you'll have to declare the local variables that you want "register"ed first.
Be consistent with your data types. If you are doing arithmetic on a mixture of data types (eg. shorts and ints, doubles and floats) then the compiler is adding implicit type conversions for each mismatch. This is wasted cpu cycles that may not be necessary.
Most of the options listed above can be used as part of normal practice without any ill effects. However if you're really trying to eke out the best performance:
- Investigate where you can (safely) disable error checking. It's not recommended, but it will save you some space and cycles.
- Hand craft portions of your code in assembler. This of course means that your code is no longer portable but where that's not an issue you may find savings here. Be aware though that there is potentially time lost moving data into and out of the registers that you have at your disposal (ie. to satisfy the register usage of your compiler). Also be aware that your compiler should be doing a pretty good job on its own. (of course there are exceptions)
As everybody else has said: profile, profile profile.
As for actual techniques, one that I don't think has been mentioned yet:
Hot & Cold Data Separation: Staying within the CPU's cache is incredibly important. One way of helping to do this is by splitting your data structures into frequently accessed ("hot") and rarely accessed ("cold") sections.
An example: Suppose you have a structure for a customer that looks something like this:
struct Customer
{
int ID;
int AccountNumber;
char Name[128];
char Address[256];
};
Customer customers[1000];
Now, lets assume that you want to access the ID and AccountNumber a lot, but not so much the name and address. What you'd do is to split it into two:
struct CustomerAccount
{
int ID;
int AccountNumber;
CustomerData *pData;
};
struct CustomerData
{
char Name[128];
char Address[256];
};
CustomerAccount customers[1000];
In this way, when you're looping through your "customers" array, each entry is only 12 bytes and so you can fit many more entries in the cache. This can be a huge win if you can apply it to situations like the inner loop of a rendering engine.
My favorite technique is to use a good profiler. Without a good profile telling you where the bottleneck lies, no tricks and techniques are going to help you.
most common techniques I encountered are:
loop unrolling
loop optimization for better cache prefetch
(i.e. do N operations in M cycles instead of NxM singular operations)
data aligning
inline functions
hand-crafted asm snippets
As for general recommendations, most of them are already sounded:
choose better algos
use profiler
don't optimize if it doesn't give 20-30% performance boost
For low-level optimization:
START_TIMER/STOP_TIMER macros from ffmpeg (clock-level accuracy for measurement of any code).
Oprofile, of course, for profiling.
Enormous amounts of hand-coded assembly (just do a wc -l on x264's /common/x86 directory, and then remember most of the code is templated).
Careful coding in general; shorter code is usually better.
Smart low-level algorithms, like the 64-bit bitstream writer I wrote that uses only a single if and no else.
Explicit write-combining.
Taking into account important weird aspects of processors, like Intel's cacheline split issue.
Finding cases where one can losslessly or near-losslessly make an early termination, where the early-termination check costs much less than the speed one gains from it.
Actually inlined assembly for tasks which are far more suited to the x86 SIMD unit, such as median calculations (requires compile-time check for MMX support).
First and foremost, use a better/faster algorithm. There is no point optimizing code that is slow by design.
When optimizing for speed, trade memory for speed: lookup tables of precomputed values, binary trees, write faster custom implementation of system calls...
When trading speed for memory: use in-memory compression
Avoid using the heap. Use obstacks or pool-allocator for identical sized objects. Put small things with short lifetime onto the stack. alloca still exists.
Pre-mature optimization is the root of all evil!
;)
As my applications usually don't need much CPU time by design, I focus on the size my binaries on disk and in memory. What I do mostly is looking out for statically sized arrays and replacing them with dynamically allocated memory where it's worth the additional effort of free'ing the memory later. To cut down the size of the binary, I look for big arrays that are initialized at compile time and put the initializiation to runtime.
char buf[1024] = { 0, };
/* becomes: */
char buf[1024];
memset(buf, 0, sizeof(buf));
This will remove the 1024 zero-bytes from the binaries .DATA section and will instead create the buffer on the stack at runtime and the fill it with zeros.
EDIT: Oh yeah, and I like to cache things. It's not C specific but depending on what you're caching, it can give you a huge boost in performance.
PS: Please let us know when your list is finished, I'm very curious. ;)
If possible, compare with 0, not with arbitrary numbers, especially in loops, because comparison with 0 is often implemented with separate, faster assembler commands.
For example, if possible, write
for (i=n; i!=0; --i) { ... }
instead of
for (i=0; i!=n; ++i) { ... }
Another thing that was not mentioned:
Know your requirements: don't optimize for situations that will unlikely or never happen, concentrate on the most bang for the buck
basics/general:
Do not optimize when you have no problem.
Know your platform/CPU...
