Vectorization of Insertion Sort using Altivec,

and an Extended Merge Sort algorithm

c 2005 Konstantinos Margaritis, [email protected]

Licensed under the LGPL
11/07/2005

Abstract
Parallelization of a sort algorithm is not always easy nor possible, or if
it is, this does not necessarily mean significant gains in performance. After
reading in Slashdot about a GPU version of Quicksort ([gpu]), I was curi-
ous if vectorization of common sorting algorithms, such as qsort, insertion
sort and merge sort could be vectorizable and adapted to be used with Al-
tivec, the PowerPC SIMD unit. The results were more than interesting, in
some cases offering speed gains of 54%!! in this paper, we will present vec-
torization techniques for all 3 of these algorithms. Note: This paper does
not yet include the quicksort vectorization.

1 Vectorizing Insertion sort

The original (scalar) Insertion sort algorithm is at worst an O(N 2 ) algorithm.
While simple in concept, it’s not considered one of the fast sorting routines so
it’s not very popular in code where performance is necessary. A very thorough
description of the algorithm can be found in [wika]. We provide here a simple
implementation against which comparisons have been made.
void inssort_c(uint8_t a[], size_t N) {
int i, j;
for(i = 1; i < N; i++) {
int value = a[i];
for (j = i-1; j >= 0 && a[j] > value; j--) {
a[j+1] = a[j];
}
a[j+1] = value;
}
}
Here, the algorithm will sort a set of N unsigned chars. The sorting will be
stable, meaning that the position of equal elements will not be changed. The
algorithm implementation remains the same in case we’re dealing with ints or
floats.
The basic concept of vectorization implicitly denotes the steps we must take
for vectorization of this algorithm. We shall devide the set of N elements in 16

1
 Figure 1: Sorting 16 sets in parallel using Altivec

sets (as an Altivec register can hold 16 chars, which can be processed in par-
allel). Again, the situation is analogous if we’re having ints or floats deviding
into 4 sets (or 8 sets if dealing with short ints).
The idea is to sort 4/8/16 sets in parallel and then merge the results. This
is actually a variant of the Shell Sort algorithm ([wikb]). How we will do the
merging stage, still remains to be explained later. First we will have to deal
with the sorting of 16 parallel columns of data. In theory this is easy, but we
must make sure we handle a vector at a time, and in as few instructions as
possible, preferably in one go. To do that, we first have to load the data into
Altivec registers. We will use two registers as the comparison will be used
between two sets, one used as the key. Obviously, this means that the Altivec
version will work only for sizes > 32 chars (which is the size of 2 vectors).

vector uint8_t va_key, va_cur,

vector unsigned char *cur;
if (length >= 32) {
// How many sets of N/16 do we have to deal with?
loops = length/16;
for(i = 1; i < loops; i++) {
// Get the key set of elements
// and load it into an Altivec vector register
cur = a + i*16;
va_key = vec_ld(0, cur);
// Compare all the previous sets to the key set.
for (j = i-1; j >= 0; j--) {
cur = a + j*16;
va_cur = vec_ld(0, cur);
// Compare the sets
}

2
 }
}

How do we compare 2 sets of 16 characters at once? Even more tricky, how

do we swap the elements between themselves in the description and imple-
mentation of the Insertion sort algorithm? Altivec comes to the rescue with
its powerful comparison and selection instructions. Using vec_cmpgt, Altivec
provides us a comparison mask of the 2 vectors, which we can use then to
select individual elements of each vector, to combine new vectors.
Let’s explain this a little more:
 The algorithm compares keys at positions i and j (initial j value is set to
i 1)
 While a[ j] > a[i] the a[ j] is moved one position down to a[ j + 1]. Decrease
j, while it’s greater than zero.
 When the first element is found that is smaller or equal to a[i], it is re-
placed by a[i].
To do this in parallel for 16 sets, we would have to do 16 comparisons at
once, and how would we handle 16 swaps, that might happen in totally dif-
ferent positions in each set? Instead of swapping elements, we create a new
vector that holds the appropriate value according to the comparison mask. So
the swapping code is equivalent to:

 Generate comparison mask for elements a[i] and a[ j].

