Threads indexing out of bounds in CUDA kernel - indexing

Question

I am running a CUDA kernel which seems to be indexing out of bounds and I can not figure out why. I get error 8 write-of-size in cuda-memcheck.
I have tried to change the number of blocks and the number of threads in each block as well as only running a fraction of all iterations needed. Here is some usefull information as well as a replicable example which gives the error:
blockSize: 128
numBlocks: 512
Nvidia GTX 970
#include <iostream>
#include <cuda_runtime_api.h>
#include <cuda.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <vector>
#include <iterator>
#include <cuda_profiler_api.h>
#include <algorithm>
#include <cmath>
#include <numeric>
#include <stdio.h>
#include <fstream>
__host__
int NchooseK(const int &N, const int &K)
{
int result = 1;
for (int i = 1; i <= K; i++)
{
result *= N - (K - i);
result /= i;
}
return result;
}
__host__
inline int get_flatten_size(const unsigned int N){
int sum = 0;
for(int i=1; i<=N ; i++){
sum +=i*NchooseK(N,i);
}
return sum;
}
__host__
std::vector<int> comb(const int &N, const int &K, const int &length)
//void comb(int N, int K, int length)
{
int k;
std::vector<int> vec(K);
std::vector<int> flatten_vec(0);
std::string bitmask(K, 1); // K leading 1's
bitmask.resize(N, 0); // N-K trailing 0's
for (int j = 0; j < length; j++) {
k = 0;
for (int i = 0; i < N; ++i) // [0..N-1] integers
{
if (bitmask[i]) {
//std::cout << i << " ";
vec[k] = i;
k++;
}
//std::cout << std::endl;
}
std::prev_permutation(bitmask.begin(), bitmask.end());
flatten_vec.insert(flatten_vec.end(), vec.begin(),vec.end());
}
return flatten_vec;
}
__host__
void get_matrix_indices(const unsigned int N, int *sub_col, int *sub_size, int *cumulative_size)
{
int size, itterator = 0;
cumulative_size[0] = 0;
std::vector<int> size_i_columns;
std::vector<int> all_columns(0);
for(int i=1; i<=N; i++){
size = NchooseK(N,i);
size_i_columns = comb(N,i,size);
for(int j=0; j<size; j++){
sub_size[itterator]=i;
cumulative_size[itterator+1]=cumulative_size[itterator]+i;
itterator++;
}
all_columns.insert(all_columns.end(),size_i_columns.begin(),size_i_columns.end());
}
//sub_col = &all_columns[0];
for(int i = 0; i < all_columns.size(); i++) sub_col[i] = all_columns[i];
}
__global__
void comb_ols(const unsigned int M, const unsigned int N, int* sub_col, int *sub_size, int* cumulative_size, const unsigned int numberOfCalculations, const unsigned int max_size){
int size;
int start_index;
int index = blockIdx.x*blockDim.x+threadIdx.x;
int stride = blockDim.x*gridDim.x;
double *sub_matrix = new double[M*(1+max_size)];
for(int i = index; i < numberOfCalculations; i+=stride){
size = sub_size[i];
start_index = cumulative_size[i];
for(int j = 0; j < size; j++){
for(int k = 0; k<M; k++){
sub_matrix[k] = 1;
}
}
}
delete [] sub_matrix;
}
And then we the main function:
int main()
{
int N = 17;
int M = 263;
const unsigned int regressors = N-1;
const unsigned int numberOfCalculations = (int) (exp2((double) regressors) - 1);
const unsigned int size_sub_col = get_flatten_size(regressors);
int blockSize =128;
int numBlocks = (numberOfCalculations + blockSize-1)/blockSize;
std::cout << "\nblockSize :" << blockSize;
std::cout << "\nnumBlocks :" << numBlocks;
std::cout << "\nblockSize*numBlocks :" << blockSize*numBlocks;
std::cout << "\nregressors :" << regressors;
std::cout << "\nNumberOfCalculations :" << numberOfCalculations;
std::cout << "\nsize_sub_col :" << size_sub_col << '\n' ;
int *sub_size, *cumulative_size, *sub_columns;
cudaMallocManaged(&sub_size, numberOfCalculations*sizeof(int));
cudaMallocManaged(&cumulative_size, (numberOfCalculations+1)*sizeof(int));
cudaMallocManaged(&sub_columns, size_sub_col*sizeof(int));
get_matrix_indices(regressors,sub_columns, sub_size, cumulative_size);
const unsigned int max_size = N*M;
cudaProfilerStart();
comb_ols<<<numBlocks, blockSize>>>(M,N,sub_columns, sub_size, cumulative_size, numberOfCalculations, max_size);
cudaProfilerStop();
cudaDeviceSynchronize();
cudaFree(sub_size);
cudaFree(cumulative_size);
cudaFree(sub_columns);
return 0;
}
I fail to see why the threads would try to access illegal memory space. The way I understood is that the matrix sub_matrix will be initilized on each thread once and then the parallel for loop happens. Thus should each thread have the necessary memory space. Am I allocating too much memory on the GPU? How is "new sub_matrix" handled here?

Answer 1

If I read your code correctly, each thread is attempting to allocate M * (1 + M*N) doubles, which is 263 * ( 1 + 263*17) = ‭1,176,136‬ doubles, or 8.97Mb of heap memory per thread. You launch 128 * 512 threads. That would mean you require 588Gb of heap space for the kernel to run successfully.
Clearly your GPU lacks that amount of memory and the out of bounds memory access comes from failures in the new call (which you can check for, BTW).
Might I suggest that something in the size calculations for the heap memory you require is wrong. Otherwise you have an extremely unrealistic problem for the GPU and will require some other approach.
Note that even if you manage to redesign things to limit the code to a feasible malloc heap memory size, you will still need, in all likelihood, to resize the malloc heap to a suitable size before running the kernel. The cudaDeviceSetLimit API can be used for this.