...know it thoroughly
know your ABI
Let the compiler do the optimization, just help it with the job.
some things that have actually helped:
Opt for size/memory:
Use bitfields for storing bools
re-use big global arrays by overlaying with a union (be careful)
Opt for speed (be careful):
use precomputed tables where possible
place critical functions/data in fast memory
Use dedicated registers for often used globals
count to-zero, zero flag is free
Difficult to summarize ...
Data structures:
Splitting of a data structure depending on case of usage is extremely important. It is common to see a structure that holds data that is accessed based on a flow control. This situation can lower significantly the cache usage.
To take into account cache line size and prefetch rules.
To reorder the members of the structure to obtain a sequential access to them from your code
Algorithms:
Take time to think about your problem and to find the correct algorithm.
Know the limitations of the algorithm you choose (a radix-sort/quick-sort for 10 elements to be sorted might not be the best choice).
Low level:
As for the latest processors it is not recommended to unroll a loop that has a small body. The processor provides its own detection mechanism for this and will short-circuit whole section of its pipeline.
Trust the HW prefetcher. Of course if your data structures are well designed ;)
Care about your L2 cache line misses.
Try to reduce as much as possible the local working set of your application as the processors are leaning to smaller caches per cores (C2D enjoyed a 3MB per core max where iCore7 will provide a max of 256KB per core + 8MB shared to all cores for a quad core die.).
The most important of all: Measure early, Measure often and never ever makes assumptions, base your thinking and optimizations on data retrieved by a profiler (please use PTU).
Another hint, performance is key to the success of an application and should be considered at design time and you should have clear performance targets.
This is far from being exhaustive but should provide an interesting base.
These days, the most important things in optimzation are:
respecting the cache - try to access memory in simple patterns, and don't unroll loops just for fun. Use arrays instead of data structures with lots of pointer chasing and it'll probably be faster for small amounts of data. And don't make anything too big.
avoiding latency - try to avoid divisions and stuff that's slow if other calculations depend on them immediately. Memory accesses that depend on other memory accesses (ie, a[b[c]]) are bad.
avoiding unpredictabilty - a lot of if/elses with unpredictable conditions, or conditions that introduce more latency, will really mess you up. There's a lot of branchless math tricks that are useful here, but they increase latency and are only useful if you really need them. Otherwise, just write simple code and don't have crazy loop conditions.
Don't bother with optimizations that involve copy-and-pasting your code (like loop unrolling), or reordering loops by hand. The compiler usually does a better job than you at doing this, but most of them aren't smart enough to undo it.
Collecting profiles of code execution get you 50% of the way there. The other 50% deals with analyzing these reports.
Further, if you use GCC or VisualC++, you can use "profile guided optimization" where the compiler will take info from previous executions and reschedule instructions to make the CPU happier.
Inline functions! Inspired by the profiling fans here I profiled an application of mine and found a small function that does some bitshifting on MP3 frames. It makes about 90% of all function calls in my applcation, so I made it inline and voila - the program now uses half of the CPU time it did before.
On most of embedded system i worked there was no profiling tools, so it's nice to say use profiler but not very practical.
First rule in speed optimization is - find your critical path.
Usually you will find that this path is not so long and not so complex. It's hard to say in generic way how to optimize this it's depend on what are you doing and what is in your power to do. For example you want usually avoid memcpy on critical path, so ever you need to use DMA or optimize, but what if you hw does not have DMA ? check if memcpy implementation is a best one if not rewrite it.
Do not use dynamic allocation at all in embedded but if you do for some reason don't do it in critical path.
Organize your thread priorities correctly, what is correctly is real question and it's clearly system specific.
We use very simple tools to analyze the bottle-necks, simple macro that store the time-stamp and index. Few (2-3) runs in 90% of cases will find where you spend your time.
And the last one is code review a very important one. In most case we avoid performance problem during code review very effective way :)
Measure performance.
Use realistic and non-trivial benchmarks. Remember that "everything is fast for small N".
Use a profiler to find hotspots.
Reduce number of dynamic memory allocations, disk accesses, database accesses, network accesses, and user/kernel transitions, because these often tend to be hotspots.
Measure performance.
In addition, you should measure performance.
Sometimes you have to decide whether it is more space or more speed that you are after, which will lead to almost opposite optimizations. For example, to get the most out of you space, you pack structures e.g. #pragma pack(1) and use bit fields in structures. For more speed you pack to align with the processors preference and avoid bitfields.
Another trick is picking the right re-sizing algorithms for growing arrays via realloc, or better still writing your own heap manager based on your particular application. Don't assume the one that comes with the compiler is the best possible solution for every application.
If someone doesn't have an answer to that question, it could be they don't know much.
It could also be that they know a lot. I know a lot (IMHO :-), and if I were asked that question, I would be asking you back: Why do you think that's important?
