I'm not sure your code is doing what you mean it to. I get totally different results when running a dot product of two gpuarrays. Did you check the output to show that it was doing what you expect? I'm actually surprised it ran at all.
As far as I know, you can't simply replace numpy arrays with gpuarrays in any numpy method, but I would love to be wrong about this... Best, Per On Thu, May 28, 2009 at 10:34 AM, Hua Wong <hua.w...@pasteur.fr> wrote: > Following Andreas remark, I replaced the following in the code : > > import pycuda.gpuarray as gpuarray > import pycuda.driver as cuda > import pycuda.autoinit > import numpy > import time > > a=numpy.float32(numpy.random.randn(4000,4000)) > b=numpy.float32(numpy.random.randn(4000,4000)) > > tic=time.time() > axb=numpy.dot(a,b) # I assume this time it is matrix multiplication, > according to numpy tutorials I've read... > toc=time.time()-tic > print toc,"CPU" > > > tic=time.time() > a_gpu = gpuarray.to_gpu(a) > b_gpu = gpuarray.to_gpu(b) > axbGPU = (numpy.dot(a_gpu,b_gpu)).get() # ditto here > toc=time.time()-tic > print toc,"GPU" > > Here are the results I get : > 2.06739115715 CPU > 0.171211004257 GPU > > It speeds up the calculation 11 times :) > But I can't try bigger matrices, I lack RAM :( > > Hua Wong a écrit : >> >> Thanks, I'm also puzzled by the results because I thought a 1e4*1e4 matrix >> was already ginormous... >> >> I expected something like a 49 time speedup like in the >> test_gpuarray_speed_random.py (size ~16000000 give a x49 speedup). >> >> So I guess I'm doing something wrong somewhere. I will check the test >> script... >> >> Getting : >> 0.46285700798 CPU >> 0.728541851044 GPU >> >> with your code on a CentOS machine, with a GTX280 and 2x quad core E5410 >> >> Per B. Sederberg a écrit : >>> >>> I modified your code slightly to make it so you are comparing apples >>> to apples a bit better and I'm getting even worse performance for the >>> GPU (GTX285 on Debian Testing): >>> >>> 0.652935028076 CPU >>> 1.61081981659 GPU >>> >>> Here's the new code, which puts the sending and receiving of the data >>> to/from the card in the loop and also has the CPU perform a float32 >>> operation just like the GPU: >>> >>> import pycuda.gpuarray as gpuarray >>> import pycuda.driver as cuda >>> import pycuda.autoinit >>> import numpy >>> import time >>> >>> a=numpy.float32(numpy.random.randn(1e4,1e4)) >>> >>> tic=time.time() >>> a_square=a*a >>> toc=time.time()-tic >>> print toc,"CPU" >>> >>> >>> tic=time.time() >>> a_gpu = gpuarray.to_gpu(a) >>> a_squared = (a_gpu*a_gpu).get() >>> toc=time.time()-tic >>> print toc,"GPU" >>> >>> It looks like you'll need to have even larger matrices before you'll >>> see a major GPU benefit, though I'm a bit surprised by these results. >>> >>> Best, >>> Per >>> >>> >>> On Thu, May 28, 2009 at 6:55 AM, Hua Wong <hua.w...@pasteur.fr> wrote: >>> >>>> >>>> Here is the results I get >>>> 0.865973949432 CPU >>>> 0.582780122757 GPU >>>> >>>> I kind of expected more... (the GPU is a GTX280) >>>> >>>> Of course, I never exclude that I did something stupid, in fact, I >>>> expect >>>> it... >>>> Is it the acceleration I should expect from this kind of matrix >>>> operation? >>>> If yes, well cool... I guess. >>>> If not, did I miss something? >>>> >>>> Here is the code I use : >>>> >>>> import pycuda.gpuarray as gpuarray >>>> import pycuda.driver as cuda >>>> import pycuda.autoinit >>>> import numpy >>>> import time >>>> >>>> a=numpy.random.randn(1e4,1e4) >>>> >>>> tic=time.time() >>>> a_square=a*a >>>> toc=time.time()-tic >>>> print toc,"CPU" >>>> >>>> a_gpu = gpuarray.to_gpu(a.astype(numpy.float32)) >>>> >>>> tic=time.time() >>>> a_squared = (a_gpu*a_gpu).get() >>>> toc=time.time()-tic >>>> print toc,"GPU" >>>> >>>> >>>> _______________________________________________ >>>> PyCuda mailing list >>>> PyCuda@tiker.net >>>> http://tiker.net/mailman/listinfo/pycuda_tiker.net >>>> >>>> >> >> > > > _______________________________________________ > PyCuda mailing list > PyCuda@tiker.net > http://tiker.net/mailman/listinfo/pycuda_tiker.net > _______________________________________________ PyCuda mailing list PyCuda@tiker.net http://tiker.net/mailman/listinfo/pycuda_tiker.net
