I'm a total newbie to OpenCL.
I'm trying to code a reduction kernel that sums along one axis for a multi-dimensional array. I have stumbled upon that code which comes from here: https://tmramalho.github.io/blog/2014/06/16/parallel-programming-with-opencl-and-python-parallel-reduce/
__kernel void reduce(__global float *a, __global float *r, __local float *b) {
uint gid = get_global_id(0);
uint wid = get_group_id(0);
uint lid = get_local_id(0);
uint gs = get_local_size(0);
b[lid] = a[gid];
barrier(CLK_LOCAL_MEM_FENCE);
for(uint s = gs/2; s > 0; s >>= 1) {
if(lid < s) {
b[lid] += b[lid+s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0) r[wid] = b[lid];
}
I don't understand the for loop part. I get that uint s = gs/2 means that we split the array in half, but then it is a complete mystery. Without understanding it, I can't really implement another version for taking the maximum of an array for instance, even less for multi-dimensional arrays.
Furthermore, as far as I understand, the reduce kernel needs to be rerun another time if "N is bigger than the number of cores in a single unit".
Could you give me further explanations on that whole piece of code? Or even guidance on how to implement it for taking the max of an array?
Complete code can be found here: https://github.com/tmramalho/easy-pyopencl/blob/master/008_localreduce.py
Your first question about the meaning of the for loop:
for(uint s = gs/2; s > 0; s >>= 1)
It means that you divide the local size gs by 2, and keep dividing by 2 (the shift part s >>= 1 is equivalent to s = s/2) while s > 0, in other words, until s = 1. This algorithm depends on your array's size being a power of 2, otherwise you'd have to deal with the excess of a power of 2 until you have reduced the whole array, or you'd have to fill your array with neutral values for the reduction until completing a power of 2 size.
Your second concern when N is bigger than the capacity of your GPU, you are right: you have to run your reduction in portions that fit and then merge the results.
Finally, when you ask for guidance on how to implement a reduction to get the max of an array, I would suggest the following:
For a simple reduction like max or sum, try using numpy, especially if you are dealing with programming the reduction by axis.
If you think that the GPU would give you an advantage, try first using pyopencl's Multidimensional Array functionality, e.g. max.
If the reduction is more math intensive, try using pyopencl's Parallel Algorithms, e.g. reduction
I think that the whole point of using pyopencl is to avoid dealing with the underlying GPU's architecture. Otherwise, it is easier to deal with CUDA or HIP directly instead of OpenCL.
I want to sort a huge array, say 10^8 entries of type X with at most N different keys, where N is ~10^2. Because I don't know the range or spacing of the elements, count sort is not an option. So my best guess so far is to use a hash map for the counts like so
std::unordered_map< X, unsigned > counts;
for (auto x : input)
counts[x]++;
This works ok-ish and is ~4 times faster than 3-way quicksort, but I'm a nervous person and it's still not fast enough.
I wonder: am I missing something? Can I make better use of the fact that N is known in advance? Or is it possible to tune the hash map to my needs?
EDIT An additional pre-condition is that the input sequence is badly sorted and the frequency of the keys is about the same.
STL implementations are often not perfect in terms of performance (no holy wars, please).
If you know a guaranteed and sensible upper on the number of unique elements (N), then you can trivially implement your own hash table of size 2^s >> N. Here is how I usually do it myself:
int size = 1;
while (size < 3 * N) size <<= 1;
//Note: at least 3X size factor, size = power of two
//count = -1 means empty entry
std::vector<std::pair<X, int>> table(size, make_pair(X(), -1));
auto GetHash = [size](X val) -> int { return std::hash<X>()(val) & (size-1); };
for (auto x : input) {
int cell = GetHash(x);
bool ok = false;
for (; table[cell].second >= 0; cell = (cell + 1) & (size-1)) {
if (table[cell].first == x) { //match found -> stop
ok = true;
break;
}
}
if (!ok) { //match not found -> add entry on free place
table[cell].first = x;
table[cell].second = 0;
}
table[cell].second++; //increment counter
}
On MSVC2013, it improves time from 0.62 secs to 0.52 secs compared to your code, given that int is used as type X.