The problem is, any a-priori notions about performance, if they are not informed by a specific situation, are guesses by definition.
I think it is important to know coding techniques for performance, but I think it is even more important to know not to use them, until diagnosis reveals that there is a problem and what it is.
Now I'm going to contradict myself and say, if you do that, you learn how to recognize the design approaches that lead to trouble so you can avoid them, and to a novice, that sounds like premature optimization.
To give you a concrete example, this is a C application that was optimized.
Great lists. I will just add one tip I didn't saw in the above lists that in some case can yield huge optimisation for minimal cost.
bypass linker
if you have some application divided in two files, say main.c and lib.c, in many cases you can just add a \#include "lib.c" in your main.c That will completely bypass linker and allow for much more efficient optimisation for compiler.
The same effect can be achieved optimizing dependencies between files, but the cost of changes is usually higher.
Sometimes Google is the best algorithm optimization tool. When I have a complex problem, a bit of searching reveals some guys with PhD's have found a mapping between this and a well-known problem and have already done most of the work.
I would recommend optimizing using more efficient algorithms and not do it as an afterthought but code it that way from the start. Let the compiler work out the details on the small things as it knows more about the target processor than you do.
For one, I rarely use loops to look things up, I add items to a hashtable and then use the hashtable to lookup the results.
For example you have a string to lookup and then 50 possible values. So instead of doing 50 strcmps, you add all 50 strings to a hashtable and give each a unique number ( you only have to do this once ). Then you lookup the target string in the hashtable and have one large switch with all 50 cases ( or have functions pointers ).
When looking up things with common sets of input ( like css rules ), I use fast code to keep track of the only possible solitions and then iterate thought those to find a match. Once I have a match I save the results into a hashtable ( as a cache ) and then use the cache results if I get that same input set later.
My main tools for faster code are:
hashtable - for quick lookups and for caching results
qsort - it's the only sort I use
bsp - for looking up things based on area ( map rendering etc )
The problem
I'm working on implementing and refining an optimization algorithm with some fairly large arrays (from tens of millions of floats and up) and using mainly Intel MKL in C (not C++, at least not so far) to squeeze out every possible bit of performance. Now I've run into a silly problem - I have a parameter that sets maxima and minima for subsets of a set of (tens of millions) of coefficients. Actually applying these maxima and minima using MKL functions is easy - I can create equally-sized vectors with the limits for every element and use V?Fmax and V?Fmin to apply them. But I also need to account for this clipping in my error metric, which requires me to count the number of elements that fall outside these constraints.
However, I can't find an MKL function that allows me to do things like counting the number of elements that fulfill some condition, the way you can create and sum logical arrays with e.g. NumPy in Python or in MATLAB. Irritatingly, when I try to google this question, I only get answers relating to Python and R.
Obviously I can just write a loop that increments a counter for each element that fulfills one of the conditions, but if there is an already optimized implementation that allows me to achieve this, I would much prefer that just owing to the size of my arrays.
Does anyone know of a clever way to achieve this robustly and very efficiently using Intel MKL (maybe with the statistics toolbox or some creative use of elementary functions?), a similarly optimized library that does this, or a highly optimized way to hand-code this? I've been racking my brain trying to come up with some out-of-the box method, but I'm coming up empty.
Note that it's necessary for me to be able to do this in C, that it's not viable for me to shift this task to my Python frontend, and that it is indeed necessary for me to code this particular subprogram in C in the first place.
Thanks!
If you were using c++, count_if from the algorithms library with an execution policy of par_unseq may parallelize and vectorize the count. On Linux at least, it typically uses Intel TBB to do this.
It's not likely to be as easy in c. Because c doesn't have concepts like templates, callables or lambdas, the only way to specialize a generic (library-provided) count()-function would be to pass a function pointer as a callback (like qsort() does). Unless the compiler manages to devirtualize and inline the callback, you can't vectorize at all, leaving you with (possibly thread parallelized) scalar code. OTOH, if you use for example gcc vector intrinsics (my favourite!), you get vectorization but not parallelization. You could try to combine the approaches, but I'd say get over yourself and use c++.
However, if you only need vectorization, you can almost certainly just write sequential code and have the compiler autovectorize, unless the predicate for what should be counted is poorly written, or your compiler is braindamaged.
For example. gcc vectorizes the code on x86 if at least sse4 instructions are available (-msse4). With AVX[2/512] (-mavx / -mavx2 / -mavx512f) you can get wider vectors to do more elements at once. In general, if you're compiling on the same hardware you will be running the program on, I'd recommend letting gcc autodetect the optimal instruction set extensions (-march=native).
