Everyone needs to sort arrays, once in a while. We think of it as a solved problem and that nothing can be further done about it in 2019, except for waiting for newer, marginally faster machines to pop-up^{1}. However, that is not the case, and if you join me in this rather long journey, we’ll end up with a replacement function for Array.Sort
, written in C# that can outperform CoreCLR’s own C++^{2} based code by a factor of 10x.
Sounds interesting? If so, down the rabbit hole we go…
As I was reading the post by Stephen Toub about Improvements in CoreCLR 3.0, it became apparent that hardware intrinsics were common to many of these, and that so many parts of CoreCLR could still be sped up with these techniques, that one thing led to another, and I decided an attempt to apply hardware intrinsics to a larger problem than I had previously done myself was in order. To see if I could rise to the challenge, I decided to take on array sorting and see how far I can go.
What I came up with eventually would become a re-write of QuickSort
/Array.Sort()
with AVX2 hardware intrinsics. Fortunately, choosing sorting, and specifically, QuickSort
, makes for a great blog post series, since:
I started googling and found an interesting paper titled: Fast Quicksort Implementation Using AVX Instructions by Shay Gueron and Vlad Krasnov. That title alone made me think this would be a walk in the park, but while promising, it wasn’t good enough as a drop-in replacement for Array.Sort
for reasons I’ll shortly get into. I ended up (or rather, still am, to be honest) having a lot of fun expanding on their ideas. I will submit a proper pull-request to start a discussion with CoreCLR devs about integrating this code into the main CoreCLR repo, but for now, let’s get in the ring and show what AVX/AVX2 intrinsics are capable of for this non-trivial problem.
Since there’s a lot to go over here, I’ve split it up into no less than 6 parts:
QuickSort
and how it compares to Array.Sort()
. If you don’t need any refresher, you can skip over it and get right down to part 2 and onwards, although I recommend skimming through, mostly because I’ve got excellent visualizations which should be in the back of everyone’s mind as we deal with vectorization & optimization later.QuickSort
is deceivingly simple.
No, it really is.
In 20 lines of C# or whatever language you can sort numbers. Lots of them, and incredibly fast. However, try and change something about it; it quickly becomes apparent just how tricky it is to improve on it without breaking any of the tenants it is built upon.
Before we discuss any of that, let’s describe QuickSort
in words, then in code:
That last point, referring to in-place sorting, sounds simple and neat, and it sure is from the perspective of the user: no additional memory allocation needs to occur regardless of how much data they’re sorting. While that’s great, I’ve spent days trying to overcome the correctness and performance challenges that arise from it, specifically in the context of vectorization. It is also essential to remain in-place since I intend for this to become a drop-in replacement for Array.Sort
.
More concretely, QuickSort
works like this:
Picking a pivot could be a mini-post in itself, but again, in the context of competing with Array.Sort
we don’t necessarily need to dive into it, we’ll copy whatever CoreCLR does, and get on with our lives.
CoreCLR uses a pretty standard scheme of median-of-three for pivot selection, which can be summed up as: “Let’s sort these 3 elements: first, middle and last, then pick the middle one of those three”.
Partitioning the array is where we spend most of the time: as we take our selected pivot value and rearrange the array segment that was handed to us such that all numbers smaller-than the pivot are in the beginning or left (in no particular order). Then comes the pivot, in its final resting position, and following it are all elements greater-than the pivot, again in no particular order amongst themselves.
After partitioning is complete, we recurse to the left and right of the pivot, as previously described.
That’s all there is: this gets millions of numbers sorted, in-place, efficiently as we know how to do 60+ years after its invention.
