I need to implement an algorithm in C++ that, when given three arrays of unequal sizes, produces triplets a,b,c (one element contributed by each array) such that max(a,b,c) - min(a,b,c) is minimized. The algorithm should produce a list of these triplets, in order of size of max(a,b,c)-min(a,b,c). The arrays are sorted.
I've implemented the following algorithm (note that I now use arrays of type double), however it runs excruciatingly slow (even when compiled using GCC with -03 optimization, and other combinations of optimizations). The dataset (and, therefore, each array) has potentially tens of millions of elements. Is there a faster/more efficient method? A significant speed increase is necessary to accomplish the required task in a reasonable time frame.
void findClosest(vector<double> vec1, vector<double> vec2, vector<double> vec3){
//calculate size of each array
int len1 = vec1.size();
int len2 = vec2.size();
int len3 = vec3.size();
int i = 0; int j = 0; int k = 0; int res_i, res_j, res_k;
int diff = INT_MAX;
int iter = 0; int iter_bound = min(min(len1,len2),len3);
while(iter < iter_bound)
while(i < len1 && j < len2 && k < len3){
int minimum = min(min(vec1[i], vec2[j]), vec3[k]);
int maximum = max(max(vec1[i], vec2[j]), vec3[k]);
//if new difference less than previous difference, update difference, store
//resultants
if(fabs(maximum - minimum) < diff){ diff = maximum-minimum; res_i = i; res_j = j; res_k = k;}
//increment minimum value
if(vec1[i] == minimum) ++i;
else if(vec2[j] == minimum) ++j;
else ++k;
}
//"remove" triplet
vec1.erase(vec1.begin() + res_i);
vec2.erase(vec2.begin() + res_j);
vec3.erase(vec3.begin() + res_k);
--len1; --len2; --len3;
++iter_bound;
}
OK, you're going to need to be clever in a few ways to make this run well.
The first thing that you need is a priority queue, which is usually implemented with a heap. With that, the algorithm in pseudocode is:
Make a priority queue for possible triples in order of max - min, then how close median is to their average.
Make a pass through all 3 arrays, putting reasonable triples for every element into the priority queue
While the priority queue is not empty:
Pull a triple out
If all three of the triple are not used:
Add triple to output
Mark the triple used
else:
If you can construct reasonable triplets for unused elements:
Add them to the queue
Now for this operation to succeed, you need to efficiently find elements that are currently unused. Doing that at first is easy, just keep an array of bools where you mark off the indexes of the used values. But once a lot have been taken off, your search gets long.
The trick for that is to have a vector of bools for individual elements, a second for whether both in a pair have been used, a third for where all 4 in a quadruple have been used and so on. When you use an element just mark the individual bool, then go up the hierarchy, marking off the next level if the one you're paired with is marked off, else stopping. This additional data structure of size 2n will require an average of marking 2 bools per element used, but allows you to find the next unused index in either direction in at most O(log(n)) steps.
The resulting algorithm will be O(n log(n)).
I am looking for a way to optimize an algorithm that I am working on. It's most repetitive and thus compute-intensive part is comparison of two sorted arrays of any size, containing unique unsigned integer (uint32_t) values in order to obtain the size of symmetric difference of them (number of elements that exist only in one of the vectors). The target machine on which the algorithm will be deployed uses Intel processors supporting AVX2, therefore I am looking for a way to perform it in-place using SIMD.
Is there a way to exploit the AVX2 instructions to obtain the size of symmetric difference of two sorted arrays of unsigned integers?
Since both arrays are sorted it should be fairly easy to implement this algorithm using SIMD (AVX2). You would just need to iterate through the two arrays concurrently, and then when you get a mismatch when comparing two vectors of 8 ints you would need to resolve the mismatch, i.e. count the differences, and get the two array indices back in phase, and continue until you get to the end of one of the arrays. Then just add the no of remaining elements in the other array, if any, to get the final count.
Unless your arrays are tiny (like <=16 elements), you need to perform merge of the two sorted arrays with additional code for dumping non-equal elements.
