I'm trying to write a trimmed mean kernel that takes as input a set of frames (~100). I'm thinking of using an insertion sort (of size ~8). This means that I'll need to read one float/ uint/ushort at a time from the input images and compare it against an 8-wide vector, shifting the elements up and inserting the new value at the correct spot (if necessary), with the largest value added to the mean.
I'm having difficulties finding a portable way of shifting the elements in the vector and inserting the new one at the correct spot. I know that AMD GPUs have ds_permute for example, but those are not portable, and I can't figure out a clever way of using arithmetic and relational operators to do it (since those operate only on their lane and AFAIK unaligned vector accesses are UB in OpenCL).
If you only have 8 items in your list then you could add some indirection and have an index table uchar[8]. You assign the pre-sorted elements values 0-7. As you perform the sort you don't rearrange those items, instead you insert their indices into the table.
To get the speedup you then need to store each index using 4 bits to that all 8 fit into a 32-bit word. Honestly, I don't think this will be faster in your case though.
float elements[8];
uint index_table = 0;
uint sorted_size = 0;
// insert elements[i]
void insert(uint i)
{
uint temp = index_table
for (j = 0; j < sorted_size ; ++j)
{
if (elements[i] < elements[temp & 0xf])
{
// Insert i
temp = (temp << 4) | i;
index_table = (index_table & (4 * j - 1)) | (temp << (4 * j));
return;
}
temp >>= 4;
}
// Insert at end
index_table |= i << 4 * sorted_size ;
}
void insertion_sort()
{
// We can skip the first iteration since the 1st element is always inserted at the start
for (sorted_size = 1; sorted_size < 8; ++sorted_size)
{
insert(sorted_size);
}
}
float ith_smallest(uint i)
{
return elements[(index_table >> 4 * i) & 0xf];
}
Related
I am learning opencl for the first time, and I am currently modifying the shortest path finding algorithm. I know that opencl usually uses the idea of parallel computing to solve problems. So I wonder if I can also use this parallel idea when I am dealing with finding the minimum value and its position in the array?
This is my previous attempt. I think that as long as the variable is the smallest, the result can be obtained regardless of whether the operation is locked or not. Unfortunately, when I use printf to view variables, although valid nodes have been judged, I can't get the correct results.
__kernel void findWay(__global int* A, __global int* B, __global int* minNode, __global int* minDis, __global int* isFinish)
{
//A: weightMatrix , B: usedNode
//dijkstra algorithm , src node is 0
size_t dst = get_global_id(1);
size_t src = get_global_id(0);
size_t vCount = get_global_size(0);
int index = dst * vCount + src;
while(isFinish[0] != vCount){
if((src == minNode[0])&&(B[dst] == 0)&&(A[index] != INT_MAX)){
A[dst*vCount] = min(A[dst*vCount + 0],A[minNode[0]*vCount + 0] + A[index]);
}
minDis[0] = INT_MAX;
barrier(CLK_GLOBAL_MEM_FENCE);
//here is the bug
if((src == 0) &&(B[dst] == 0)){
if(minDis[0] > A[index]){
minDis[0] = A[index];
minNode[0] = dst;
}
}
//=========
barrier(CLK_GLOBAL_MEM_FENCE);
B[minNode[0]] = 1;
if(index == 0){
isFinish[0]++;
}
}
}
In the end, I can only use a normal way to achieve this operation.
if((src == 0) &&(dst == 0)){
for(int i = 0 ; i < vCount ;i++){
if(B[i] == 0 && minDis[0] > A[i *vCount]){
minDis[0] = A[i*vCount];
minNode[0] = i;
}
}
I would like to ask about this search process, can the looping step be omitted?
Horizontal operations on the parallelized array are difficult. The general approach to them is binary-tree-like kernel passes. Start with the original array, make each GPU thread load 2 neighboring elements and choose the smaller one, write that in the same array to position of the first of the two elements. Next kernel loads two elements from the list of every second element, compares the two, writes the smaller one in the first position of the two. Repeat until there is only one element left.