Note that in the provided code, the conditions should not use short-circuiting or (||), because then the read from the max-vector is semantically forbidden if the comparison with the min-vector was already true for the current element, severely hindering vectorization (though avx512 could potentially vectorize this with somewhat catastrophic slowdown).
I'm pretty sure gcc is not nearly optimal in the code it generates for avx512, since it could do the k-reg (mask register) or in the mask registers with kor[b/w/d/q], but maybe somebody with more experience in avx512 (*cougth* Peter Cordes *cough*) could weigh in on that.
MKL doesn't provide such functions but You may try to check another performance library - IPP which contains a set of threshold functions that could be useful to your case. Please refer to the IPP Developer Reference to check more details - https://software.intel.com/content/www/us/en/develop/documentation/ipp-dev-reference/top/volume-1-signal-and-data-processing/essential-functions/conversion-functions/threshold.html
As part of a bigger code, I have a CUDA RK4 solver that integrates a large number of ODEs (Can be 1000+) in parallel. One step of this operation is calculating 'xdot', which is different for each equation (or data element). As of now, I have a switch-case branching setup to calculate the value for each data element in the kernel. All the different threads use the same 3-6 data elements to calculate their output, but in a different way. For example, for thread 1, it could be
xdot = data[0]*data[0] + data[1];
while for thread 2 it could be,
xdot = -2*data[0] + data[2];
and so on.
So if I have a hundred data elements, the execution path is different for each of them.
Is there any way to avoid/decrease the thread-divergence penalty in such a scenario?
Would running only one thread per block be of any help ?
Running one thread per block simply nulls 31/32 threads in the single warp you launch and wastes a lot of cycles and opportunities to hide latency. I would never recommend it, no matter how much branch divergence penalty your code incurred.
Your application sounds pretty orthognal to the basic CUDA programming paradigm and there really isn't going to be much you can do to avoid branch divergence penalties. One approach which could slightly improve things would be to perform some prior analysis of the expressions for each equation and group those with common arithmetic terms together. Recent hardware can run a number of kernels simultaneously, so it might be profitable to group calculations sharing like terms into different kernels and launch them simultaneously, rather than a single large kernel. CUDA supports C++ templating, and that can be a good way of generating a lot of kernel code from a relatively narrow base and make a lot of logic statically evaluable, which can help the compiler. But don't expect miracles - your problem is probay better suited to a different architecture than the GPU (Intel's Xeon Phi, for example).
A matrix is created in processor 0 and scattered to other processors. A matrix is a symmetric dense matrix. That's why it is initialized in processor 0.
A matrix is created in this way:
A=malloc(sizeof(double)*N*N);
for (i=0; i<N; i++)
for(j=0; j<N; j++)
A(i,j)=rand()%10; // The code will be changed.
A(i,j) is defined as:
#define A(i,j) A[i*N+j]
and N has to be 100,000 to test the algorithm.
The problem here is: if N=100,000 then the memory needed is approximately 76GB. What do you suggest to store the A matrix?
PS: Algorithm works very well when N<20.000 and the cluster is a distrubed memory system(2GB RAM per processor)
If you are doing this, as stated in comments, to do a scaling test, then Oli Charlesworth is completely right; anything you do is going to make this an apples-to-oranges comparison, because your node doesn't have 76GB to use. Which is fine; one of the big reasons to use MPI is to tackle problems that couldn't fit on one node. But by trying to shoehorn 76GB of data onto one processor, the comparison you're doing isn't going to make any sense. As mentioned by both Oli Charlesworth and caf, through various methods you can use disk instead of RAM, but then your 1 processor answer is going not going to be directly comparable to the fits-in-RAM numbers you get from larger number of nodes, so you're going to be going to a lot of work to get a number which won't actually mean anything.
If you want scaling results on this sort of problem, you either start with the lowest number of nodes that the problem does fit on, and take data at increasing numbers of processors, or you do weak scaling, rather than strong scaling tests -- you keep the work-per-processor constant while scaling up the number of processors, rather than the total work being constant.
Incidentally, however you do the measurements, you'll end up with better results if, as Oli Charlesworth suggests, you have each procesor generate its own data rather than have a serial bottleneck by having rank 0 do the generation of the matrix and then have all the processors receive their parts.
If you are programming on a POSIX system with sufficient virtual address space (which in practice will mean a 64 bit system), you can use mmap().
Either create an anonymous mapping of the required size (this will be swap-backed, which will mean you'll need at least 76GB of swap), or create a real file of the required size and map that.
The file-backed solution has the advantage that if your cluster has a shared file system, you don't need to explicitly transfer the matrix to each processor - you can simply msync() it after creating it, and then map the right region on each processor.
If you can switch to C++, you might look into STXXL, which is an STL implementation specifically designed for huge datasets, with transparent disk-backed support, etc.