barbagroup / gemsfmm Goto Github PK

Dear Rio,
Can you let me know, if you tried to run gemsfmm code with more or less modern version of cuda sdk, because when I tried to compile the existing code after it "trivial" modifications, I have
Kernel execution failed in file 'gpukernel_p3.cu' in line xxx : invalid device function.

This line corresponds to lines 605 and 606 of your original code:
p2p_kernel<<< grid, block >>>(deviceOffset,devicePosTarget,devicePosSource,deviceAccel);
CUT_CHECK_ERROR("Kernel execution failed");

Thank you in advance!

I am trying to build a pytorch interface based on this implementation and might have found a possible bug?

Note that in order to get the code to work 10 years after it has been written, I have changed some things but mostly only removed the SDK dependent checks but most of the remaining code remains the same.

I have encountered the following problem:

In the function m2l_kernel (line 322) a variable "jnk" is declared on line 429.
jnk=(j+n)*(j+n)+j+n;

jnk is used to access an array of size numExpansion2 (10**2) but jnk oftentimes exceeds 100 which results in an illegal memory access.

Is this a bug or am I doing something wrong?

The timing parts (tic, toc) use things only found *nix systems.  
http://stackoverflow.com/questions/2494356/how-to-use-gettimeofday-or-something-
equivalent-with-visual-studio-c-2008

"int neo[numBoxIndexFull];" in cpukernel.cpp wouldn't work, as my compiler at 
least didn't think numBoxIndexFull was a constant expression.

required for constants.h (to get at M_PI)
#define _USE_MATH_DEFINES
#include <math.h>

"complex" type didn't work until I explicitly specified std::complex.  
Conversely, min and max didn't work until I removed the std:: part.

plenty of compiler specific stuff in sse.h

by the time I gave up (at least for a first attack), I was experiencing issues 
in FmmSystem::sort.  Specifically,  
"sortValueBuffer[sortIndexBuffer[sortValue[i]]] = sortValue[i];".  the index 
into sortValueBuffer I think was out of bounds.  No clue if it is a bug or just 
something I broke while trying to get other things working...

If/When I return to this and get it working, I'll submit my modified code for 
others to benefit.  Though it'll probably be greatly reduced in functionality, 
focused on how I plan to use it...  eg I'll remove get_time rather than fixing 
it.

Original issue reported on code.google.com by [email protected] on 28 Aug 2012 at 11:08