Bonus trivia points for those who are still here with me: Tony Hoare, who invented QuickSort back in the early 60s also took responsibility for inventing the null
pointer concept. So I guess there really is no good without evil in this world.
void QuickSort(int[] items) => QuickSort(items, 0, items.Length - 1);
void QuickSort(int[] items, int left, int right)
{
if (left == right) return;
int pivot = PickPivot(items, left, right);
int pivotPos = Partition(items, pivot, left, right);
QuickSort(items, left, pivotPos);
QuickSort(items, pivotPos + 1, right);
}
int PickPivot(int[] items, int left, int right)
{
var mid = left + ((right - left) / 2);
SwapIfGreater(ref items[left], ref items[mid]);
SwapIfGreater(ref items[left], ref items[right]);
SwapIfGreater(ref items[mid], ref items[right]);
var pivot = items[mid];
}
int Partition(int[] array, int pivot, int left, int right)
{
while (left <= right) {
while (array[left] < pivot) left++;
while (array[right] > pivot) right--;
if (left <= right) {
var t = array[left];
array[left++] = array[right];
array[right--] = t;
}
}
return left;
}
I did say it is deceptively simple, and grasping how QuickSort really works sometimes feels like trying to lift sand through your fingers, so I’ve decided to include two more visualizations of QuickSort, which are derivatives of the amazing work done by Michael Bostock (@mbostock) with d3.js.
One thing that we have to keep in mind is that the same data is partitioned over-and-over again, many times, with ever-shrinking partition sizes until we end up having a partition size of 2 or 3, in which case we can trivially sort the partition as-is and return.
To help see this better, we’ll use this way of visualizing arrays in QuickSort:
Here, we see an unsorted array of 200 elements (in the process of getting sorted).
The different sticks represent numbers in the [-45°..+45°] range, and the angle of the stick represents that value, as I hope it is easy to discern.
We represent the pivots with two colors:
Our ultimate goal is to go from the image above to the image below:
What follows is a static (e.g., non-animated) visualization that shows how pivots are randomly selected at each level of recursion and how, by the next step, the unsorted segments around them become partitioned until we finally have a completely sorted array. Here is how the whole thing looks:
These visuals are auto-generated in Javascript + d3.js, so feel free to hit that “Reload” button if you feel you want to see a new set of random sticks sorted.
I encourage you to look at this and try to explain to yourself what QuickSort “does” here, at every step. What you can witness here is the interaction between pivot selection, where it “lands” in the next recursion level (or row), and future pivots to its left and right and in the next levels of recursion. We also see how, with every level of recursion, the partition sizes decrease in until finally, every element is a pivot, which means sorting is complete.
While the above visualization really does a lot to help understand how QuickSort
works, I also wanted to leave you with an impression of the total amount of work done by QuickSort
, here is an animation of the whole process as it goes over the same array, slowly and recursively going from an unsorted mess to a completely sorted array:
We can witness just how many comparisons and swap operations need to happen for a 200 element QuickSort
to complete successfully. There’s genuinely a lot of work that needs to happen per element (when considering how we re-partition virtually all elements again and again) for the whole thing to finish.
It’s important to note that Array.Sort
uses a couple of more tricks to get better performance. I would be irresponsible if I didn’t mention those since in the later posts, I borrow at least one idea from its play-book, and improve upon it with intrinsics.
Array.Sort
isn’t QuickSort
; it is a variation on it called Introspective Sort invented by David Musser in 1997. What it roughly does is combine QuickSort, HeapSort, and Insertion Sort, and switch dynamically between them according to the recursion depth and the partition size. This last trick, where it switches to using “Insertion Sort” on small partitions, is critical, both for the general case and also for intrinsics/vectorization. It is beneficial because it replaces (up to) the last 4 levels of recursion (for partition sizes <= 16
) with a single call to an insertion sort. This trick reduces the overhead associated with recursion with simpler loop-based code, which ultimately runs faster for such small partitions.
As mentioned, I ended up borrowing this idea for my code as the issues around smaller partition sizes are exacerbated when using vectorized intrinsics in the following posts.
With all this new information, this is a good time to measure how a couple of different scalar (e.g. non-vectorized) versions compare to Array.Sort
. I’ll show some results generated using BenchmarkDotNet (BDN) with:
Array.Sort()
as the baseline.Scalar
as the code I’ve just presented above. This version is using regular/safe C#. With this version, every time we access an array element, the JIT inserts bounds-check machine code around our actual access that ensures the CPU does not read/write outside the memory region owned by the array.ScalarUnmanaged
as an alternative/faster version to Scalar
where:
unmanaged
constraint, neat!).InsertionSort
(again, copy-pasted from CoreCLR) when below 16 elements, just like Array.Sort
does.I’ve prepared this last version to show that with unsafe code + InsertionSort
, we can remove most of the performance gap between C# and C++ for this type of code, which mainly stems from bounds-checking, that the JIT cannot elide for these sort of random-access patterns.
