Description
As your final CS194 homework, you’ll be implementing radix sort on GPUs using OpenCL. While
this might sound scary, you’ve already implemented the key kernel in data parallel radix sort, prefix scan,
in last week’s homework. After implementing radix sort, you can truly call yourself a GPU programmer.
1 Background on radix sort
You’ll be implementing a least significant digit (LSD) version of radix sort. In this version of radix
sort, we’ll process keys starting with the least significant bit (bit 0) and proceed to the most significant
bit (bit 31).
At each stage (bit) of the LSB radix sort, we place keys into two buckets. One bucket holds keys
with the current bit set while the other bucket holds keys with bit unset. While we split the keys into
two separate bucks, the key to the correctness of radix is that the original order of the keys (within a
bucket) is preserved. The split computation is performed 32 times (for the case of 32-bit) integers to
fully sort the array.
1 for (int32 t k = 0; k < 3 2; k++)
2 {
3 /⇤ compute en tr ies with set lsb ⇤/
4 for (int32 t i = 0; i < n ; i++)
5 temp [ i ] = ( in [ i ] >> k) & 0x1 ;
6
7 /⇤ compute ind ices of elements with 0 in lsb ⇤/
8 scan ( temp , zeros , n , 0 );
9 /⇤ compute ind ices of elements with 1 in lsb ⇤/
10 sc an ( temp , ones , n , 1 ) ;
11
12 /⇤ scatter partially ordered output ⇤/
13 for (int32 t i = 0; i < n ; i++)
14 {
15 i f ( temp [ i ] )
16 id x = z e r o s [ n1] + ones [ i ] 1 ;
17 else
18 id x = z e r o s [ i ] 1 ;
19 out [ id x ] = i n [ i ] ;
20 }
21
22 /⇤ Not an inplace sort ,
23 ⇤ need a pointer swap ⇤/
24 tP t r = i n ;
25 i n = out ;
26 out = tP t r ;
27 }
Listing 1: An approach for implementing data-parallel radix sort
Listing 1 presents the core of the data-parallel radix sort algorithm. Line 1 iterates over the 32 bits
in an int32 t while lines 4 and 5 compute if each key has the current bit set. Lines 8 and 10 compute
an additive scan on the set bit temporary array in order to find the output index.
Finally, we need to
merge the two buckets back into a partially sorted array. The address computation for keys with a zero
in the current bit under inspection is straightforward and only requires accessing the zero scan array.
The set bit index requires using the last entry of the zero array as an o↵set (think about why). Finally,
radix sort is an out-of-place algorithm, so lines 24 through 26 handle double bu↵ering.
2 Your task
For this assignment, you’ll need to implement a fully functioning GPU radix sort. We’ll provide the
core of the scan primitive, but you’ll need to add a little code to the scan kernel though (from HW5).
Your primary challenge is to implement the rest of the algorithm. This should require implementing just
one new kernel then hooking everything appropriately. The starter code includes a correctness checking
routine too, so you’ll know when your code works.
2.1 Submit your code
We want your commented radix sort code. Make sure the comments explain how your code works
and any design decisions you made when implementing radix sort.
2.2 How does it perform?
Our correctness testing code calls the C standard library implementation of quicksort. We also record
the CPU runtime of quicksort and the GPU runtime of radix sort. If you use the included benchmarking
script (run with the command “bash ./bench.sh”), you will be rewarded with a CSV (comma-separated
values) file that contains performance data for both CPU and GPU implementations. Please plot these
results and explain the performance behavior.
In particular, what array length is required for better
performance on the GPU than the CPU? We want both your graph and a short paragraph of analysis
included in a PDF document submitted with your code.