I will illustrate it beloe. I mark values that are not touched by the kernel anymore with *.
original array: 5|2|1|6|9|3|4|8
after 1st kernel pass: 2 *|1 *|3 *|4 *
after 2nd kernel pass: 1 * * *|3 * * *
after 3nd kernel pass: 1 * * * * * * *
smallest element is 1.
I'm working on a problem in which I would greatly benefit from being able to load vectors that are saved in disk dynamically inside a loop as this allows me to skip calculating the vectors on the fly (in my actual process one vector is used many times and the collection of vectors as a matrix is too big to have in memory all at once). As a simplified example, lets say that we have the vectors stored in a directory with path prefix (each in its own file). The names of these files are vec0.txt, vec1.txt, vec2.txt, ... etc. We wish to sum all the numbers of all specified vectors in the inclusive range start-end. The size of all vectors is known and is always the same. I thought of something like:
library(Rcpp)
cppFunction('int sumvectors(int start, int end, string prefix, int size) {
int i;
int j;
int arr[size];
int sum=0;
for (i=start; i <= end; i++) {
// Here you would construct the path to the file paste0(prefix, vec, i, ".txt")
// Then load it and put it into an array
for (j=0; j <= size; j++) {
sum+=arr[j];
}
}
return sum;
}')
Is something like this even possible? I'm ok at R but never worked with C or C++ so I don't really even know if this is even doable with Rcpp
Yes, this is certainly possible. If your numbers are written in plain text files separated by spaces like this:
C://Users/Administrator/vec1.txt
5.1 21.4 563 -21.2 35.6
C://Users/Administrator/vec2.txt
3 6 8 7 10 135
Then you can write the following function:
cppFunction("
std::vector<float> read_floats(const std::string& path)
{
std::vector<float> result;
for(int i = 1; i < 3; ++i)
{
std::string file_path = path + std::to_string(i) + \".txt\";
std::ifstream myfile(file_path.c_str(), std::ios_base::in);
float a, vec_sum = 0;
std::vector<float> vec;
while(myfile >> a)
{
vec.push_back(a);
}
for(std::vector<float>::iterator it = vec.begin(); it != vec.end(); ++it)
{
vec_sum += *it;
}
result.push_back(vec_sum);
}
return result;
}", include = c("#include<string>", "#include<fstream>", "#include<vector>"))
Which creates an R function that allows you to do this:
read_floats("c:/Users/Administrator/vec")
#> [1] 603.9 169.0
Which you can confirm is the sum of the numbers in each file.
Given a bag with a maximum of 100 chips,each chip has its value written over it.
Determine the most fair division between two persons. This means that the difference between the amount each person obtains should be minimized. The value of a chips varies from 1 to 1000.
Input: The number of coins m, and the value of each coin.
Output: Minimal positive difference between the amount the two persons obtain when they divide the chips from the corresponding bag.
I am finding it difficult to form a DP solution for it. Please help me.
Initially I had to tried it as a Non DP solution.Actually I havent thought of solving it using DP. I simply sorted the value array. And assigned the largest value to one of the person, and incrementally assigned the other values to one of the two depending upon which creates minimum difference. But that solution actually didnt work.
I am posting my solution here :
bool myfunction(int i, int j)
{
return(i >= j) ;
}
int main()
{
int T, m, sum1, sum2, temp_sum1, temp_sum2,i ;
cin >> T ;
while(T--)
{
cin >> m ;
sum1 = 0 ; sum2 = 0 ; temp_sum1 = 0 ; temp_sum2 = 0 ;
vector<int> arr(m) ;
for(i=0 ; i < m ; i++)
{
cin>>arr[i] ;
}
if(m==1 )
{
if(arr[0]%2==0)
cout<<0<<endl ;
else
cout<<1<<endl ;
}
else {
sort(arr.begin(), arr.end(), myfunction) ;
// vector<int> s1 ;
// vector<int> s2 ;
for(i=0 ; i < m ; i++)
{
temp_sum1 = sum1 + arr[i] ;
temp_sum2 = sum2 + arr[i] ;
if(abs(temp_sum1 - sum2) <= abs(temp_sum2 -sum1))
{
sum1 = sum1 + arr[i] ;
}
else
{
sum2 = sum2 + arr[i] ;
}
temp_sum1 = 0 ;
temp_sum2 = 0 ;
}
cout<<abs(sum1 -sum2)<<endl ;
}
}
return 0 ;
}
what i understand from your question is you want to divide chips in two persons so as to minimize the difference between sum of numbers written on those.