Note that for this series, We’ll benchmark each sorting method with various array sizes (BDN parameter: N
): $10^i_{i=1\cdots7}$. I’ve added a custom column to the BDN column to the report: Time / N
. This represents the time spent sorting per element in the array, and as such, very useful to compare the results on a more uniform scale.
Here are the results in the form of charts showing: (1) the ratio (scaling) of various implementations being compared, (2) the time spent sorting a single element in an array of N elements (Time / N) and finally (3..) various BDN results in table form if you’re more into tables:
Method Name |
Size^{} |
Time / Element (ns)^{} |
Scaling^{} | Measurements^{} |
---|
Method Name |
Size^{} |
Max Depth ^{} |
# Partitions ^{} |
Avg. Small Sorts Size ^{} |
# Scalar Compares ^{} |
# Small Sort Compares ^{} |
---|
Surprisingly^{3}, the unmanaged C# version is running slightly faster than Array.Sort
, but with one caveat: it only outperforms the C++ version for large inputs. Otherwise, everything is as expected: The purely Scalar
variant is just slow, and the Unamanged
one mostly is on par with Array.Sort
.
These C# implementations were written to verify that we can get to Array.Sort
like performance in C#, and they do just that. Running 5% faster for some input sizes will not cut it for me; I want it much faster. Another important reason for re-implementing these basic versions is that we can now sprinkle statistics-collecting-code magic fairy dust^{4} on them so that we could now dig into some high-level sorting statistics that will assist us in deciphering and comparing future results and implementations. I’ve made this available in the Scalar Stats
tab above, where we can see, per each N
value (with some notes):
Array.Sort
switches to InsertionSort
for the last couple of recursion levels, the maximal depth is smaller.InsertionSort
for the Unmanaged
variant).
Scalar
version obviously doesn’t have any of this, so it’s just 0.Unmanaged
version, we see a consistent value of 9.x: Given that we special case 1,2,3 in the code and 16 is the upper limit, 9.x seems like a reasonable result here.if
or a while
or any other branch operation whose condition depended on unsorted user data!Unmanged
variant, the number is even higher (well only surprising if you don’t know anything about how InsertionSort
works…) and yet this version seems to run faster… I have an entire post dedicated just to this part of the problem in this series, so let’s just make note of this for now, but already we see peculiar thingsScalar
variant, but we can see that a large part of those compares are now coming from those last few levels of recursion that were converted to InsertionSort
…Some of these statistics will remain pretty much the same for the rest of this series, regardless of what we do next in future versions, while others radically change; We’ll observe and make use of these as key inputs in helping us to figure out how/why something worked, or not!
We’ve spent quite some time polishing our foundations concerning QuickSort
. I know lengthy introductions are somewhat dull, but I think time spent on this post will pays off when we next encounter our actual implementation in the 3^{rd} post.
Before that, we need to pick up some knowledge that is specific to vectorized intrinsics and introduce a few select intrinsics we’ll be using, so, this is an excellent time to break off this post, grab a fresh cup of coffee and head to the next post.
Which is increasingly taking more and more time to happen, due to Dennard scaling and the slow-down of Moore’s law… ↩
Since CoreCLR 3.0 was release, a PR to provide a span based version of this has been recently merged into the 5.0 master branch, but I’ll ignore this for the time being as it doesn’t seem to matter in this context. ↩
Believe it or not, I pretty much wrote every other version features in this series before I wrote the Unmanaged
one, so I really was quite surprised that it ended up being slightly faster that Array.Sort
↩
I have a special build configuration called Stats
which compiles in a bunch of calls into various conditionally compiled functions that bump various counters, and finally, dump it all to json and it eventually makes it all the way into these posts (if you dig deep you can get the actual json files :) ↩