max and min for float32 cupy array may be slow

[xu3kev]

• Conditions (you can just paste the output of python -c 'import cupy; cupy.show_config()')
  • CuPy version = commit 2146ce2
  • OS/Platform = Ubuntu 16.04/ V100
  • CUDA version = 10.0
• Code to reproduce

import cupy as cp
from contextlib import contextmanager

@contextmanager
def sync_time(name):
    start = cp.cuda.Event()
    end = cp.cuda.Event()
    start.record()
    start.synchronize()
    yield
    end.record()
    end.synchronize()
    t = cp.cuda.get_elapsed_time(start,end)
    print("{} : {} ms".format(name,t))

x = cp.random.normal(size=((400, 32, 28, 28))).astype(cp.float32)

with sync_time("cupy"):
    for i in range(1000):
        x.max()

x = cp.asnumpy(x) #move to cpu

with sync_time("numpy"):
    for i in range(1000):
        x.max()

• Results

cupy : 8457.2451171875 ms
numpy : 3005.8154296875 ms

In this case, Cupy is slower then Numpy.

[anaruse]

"reduction" operations including "max" and "min" in cupy are currently implemented in rather general way and are not so optimized in terms of performance, as far as I know. You may be able to get better performance by using some reduction implementation in cuDNN, CUB or Thrust, though those are not used in cupy for now.

CuPy team: Is anyone already working on performance improvement of "reduction" operations? I'm considering to speed up cupy reductions with CUB. Is there any concerns on use of CUB in cupy?

[samrere]

Hi, as of the latest version of cupy, the max function is still much slower than its numpy equivalence:

cupy : 8153.27001953125 ms
numpy : 2700.735595703125 ms

May I please get more information on what may be a workaround? Thanks

[asi1024]

#6549 resolved this issue.

cupy : 49.38854217529297 ms
numpy : 3057.1181640625 ms