If understanding is correct, then potentially you can follow below approach to arrive at solution.
Sort the values array i.e. int values[100]
Start adding elements from both ends of array in for loop i.e. for(i=0; j=values.length;i<j;i++,j--)
Odd numbered iteration sum belongs to one person & even numbered sum to other person
run the loop till i < j
now, the difference between two sums obtained in odd & even iterations should be minimum as array was sorted earlier.
If my understanding of the question is correct, then this solution should resolve your problem.
Reflect as appropriate.
Thanks
Ravindra
Have you tried the latest Codility test?
I felt like there was an error in the definition of what a K-Sparse number is that left me confused and I wasn't sure what the right way to proceed was. So it starts out by defining a K-Sparse Number:
In the binary number "100100010000" there are at least two 0s between
any two consecutive 1s. In the binary number "100010000100010" there
are at least three 0s between any two consecutive 1s. A positive
integer N is called K-sparse if there are at least K 0s between any
two consecutive 1s in its binary representation. (My emphasis)
So the first number you see, 100100010000 is 2-sparse and the second one, 100010000100010, is 3-sparse. Pretty simple, but then it gets down into the algorithm:
Write a function:
class Solution { public int sparse_binary_count(String S,String T,int K); }
that, given:
string S containing a binary representation of some positive integer A,
string T containing a binary representation of some positive integer B,
a positive integer K.
returns the number of K-sparse integers within the range [A..B] (both
ends included)
and then states this test case:
For example, given S = "101" (A = 5), T = "1111" (B=15) and K=2, the
function should return 2, because there are just two 2-sparse integers
in the range [5..15], namely "1000" (i.e. 8) and "1001" (i.e. 9).
Basically it is saying that 8, or 1000 in base 2, is a 2-sparse number, even though it does not have two consecutive ones in its binary representation. What gives? Am I missing something here?
Tried solving that one. The assumption that the problem makes about binary representations of "power of two" numbers being K sparse by default is somewhat confusing and contrary.
What I understood was 8-->1000 is 2 power 3 so 8 is 3 sparse. 16-->10000 2 power 4 , and hence 4 sparse.
Even we assume it as true , and if you are interested in below is my solution code(C) for this problem. Doesn't handle some cases correctly, where there are powers of two numbers involved in between the two input numbers, trying to see if i can fix that:
int sparse_binary_count (const string &S,const string &T,int K)
{
char buf[50];
char *str1,*tptr,*Sstr,*Tstr;
int i,len1,len2,cnt=0;
long int num1,num2;
char *pend,*ch;
Sstr = (char *)S.c_str();
Tstr = (char *)T.c_str();
str1 = (char *)malloc(300001);
tptr = str1;
num1 = strtol(Sstr,&pend,2);
num2 = strtol(Tstr,&pend,2);
for(i=0;i<K;i++)
{
buf[i] = '0';
}
buf[i] = '\0';
for(i=num1;i<=num2;i++)
{
str1 = tptr;
if( (i & (i-1))==0)
{
if(i >= (pow((float)2,(float)K)))
{
cnt++;
continue;
}
}
str1 = myitoa(i,str1,2);
ch = strstr(str1,buf);
if(ch == NULL)
continue;
else
{
if((i % 2) != 0)
cnt++;
}
}
return cnt;
}
char* myitoa(int val, char *buf, int base){
int i = 299999;
int cnt=0;
for(; val && i ; --i, val /= base)
{
buf[i] = "0123456789abcdef"[val % base];
cnt++;
}
buf[i+cnt+1] = '\0';
return &buf[i+1];
}
There was an information within the test details, showing this specific case. According to this information, any power of 2 is considered K-sparse for any K.
