Branchless Programming: Difference between revisions
Created page with "The idea of branchless programming is to avoid <code>if</code> and <code>for</code> statements. Typically, these statements require the CPU to perform branch prediction. If..." |
|||
(17 intermediate revisions by the same user not shown) | |||
Line 3: | Line 3: | ||
If the CPU predicts incorrectly, it will have to discard all pipelined results and go down the other route. | If the CPU predicts incorrectly, it will have to discard all pipelined results and go down the other route. | ||
Furthermore warps on GPUs must all go through both branches if there are threads which go down each branch. | Furthermore warps on GPUs must all go through both branches if there are threads which go down each branch. | ||
In general, I consider avoiding if statements to be a micro-optimization. | In general, I consider avoiding if statements to be a micro-optimization. | ||
Line 14: | Line 13: | ||
==Switch Statements== | ==Switch Statements== | ||
Technically, switch statements are still branches. | Technically, switch statements are still branches. | ||
However, they can lead to a negligible performance gains if there are lots of cases. | However, they can lead to a negligible performance gains if there are lots of cases. | ||
Only convert if to | Only convert if statements to switch statements when they are significantly more readable in your scenario. | ||
==Polymorphism== | ==Polymorphism== | ||
Line 31: | Line 30: | ||
This is one way to simulate switch statements. | This is one way to simulate switch statements. | ||
Technically these do not have branching but will have a memory read which could be just as slow as a single branch. | Technically these do not have branching but will have a memory read which could be just as slow as a single branch. | ||
< | <syntaxhighlight lang="python"> | ||
# Cache the following at the start of your program or in the constructor | # Cache the following at the start of your program or in the constructor | ||
# Assume we have some functions batch_norm(), group_norm() | # Assume we have some functions batch_norm(), group_norm() | ||
Line 39: | Line 38: | ||
"identity": identity | "identity": identity | ||
} | } | ||
# At runtime, you can do: | |||
x = my_cache[operation](x) | x = my_cache[operation](x) | ||
</ | </syntaxhighlight> | ||
==Booleans== | |||
Multiplications by booleans can be used instead of switch statements. | |||
Rather than doing | Rather than doing | ||
< | <syntaxhighlight lang="cpp"> | ||
if age > 10 | if (age > 10) { | ||
x = | x += 10; | ||
else | } else { | ||
x = | x -= 10; | ||
</ | } | ||
</syntaxhighlight> | |||
you can do | you can do | ||
< | <syntaxhighlight lang="cpp"> | ||
x = x + (20 * (age > 10) - 10); | |||
// or | // or | ||
y = age > 10 | y = age > 10; | ||
x = x + 10 * y - 10 * ( | x = x + 10 * y - 10 * !y; | ||
</ | </syntaxhighlight> | ||
Similarly, instead of doing | |||
<syntaxhighlight lang="cpp"> | |||
if (age > 10) { | |||
x++; | |||
} | |||
</syntaxhighlight> | |||
you can just write <code>x += age>10;</code>. | |||
These are micro optimizations but they | These are micro optimizations but they could be useful for speeding up GPU shaders in some cases. | ||
In high-level programming languages such as Python, these tricks can actually slow down your code<ref name="branchless-slower-python">https://stackoverflow.com/questions/72118249/why-are-the-branchless-and-built-in-functions-slower-in-python</ref>. | |||
==Resources== | ==Resources== | ||
* [https://francescocirillo.com/pages/anti-if-campaign Anti If Campaign] | * [https://francescocirillo.com/pages/anti-if-campaign Anti If Campaign] | ||
* [https://code.joejag.com/2016/anti-if-the-missing-patterns.html Anti-If: The missing patterns] | |||
==References== | |||
<references /> |
Latest revision as of 19:58, 16 November 2023
The idea of branchless programming is to avoid if
and for
statements.
Typically, these statements require the CPU to perform branch prediction.
If the CPU predicts incorrectly, it will have to discard all pipelined results and go down the other route.
Furthermore warps on GPUs must all go through both branches if there are threads which go down each branch.
In general, I consider avoiding if statements to be a micro-optimization. You should just use if statements wherever the optimized alternative is less readable. Typically, this is fine for CPUs which are good at branch prediction.
Below are a few tricks to avoiding branches and if statements. In general, I would not refractor to improve performance through these micro-optimizations, only to improve readability.
Switch Statements
Technically, switch statements are still branches.
However, they can lead to a negligible performance gains if there are lots of cases.
Only convert if statements to switch statements when they are significantly more readable in your scenario.
Polymorphism
Rather than do
def speak(): if self.type == "bird": print("Chirp") elif self.type == "cat": print("Meow")
create separate classes for bird and cat each overriding the speak method.
Hash Maps
This is one way to simulate switch statements. Technically these do not have branching but will have a memory read which could be just as slow as a single branch.
# Cache the following at the start of your program or in the constructor
# Assume we have some functions batch_norm(), group_norm()
my_cache = {
"batch_norm": batch_norm,
"group_norm": group_norm,
"identity": identity
}
# At runtime, you can do:
x = my_cache[operation](x)
Booleans
Multiplications by booleans can be used instead of switch statements.
Rather than doing
if (age > 10) {
x += 10;
} else {
x -= 10;
}
you can do
x = x + (20 * (age > 10) - 10);
// or
y = age > 10;
x = x + 10 * y - 10 * !y;
Similarly, instead of doing
if (age > 10) {
x++;
}
you can just write x += age>10;
.
These are micro optimizations but they could be useful for speeding up GPU shaders in some cases. In high-level programming languages such as Python, these tricks can actually slow down your code[1].