If the size of symmetric difference is expected to be very small, then use the method described by PaulR.
If the size is expected to be high (like 10% of total number of elements), then you will have real trouble with vectorizing it. It is much easier to optimize scalar solution.
After writing several versions of code, the fastest one for me is:
int Merge3(const int *aArr, int aCnt, const int *bArr, int bCnt, int *dst) {
int i = 0, j = 0, k = 0;
while (i < aCnt - 32 && j < bCnt - 32) {
for (int t = 0; t < 32; t++) {
int aX = aArr[i], bX = bArr[j];
dst[k] = (aX < bX ? aX : bX);
k += (aX != bX);
i += (aX <= bX);
j += (aX >= bX);
}
}
while (i < aCnt && j < bCnt) {
... //use simple code to merge tails
The main optimizations here are:
Perform merging iterations in blocks (32 iterations per block). This allows to simplify stop criterion from (i < aCnt && j < bCnt) to t < 32. This can be done for most of the elements, and the tails can be processed with slow code.
Perform iterations in branchless fashion. Note that ternary operator is compiled into cmov instruction, and comparisons are compiled into setXX instructions, so there are no branches in the loop body. The output data is stored with the well-known trick: write all elements, but increase pointer only for the valid ones.
What else I have tried:
(no vectorization) perform (4 + 4) bitonic merge, check consecutive elements for duplicates, move pointers so that 4 min elements (in total) are skipped:
gets 4.95ns vs 4.65ns --- slightly worse.
(fully vectorized) compare 4 x 4 elements pairwise, extract comparison results into 16-bit mask, pass it through perfect hash function, use _mm256_permutevar8x32_epi32 with 128-entry LUT to get sorted 8 elements, check consecutive elements for duplicates, use _mm_movemask_ps + 16-entry LUT + _mm_shuffle_epi8 to store only unique elements among minimal 4 elements: gets 4.00ns vs 4.65ns --- slightly better.
Here is the file with solutions and file with perfect hash + LUT generator.
P.S. Note that similar problem for intersection of sets is solved here. The solution is somewhat similar to what I outlined as point 2 above.
This question already has answers here:
Gray code increment function
(4 answers)
Closed 8 years ago.
Let's say i have n integers in an array a, and i want to iterate through all possible subsets of these integers, find the sum, and then do something with it.
What i immedieatelly did, was to create a bit field b, which indicated which numbers were included in the subset, and iterate through its possible values using ++b. Then, to compute the sum in each step, i had to iterate through all bits like this:
int sum = 0;
for (int i = 0; i < n; i++)
if (b&1<<i)
sum += a[i];
Then i realized that if i iterated through the possible values of b in a Gray code order, so that each time only a single bit is flipped, i wouldn't have to reconstruct the sum completely, but only needed to add or subtract the single value that is being added or removed from the subset. It should work like this:
int sum = 0;
int whichBitToFlip = 0;
bool isBitSet = false;
for (int k = 0; whichBitToFlip < n; k++) {
sum += (isBitSet ? -1 : 1)*a[whichBitToFlip];
// do something with sum here
whichBitToFlip = ???;
bool isBitSet = ???;
}
But i can't figure out how to directly and efficiently compute whichBitToFlip. The desired values are basically sequence A007814. I know that i can compute the Gray code using the formula (k>>1)^k and xor it with the previous one, but then i need to find the position of the changed bit, which might not be much faster.
So is there any better way to determine these values (index of flipped bit), preferably without a cycle, faster than recomputing the whole sum (of at most 64 values) every time?
To convert a bitmask to a bit index, you can use the ffs function (if you have one), which corresponds to a machine opcode on some machines.
Otherwise, the bit changed in the gray code corresponds to the ruler function:
0, 1, 0, 2, 0, 1, 0, 3, 0, 1...
for which there is a simple recursion. You can simulate the recursion with a stack (it will have maximum depth O(log N), so it's not much space), but probably ffs is a lot faster.
