CODEWITHSUNDEEP
pythonperformancenumpyscipy
user2675516
user2675516

Reputation:

Why scipy.distance.cdist has a big performance difference between using float32 (slower) and float64 (faster)?

Why scipy.distance.cdist has a big performance difference between using float32 and float64?

from scipy.spatial import distance
import numpy as np
import time

a_float32 = np.empty((1000000, 512), dtype=np.float32)
b_float32 = np.empty((1, 512), dtype=np.float32)
a_float64 = np.empty((1000000, 512), dtype=np.float64)
b_float64 = np.empty((1, 512), dtype=np.float64)

t1 = time.time()
for i in range(100):
    distance.cdist(a_float32, b_float32, "sqeuclidean")
t2 = time.time()
print(t2-t1)
t1 = time.time()
for i in range(100):
    distance.cdist(a_float64, b_float64, 'sqeuclidean')
t2 = time.time()
print(t2-t1)

On my computer, this code produces a timing of 130.6998474597931 for float32 and 22.450339794158936 for float64, a whopping 6x difference. What is the reason for such a big difference with float64 being faster than float32?

However the gap seems to become smaller and smaller if instead of b being np.empty((1, 512)), b is np.empty((k, 512)) with k (>1) becoming larger and larger. For example, when k=5, I get 222.25975680351257 for float32 and 110.36117148399353 for float64 (2x). Why the gap becomes smaller when k becomes larger?

Upvotes: 4

Views: 2671

Answers (1)

Warren Weckesser
Warren Weckesser

Reputation: 114921

The underlying C code that actually does the computation is implemented using C double variables, which are 64 bit floating point values. When you pass in arrays of np.float32, the data must be copied.

For the second part of your question: larger k means more work, so the overhead of copying the data is a smaller fraction of the overall time.

Upvotes: 6

Related Questions