 Generate first result using the comparison mask and the two elements as
operators (eg. select a[i] if mask is 1, or a[ j] if mask is 0).
 Generate second result using the same comparison mask but reversed the
order element (ie. select a[ j] if mask is 1, a[i] if mask is 0).
 Store the first result in a[ j]

Now this repeated 16 times over a vector is what Altivec does. What is
the benefit? The benefit is that this code does not use any branches and it
can be pre-calculated in the pipeline by the processor, giving very good per-
formance results. Plus it is constant, meaning it will take exactly the same
number of instructions regardless the data given. So the initial scalar loop can
be replaced by the following Altivec instructions (the following code sorts an
array of chars):

int vecinssort_c(uint8_t a[], int length) {

vector uint8_t va_key, va_cur, va_next, va_first, va_second;
vector bool char va_cmpmask;
vector uint8_t *cur;
int i, j, k;
if (length >= 32) {
// How many sets of N/16 do we have to deal with?
loops = length/16;
for(i = 1; i < loops; i++) {

3
 // Get the key set of elements
// and load it into an Altivec vector register
cur = a + i*16;
va_key = vec_ld(0, cur);
// Compare all the previous sets to the key set.
for (j = i-1; j >= 0; j--) {
cur = a + j*16;
/* Load current and next 16-byte sets
into Altivec registers
*/
va_cur = vec_ld(0, cur);
va_next = vec_ld(16, cur);

/* Generate the comparison mask between

the current vector and the key vector
*/
va_cmpmask = vec_cmpgt(va_cur, va_key);

/* And construct the first and second

result vectors to replace the 2 16-byte sets
*/
va_first = vec_sel(va_cur, va_key, va_cmpmask);
va_second = vec_sel(va_next, va_cur, va_cmpmask);

// Store the vectors to their appropriate

// positions.
vec_st(va_first, 0, cur);
vec_st(va_second, 16, cur);
}
}
}
}

The result will be 16 sorted sets of N =16 elements. As the algorithm is still
basically an Insertion Sort the time required is still O(N 0 2 ) but this time N 0 =
N =16. So for example for N = 64 elements, the original algorithm would do
N 2 = 4096 steps to execute, while the Altivec version would do just N 0 = 4,
N 0 2 = 16 steps!!
So is this over? No, because the result is still not what we want, we have to
merge the remaining sets to a final sorted set. For that reason we have to use
an extended version of Merge Sort.

2 Extending/Vectorizing Merge sort

Why extended? Well, the original Merge Sort algorithm copes for 2 sorted sets,
while we need to merge 4/8/16 sets of already sorted elements.
The process is quite quite generic (i.e. not Altivec specific), with a similar
concept to the 2-set case, but obviously extended to cover more (at first we will

4
 Figure 2: Merging the 16 sets using extended Merge sort

initially cover N cases, optimizing afterwards for 4/8/16 sets). Let’s begin by
describing the algorithm, assuming we have N sets of sorted data:

1. Initialize a indices[] array of integers, with size N. Set all elements to 0.

2. Initialize an flag[] array of booleans, size N. Set all elements to 0.
3. Initialize the result[] array to store the final sorted data (which has size
N  columnsize).
4. Get the next N-tuple of data which we want to sort. The N-tuple will be
comprised of the elements in the set, each in the position indicated by the
indices[] array.

5. Find the minimum value(s) in the N-tuple and note its position(s), pos.
6. Insert the value(s) to the next available slot in the result[] array.
7. Increase indices[pos] by 1 (for all elements if more than one is found).
8. Set flag[pos] to 1.
9. Go back to 5 until all elements in flag[] are set to 1.
10. Go back to 4 while there are remaining sets.

