64 lines
19 KiB
Markdown
64 lines
19 KiB
Markdown
# Optimization
|
|
|
|
Optimization means making a program more efficient in terms of consumption of some computing resource or by any similar metric, commonly aiming for greater execution speed or lower memory usage (but also e.g. lower power consumption, lower network usage etc.) while preserving how the program functions externally. Unlike [refactoring](refactoring.md), which aims primarily for a better readability of source code, optimization changes the inner behavior of the executed program to a more optimal one.
|
|
|
|
## General Tips'N'Tricks
|
|
|
|
These are mainly for [C](c.md), but may be usable in other languages as well.
|
|
|
|
- **Tell your compiler to actually optimize** (`-O3`, `-Os` flags etc.). Also check out further compiler flags that may help you turn off unnecessary things you don't need, AND try out different compilers, some may just produce better code.
|
|
- **[gprof](gprof.md) is a utility you can use to profile your code**.
|
|
- **`<stdint.h>` has fast type nicknames**, types such as `uint_fast32_t` which picks the fastest type of at least given width on given platform.
|
|
- **Actually measure the performance** to see if your optimizations work or not. Sometimes things behave counterintuitively and you end up making your program perform worse by trying to optimize it! Also make sure that you MEASURE THE PERFORMANCE CORRECTLY, many beginners for example just try to measure run time of a single simple function call which doesn't really work, you want to try to measure something like a million of such function calls in a loop and then average the time.
|
|
- **Keywords such as `inline`, `static` and `const` can help compiler optimize well**.
|
|
- **Optimize the [bottlenecks](bottleneck.md)!** Optimizing in the wrong place is a complete waste of time. If you're optimizing a part of code that's taking 1% of your program's run time, you will never speed up your program by more than that 1% even if you speed up the specific part by 10000%. Bottlenecks are usually inner-most loops of the main program loop, you can identify them with [profiling](profiling.md). Generally initialization code that runs only once in a long time doesn't need much optimization -- no one is going to care if a program starts up 1 millisecond faster (but of course in special cases such as launching many processes this may start to matter).
|
|
- **You can almost always trade space (memory usage) for time (CPU demand) and vice versa** and you can also fine-tune this. You typically gain speed by precomputation (look up tables, more demanding on memory) and memory with [compression](compression.md) (more demanding on CPU).
|
|
- **[Static](static.md) things are faster and smaller than [dynamic](dynamic.md) things.** This means that things that are somehow fixed/unchangeable are better in terms of performance (and usually also safer and better testable) than things that are allowed to change during [run time](runtime.md) -- for example calling a function directly (e.g. `myVar = myFunc();`) is both faster and requires fewer instructions than calling a function by pointer (e.g. `myVar = myFuncPointer();`): the latter is more flexible but for the price of performance, so if you don't need flexibility (dynamic behavior), use static behavior. This also applies to using [constants](constant.md) (faster/smaller) vs [variables](variable.md), static vs dynamic [typing](typing.md), normal vs dynamic [arrays](array.md) etc.
|
|
- **Be smart, use [math](math.md)**, for example simplify [expressions](expression.md) using known formulas, minimize [logic circuits](logic_circuit.md) etc. Example: let's say you want to compute the radius of a zero-centered [bounding sphere](bounding_sphere.md) of an *N*-point [point cloud](point_cloud.md). Naively you might be computing the Euclidean distance (*sqrt(x^2 + y^2 + z^2)*) to each point and taking a maximum of them, however you can just find the maximum of squared distances (*x^2 + y^2 + z^2*) and return a square root of that maximum. This saves you a computation of *N - 1* square roots.
|
|
- **Learn about [dynamic programming](dynamic_programming.md)**.
|
|
- **Avoid branches (ifs)** if you can (remember [ternary operators](ternary_operator.md), loop conditions etc. are branches as well). They break prediction in CPU pipelines and instruction preloading and are often source of great performance losses. Don't forget that you can many times compare and use the result of operations without using any branching (e.g. `x = (y == 5) + 1;` instead of `x = (y == 5) ? 2 : 1;`).
|
|
- **Use iteration instead of [recursion](recursion.md)** if possible (calling a function costs something).
