Why is it very slow when I use jax.vmap on jax.numpy.convolve.(On CPU)

[xinyuxiao113]

I find a strange phenomenon: the time cost of jax.numpy.convolve(vmap on CPU) with respect to batch quantity is increasing much faster than linearly.

import jax, time, jax.random as rd, jax.numpy as jnp, matplotlib.pyplot as plt

n = list(range(10,100,10))
time_list = []

for k in n:
    x0 = rd.normal(rd.PRNGKey(0),(k,1000))
    y0 = rd.normal(rd.PRNGKey(1),(k,1000))
    t0 = time.time()
    y = jax.vmap(jnp.convolve)(x0,y0)
    t1 = time.time()
    time_list.append(t1 - t0)

plt.plot(n, time_list)

I can even use a for loop instead of jax.vmap to get a linearly increasing time consumption. So what's wrong with jax.vmap here? By the way, this problem is not exist on GPU.