DEV Community

Cover image for 5 (Extreme) Performance Tips in C# 🔥
ByteHide
ByteHide

Posted on • Edited on • Originally published at bytehide.com

5 (Extreme) Performance Tips in C# 🔥

This article is transcribed from Bartosz Adamczewski’s video on LevelUp’s Youtube channel about 5 (Extreme) Performance Tips in C#.

This transcription is made on behalf of ByteHide, we have no association with the author, we just found it super interesting, the team thanks him.

In this article, you will learn 5 different performance tricks that you can do in C#.

They are called EXTREME because, personally, I have not found information like this on the Internet. Yes, there are different performance tricks, but not the ones I’m going to show you here.


Let’s start with a very simple example where we’re gonna have a sum of odd elements, so what we can do here is we can take the array element and they we can check if it’s divisible by two and, if it’s not, then we’re gonna add that element because that element is odd and we can turn the result.

private static int SumOdd(int[] array)
{
  int counter = 0;
  for (int i = 0; i < array.Length; i++)
  {
    var element = array[i];
    if (element % 2 != 0)
    counter += element;
  }
  return counter;
}
Enter fullscreen mode Exit fullscreen mode

So, if we run it in our simple measuring procedure this is going to take quite a while to compute because we’re going to compute it on 40 million elements, so the average result is around 240–250 milliseconds.

Image description

If we look at this function…

Is there something that we can do in order to be to actually optimized? 🤔

It’s there a way for us to have a faster version of this function turns out that there is? 🤔


1. Bit Tricks

We can replace more expensive elements with least expensive elements and in this case I’m talking about the modulus operation. Turns out that the modulus operation can be extremely expensive but the good news is that the jit compiler automatically uses a shift left on the module, so already we have a better implementation but, we can still do better because we have a proper context that the compiler doesn’t and there is a way to make it even faster.

sum = SumOdd_Bit(array);
Enter fullscreen mode Exit fullscreen mode

What we can do is we can do a and operation with one and that will effectively test if our element is out and if it’s odd it’s going to be reduced to just a single value one.

private static int SumOdd(int[] array)
{
  int counter = 0;
  for (int i = 0; i < array.Length; i++)
  {
    var element = array[i];
    if ((element & 1) == 1)
    counter += element;
  }
  return counter;
}
Enter fullscreen mode Exit fullscreen mode

Now we can check is if this got faster 👇

Image description

Now it takes slightly faster because it’s 217 milliseconds from 240 milliseconds* and we got a improvement.


2. Branch Elimination

We can do a branch-free version of that some odd function. We can do it because we already had a kind of an operation that did it because this element and one operation will return one if the element is odd, otherwise it’s going to return zero. That already allows us to eliminate the branch and we can do a multiplication or by the element.

sum = SumOdd_BranchFree(array);
Enter fullscreen mode Exit fullscreen mode

If this is one that means the element is odd, we’re gonna multiply that by one. We’re gonna have an element, otherwise, we’re just going to multiply by zero and have zero. Branch elimination is interesting for a two of reasons:

  • First of all, there’s certain data sets and data workflows that you can have in an application where the data is very extreme that the branch predictor cannot do a good enough job well.

  • The second reason is because you want to have a stable performance, because like i said, branch prediction depends on the data and you can have super fast function because you have predictable data but, on the other that might get a bit slower. That’s why you might want to consider that.

Of course you have to keep in mind that branch prediction is expensive, because all of the things that you have to do to eliminate the branch can be expensive like bit hacking tricks.

private static int SumOdd_BranchFree(int[] array)
{
  int counter = 0;
  for (int i = 0; i < array.Length; i++)
  {
    var element = array[i];
    var odd = element & 1; 
    counter += (odd * element);
  }
  return counter;
}
Enter fullscreen mode Exit fullscreen mode

Let’s check the performance 👇

Image description

It took 43 milliseconds which is a big improvement from 217 milliseconds and that’s really good.


3. Instruction Parallelism

Since we did branch elimination already, what we can do now is we can do a instruction level parallelism here. Instruction level parallelism means that modern cpus usually can do multiple things at the same time provided that there’s no data hazards between different sort of elements and that these instructions that they execute can really be executed on multiple ports.

sum = SumOdd_BranchFree_Parallel(array);
Enter fullscreen mode Exit fullscreen mode

We effectively duplicated our counter to not have data hazards and now we can do certain operations at the same time, for example: and operation can be done four times per one cpu cycle and in order to be able to figure out if you can benefit from these improvements.

private static int SumOdd_BranchFree_Parallel(int[] array)
{
  int counterA = 0;
  int counterB = 0;
  for (int i = 0; i < array.Length; i += 2)
  {
    var elementA = array[i];
    var elementB = array[i + 1];
    var oddA = elementA & 1;
    var oddB = elementB & 1;
    counter += (oddA * elementA);
    counter += (oddB * elementB);
  }
  return counterA + counterB;
}
Enter fullscreen mode Exit fullscreen mode

Let’s test the performance of this version 👇

Image description

Now 39 milliseconds. It’s slightly faster but only slightly and the reason might be that we had a multiplication here.


