python如何使用gpu_Python GPU编程之加速篇

import sys

import numpy as np

from scipy.signal import fftconvolve

from scipy import misc, ndimage

from matplotlib import pyplot as plt

from numbapro.cudalib import cufft

from numbapro import cuda, vectorize

from timeit import default_timer as timer

@vectorize(['complex64(complex64, complex64)'], target='gpu')

#目标平台是64位机器且拥有GPU

def vmult(a, b):

return a * b

def best_grid_size(size, tpb):

bpg = np.ceil(np.array(size, dtype=np.float) / tpb).astype(np.int).tolist()

return tuple(bpg)

def main():

# 构建过滤器

laplacian_pts = '''

-4 -1 0 -1 -4

-1 2 3 2 -1

0 3 4 3 0

-1 2 3 2 -1

-4 -1 0 -1 -4

'''.split()

laplacian = np.array(laplacian_pts, dtype=np.float32).reshape(5, 5)

# 构建图像

try:

filename = sys.argv[1]

image = ndimage.imread(filename, flatten=True).astype(np.float32)

except IndexError:

image = misc.lena().astype(np.float32)

print("Image size: %s" % (image.shape,))

response = np.zeros_like(image)

response[:5, :5] = laplacian

# CPU

ts = timer()

cvimage_cpu = fftconvolve(image, laplacian, mode='same')

te = timer()

print('CPU: %.2fs' % (te - ts))

# GPU

threadperblock = 32, 8

blockpergrid = best_grid_size(tuple(reversed(image.shape)), threadperblock)

print('kernel config: %s x %s' % (blockpergrid, threadperblock))

# cuFFT系统，触发器初始化.

# 对于小数据集来说，效果更明显.

# 不应该计算这里浪费的时间

cufft.FFTPlan(shape=image.shape, itype=np.complex64, otype=np.complex64)

# 开始GPU运行计时

ts = timer()

image_complex = image.astype(np.complex64)

response_complex = response.astype(np.complex64)

d_image_complex = cuda.to_device(image_complex)

d_response_complex = cuda.to_device(response_complex)

cufft.fft_inplace(d_image_complex)

cufft.fft_inplace(d_response_complex)

vmult(d_image_complex, d_response_complex, out=d_image_complex)

cufft.ifft_inplace(d_image_complex)

cvimage_gpu = d_image_complex.copy_to_host().real / np.prod(image.shape)

te = timer()

print('GPU: %.2fs' % (te - ts))

# 绘制结果

plt.subplot(1, 2, 1)

plt.title('CPU')

plt.imshow(cvimage_cpu, cmap=plt.cm.gray)

plt.axis('off')

plt.subplot(1, 2, 2)

plt.title('GPU')

plt.imshow(cvimage_gpu, cmap=plt.cm.gray)

plt.axis('off')

plt.show()

if __name__ == '__main__':

main()

Image size: (512L, 512L)

CPU: 0.66s

kernel config: (16, 64) x (32, 8)

GPU: 0.09s

[Finished in 61.4s]

Numpy提供常见的数学函数，包含许多有用的数学库

Scipy是python下的数值计算库，和Numpy一样是科学计算不可或缺的库

Matplotlib是用以绘制二维图形的Python模块

Numbapro是CUDA提供的专用库

timeit是计时工具