|
|
- **Use [good enough](good_enough.md) [approximations](approximation.md) instead of completely accurate calculations**, e.g. taxicab distance instead of Euclidean distance, capsule shape to represent the player's collision shape rather than the 3D model's mesh etc. With a physics engine instead of running the simulation at the same FPS as rendering, you may just run it at half and [interpolate](interpolation.md) between two physics states at every other frame. Nice examples can also be found in [computer graphics](graphics.md), e.g. some [software renderers](sw_rendering.md) use perspective-correct texturing only for large near triangles and cheaper affine texturing for other triangles, which mostly looks OK.
|
|
- **Use quick opt-out conditions**: many times before performing some expensive calculation you can quickly check whether it's even worth performing it and potentially skip it. For example in physics [collision detections](collision_detection.md) you may first quickly check whether the bounding spheres of the bodies collide before running an expensive precise collision detection -- if bounding spheres of objects don't collide, it is not possible for the bodies to collide and so we can skip further collision detection.
|
|
- **Operations on static data can be accelerated with accelerating structures** ([look-up tables](lut.md) for functions, [indices](indexing.md) for database lookups, spatial grids for collision checking, various [trees](tree.md) ...).
|
|
- **Use powers of 2** (1, 2, 4, 8, 16, 32, ...) whenever possible, this is efficient thanks to computers working in [binary](binary.md). Not only may this help nice utilization and alignment of memory, but mainly multiplication and division can be optimized by the compiler to mere bit shifts which is a tremendous speedup.
|
|
- **Memory alignment usually helps speed**, i.e. variables at "nice addresses" (usually multiples of the platform's native integer size) are faster to access, but this may cost some memory (the gaps between aligned data).
|
|
- **Write [cache-friendly](cache-friendly.md) code** (minimize long jumps in memory).
|
|
- **Compare to [0](zero.md) rather than other values**. There's usually an instruction that just checks the zero flag which is faster than loading and comparing two arbitrary numbers.
|
|
- **Use [bit tricks](bit_hack.md)**, hacks for manipulating binary numbers in clever ways only using very basic operations without which one might naively write complex inefficient code with loops and branches. Example of a simple bit trick is checking if a number is power of two as `!(x & (x - 1)) && x`.
|
|
- **Consider moving computation from run time to compile time**. E.g. if you make a resolution of your game constant (as opposed to a variable), the compiler will be able to partially precompute expressions with the display dimensions and so speed up your program (but you won't be able to dynamically change resolution).
|
|
- On some platforms such as [ARM](arm.md) the first **arguments to a function may be passed via registers**, so it may be better to have fewer parameters in functions.
|
|
- **Optimize when you already have a working code**. As [Donald Knuth](knuth.md) put it: "premature optimization is the root of all evil". Nevertheless you should get used to simple nobrainer efficient patterns by default and just write them automatically.
|
|
- **Use your own [caches](cache.md) where they help**, for example if you're frequently working with some database item you better pull it to memory and work with it there, then write it back once you're done (as opposed to communicating with the DB there and back).
|
|
- **[Single compilation unit](single_compilation_unit.md) (one big program without [linking](linking.md)) can help compiler optimize better** because it can see the whole code at once, not just its parts. It will also make your program compile faster.
|
|
- Search literature for **algorithms with better [complexity class](complexity_class.md)** ([sorts](sorting.md) are a nice example).
|
|
- For the sake of simple computers such as [embedded](embedded.md) platforms **avoid [floating point](floating_point.md)** as that is often painfully slowly emulated in software. Use [fixed point](fixed_point.md), or at least offer it as a [fallback](fallback.md). This also applies to other hardware requirements such as [GPU](gpu.md) or sound cards: while such hardware accelerates your program on computers that have the hardware, making use of it may lead to your program being slower on computers that lack it.
|
|
- **[Early branching](early_branching.md) can create a speed up** (instead of branching inside the loop create two versions of the loop and branch in front of them). This is a kind of space-time tradeoff.