Also, we can choose a faster hash function. Note however, that the choice of hash function depends heavily on the properties of the input. Let's take Knuth's multiplicative hash:
auto GetHash = [size](X val) -> int { return (val*2654435761) & (size-1); };
It further improves time to 0.34 secs.
As a conclusion: do you really want to reimplement standard data structures to achieve a 2X speed boost?
Notes: Speedup may be entirely different on another compiler/machine. You may have to do some hacks if your type X is not POD.
Counting sort really would by best, but isnt applicable due to unknown range or spacing.
Seems to be easily parallelized with fork-join, e.g. boost::thread.
You could also try a more efficient, handrolled hashmap. Unorded_map typically uses linked lists to counter potentially bad hash functions. The memory overhead of linked lists may hurt performance if the hashtable doesnt fit into L1 cache. Closed Hashing may use less memory. Some hints for optimizing:
Closed Hashing with linear probing and without support for removal
power of two sized hashtable for bit shifting instead of modulo (division requires multiple cycles and there is only one hardware divider per core)
Low LoadFactor (entries through size) to minimize collisions. Thats a tradeof between memory usage and number of collisions. A LoadFactor over 0.5 should be avoided. A hashtable-size of 256 seems suitable for 100 entries.
cheapo hash function. You havent shown the type of X, so perhaps a cheaper hash function could outweigh more collisions.
I would look to store items in a sorted vector, as about 100 keys, would mean inserting into the vector would only occur 1 in 10^6 entries. Lookup would be processor efficient bsearch in vector
I'm using CUBLAS (Cuda Blas libraries) for matrix operations.
Is possible to use CUBLAS to achieve the exponentiation/root mean square of a matrix items?
I mean, having the 2x2 matrix
1 4
9 16
What I want is a function to elevate to a given value e.g. 2
1 16
81 256
and calculating the root mean square e.g.
1 2
3 4
Is this possible with CUBLAS? I can't find a function suitable to this goal, but I'll ask here first to begin coding my own kernel.
So this may well be something you do have to implement yourself, because the library won't do it for you. (There's probably some way to implement it some of it in terms of BLAS level 3 routines - certainly the squaring of the matrix elements - but it would involve expensive and otherwise unnecessary matrix-vector multiplications. And I still don't know how you'd do the squareroot operation). The reason is that these operations aren't really linear-algebra procedures; taking the square root of each matrix element doesn't really correspond to any fundamental linear algebra operation.
The good news is that these elementwise operations are very simple to implement in CUDA. Again, there are lots of tuning options one could play with for best performance, but one can get started fairly easily.
As with the matrix addition operations, you'll be treating the NxM matricies here as (N*M)-length vectors; the structure of the matrix doesn't matter for these elementwise operations. So you'll be passing in a pointer to the first element of the matrix and treating it as a single list of N*M numbers. (I'm going to assume you're using floats here, as you were talking about SGEMM and SAXPY earlier.)
The kernel, the actual bit of CUDA code which implements the operation, is quite simple. For now, each thread will compute the square (or squareroot) of one array element. (Whether this is optimal or not for performance is something you could test). So the kernels would look like the following. I'm assuming you're doing something like B_ij = (A_ij)^2; if you wanted to do the operation inplace, eg A_ij = (A_ij)^2, you could do that, too:
__global__ void squareElements(float *a, float *b, int N) {
/* which element does this compute? */
int tid = blockDim.x * blockIdx.x + threadIdx.x;
/* if valid, squre the array element */
if (tid < N)
b[tid] = (a[tid]*a[tid]);
}
__global__ void sqrtElements(float *a, float *b, int N) {
/* which element does this compute? */
int tid = blockDim.x * blockIdx.x + threadIdx.x;
/* if valid, sqrt the array element */
if (tid < N)
b[tid] = sqrt(a[tid]); /* or sqrtf() */
}
Note that if you're ok with very slightly increased error, the 'sqrtf()' function which has maximum error of 3 ulp (units in the last place) is significantly faster.
How you call these kernels will depend on the order in which you're doing things. If you've already made some CUBLAS calls on these matricies, you'll want to use them on the arrays which are already in GPU memory.
This is a long text. Please bear with me. Boiled down, the question is: Is there a workable in-place radix sort algorithm?
Preliminary
I've got a huge number of small fixed-length strings that only use the letters “A”, “C”, “G” and “T” (yes, you've guessed it: DNA) that I want to sort.