(By the way, even if you were to count bits one at a time from right-to-left, the increment function would be O(1) on average because the total number of trailing 0s in the integers from 1 to 2k is 2k-1.)
So i came up with this:
int sum = 0;
unsigned long grayPos = 0;
int graySign = 1;
for (uint64 k = 2; grayPos < n; k++) {
sum += graySign*a[grayPos];
// Do something with sum
#ifdef _M_X64
grayPos = n;
_BitScanForward64(&grayPos, k);
#else
for (grayPos = 0; !(k&1ull<<grayPos); grayPos++);
#endif
graySign = 2-(k>>grayPos&0x3);
}
It works really well, brought down the execution time (in comparison to always recomputing the whole sum) from 254 to only 7 seconds for n = 32. I also found that counting trailing zeroes with the for cycle is only slightly (~15%) slower than using _BitScanForward64 for the reasons mentioned by rici. So thanks.
The full problem statement is here. Suppose we have a double ended queue of known values. Each turn, we can take a value out of one or the other end and the values still in the queue increase as value*turns. The goal is to find maximum possible total value.
My first approach was to use straightforward top-down DP with memoization. Let i,j denote starting, ending indexes of "subarray" of array of values A[].
A[i]*age if i == j
f(i,j,age) =
max(f(i+1,j,age+1) + A[i]*age , f(i,j-1,age+1) + A[j]*age)
This works, however, proves to be too slow, as there are superfluous stack calls. Iterative bottom-up should be faster.
Let m[i][j] be the maximum reachable value of the "subarray" of A[] with begin/end indexes i,j. Because i <= j, we care only about the lower triangular part.
This matrix can be built iteratively using the fact that m[i][j] = max(m[i-1][j] + A[i]*age, m[i][j-1] + A[j]*age), where age is maximum on the diagonal (size of A[] and linearly decreases as A.size()-(i-j).
My attempt at implementation meets with bus error.
Is the described algorithm correct? What is the cause for the bus error?
Here is the only part of the code where the bus error might occur:
for(T j = 0; j < num_of_treats; j++) {
max_profit[j][j] = treats[j]*num_of_treats;
for(T i = j+1; i < num_of_treats; i++)
max_profit[i][j] = max( max_profit[i-1][j] + treats[i]*(num_of_treats-i+j),
max_profit[i][j-1] + treats[j]*(num_of_treats-i+j));
}
for(T j = 0; j < num_of_treats; j++) {
Inside this loop, j is clearly a valid index into the array max_profit. But you're not using just j.
The bus error is caused by trying to access array via negative index when j=0 and i=1 as I should have noticed during the debugging. The algorithm is wrong as well. First, the relationship used to construct the max_profit[][] array should is
max_profit[i][j] = max( max_profit[i+1][j] + treats[i]*(num_of_treats-i+j),
max_profit[i][j-1] + treats[j]*(num_of_treats-i+j));
Second, the array must by filled diagonally, so that max_profit[i+1][j] and max_profit[i][j-1] is already computed with exception of the main diagonal.
Third, the data structure chosen is extremely inefficient. I am using only half of the space allocated for max_profit[][]. Plus, at each iteration, I only need the last computed diagonal. An array of size num_of_treats should suffice.
Here is a working code using this improved algorithm. I really like it. I even used bit operators for the first time.
I use C++ and CUDA/C and want to write code for a specific problem and I ran into a quite tricky reduction problem.
My experience in parallel programming isn't negligible but quite limited and I cannot totally forsee the specificity of this problem.
I doubt there is a convenient or even "easy" way to handle the problems I am facing but perhaps I am wrong.
If there are any resources (i.e. articles, books, web-links, ...) or key-words covering this or similar problems, please let me know.
I tried to generalize the whole case as good as possible and keep it abstract instead of posting too much code.
The Layout ...
I have a system of N inital elements and N result elements. (I'll use N=8 for example but N can be any integral value greater than three.)
static size_t const N = 8;
double init_values[N], result[N];
I need to calculate almost every (not all i'm afraid) unique permutation of the init-values without self-interference.