|
|
- **Division can be replaced by multiplication by [reciprocal](reciprocal.md)**, i.e. *x / y = x * 1/y*. The point is that multiplication is usually faster than division. This may not help us when performing a single division by variable value (as we still have to divide 1 by *y*) but it does help when we need to divide many numbers by the same variable number OR when we know the divisor at compile time; we save time by precomputing the reciprocal before a loop or at compile time. Of course this can also easily be done with [fixed point](fixed_point.md) and integers!
|
|
- **Consider the difference between logical and bitwise operators!** For example [AND](and.md) and [OR](or.md) boolean functions in C have two variants, one bitwise (`&` and `|`) and one logical (`&&` and `||`) -- they behave a bit differently but sometimes you may have a choice which one to use, then consider this: bitwise operators usually translate to only a single fast (and small) instruction while the logical ones usually translate to a branch (i.e. multiple instructions with potentially slow jumps), however logical operators may be faster because they are evaluated as [short circuit](short_circuit_eval.md) (e.g. if first operand of OR is true, second operand is not evaluated at all) while bitwise operators will evaluate all operands.
|
|
- **Consider the pros and cons of using indices vs pointers**: When working with arrays you usually have the choice of using either pointers or indices, each option has advantages and disadvantages; working with pointers may be faster and produce smaller code (fewer instructions), but array indices are portable, may be smaller and safer. E.g. imagine you store your game sprites as a continuous array of images in RAM and your program internally precomputes a table that says where each image starts -- here you can either use pointers (which say directly the memory address of each image) or indices (which say the offset from the start of the big image array): using indices may be better here as the table may potentially be smaller (an index into relatively small array doesn't have to be able to keep any possible memory address) and the table may even be stored to a file and just loaded next time (whereas pointers can't because on next run the memory addresses may be different), however you'll need a few extra instructions to access any image (adding the index to the array pointer), which will however most definitely be negligible.
|
|
- **Reuse variables to save space**. A warning about this one: readability may suffer, mainstreamers will tell you you're going against "good practice", and some compilers may do this automatically anyway. Be sure to at least make this clear in your comments. Anyway, on a lower level and/or with dumber compilers you can just reuse variables that you used for something else rather than creating a new variable that takes additional RAM; of course a prerequisite for "merging" variables is that the variables aren't used at the same time.
|
|
- **To save memory use [compression](compression.md) techniques.** Compression doesn't always have to mean you use a typical compression algorithm such as [jpeg](jpg.md) or [LZ77](lz77.md), you may simply just throw in a few compression techniques such as [run length](run_length.md) or word dictionaries into your data structures. E.g. in [Anarch](anarch.md) maps are kept small by consisting of a small dictionary of tile definitions and map cells referring to this dictionary (which makes the cells much smaller than if each one held a complete tile definition).
|
|
- **What's fast on one platform may be slow on another**. This depends on the instruction set as well as on compiler, operating system, available hardware, [driver](driver.md) implementation and other details. In the end you always need to test on the specific platform to be sure about how fast it will run. A good approach is to optimize for the weakest platform you want to support -- if it runs fasts on a weak platform, a "better" platform will most likely still run it fast.
|
|
- **Prefer preincrement over postincrement** (typically e.g. in a for loop), i.e. rather do `++i` than `i++` as the latter is a bit more complex and normally generates more instructions.
|
|
- **Mental calculation tricks**, e.g. multiplying by one less or more than a power of two is equal to multiplying by power of two and subtracting/adding once, for example *x * 7 = x * 8 - x*; the latter may be faster as a multiplication by power of two (bit shift) and addition/subtraction may be faster than single multiplication, especially on some primitive platform without hardware multiplication. However this needs to be tested on the specific platform. Smart compilers perform these optimizations automatically, but not every compiler is high level and smart.
|
|
- **Else should be the less likely branch**, try to make if conditions so that the if branch is the one with higher probability of being executed -- this can help branch prediction.
|
|
- Similarly **order if-sequences and switch cases from most probable**: If you have a sequences of ifs such as `if (x) ... else if (y) ... else if (z) ...`, make it so that the most likely condition to hold gets checked first, then second most likely etc. Compiler most likely can't know the probabilities of the conditions so it can't automatically help with this. Do the same with the `switch` statement -- even though switch typically gets compiled to a table of jump addresses, in which case order of the cases doesn't matter, it may also get compiled in a way similar to the if sequence (e.g. as part of size optimization if the cases are sparse) and then it may matter again.
