Compare Numpy and Cupy performance

meas_numpy_cupy_performance.py

# based on https://gist.github.com/fukatani/4702aa05aed255cd25f42e77d0a22e37
import time

import cupy
import numpy

cnt = 100
N = 10

def meas_cupy(func, operand):
    stream = cupy.cuda.Stream.null
    start = stream.record()
    for i in range(cnt):
        func(operand, xp=cupy)
    end = stream.record()
    end.synchronize()
    elapsed = cupy.cuda.get_elapsed_time(start, end) / cnt
    return elapsed

def meas_numpy(func, operand):
    start = time.time()
    for i in range(cnt):
        func(operand, xp=numpy)
    end = time.time()
    elapsed = (end - start) * 1000 / cnt
    return elapsed

def meas_numpy_cupy(title, func, shapes):
    meas_results = []
    print()
    print("===", title, "===")
    print("Array size\tNumpy\tCupy")
    for shape in shapes:
        A = cupy.random.rand(*shape)
        a = A.get()
        cupy_elapsed = meas_cupy(func, A)
        numpy_elapsed = meas_numpy(func, a)
        meas_results.append((a.size, numpy_elapsed, cupy_elapsed))
    for result in meas_results:
        print("{0:10d}\t{1:.3f}\t{2:.3f}".format(*result))

# ----------------------------------------

shapes_ = ((N, N), (N, N, N, N), (N, N, N, N, N, N), (N, N, N, N, N, N, N))

for title, func in (("Array add", "A+A"),
                    ("Array sub", "A-A"),
                    ("Array sum", "A.sum()"),
                    ("Array argmax", "A.argmax()"),
                    ("Array sort", "xp.sort(A)")):
    meas_numpy_cupy(title, lambda A, xp: eval(func), shapes_)

shapes_ = ((N, N), (N, N, N, N), (N, N, N, N, N, N))

for title, func in (("Array tensordot", "xp.tensordot(A, A)"),
                    ("Array matmul", "xp.matmul(A, A)"),
                    ("Array einsum", "xp.einsum('...i, ...j->...ij', A, A)"),
                    ("Array transpose", "xp.moveaxis(A, 0, -1)"),
                    ("Array sin", "xp.sin(A)")):
    meas_numpy_cupy(title, lambda A, xp: eval(func), shapes_)

shapes_ = ((N, N), (N * N, N * N), (N * N * N, N * N * N))

for title, func in (("Array eigenvalue", "xp.linalg.eigh(A)"),
                    ("Array inv", "xp.linalg.inv(A)")):
    meas_numpy_cupy(title, lambda A, xp: eval(func), shapes_)

results.txt

# results for Dell XPS 2020

# CPU:	Intel(R) Core(TM) i7-10750H CPU @ 2.60GHz
# RAM:	16.0 GB (15.8 GB usable)
# GPU: NVIDIA GeForce GTX 1650 Ti
# Windows 11 Pro 21H2
# CUDA 11.6
# numpy==1.18.5
# Python 3.7.7


=== Array add ===
Array size      Numpy   Cupy
       100      0.010   0.085
     10000      0.020   0.045
   1000000      1.860   0.102
  10000000      18.062  0.951

=== Array sub ===
Array size      Numpy   Cupy
       100      0.010   0.129
     10000      0.020   0.045
   1000000      1.780   0.097
  10000000      18.374  0.918

=== Array sum ===
Array size      Numpy   Cupy
       100      0.010   0.145
     10000      0.020   0.051
   1000000      0.430   0.666
  10000000      12.210  5.490

=== Array argmax ===
Array size      Numpy   Cupy
       100      0.010   0.112
     10000      0.020   0.042
   1000000      0.540   1.239
  10000000      6.334   12.349

=== Array sort ===
Array size      Numpy   Cupy
       100      0.010   0.203
     10000      0.130   0.180
   1000000      10.643  3.280
  10000000      108.295 39.788

=== Array tensordot ===
Array size      Numpy   Cupy
       100      0.040   0.163
     10000      0.210   24.793
   1000000      559.860 177.300

=== Array matmul ===
Array size      Numpy   Cupy
       100      0.020   0.204
     10000      0.050   0.224
   1000000      4.913   4.124

=== Array einsum ===
Array size      Numpy   Cupy
       100      0.020   0.235
     10000      0.130   0.160
   1000000      23.745  0.566

=== Array transpose ===
Array size      Numpy   Cupy
       100      0.020   0.012
     10000      0.020   0.019
   1000000      0.020   0.014

=== Array sin ===
Array size      Numpy   Cupy
       100      0.010   0.138
     10000      0.080   0.046
   1000000      6.185   1.396

=== Array eigenvalue ===
Array size      Numpy   Cupy
       100      0.200   10.206
     10000      8.662   5.830
   1000000      400.888 102.044

=== Array inv ===
Array size      Numpy   Cupy
       100      0.220   0.699
     10000      5.722   1.975
   1000000      92.929  46.838