This means calculation f(init_values[0],init_values[1]), f(init_values[0],init_values[2]), ..., f(init_values[0],init_values[N-1]), f(init_values[1],init_values[2]), ..., f(init_values[1],init_values[N-1]), ... and so on.
This is in fact a virtual triangular matrix which has the shape seen in the following illustration.
P 0 1 2 3 4 5 6 7
|---------------------------------------
0| x
|
1| 0 x
|
2| 1 2 x
|
3| 3 4 5 x
|
4| 6 7 8 9 x
|
5| 10 11 12 13 14 x
|
6| 15 16 17 18 19 20 x
|
7| 21 22 23 24 25 26 27 x
Each element is a function of the respective column and row elements in init_values.
P[i] (= P[row(i)][col(i]) = f(init_values[col(i)], init_values[row(i)])
i.e.
P[11] (= P[5][1]) = f(init_values[1], init_values[5])
There are (N*N-N)/2 = 28 possible, unique combinations (Note: P[1][5]==P[5][1], so we only have a lower (or upper) triangular matrix) using the example N = 8.
The basic problem
The result array is computed from P as a sum of the row elements minus the sum of the respective column elements.
For example the result at position 3 will be calculated as a sum of row 3 minus the sum of column three.
result[3] = (P[3]+P[4]+P[5]) - (P[9]+P[13]+P[18]+P[24])
result[3] = sum_elements_row(3) - sum_elements_column(3)
I tried to illustrate it in a picture with N = 4.
As a consequence the following is true:
N-1 operations (potential concurrent writes) will be performed on each result[i]
result[i] will have N-(i+1) writes from subtractions and i additions
Outgoing from each P[i][j] there will be a subtraction to r[j] and a addition to r[i]
This is where the main problems come into place:
Using one thread to compute each P and updating the result directly will result in multiple kernels trying to write to the same result location (N-1 threads each).
Storing the whole matrix P for a subsequent reduction step on the other hand is very expensive in terms of memory consumption and therefore impossible for very large systems.
The idea of having a unqiue, shared result vector for each thread-block is impossible, too.
(N of 50k makes 2.5 billion P elements and therefore [assuming a maximum number of 1024 threads per block] a minimal number of 2.4 million blocks consuming over 900GiB of memory if each block has its own result array with 50k double elements.)
I think I could handle reduction for a more static behaviour but this problem is rather dynamic in terms of potential concurrent memory write-access.
(Or is it possible to handle it by some "basic" type of reduction?)
Adding some complications ...
Unfortunatelly, depending on (arbitrary user) input, which is independant of the initial values, some elements of P need to be skipped.
Let's assume we need to skip permutations P[6], P[14] and P[18]. Therefore we have 24 combinations left, which need to be calculated.
How to tell the kernel which values need to be skipped?
I came up with three approaches, each having notable downsides if N is very large (like several ten thousands of elements).
1. Store all combinations ...
... with their respective row and column index struct combo { size_t row,col; };, that need to be calculated in a vector<combo> and operate on this vector. (used by the current implementation)
std::vector<combo> elements;
// somehow fill
size_t const M = elements.size();
for (size_t i=0; i<M; ++i)
{
// do the necessary computations using elements[i].row and elements[i].col
}
This solution consumes is consuming lots of memory since only "several" (may even be ten thousands of elements but that's not much in contrast to several billion in total) but it avoids
indexation computations
finding of removed elements
for each element of P which is the downside of the second approach.
2. Operate on all elements of P and find removed elements
If I want to operate on each element of P and avoid nested loops (which i couldn't reproduce very well in cuda) I need to do something like this:
size_t M = (N*N-N)/2;
for (size_t i=0; i<M; ++i)
{
// calculate row indices from `i`
double tmp = sqrt(8.0*double(i+1))/2.0 + 0.5;
double row_d = floor(tmp);
size_t current_row = size_t(row_d);
size_t current_col = size_t(floor(row_d*(ict-row_d)-0.5));
// check whether the current combo of row and col is not to be removed
if (!removes[current_row].exists(current_col))
{
// do the necessary computations using current_row and current_col
}
}
The vector removes is very small in contrast to the elements vector in the first example but the additional computations to obtain current_row, current_col and the if-branch are very inefficient.