At the moment, I use std::sort which uses introsort in all common implementations of the STL. This works quite well. However, I'm convinced that radix sort fits my problem set perfectly and should work much better in practice.
Details
I've tested this assumption with a very naive implementation and for relatively small inputs (on the order of 10,000) this was true (well, at least more than twice as fast). However, runtime degrades abysmally when the problem size becomes larger (N > 5,000,000).
The reason is obvious: radix sort requires copying the whole data (more than once in my naive implementation, actually). This means that I've put ~ 4 GiB into my main memory which obviously kills performance. Even if it didn't, I can't afford to use this much memory since the problem sizes actually become even larger.
Use Cases
Ideally, this algorithm should work with any string length between 2 and 100, for DNA as well as DNA5 (which allows an additional wildcard character “N”), or even DNA with IUPAC ambiguity codes (resulting in 16 distinct values). However, I realize that all these cases cannot be covered, so I'm happy with any speed improvement I get. The code can decide dynamically which algorithm to dispatch to.
Research
Unfortunately, the Wikipedia article on radix sort is useless. The section about an in-place variant is complete rubbish. The NIST-DADS section on radix sort is next to nonexistent. There's a promising-sounding paper called Efficient Adaptive In-Place Radix Sorting which describes the algorithm “MSL”. Unfortunately, this paper, too, is disappointing.
In particular, there are the following things.
First, the algorithm contains several mistakes and leaves a lot unexplained. In particular, it doesn’t detail the recursion call (I simply assume that it increments or reduces some pointer to calculate the current shift and mask values). Also, it uses the functions dest_group and dest_address without giving definitions. I fail to see how to implement these efficiently (that is, in O(1); at least dest_address isn’t trivial).
Last but not least, the algorithm achieves in-place-ness by swapping array indices with elements inside the input array. This obviously only works on numerical arrays. I need to use it on strings. Of course, I could just screw strong typing and go ahead assuming that the memory will tolerate my storing an index where it doesn’t belong. But this only works as long as I can squeeze my strings into 32 bits of memory (assuming 32 bit integers). That's only 16 characters (let's ignore for the moment that 16 > log(5,000,000)).
Another paper by one of the authors gives no accurate description at all, but it gives MSL’s runtime as sub-linear which is flat out wrong.
To recap: Is there any hope of finding a working reference implementation or at least a good pseudocode/description of a working in-place radix sort that works on DNA strings?
Well, here's a simple implementation of an MSD radix sort for DNA. It's written in D because that's the language that I use most and therefore am least likely to make silly mistakes in, but it could easily be translated to some other language. It's in-place but requires 2 * seq.length passes through the array.
void radixSort(string[] seqs, size_t base = 0) {
if(seqs.length == 0)
return;
size_t TPos = seqs.length, APos = 0;
size_t i = 0;
while(i < TPos) {
if(seqs[i][base] == 'A') {
swap(seqs[i], seqs[APos++]);
i++;
}
else if(seqs[i][base] == 'T') {
swap(seqs[i], seqs[--TPos]);
} else i++;
}
i = APos;
size_t CPos = APos;
while(i < TPos) {
if(seqs[i][base] == 'C') {
swap(seqs[i], seqs[CPos++]);
}
i++;
}
if(base < seqs[0].length - 1) {
radixSort(seqs[0..APos], base + 1);
radixSort(seqs[APos..CPos], base + 1);
radixSort(seqs[CPos..TPos], base + 1);
radixSort(seqs[TPos..seqs.length], base + 1);
}
}
Obviously, this is kind of specific to DNA, as opposed to being general, but it should be fast.
Edit:
I got curious whether this code actually works, so I tested/debugged it while waiting for my own bioinformatics code to run. The version above now is actually tested and works. For 10 million sequences of 5 bases each, it's about 3x faster than an optimized introsort.
I've never seen an in-place radix sort, and from the nature of the radix-sort I doubt that it is much faster than a out of place sort as long as the temporary array fits into memory.
Reason:
The sorting does a linear read on the input array, but all writes will be nearly random. From a certain N upwards this boils down to a cache miss per write. This cache miss is what slows down your algorithm. If it's in place or not will not change this effect.
