Numpy VS. Cupy

meas_numpy_cupy_performance.py

import time

import cupy
import numpy


cnt = 100
N = 10
shapes = ((N, N), (N, N, N, N), (N, N, N, N, N, N), (N, N, N, N, N, N, N))


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(func):
    meas_results = []
    print("Array size,Numpy,Cupy")
    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},{1},{2}".format(*result))


def array_add(A, xp):
    return A + A

print("Array add")
meas_numpy_cupy(array_add)

def array_sub(A, xp):
    return A - A

print("Array sub")
meas_numpy_cupy(array_sub)

def array_sum(A, xp):
    return A.sum()

print("Array sum")
meas_numpy_cupy(array_sum)

def array_argmax(A, xp):
    return A.argmax()

print("Array argmax")
meas_numpy_cupy(array_argmax)

def array_sort(A, xp):
    return xp.sort(A)

print("Array sort")
meas_numpy_cupy(array_sort)

shapes = ((N, N), (N, N, N, N), (N, N, N, N, N, N))

def array_tensordot(A, xp):
    return xp.tensordot(A, A)

print("Array tensordot")
meas_numpy_cupy(array_tensordot)

def array_matmul(A, xp):
    return xp.matmul(A, A)

print("Array matmul")
meas_numpy_cupy(array_matmul)

def array_einsum(A, xp):
    return xp.einsum("...i, ...j->...ij", A, A)

print("Array einsum")
meas_numpy_cupy(array_einsum)

def array_moveaxis(A, xp):
    return xp.moveaxis(A, 0, -1)

print("Array transpose")
meas_numpy_cupy(array_moveaxis)

def array_sin(A, xp):
    return xp.sin(A)

print("Array sin")
meas_numpy_cupy(array_sin)

shapes = ((N, N), (N * N, N * N), (N * N * N, N * N * N))

def array_eigh(A, xp):
    return xp.linalg.eigh(A)

print("Array eigenvalue")
meas_numpy_cupy(array_eigh)

def array_inv(A, xp):
    return xp.linalg.inv(A)

print("Array inv")
meas_numpy_cupy(array_inv)

Hi, I like your script, so I made a more abstract version. Hope you'll like it: https://gist.github.com/magnium/cf96160d248a79f9463439695a7748e8