Numpy(np.exp)とPyCUDAで指数関数計算の速度比較

import pycuda.autoinit
import pycuda.driver as drv
from pycuda import gpuarray, cumath
from pycuda.elementwise import ElementwiseKernel
from pycuda.compiler import SourceModule
import numpy as np

start = drv.Event()
end = drv.Event()

kernel = ElementwiseKernel(
   "double *a,double *b",
   "b[i] = exp(a[i]);")

mod = SourceModule("""
  __global__ void gexp(double *a,double *b,int n)
  {
    int i = threadIdx.x + blockIdx.x * blockDim.x;
    while (i < n) {
    b[i] = exp(a[i]);
    i += blockDim.x * gridDim.x;
    }
  }
  """)
knl = mod.get_function("gexp")

results = []
for N in [10**4, 10**5, 10**6, 10**7, 10**8]:
    a = 2*np.ones(N,dtype=np.float64)
    a_gpu = gpuarray.to_gpu(a)
    b_gpu = gpuarray.zeros_like(a_gpu)
    c_gpu = gpuarray.to_gpu(a)
    d_gpu = gpuarray.zeros_like(c_gpu)
    
    start.record()
    np.exp(a)
    end.record()
    end.synchronize()
    sec1 = start.time_till(end)*1e-3
    print ("Numpy",sec1)
    
    start.record() # start timing
    kernel(a_gpu,b_gpu)
    end.record() # end timing
    end.synchronize()
    sec2 = start.time_till(end)*1e-3
    print ("Kernel",sec2)
    print (np.allclose(np.exp(a),b_gpu.get()))
    
    start.record() # start timing
    knl(c_gpu,d_gpu,block=(1024,1,1),grid=(N//1024,1,1))
    end.record() # end timing
    end.synchronize()
    sec3 = start.time_till(end)*1e-3
    print ("knl",sec3)
    print (np.allclose(np.exp(a),d_gpu.get()))
    
    start.record()
    cumath.exp(a_gpu)
    end.record()
    end.synchronize()
    sec4 = start.time_till(end)*1e-3
    print ("Cumath", sec4)
    #print (np.allclose(np.exp(a),cumath.exp(a_gpu))
    results.append([N,sec1,sec2,sec3,sec4])

import matplotlib.pyplot as plt
results = np.array(results)
legends = []
nH = results[:5, 0:1]
rows = results[:5,1:6]
plt.semilogx(nH,rows, 'o-')
legends += ['' + s for s in ['numpy','wise','source','cumath']]
plt.rcParams['figure.figsize'] = 18, 10
plt.rcParams["font.size"] = "20"
plt.ylabel('Seconds')
plt.xlabel('Value of N')
plt.legend(legends);