I know that this will not answer your question directly, but if sorting is a bottleneck you may want to have a look at near sorting algorithms as a preprocessing step (the wiki-page on the soft-heap may get you started).
That could give a very nice cache locality boost. A text-book out-of-place radix sort will then perform better. The writes will still be nearly random but at least they will cluster around the same chunks of memory and as such increase the cache hit ratio.
I have no idea if it works out in practice though.
Btw: If you're dealing with DNA strings only: You can compress a char into two bits and pack your data quite a lot. This will cut down the memory requirement by factor four over a naiive representation. Addressing becomes more complex, but the ALU of your CPU has lots of time to spend during all the cache-misses anyway.
You can certainly drop the memory requirements by encoding the sequence in bits.
You are looking at permutations so, for length 2, with "ACGT" that's 16 states, or 4 bits.
For length 3, that's 64 states, which can be encoded in 6 bits. So it looks like 2 bits for each letter in the sequence, or about 32 bits for 16 characters like you said.
If there is a way to reduce the number of valid 'words', further compression may be possible.
So for sequences of length 3, one could create 64 buckets, maybe sized uint32, or uint64.
Initialize them to zero.
Iterate through your very very large list of 3 char sequences, and encode them as above.
Use this as a subscript, and increment that bucket.
Repeat this until all of your sequences have been processed.
Next, regenerate your list.
Iterate through the 64 buckets in order, for the count found in that bucket, generate that many instances of the sequence represented by that bucket.
when all of the buckets have been iterated, you have your sorted array.
A sequence of 4, adds 2 bits, so there would be 256 buckets.
A sequence of 5, adds 2 bits, so there would be 1024 buckets.
At some point the number of buckets will approach your limits.
If you read the sequences from a file, instead of keeping them in memory, more memory would be available for buckets.
I think this would be faster than doing the sort in situ as the buckets are likely to fit within your working set.
Here is a hack that shows the technique
#include <iostream>
#include <iomanip>
#include <math.h>
using namespace std;
const int width = 3;
const int bucketCount = exp(width * log(4)) + 1;
int *bucket = NULL;
const char charMap[4] = {'A', 'C', 'G', 'T'};
void setup
(
void
)
{
bucket = new int[bucketCount];
memset(bucket, '\0', bucketCount * sizeof(bucket[0]));
}
void teardown
(
void
)
{
delete[] bucket;
}
void show
(
int encoded
)
{
int z;
int y;
int j;
for (z = width - 1; z >= 0; z--)
{
int n = 1;
for (y = 0; y < z; y++)
n *= 4;
j = encoded % n;
encoded -= j;
encoded /= n;
cout << charMap[encoded];
encoded = j;
}
cout << endl;
}
int main(void)
{
// Sort this sequence
const char *testSequence = "CAGCCCAAAGGGTTTAGACTTGGTGCGCAGCAGTTAAGATTGTTT";
size_t testSequenceLength = strlen(testSequence);
setup();
// load the sequences into the buckets
size_t z;
for (z = 0; z < testSequenceLength; z += width)
{
int encoding = 0;
size_t y;
for (y = 0; y < width; y++)
{
encoding *= 4;
switch (*(testSequence + z + y))
{
case 'A' : encoding += 0; break;
case 'C' : encoding += 1; break;
case 'G' : encoding += 2; break;
case 'T' : encoding += 3; break;
default : abort();
};
}
bucket[encoding]++;
}
/* show the sorted sequences */
for (z = 0; z < bucketCount; z++)
{
while (bucket[z] > 0)
{
show(z);
bucket[z]--;
}
}
teardown();
return 0;
}
If your data set is so big, then I would think that a disk-based buffer approach would be best:
sort(List<string> elements, int prefix)
if (elements.Count < THRESHOLD)
return InMemoryRadixSort(elements, prefix)
else
return DiskBackedRadixSort(elements, prefix)
DiskBackedRadixSort(elements, prefix)
DiskBackedBuffer<string>[] buckets
foreach (element in elements)
buckets[element.MSB(prefix)].Add(element);
List<string> ret
foreach (bucket in buckets)
ret.Add(sort(bucket, prefix + 1))
return ret
I would also experiment grouping into a larger number of buckets, for instance, if your string was:
GATTACA
the first MSB call would return the bucket for GATT (256 total buckets), that way you make fewer branches of the disk based buffer. This may or may not improve performance, so experiment with it.
