On Freitag 06 Februar 2009, you wrote: > Hello, > > thanks for good news. The 'if __name__ == "__main__":' > clause with my tests could be probably removed.
It doesn't hurt, I'd prefer to leave it in. > Also note that I removed the dtype argument of rand function. I've tacked the dtype argument back on, but made it raise an error if != float32. Moderately pointless, but at least it matches the docs now. :) > Kernel normalizes raw random bits from md5 to float32, > so other dtypes make no sense. Well, one could have slightly > different kernels (without normalization) for all integer > types ... Agreed that that would be easy. I don't need it right now, so I'll wait for somebody to send a patch. > Do you have any idea how to use all 16 bytes of md5? I mean > how to modify the block/grid sizes to have one thread for 4 > elements? You could do it like this: [123456..][123456...] where thread 1 writes the first 4 bytes to the first spot labelled 1, the second to the second, etc. Just need to check for out of bounds and get your rounding right. > Also as I tried the git version I found that I have a compilation > problem since commit: > > fab598676dbb4c43f53ad1c306a7880061bf6518 > Make memory pool less wasteful: Use quasi-floating point values as bin nr. > > x86_64-pc-linux-gnu-g++ -pthread -fPIC -Isrc/cpp > -I/opt/cuda/include/ > -I/usr/lib64/python2.5/site-packages/numpy/core/include > -I/usr/include/python2.5 -c -Isrc/cpp -I/opt/cuda/include/ > -I/usr/lib64/python2.5/site-packages/numpy/core/include > -I/usr/include/python2.5 -c src/wrapper/tools.cpp -o > build/temp.linux-x86_64-2.5/src/wrapper/tools.o > src/wrapper/tools.cpp: In member function 'typename > Allocator::pointer<unnamed>::memory_pool<Allocator>::allocate(typename > Allocator::size_type) [with Allocator = > <unnamed>::cuda_allocator]': > src/wrapper/tools.cpp:348: instantiated from here > src/wrapper/tools.cpp:220: error: no match for 'operator==' > in '<unnamed>::memory_pool<Allocator>::bin_number [with > Allocator = <unnamed>::cuda_allocator](alloc_sz) == bin' > error: command 'x86_64-pc-linux-gnu-g++' failed with exit > status 1 > make: *** [all] Error 1 > > I have boost 1.34.1, numpy 1.2.1, gcc 4.1.2, python 2.5.2 and > cuda 2.1 Ah great, gcc 4.1 being fabulous again. I tried this gcc here: gcc (GCC) 4.1.2 20071124 (Red Hat 4.1.2-42) and I didn't have a problem. (Important note: http://is.gd/iFIp) Since I can't reproduce it, can you try and see what's wrong there? Also, the file src/wrapper/tools.cpp hasn't been in use for quite a while, maybe your checkout went bad somehow? Andreas
signature.asc
Description: This is a digitally signed message part.
_______________________________________________ PyCuda mailing list [email protected] http://tiker.net/mailman/listinfo/pycuda_tiker.net