|
|
- **You can save space by "squeezing" variables** -- this is a space-time tradeoff, it's a no brainer but nubs may be unaware of it -- for example you may store 2 4bit values in a single `char` variable (8bit data type), one in the lower 4bits, one in the higher 4bits (use bit shifts etc.). So instead of 16 memory-aligned booleans you may create one `int` and use its individual bits for each boolean value. This is useful in environments with extremely limited RAM such as 8bit Arduinos.
|
|
- **Consider [lazy evaluation](lazy_evaluation.md)** (only evaluate what's actually needed).
|
|
- **You can optimize critical parts of code in [assembly](assembly.md)**, i.e. manually write the assembly code that takes most of the running time of the program, with as few and as inexpensive instructions as possible (but beware, popular compilers are very smart and it's often hard to beat them). But note that such code loses [portability](portability.md)! So ALWAYS have a C (or whatever language you are using) [fallback](fallback.md) code for other platforms, use [ifdefs](ifdef.md) to switch to the fallback version on platforms running on different assembly languages.
|
|
- **Loop unrolling/splitting/fusion, function inlining etc.**: there are optimizations that are usually done by high level languages at [assembly](assembly.md) level (e.g. loop unrolling physically replaces a loop by repeated commands which gains speed but also makes the program bigger). However if you're writing in assembly or have a dumb compiler (or are even writing your own) you may do these manually, e.g. with macros/templates etc. Sometimes you can hint a compiler to perform these optimizations, so look this up.
|
|
- **[Parallelism](parallelism.md) ([multithreading](multithreading.md), [compute shaders](compute_shader.md), ...) can astronomically accelerate many programs**, it is one of the most effective techniques of speeding up programs -- we can simply perform several computations at once and save a lot of time -- but there are a few notes. Firstly not all problems can be parallelized, some problem are sequential in nature, even though most problems can probably be parallelized to some degree. Secondly it is hard to do, opens the door for many new types of bugs, requires hardware support (software simulated parallelism can't work here of course) and introduces [dependencies](dependency.md); in other words it is huge [bloat](bloat.md), we don't recommend parallelization unless a very, very good reason is given. Optional use of [SIMD](simd.md) instructions can be a reasonable midway to going full parallel computation.
|
|
- **Specialized hardware (e.g. a [GPU](gpu.md)) astronomically accelerates programs**, but as with the previous point, portablity and simplicity greatly suffers, your program becomes bloated and gains dependencies, always consider using specialized hardware and offer software fallbacks.
|
|
|
|
## When To Actually Optimize?
|
|
|
|
Nubs often ask this and this can also be a very nontrivial question. Generally fine, sophisticated optimization should come as one of the last steps in development, when you actually have a working thing. These are optimizations requiring significant energy/time to implement -- you don't want to spend resources on this at the stage when they may well be dropped in the end, or they won't matter because they'll be outside the bottleneck. However there are two "exceptions".
|
|
|
|
The highest-level optimization is done as part of the initial design of the program, before any line of code gets written. This includes the choice of data structures and mathematical models you're going to be using, the very foundation around which you'll be building your castle. This happens in your head at the time you're forming an idea for a program, e.g. you're choosing between [server-client](server_client.md) or [P2P](p2p.md), [monolithic or micro kernel](kernel.md), [raytraced](ray_tracing.md) or [rasterized](rasterization.md) graphics etc. These choices affect greatly the performance of your program but can hardly be changed once the program is completed, so they need to be made beforehand. **This requires wide knowledge and experience** as you work by intuition.
|
|
|
|
Another kind of optimization done during development is just automatically writing good code, i.e. being familiar with specific patterns and using them without much thought. For example if you're computing some value inside a loop and this value doesn't change between iterations, you just automatically put computation of that value **before** the loop. Without this you'd simply end up with a shitty code that would have to be rewritten line by line at the end. Yes, compilers can often do this simple kind of optimization for you, but you don't want to rely on it.
|
|
|
|
## See Also
|
|
|
|
- [bit hacks](bit_hack.md) |