I'm going to go out on a limb and suggest you switch to a heap/heapsort implementation. This suggestion comes with some assumptions:
You control the reading of the data
You can do something meaningful with the sorted data as soon as you 'start' getting it sorted.
The beauty of the heap/heap-sort is that you can build the heap while you read the data, and you can start getting results the moment you have built the heap.
Let's step back. If you are so fortunate that you can read the data asynchronously (that is, you can post some kind of read request and be notified when some data is ready), and then you can build a chunk of the heap while you are waiting for the next chunk of data to come in - even from disk. Often, this approach can bury most of the cost of half of your sorting behind the time spent getting the data.
Once you have the data read, the first element is already available. Depending on where you are sending the data, this can be great. If you are sending it to another asynchronous reader, or some parallel 'event' model, or UI, you can send chunks and chunks as you go.
That said - if you have no control over how the data is read, and it is read synchronously, and you have no use for the sorted data until it is entirely written out - ignore all this. :(
See the Wikipedia articles:
Heapsort
Binary heap
"Radix sorting with no extra space" is a paper addressing your problem.
Performance-wise you might want to look at a more general string-comparison sorting algorithms.
Currently you wind up touching every element of every string, but you can do better!
In particular, a burst sort is a very good fit for this case. As a bonus, since burstsort is based on tries, it works ridiculously well for the small alphabet sizes used in DNA/RNA, since you don't need to build any sort of ternary search node, hash or other trie node compression scheme into the trie implementation. The tries may be useful for your suffix-array-like final goal as well.
A decent general purpose implementation of burstsort is available on source forge at http://sourceforge.net/projects/burstsort/ - but it is not in-place.
For comparison purposes, The C-burstsort implementation covered at http://www.cs.mu.oz.au/~rsinha/papers/SinhaRingZobel-2006.pdf benchmarks 4-5x faster than quicksort and radix sorts for some typical workloads.
You'll want to take a look at Large-scale Genome Sequence Processing by Drs. Kasahara and Morishita.
Strings comprised of the four nucleotide letters A, C, G, and T can be specially encoded into Integers for much faster processing. Radix sort is among many algorithms discussed in the book; you should be able to adapt the accepted answer to this question and see a big performance improvement.
You might try using a trie. Sorting the data is simply iterating through the dataset and inserting it; the structure is naturally sorted, and you can think of it as similar to a B-Tree (except instead of making comparisons, you always use pointer indirections).
Caching behavior will favor all of the internal nodes, so you probably won't improve upon that; but you can fiddle with the branching factor of your trie as well (ensure that every node fits into a single cache line, allocate trie nodes similar to a heap, as a contiguous array that represents a level-order traversal). Since tries are also digital structures (O(k) insert/find/delete for elements of length k), you should have competitive performance to a radix sort.
I would burstsort a packed-bit representation of the strings. Burstsort is claimed to have much better locality than radix sorts, keeping the extra space usage down with burst tries in place of classical tries. The original paper has measurements.
It looks like you've solved the problem, but for the record, it appears that one version of a workable in-place radix sort is the "American Flag Sort". It's described here: Engineering Radix Sort. The general idea is to do 2 passes on each character - first count how many of each you have, so you can subdivide the input array into bins. Then go through again, swapping each element into the correct bin. Now recursively sort each bin on the next character position.
Radix-Sort is not cache conscious and is not the fastest sort algorithm for large sets.
You can look at:
ti7qsort. ti7qsort is the fastest sort for integers (can be used for small-fixed size strings).
Inline QSORT
String sorting
You can also use compression and encode each letter of your DNA into 2 bits before storing into the sort array.
dsimcha's MSB radix sort looks nice, but Nils gets closer to the heart of the problem with the observation that cache locality is what's killing you at large problem sizes.
I suggest a very simple approach:
Empirically estimate the largest size m for which a radix sort is efficient.
Read blocks of m elements at a time, radix sort them, and write them out (to a memory buffer if you have enough memory, but otherwise to file), until you exhaust your input.
Mergesort the resulting sorted blocks.