(Remember we're still talking about billions of evaluations.)
3. Operate on all elements of P and remove elements afterwards
Another idea I had was to calculate all valid and invalid combinations independently.
But unfortunately, due to summation errors the following statement is true:
calc_non_skipped() != calc_all() - calc_skipped()
Is there a convenient, known, high performance way to get the desired results from the initial values?
I know that this question is rather complicated and perhaps limited in relevance. Nevertheless, I hope some illuminative answers will help me to solve my problems.
The current implementation
Currently this is implemented as CPU Code with OpenMP.
I first set up a vector of the above mentioned combos storing every P that needs to be computed and pass it to a parallel for loop.
Each thread is provided with a private result vector and a critical section at the end of the parallel region is used for a proper summation.
First, I was puzzled for a moment why (N**2 - N)/2 yielded 27 for N=7 ... but for indices 0-7, N=8, and there are 28 elements in P. Shouldn't try to answer questions like this so late in the day. :-)
But on to a potential solution: Do you need to keep the array P for any other purpose? If not, I think you can get the result you want with just two intermediate arrays, each of length N: one for the sum of the rows and one for the sum of the columns.
Here's a quick-and-dirty example of what I think you're trying to do (subroutine direct_approach()) and how to achieve the same result using the intermediate arrays (subroutine refined_approach()):
#include <cstdlib>
#include <cstdio>
const int N = 7;
const float input_values[N] = { 3.0F, 5.0F, 7.0F, 11.0F, 13.0F, 17.0F, 23.0F };
float P[N][N]; // Yes, I'm wasting half the array. This way I don't have to fuss with mapping the indices.
float result1[N] = { 0.0F, 0.0F, 0.0F, 0.0F, 0.0F, 0.0F, 0.0F };
float result2[N] = { 0.0F, 0.0F, 0.0F, 0.0F, 0.0F, 0.0F, 0.0F };
float f(float arg1, float arg2)
{
// Arbitrary computation
return (arg1 * arg2);
}
float compute_result(int index)
{
float row_sum = 0.0F;
float col_sum = 0.0F;
int row;
int col;
// Compute the row sum
for (col = (index + 1); col < N; col++)
{
row_sum += P[index][col];
}
// Compute the column sum
for (row = 0; row < index; row++)
{
col_sum += P[row][index];
}
return (row_sum - col_sum);
}
void direct_approach()
{
int row;
int col;
for (row = 0; row < N; row++)
{
for (col = (row + 1); col < N; col++)
{
P[row][col] = f(input_values[row], input_values[col]);
}
}
int index;
for (index = 0; index < N; index++)
{
result1[index] = compute_result(index);
}
}
void refined_approach()
{
float row_sums[N];
float col_sums[N];
int index;
// Initialize intermediate arrays
for (index = 0; index < N; index++)
{
row_sums[index] = 0.0F;
col_sums[index] = 0.0F;
}
// Compute the row and column sums
// This can be parallelized by computing row and column sums
// independently, instead of in nested loops.
int row;
int col;
for (row = 0; row < N; row++)
{
for (col = (row + 1); col < N; col++)
{
float computed = f(input_values[row], input_values[col]);
row_sums[row] += computed;
col_sums[col] += computed;
}
}
// Compute the result
for (index = 0; index < N; index++)
{
result2[index] = row_sums[index] - col_sums[index];
}
}
void print_result(int n, float * result)
{
int index;
for (index = 0; index < n; index++)
{
printf(" [%d]=%f\n", index, result[index]);
}
}
int main(int argc, char * * argv)
{
printf("Data reduction test\n");
direct_approach();
printf("Result 1:\n");
print_result(N, result1);
refined_approach();
printf("Result 2:\n");
print_result(N, result2);
return (0);
}
Parallelizing the computation is not so easy, since each intermediate value is a function of most of the inputs. You can compute the sums individually, but that would mean performing f(...) multiple times. The best suggestion I can think of for very large values of N is to use more intermediate arrays, computing subsets of the results, then summing the partial arrays to yield the final sums. I'd have to think about that one when I'm not so tired.