You can solve this simply by binary operations on integers. You are even able to tell, that you will find no K-sparse integers bigger than some specific integer and lower than (or equal to) integer represented by T.
As far as I can see, you must pay also a lot of attention to the performance, as there are sometimes hundreds of milions of integers to be checked.
My own solution, written in Python, working very efficiently even on large ranges of integers and being successfully tested for many inputs, has failed. The results were not very descriptive, saying it does not work as required within question (although it meets all the requirements in my opinion).
/////////////////////////////////////
solutions with bitwise operators:
no of bits per int = 32 on 32 bit system,check for pattern (for K=2,
like 1001, 1000) in each shift and increment the count, repeat this
for all numbers in range.
///////////////////////////////////////////////////////
int KsparseNumbers(int a, int b, int s) {
int nbits = sizeof(int)*8;
int slen = 0;
int lslen = pow(2, s);
int scount = 0;
int i = 0;
for (; i < s; ++i) {
slen += pow(2, i);
}
printf("\n slen = %d\n", slen);
for(; a <= b; ++a) {
int num = a;
for(i = 0 ; i < nbits-2; ++i) {
if ( (num & slen) == 0 && (num & lslen) ) {
scount++;
printf("\n Scount = %d\n", scount);
break;
}
num >>=1;
}
}
return scount;
}
int main() {
printf("\n No of 2-sparse numbers between 5 and 15 = %d\n", KsparseNumbers(5, 15, 2));
}
Project Euler problem 14:
The following iterative sequence is
defined for the set of positive
integers:
n → n/2 (n is even) n → 3n + 1 (n is
odd)
Using the rule above and starting with
13, we generate the following
sequence: 13 → 40 → 20 → 10 → 5 → 16 →
8 → 4 → 2 → 1
It can be seen that this sequence
(starting at 13 and finishing at 1)
contains 10 terms. Although it has not
been proved yet (Collatz Problem), it
is thought that all starting numbers
finish at 1.
Which starting number, under one
million, produces the longest chain?
My first instinct is to create a function to calculate the chains, and run it with every number between 1 and 1 million. Obviously, that takes a long time. Way longer than solving this should take, according to Project Euler's "About" page. I've found several problems on Project Euler that involve large groups of numbers that a program running for hours didn't finish. Clearly, I'm doing something wrong.
How can I handle large groups of numbers quickly?
What am I missing here?
Have a read about memoization. The key insight is that if you've got a sequence starting A that has length 1001, and then you get a sequence B that produces an A, you don't to repeat all that work again.
This is the code in Mathematica, using memoization and recursion. Just four lines :)
f[x_] := f[x] = If[x == 1, 1, 1 + f[If[EvenQ[x], x/2, (3 x + 1)]]];
Block[{$RecursionLimit = 1000, a = 0, j},
Do[If[a < f[i], a = f[i]; j = i], {i, Reverse#Range#10^6}];
Print#a; Print[j];
]
Output .... chain length´525´ and the number is ... ohhhh ... font too small ! :)
BTW, here you can see a plot of the frequency for each chain length
Starting with 1,000,000, generate the chain. Keep track of each number that was generated in the chain, as you know for sure that their chain is smaller than the chain for the starting number. Once you reach 1, store the starting number along with its chain length. Take the next biggest number that has not being generated before, and repeat the process.
This will give you the list of numbers and chain length. Take the greatest chain length, and that's your answer.