Mergesort is the most cache-friendly sorting algorithm I'm aware of: "Read the next item from either array A or B, then write an item to the output buffer." It runs efficiently on tape drives. It does require 2n space to sort n items, but my bet is that the much-improved cache locality you'll see will make that unimportant -- and if you were using a non-in-place radix sort, you needed that extra space anyway.
Please note finally that mergesort can be implemented without recursion, and in fact doing it this way makes clear the true linear memory access pattern.
First, think about the coding of your problem. Get rid of the strings, replace them by a binary representation. Use the first byte to indicate length+encoding. Alternatively, use a fixed length representation at a four-byte boundary. Then the radix sort becomes much easier. For a radix sort, the most important thing is to not have exception handling at the hot spot of the inner loop.
OK, I thought a bit more about the 4-nary problem. You want a solution like a Judy tree for this. The next solution can handle variable length strings; for fixed length just remove the length bits, that actually makes it easier.
Allocate blocks of 16 pointers. The least significant bit of the pointers can be reused, as your blocks will always be aligned. You might want a special storage allocator for it (breaking up large storage into smaller blocks). There are a number of different kinds of blocks:
Encoding with 7 length bits of variable-length strings. As they fill up, you replace them by:
Position encodes the next two characters, you have 16 pointers to the next blocks, ending with:
Bitmap encoding of the last three characters of a string.
For each kind of block, you need to store different information in the LSBs. As you have variable length strings you need to store end-of-string too, and the last kind of block can only be used for the longest strings. The 7 length bits should be replaced by less as you get deeper into the structure.
This provides you with a reasonably fast and very memory efficient storage of sorted strings. It will behave somewhat like a trie. To get this working, make sure to build enough unit tests. You want coverage of all block transitions. You want to start with only the second kind of block.
For even more performance, you might want to add different block types and a larger size of block. If the blocks are always the same size and large enough, you can use even fewer bits for the pointers. With a block size of 16 pointers, you already have a byte free in a 32-bit address space. Take a look at the Judy tree documentation for interesting block types. Basically, you add code and engineering time for a space (and runtime) trade-off
You probably want to start with a 256 wide direct radix for the first four characters. That provides a decent space/time tradeoff. In this implementation, you get much less memory overhead than with a simple trie; it is approximately three times smaller (I haven't measured). O(n) is no problem if the constant is low enough, as you noticed when comparing with the O(n log n) quicksort.
Are you interested in handling doubles? With short sequences, there are going to be. Adapting the blocks to handle counts is tricky, but it can be very space-efficient.
While the accepted answer perfectly answers the description of the problem, I've reached this place looking in vain for an algorithm to partition inline an array into N parts. I've written one myself, so here it is.
Warning: this is not a stable partitioning algorithm, so for multilevel partitioning, one must repartition each resulting partition instead of the whole array. The advantage is that it is inline.
The way it helps with the question posed is that you can repeatedly partition inline based on a letter of the string, then sort the partitions when they are small enough with the algorithm of your choice.
function partitionInPlace(input, partitionFunction, numPartitions, startIndex=0, endIndex=-1) {
if (endIndex===-1) endIndex=input.length;
const starts = Array.from({ length: numPartitions + 1 }, () => 0);
for (let i = startIndex; i < endIndex; i++) {
const val = input[i];
const partByte = partitionFunction(val);
starts[partByte]++;
}
let prev = startIndex;
for (let i = 0; i < numPartitions; i++) {
const p = prev;
prev += starts[i];
starts[i] = p;
}
const indexes = [...starts];
starts[numPartitions] = prev;
let bucket = 0;
while (bucket < numPartitions) {
const start = starts[bucket];
const end = starts[bucket + 1];
if (end - start < 1) {
bucket++;
continue;
}
let index = indexes[bucket];
if (index === end) {
bucket++;
continue;
}
let val = input[index];
let destBucket = partitionFunction(val);
if (destBucket === bucket) {
indexes[bucket] = index + 1;
continue;
}
let dest;
do {
dest = indexes[destBucket] - 1;
let destVal;
let destValBucket = destBucket;
while (destValBucket === destBucket) {
dest++;
destVal = input[dest];
destValBucket = partitionFunction(destVal);
}
input[dest] = val;
indexes[destBucket] = dest + 1;
val = destVal;
destBucket = destValBucket;
} while (dest !== index)
}
return starts;
}