4. Bounds Checking

Tip number four would be to eliminate all of the bounce checks, because the previous method had a lot of bounce checking of the array, although, we sort of fulfill almost the correct signature not to have any bounce checks but, if we’re going to change the signature of i, we’re gonna get two elements of the array.

sum = SumOdd_BranchFree_Parallel_NoChecks(array);
Enter fullscreen mode Exit fullscreen mode

There are a couple of ways of eliminating them but, one of them that’s the simplest one it’s not the best one mind you. Is to just have a fixed pointer to that array and then, basically, we’re gonna convert that to the end pointer and gonna we’re gonna take that pointer and access the elements.

private static int SumOdd_BranchFree_Parallel_NoChecks(int[] array)
{
  int counterA = 0;
  int counterB = 0;
  fixed (int* data = &array[0])
  {
    var p = (int*)data;
for (var i = 0; i < array.Length; i += 2)
    {
      var elementA += (p[0] & 1) * p[0];
      var elementB += (p[1] & 1) * p[0];
      p += 2;
    }
  return counterA + counterB;
}
Enter fullscreen mode Exit fullscreen mode

Let’s measure the performance of this version 👇

Image description

Better, 32 milliseconds from 39 milliseconds. We got a improvement.


5. Maximize Ports

This tip would be, if we know all of these things now, we can do a better job with ports. We can get another pointer to our data and the first pointer is going to be loaded to registers but the second won’t be. That effectively eliminates the need to have just a single multiplication per operation.

sum = SumOdd_BranchFree_Parallel_NoChecks_BetterPorts(array);
Enter fullscreen mode Exit fullscreen mode

Although, we’re still constrained by loads as we can do only two loads per cycle and we have eight loads here, still the multiplication would be the term the the biggest factor in performance degradation here and we can check if this is really true.

private static int SumOdd_BranchFree_Parallel_NoChecks_BetterPorts(int[] array)
{
  int counterA = 0;
  int counterB = 0;
  int counterC = 0;
  int counterD = 0;
  fixed (int* data = &array[0])
  {
    var p = (int*)data;
    var n = (int*)data;
for (var i = 0; i < array.Length; i += 4)
    {
      counterA += (n[0] & 1) * p[0];
      counterB += (n[1] & 1) * p[1];
      counterC += (n[2] & 1) * p[2];
      counterD += (n[3] & 1) * p[3];
      p += 4;
      n += 4;
    }
  return counterA + counterB + counterC + counterD;
}
Enter fullscreen mode Exit fullscreen mode

Let’s check it out and let’s run this version👇

Image description

It took 25.6 milliseconds which is the fastest version.


Just to show you in how it looks in a nice graph, you can see the ports version is slightly faster than the no checks with four parts version and from the first tip to the last tip we have a performance benefit by almost a factor of 10.

Image description

Top comments (11)

Collapse
 
badamczewski01 profile image
Info Comment hidden by post author - thread only accessible via permalink
Bartosz Adamczewski

This article is copied wholesale from my youtube video: youtube.com/watch?v=Tb2Fx9qku_o

If you're going to copy all of the things 1to1, then at least you should credit the author.

Collapse
 
kenrogoway profile image
Ken Rogoway

Overall a good article, but the example code has a lot of bugs. So it is hard to know if what is shown was used for your performance tests. If it was then your speed improvements may have more to do with optimizations for invariants. For instance in #3 you add to “counter”, but return the sum of counterA and counterB which are both zero. Similar issue in #4, and it also uses the wrong value for the B portion (you’re using p[0] for both A and B sums.

Collapse
 
denisvlah profile image
Denis Vlah

Found how to eliminate multiplication. It makes the code just a little bit faster compared to SumOdd_BranchFree but this technique performs worse than multiplication methods with parallel execution.
The code is the following:

private static int SumOdd_BranchFree_NoMultiply(int[] array)
        {
            int mask = -1; // in binary format is 11111111111111111111111111111111
            int counter = 0;
            for (int i = 0; i < array.Length; i++)
            {
                var element = array[i];
                var odd = element & 1;
                var elementMask = ~(mask + odd); // will cause int overflow if odd == 1 
                counter += (element & elementMask);
            }
            return counter;
        }
Enter fullscreen mode Exit fullscreen mode
Collapse
 
cantak profile image
cantak • Edited

Here's one more option to consider. It's twice as fast as the best one above and uses the bitwise step in actual parallel processing. It probably could have further optimizations from above incorporated into it also. With this you'd call ParallelLoop(array, numberOfTasksYouWantForYourMachine).

Also, I had never thought of bit checking before and have always used mod. I will certainly use this in the future. Nice article!

private static long AddViaGap(long[] array, int gap, int offset) {
long counter = 0;
for (int i = offset; i < array.Length; i += gap) {
var element = array[i];
var odd = element & 1;
counter += (odd * element);
}
return counter;
}

private static long ParallelLoop(long[] array, int taskCount) {
var tasks = new Task[taskCount];
for (var counter = 0; counter < taskCount; counter++) {
var currentCounter = counter;
tasks[counter] = Task.Run(() => AddViaGap(array, taskCount, currentCounter));
}
Task.WaitAll(tasks);
return tasks.Sum(o => o.Result);
}

Collapse
 
rosskillen profile image
rosskillen

Good tips, thanks. Not strong on C# myself but surely you've introduced a bug at stage 3, e.g. if your array has an odd length? I.e. elementB won't exist.

Collapse
 
semidewi profile image
Semidewi

You could get rid of the multiplication by defining a array with length 2.

First entry at Index 0 is 0
Second entry at index 1 is -1

Instead of multiplying value by 1 or 0, you could simply do a Bitwise AND Operation with the value at Index 1 or 0.

Collapse
 
dariojavierrick profile image
Dario Javier Rick

great article! thanks for sharing!

Collapse
 
bytehide profile image
ByteHide

🔒🔒🔒🔒🔒

Collapse
 
marylynch profile image
Maryanne

Nice article.

Collapse
 
bytehide profile image
ByteHide

☕☕☕☕☕

Some comments have been hidden by the post's author - find out more