I have no idea if what you propose would work well. You can make a new OP that use pycuda for the computation. We do that for our fft op in the new back-end:
https://github.com/Theano/Theano/blob/master/theano/gpuarray/fft.py On Sat, Feb 11, 2017 at 6:45 AM Kiuhnm Mnhuik <[email protected]> wrote: > What do you mean by "reusing a row"? If each core does one and only one > reduction on a single row, therefore there shouldn't be any reuse. > I mean that one core of the GPU accesses and reduces one and only one > specific row: > > *** row1 *** <----- just core 1 > *** row2 *** <----- just core 2 > *** row3 *** <----- just core 3 > *** row4 *** <----- just core 4 > *** row5 *** <----- just core 5 > > This makes sense because there are so many rows that all cores can run in > parallel, each one working on its own row. > > Reductions aren't usually a bottleneck, but I'm doing something quite > unusual. > > Can I use pyCuda to work *directly* on Theano data already allocated on > the GPU? This might be my only option. I can't copy or move the data back > to the CPU or it'll kill performance. > > > On Friday, February 10, 2017 at 7:37:57 PM UTC+1, nouiz wrote: > > X+Y is trivially parallelisable. Bug not X.sum(axis=1). I'm pretty sure we > do something sensible. I check the code and it is the case. > > Reduction isn't trivially parallelisable. This is way it get less speed > up. When we reuse a row, we can't parallelize it as much as when adding 2 > matrix. > But in all cases, in a real model, it shouldn't make a difference, > reduction aren't bottleneck normally. If you have such case, I would like > to see a profile that show this. > > Fred > > On Tue, Feb 7, 2017 at 6:28 PM Kiuhnm Mnhuik <[email protected]> wrote: > > Hi Fred, > > I'm talking about the GPU. With a 10000x1000 matrix X, X.sum(axis=1) is 10 > times slower than X + Y, where Y is another matrix of the same shape, > according to my tests. > I suspect that you're reducing each row using some O(logn) algorithm which > makes sense when one needs to reduce a single long vector. But in this > case, shouldn't we assign each row to a single core of the GPU and reduce > the row as we would do on the CPU? The parallelism would result from having > so many rows. > Of course, if the matrix had just 10 rows this algorithm would be very > slow, but with 10000 rows it should be faster than what you're doing right > now. It might be almost as fast as doing X + Y. > I'm speculating since I've never looked into CUDA programming (it's on my > TODO list!). > > > On Tuesday, February 7, 2017 at 10:49:47 PM UTC+1, nouiz wrote: > > Hi, > > for the gpu, int are only supported in the new gpu back-end > (device=cuda*). In the old back-end, they would end up being on CPU. This > is why at many places it is told to not use int on the GPU. But it isn't > true with the new back-end. > > For the reduction being slow, we didn't parallelize it on the CPU. It > wasn't a bottleneck on the CPU and we don't have much time to optimize the > CPU. So I would recommand to time your real model on the CPU before > spending much time thinking about the parallel reduction on CPU as it is > probably not a problem. > > > > > Fred > > On Mon, Feb 6, 2017 at 8:11 PM Kiuhnm Mnhuik <[email protected]> wrote: > > Reductions are quite slow. Without the final reduction I get a 100x speed > up. > Why is Y.sum(axis=1) so slow? I think that if each core handled a single > row it'd be 10 times faster for matrices with many rows like in this case. > Theano is probably using an O(logn) algorithm which is only useful when > one needs to reduce a single but long vector. > Can you confirm? > > > On Tuesday, February 7, 2017 at 12:37:02 AM UTC+1, Kiuhnm Mnhuik wrote: > > I tried the following code: > > def test_speed(): > print('Computing X and X2...', end='', flush=True) > X_np = np.random.uniform(0, 100, size=(10000, 1000)).astype(floatX) > X2_np = np.random.uniform(0, 100, size=(10000, > 1000)).astype(floatX) > print('done!', flush=True) > > print('Moving X and X2 to the GPU...', end='', flush=True) > X = theano.shared(X_np) > X2 = theano.shared(X2_np) > print('done!', flush=True) > > print('Building the graph...', end='', flush=True) > Y = X > for _ in range(100): > # Y = Y * (Y <= X2) > Y = Y * (Y - X2) > Y.sum(axis=1) > print('done!', flush=True) > > print('compiling...', end='', flush=True) > f = theano.function([], Y) > print('done!', flush=True) > > import time > t = time.clock() > f() > print(time.clock() - t) > > Note that there is a line with '<=' and another with '-' in the loop. > They're exclusive. Here are the timings in seconds: > > CPU GPU > '-' 0.21 0.016 > <= 0.39 0.019 > > I'd say I don't need to worry about using comparisons. > > On Monday, February 6, 2017 at 1:20:13 PM UTC+1, Kiuhnm Mnhuik wrote: > > I'm using Theano 0.9.0b1 with the new back-end. > Should I use float32 for everything (even for bool masks) for maximum > speed on GPU (GTX 970)? > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > > > For more options, visit https://groups.google.com/d/optout. > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
