Getting strange result in simple multiplication

Accelerated Computing CUDA CUDA Programming and Performance
#1

I have define M = 250 and N = 250 and try to compute int bigsize = (M-1)(N-1)(M-1)(N-1); but getting strange result in console ,like
M-1 = 249 N-1 = 249 bigsize = -450843295
What is the problem here.
when I multiply (M-1)(N-1)*(M-1), I get 15438249, which is correct.
But when I include math.h and calculate int bigsize = pow(N-1, 2)*pow(M-1, 2), I am getting bigsize = 2147483647
but it should be 3844124001.

Thanks in advance

#2

Google “integer overflow”.

#3

For M=250 and N=250, the expression (M-1)(N-1)(M-1)*(N-1) mathematically results in a value between 231 and 232, which, however, is too large to be representable in a 32-bit signed ‘int’. The result of signed integer overflow is undefined, it could be anything.

Here you could get by using an unsigned ‘int’:

#define M (250U)
#define N (250U)
unsigned int bigsize = (M-1)*(N-1)*(M-1)*(N-1);

For a general case, it would be best to perform such computations with ‘size_t’ operands, as you could easily overflow the range of 32-bit unsigned integers as well. Side remark: You would never want to invoke pow() to simply square operands.

#4

But
unsigned int linsizeA2 = (N-1)(M-1)(N-1)(M-1);
checkCuda(cudaMalloc((void**) &A2l, linsizeA2sizeof(double)));

giving error like that
$ cuda-memcheck ./hesttry2 |more
GPUassert: out of memory hestry2.cu 931

free memory = 11926175744 total memory = 12079136768

========= CUDA-MEMCHECK
========= Program hit cudaErrorMemoryAllocation (error 2) due to “out of memory” on CUDA API call to cudaMalloc.
========= Saved host backtrace up to driver entry point at error
========= Host Frame:/usr/lib64/libcuda.so.1 [0x2e40d3]
========= Host Frame:./hesttry2 [0x43549]
========= Host Frame:./hesttry2 [0x52f9]
========= Host Frame:/lib64/libc.so.6 (__libc_start_main + 0xfd) [0x1ed1d]
========= Host Frame:./hesttry2 [0x25b9]

========= ERROR SUMMARY: 1 error

I have tried with size_t instead of unsigned int but getting same result

Now 11926175744/((N-1)(M-1)(N-1)*(M-1)) = 3.10 but double variable would take 8 byte, is this the reason or am I making some another mistake?

#5

You’re out of memory.

2492492492498 = 30752992008

That number is a little over 30GB

Probably your GPU does not have 30GB of memory.

Also note that you still haven’t handled the calculation correctly. If your final result is 30GB (i.e. larger than 2^32) you should use a 64-bit type such as size_t (on a 64-bit platform) instead of this:

linsizeA2*sizeof(double)

But this wouldn’t “resolve” your issue because the request to allocate 30GB of memory is going to fail on any currently available CUDA GPU that I am aware of.

#6

Thank you very much for help.