This looks even easier if you see it in code:

#define INDEX(n,x,y) ((x)*(n) +(y))

int mergesort_c(uint8_t a[], size_t sets, size_t columnsize) {

// Set up variables

5
uint32_t flag = 0, remsets = sets, length = sets*columnsize;
size_t i, j, set = 0, index = 0;
uint8_t min, cur;

// Initialize and zero the counters[] array.

size_t counters[sets];
memset(counters, 0, sets*sizeof(size_t));
// Allocate the memory for the final sorted array.
uint8_t *target = malloc(length*sizeof(uint8_t));
if (target == NULL) {
printf("Error: mergesort_c(): malloc() \\
failed to allocate %d bytes\n", length);
exit(10);
}

// While we have remaining sets to sort, loop

while (remsets) {
// set up the initial min value to the max value,
// so that it gets reset immediately.
min = UCHAR_MAX;

// Loop over all the sets

for (j=0; j < sets; j++) {
// Check if this particular set has been finished
// (its bit is set) otherwise consider it when
// searching for the min value.
if (!((1 << j) & flag)) {
cur = a[INDEX(sets, counters[j], j)];

if (min >= cur) {

min = cur;
set = j;
}
}
}
// Insert the min value to the next available position
// in the final array
target[index++] = min;
// Increase the counter for this particular set
counters[set]++;
// If this counter exceeds the columnsize, then
// we have finished the set. Set its bit in
// the flag variable to 1 and decrease remsets.
if (counters[set] >= columnsize) {
flag |= (1 << set);
remsets--;
}
}
// Copy the target array to our given array and free target
memcpy(a, target, length);

6
 Size It. (Alticec) It. (Scalar) Ratio
8 41 14 0.33
16 86 59 0.75
32 188 286 1.10
64 440 917 1.33
128 1136 4335 2.25
256 3296 16019 2.67
512 10688 66644 3.24
1024 37760 262649 3.50
2048 141056 1064615 3.73
4096 544256 4237939 3.81
8192 2137088 16521839 3.76

Figure 3: Insertion Sort: Altivec versus Scalar (ints)

if (target)
free(target);

return;
}

You may have noticed that the code is scalar and with nothing Altivec spe-
cific. It is possible to do some Altivec optimizations in this one, especially with
regards to finding the minimum value, as Altivec provides such a function
(vec_min) that will find the minimum value in a vector in just 1 CPU cycle!
This will follow in a next revision of this paper.
Also worth mentioning, we have used a 32-bit integer and setting bits to
denote that a set has been completed. This was done for performance reasons,
but it also presents a limitation. We can’t use more than 32 sets of sorted data
to merge. But that’s quite acceptable for our purposes, as we only need 16 sets
maximum.
This Extended Merge Sort algorith, like the original algorithm is not an in-
place sorting algorithm, that is, the data get sorted in another array and then
copied back to the original array. We have used memcpy() for the copying pro-
cess, as we believe it to be faster than other methods (eg. per-element copying),
but there is of course no other particular reason or this choice. In particular,
since this function has already been vectorized for Altivec, we might benefit
from that as well.

2.1 Performance Results

Here we will present the results of using our Insertion sort for unsigned inte-
gers and characters (ie 32-bit and 8-bit integers) for 100 loops. Figure 3 presents
the case with 32-bit integers, while figure 4 presents the same results for 8-bit
unsigned characters. As you can see, the Altivec version gives really impres-
sive results! Of course this has to do with the advanced Merge Sort and the
fact that for large sizes, the chance of finding the same keys increases. This is
obvious for chars, as there are only 255 available values for a char.

7
 Size It. (Alticec) It. (Scalar) Ratio
128 2048 4199 1.00
256 4096 16429 2.00
512 8192 67361 4.25
1024 16384 259680 7.33
2048 32768 1036143 13.79
4096 65536 4116630 20.27
8192 131072 16996544 29.85
16384 262144 67815998 37.50
32768 524288 265531367 42.12

Figure 4: Insertion Sort: Altivec versus Scalar (chars)

3 Conclusions
Altivec is really a very powerful tool but so far its use is really mainly centered
around Multimedia and or Linear Algebra scientific applications. We strongly
believe that Altivec can be of real benefit to generic system-wide OS usage
as well. We will show eventually with a series of papers like this one that it
can be used for other generic applications, like data manipulation, other sort
algorithms, etc.

References
[gpu] Gpusort.
[wika] Shell sort.
[wikb] Shell sort.