I'll make some code to clarify.
public static long nextInChain(long n) {
if (n==1) return 1;
if (n%2==0) {
return n/2;
} else {
return (3 * n) + 1;
}
}
public static void main(String[] args) {
long iniTime=System.currentTimeMillis();
HashSet<Long> numbers=new HashSet<Long>();
HashMap<Long,Long> lenghts=new HashMap<Long, Long>();
long currentTry=1000000l;
int i=0;
do {
doTry(currentTry,numbers, lenghts);
currentTry=findNext(currentTry,numbers);
i++;
} while (currentTry!=0);
Set<Long> longs = lenghts.keySet();
long max=0;
long key=0;
for (Long aLong : longs) {
if (max < lenghts.get(aLong)) {
key = aLong;
max = lenghts.get(aLong);
}
}
System.out.println("number = " + key);
System.out.println("chain lenght = " + max);
System.out.println("Elapsed = " + ((System.currentTimeMillis()-iniTime)/1000));
}
private static long findNext(long currentTry, HashSet<Long> numbers) {
for(currentTry=currentTry-1;currentTry>=0;currentTry--) {
if (!numbers.contains(currentTry)) return currentTry;
}
return 0;
}
private static void doTry(Long tryNumber,HashSet<Long> numbers, HashMap<Long, Long> lenghts) {
long i=1;
long n=tryNumber;
do {
numbers.add(n);
n=nextInChain(n);
i++;
} while (n!=1);
lenghts.put(tryNumber,i);
}
Suppose you have a function CalcDistance(i) that calculates the "distance" to 1. For instance, CalcDistance(1) == 0 and CalcDistance(13) == 9. Here is a naive recursive implementation of this function (in C#):
public static int CalcDistance(long i)
{
if (i == 1)
return 0;
return (i % 2 == 0) ? CalcDistance(i / 2) + 1 : CalcDistance(3 * i + 1) + 1;
}
The problem is that this function has to calculate the distance of many numbers over and over again. You can make it a little bit smarter (and a lot faster) by giving it a memory. For instance, lets create a static array that can store the distance for the first million numbers:
static int[] list = new int[1000000];
We prefill each value in the list with -1 to indicate that the value for that position is not yet calculated. After this, we can optimize the CalcDistance() function:
public static int CalcDistance(long i)
{
if (i == 1)
return 0;
if (i >= 1000000)
return (i % 2 == 0) ? CalcDistance(i / 2) + 1 : CalcDistance(3 * i + 1) + 1;
if (list[i] == -1)
list[i] = (i % 2 == 0) ? CalcDistance(i / 2) + 1: CalcDistance(3 * i + 1) + 1;
return list[i];
}
If i >= 1000000, then we cannot use our list, so we must always calculate it. If i < 1000000, then we check if the value is in the list. If not, we calculate it first and store it in the list. Otherwise we just return the value from the list. With this code, it took about ~120ms to process all million numbers.
This is a very simple example of memoization. I use a simple list to store intermediate values in this example. You can use more advanced data structures like hashtables, vectors or graphs when appropriate.
Minimize how many levels deep your loops are, and use an efficient data structure such as IList or IDictionary, that can auto-resize itself when it needs to expand. If you use plain arrays they need to be copied to larger arrays as they expand - not nearly as efficient.
This variant doesn't use an HashMap but tries only to not repeat the first 1000000 numbers. I don't use an hashmap because the biggest number found is around 56 billions, and an hash map could crash.
I have already done some premature optimization. Instead of / I use >>, instead of % I use &. Instead of * I use some +.
void Main()
{
var elements = new bool[1000000];
int longestStart = -1;
int longestRun = -1;
long biggest = 0;
for (int i = elements.Length - 1; i >= 1; i--) {
if (elements[i]) {
continue;
}
elements[i] = true;
int currentStart = i;
int currentRun = 1;
long current = i;
while (current != 1) {
if (current > biggest) {
biggest = current;
}
if ((current & 1) == 0) {
current = current >> 1;
} else {
current = current + current + current + 1;
}
currentRun++;
if (current < elements.Length) {
elements[current] = true;
}
}
if (currentRun > longestRun) {
longestStart = i;
longestRun = currentRun;
}
}
Console.WriteLine("Longest Start: {0}, Run {1}", longestStart, longestRun);
Console.WriteLine("Biggest number: {0}", biggest);
}