To cope with the skip issue: If it's a simple matter of "don't use input values x, y, and z", you can store x, y, and z in a do_not_use array and check for those values when looping to compute the sums. If the values to be skipped are some function of row and column, you can store those as pairs and check for the pairs.
Hope this gives you ideas for your solution!
Update, now that I'm awake: Dealing with "skip" depends a lot on what data needs to be skipped. Another possibility for the first case - "don't use input values x, y, and z" - a much faster solution for large data sets would be to add a level of indirection: create yet another array, this one of integer indices, and store only the indices of the good inputs. F'r instance, if invalid data is in inputs 2 and 5, the valid array would be:
int valid_indices[] = { 0, 1, 3, 4, 6 };
Interate over the array valid_indices, and use those indices to retrieve the data from your input array to compute the result. On the other paw, if the values to skip depend on both indices of the P array, I don't see how you can avoid some kind of lookup.
Back to parallelizing - No matter what, you'll be dealing with (N**2 - N)/2 computations
of f(). One possibility is to just accept that there will be contention for the sum
arrays, which would not be a big issue if computing f() takes substantially longer than
the two additions. When you get to very large numbers of parallel paths, contention will
again be an issue, but there should be a "sweet spot" balancing the number of parallel
paths against the time required to compute f().
If contention is still an issue, you can partition the problem several ways. One way is
to compute a row or column at a time: for a row at a time, each column sum can be
computed independently and a running total can be kept for each row sum.
Another approach would be to divide the data space and, thus, the computation into
subsets, where each subset has its own row and column sum arrays. After each block
is computed, the independent arrays can then be summed to produce the values you need
to compute the result.
This probably will be one of those naive and useless answers, but it also might help. Feel free to tell me that I'm utterly and completely wrong and I have misunderstood the whole affair.
So... here we go!
The Basic Problem
It seems to me that you can define you result function a little differently and it will lift at least some contention off your intermediate values. Let's suppose that your P matrix is lower-triangular. If you (virtually) fill the upper triangle with the negative of the lower values (and the main diagonal with all zeros,) then you can redefine each element of your result as the sum of a single row: (shown here for N=4, and where -i means the negative of the value in the cell marked as i)
P 0 1 2 3
|--------------------
0| x -0 -1 -3
|
1| 0 x -2 -4
|
2| 1 2 x -5
|
3| 3 4 5 x
If you launch independent threads (executing the same kernel) to calculate the sum of each row of this matrix, each thread will write a single result element. It seems that your problem size is large enough to saturate your hardware threads and keep them busy.
The caveat, of course, is that you'll be calculating each f(x, y) twice. I don't know how expensive that is, or how much the memory contention was costing you before, so I cannot judge whether this is a worthwhile trade-off to do or not. But unless f was really really expensive, I think it might be.
Skipping Values
You mention that you might have tens of thousands elements of the P matrix that you need to ignore in your calculations (effectively skip them.)
To work with the scheme I've proposed above, I believe you should store the skipped elements as (row, col) pairs, and you have to add the transposed of each coordinate pair too (so you'll have twice the number of skipped values.) So your example skip list of P[6], P[14] and P[18] becomes P(4,0), P(5,4), P(6,3) which then becomes P(4,0), P(5,4), P(6,3), P(0,4), P(4,5), P(3,6).
Then you sort this list, first based on row and then column. This makes our list to be P(0,4), P(3,6), P(4,0), P(4,5), P(5,4), P(6,3).
If each row of your virtual P matrix is processed by one thread (or a single instance of your kernel or whatever,) you can pass it the values it needs to skip. Personally, I would store all these in a big 1D array and just pass in the first and last index that each thread would need to look at (I would also not store the row indices in the final array that I passed in, since it can be implicitly inferred, but I think that's obvious.) In the example above, for N = 8, the begin and end pairs passed to each thread will be: (note that the end is one past the final value needed to be processed, just like STL, so an empty list is denoted by begin == end)
Thread 0: 0..1
Thread 1: 1..1 (or 0..0 or whatever)
Thread 2: 1..1
Thread 3: 1..2
Thread 4: 2..4
Thread 5: 4..5
Thread 6: 5..6
Thread 7: 6..6
Now, each thread goes on to calculate and sum all the intermediate values in a row. While it is stepping through the indices of columns, it is also stepping through this list of skipped values and skipping any column number that comes up in the list. This is obviously an efficient and simple operation (since the list is sorted by column too. It's like merging.)
Pseudo-Implementation
I don't know CUDA, but I have some experience working with OpenCL, and I imagine the interfaces are similar (since the hardware they are targeting are the same.) Here's an implementation of the kernel that does the processing for a row (i.e. calculates one entry of result) in pseudo-C++:
double calc_one_result (
unsigned my_id, unsigned N, double const init_values [],
unsigned skip_indices [], unsigned skip_begin, unsigned skip_end
)
{
double res = 0;
for (unsigned col = 0; col < my_id; ++col)
// "f" seems to take init_values[column] as its first arg
res += f (init_values[col], init_values[my_id]);
for (unsigned row = my_id + 1; row < N; ++row)
res -= f (init_values[my_id], init_values[row]);
// At this point, "res" is holding "result[my_id]",
// including the values that should have been skipped
unsigned i = skip_begin;
// The second condition is to check whether we have reached the
// middle of the virtual matrix or not
for (; i < skip_end && skip_indices[i] < my_id; ++i)
{
unsigned col = skip_indices[i];
res -= f (init_values[col], init_values[my_id]);
}
for (; i < skip_end; ++i)
{
unsigned row = skip_indices[i];
res += f (init_values[my_id], init_values[row]);
}
return res;
}
Note the following:
The semantics of init_values and function f are as described by the question.
This function calculates one entry in the result array; specifically, it calculates result[my_id], so you should launch N instances of this.
The only shared variable it writes to is result[my_id]. Well, the above function doesn't write to anything, but if you translate it to CUDA, I imagine you'd have to write to that at the end. However, no one else writes to that particular element of result, so this write will not cause any contention of data race.
The two input arrays, init_values and skipped_indices are shared among all the running instances of this function.
All accesses to data are linear and sequential, except for the skipped values, which I believe is unavoidable.
skipped_indices contain a list of indices that should be skipped in each row. It's contents and structure are as described above, with one small optimization. Since there was no need, I have removed the row numbers and left only the columns. The row number will be passed into the function as my_id anyways and the slice of the skipped_indices array that should be used by each invocation is determined using skip_begin and skip_end.
For the example above, the array that is passed into all invocations of calc_one_result will look like this:[4, 6, 0, 5, 4, 3].
As you can see, apart from the loops, the only conditional branch in this code is skip_indices[i] < my_id in the third for-loop. Although I believe this is innocuous and totally predictable, even this branch can be easily avoided in the code. We just need to pass in another parameter called skip_middle that tells us where the skipped items cross the main diagonal (i.e. for row #my_id, the index at skipped_indices[skip_middle] is the first that is larger than my_id.)
In Conclusion
I'm by no means an expert in CUDA and HPC. But if I have understood your problem correctly, I think this method might eliminate any and all contentions for memory. Also, I don't think this will cause any (more) numerical stability issues.
The cost of implementing this is:
Calling f twice as many times in total (and keeping track of when it is called for row < col so you can multiply the result by -1.)
Storing twice as many items in the list of skipped values. Since the size of this list is in the thousands (and not billions!) it shouldn't be much of a problem.
Sorting the list of skipped values; which again due to its size, should be no problem.
(UPDATE: Added the Pseudo-Implementation section.)