我是一个初学者，试图学习如何使用GPU来执行高速计算 . 我正在尝试使用GPU实现一个简单的FFT程序 . 下面是我用于使用CPU内核计算FFT的程序 .

from time import time as timer

import numpy as np

import matplotlib.pyplot as plt

winsize=512

shift=16

my_cmap='gray_r'

Fs = 8000

f = 1000

sample =200000

x = np.arange(sample)

y = np.sin(2 * np.pi * f * x / Fs)

data_len=len(y)

window_func=np.blackman(winsize)

fftdata=np.zeros((0,int(winsize/2)))

startime=timer()

for frame in range(0, data_len, shift):

#==============================================================================

# if frame>0:

# break

#==============================================================================

kiri=y[frame:frame+winsize]

if len(kiri) != winsize:

break

windata = window_func * kiri

fftframe=np.fft.fft(windata,n=winsize)

magframe=np.abs(fftframe)**2

powerframe=np.log10(magframe):int(winsize/2)].reshape((1,int(winsize/2)))

fftdata=np.append(fftdata,powerframe,axis=0)

endtime=timer()-startime

fftdata=np.asarray(fftdata)

fftfrq=np.fft.fftfreq(winsize,d=1/Fs)[:int(winsize/2)]

print("CPU runtime:",endtime,"sec")

现在，下图是使用 imshow() 函数绘制时的输出作为频谱图：

时序输出如下：

CPU runtime: 65.02100014686584 sec

现在我重写上面的程序，使用我的PC的GPU，即使用Anaconda提供的pyculib和numba软件包的Quadro K2200 .

from time import time as timer

import numpy as np

import pyculib.fft

from numba import cuda

import matplotlib.pyplot as plt

winsize=512

shift=16

Fs = 8000

f = 1000

sample =200000

t = np.arange(sample,dtype=np.float64)

y = np.sin(2 * np.pi * f * t / Fs)

my_cmap='gray_r'

data_len=len(y)

window_func=np.blackman(winsize)

fftdata_gpu=np.zeros((0,int(winsize/2)))

startime=timer()

for frame in range(0, data_len, shift):

# =============================================================================

# if frame>0:

# break

# =============================================================================

kiri=y[frame:frame+winsize]

if len(kiri) != winsize:

break

windata = window_func * kiri

fftframe_gpu = np.zeros(winsize, np.complex128)

d_xf_gpu = cuda.to_device(fftframe_gpu)

pyculib.fft.fft(windata.astype(np.complex128),d_xf_gpu)

d_xf_gpu.copy_to_host(fftframe_gpu)

magframe_gpu=np.abs(fftframe_gpu)**2

powerframe_gpu=np.log10(magframe_gpu)[:int(winsize/2)].reshape((1,int(winsize/2)))

fftdata_gpu=np.append(fftdata_gpu,powerframe_gpu,axis=0)

endtime=timer()-startime

fftdata_gpu=np.asarray(fftdata_gpu)

print("GPU runtime:",endtime,"sec")

运行上述程序时的时序输出显示GPU实现实际上花费了30秒 .

GPU runtime: 92.87200021743774 sec

我猜测这是因为我反复将阵列复制到设备上，并为每一帧取回它 . 有没有更好的方法来实现这个？我真的很想知道我在这里做错了什么 .

下面我粘贴GPU实现的输出 .

Edit:Adding the profiling results

对于CPU代码：前20个函数调用根据累计时间排序

569255 function calls (563691 primitive calls) in 64.792 seconds

Ordered by: cumulative time

List reduced from 2594 to 20 due to restriction <20>

ncalls tottime percall cumtime percall filename:lineno(function)

296/1 0.009 0.000 64.793 64.793 {built-in method builtins.exec}

1 11.489 11.489 64.793 64.793 cuda_fft_tr1_cpu.py:6()

12469 0.037 0.000 52.145 0.004 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numpy\lib\function_base.py:5095(append)

12469 52.093 0.004 52.093 0.004 {built-in method numpy.core.multiarray.concatenate}

12469 0.073 0.000 0.622 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numpy\fft\fftpack.py:102(fft)

333/2 0.002 0.000 0.493 0.247 :966(_find_and_load)

333/2 0.001 0.000 0.493 0.246 :936(_find_and_load_unlocked)

323/3 0.002 0.000 0.491 0.164 :651(_load_unlocked)

272/3 0.001 0.000 0.491 0.164 :672(exec_module)

432/3 0.000 0.000 0.490 0.163 :211(_call_with_frames_removed)

12469 0.073 0.000 0.415 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numpy\fft\fftpack.py:47(_raw_fft)

356/24 0.000 0.000 0.336 0.014 {built-in method builtins.__import__}

1 0.000 0.000 0.232 0.232 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\matplotlib\pyplot.py:17()

1445/628 0.001 0.000 0.220 0.000 :997(_handle_fromlist)

1 0.000 0.000 0.141 0.141 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numpy\__init__.py:106()

12469 0.139 0.000 0.139 0.000 {built-in method numpy.fft.fftpack_lite.cfftf}

328 0.003 0.000 0.122 0.000 :870(_find_spec)

310 0.000 0.000 0.117 0.000 :1149(find_spec)

310 0.002 0.000 0.117 0.000 :1117(_get_spec)

1 0.000 0.000 0.115 0.115 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\matplotlib\__init__.py:101()

对于GPU代码：基于累积时间进行排序

5689881 function calls (5642977 primitive calls) in 94.179 seconds

Ordered by: cumulative time

List reduced from 4373 to 20 due to restriction <20>

ncalls tottime percall cumtime percall filename:lineno(function)

590/1 0.019 0.000 94.207 94.207 {built-in method builtins.exec}

1 12.080 12.080 94.207 94.207 cuda_fft_tr1_gpu.py:6()

12469 0.046 0.000 51.752 0.004 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numpy\lib\function_base.py:5095(append)

12469 51.679 0.004 51.679 0.004 {built-in method numpy.core.multiarray.concatenate}

62345/49876 0.131 0.000 22.880 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\devices.py:209(_require_cuda_context)

12469 0.111 0.000 20.875 0.002 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\fft\api.py:190(fft)

49876 17.129 0.000 17.171 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\utils\libutils.py:40(wrapped)

12469 0.176 0.000 15.354 0.001 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\fft\api.py:38(__init__)

12469 0.284 0.000 15.046 0.001 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\fft\binding.py:207(many)

37407 0.151 0.000 7.634 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\devicearray.py:451(auto_device)

24938 0.108 0.000 4.707 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\devicearray.py:422(from_array_like)

12469 0.043 0.000 4.644 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\fft\api.py:134(forward)

24938 0.356 0.000 4.599 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\devicearray.py:58(__init__)

87290 4.156 0.000 4.398 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\driver.py:284(safe_cuda_api_call)

12469 0.035 0.000 4.292 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\api.py:26(to_device)

12469 0.064 0.000 3.468 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\pyculib\fft\api.py:86(_prepare)

24938 0.027 0.000 3.404 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\api.py:275(_auto_device)

24938 0.114 0.000 2.452 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\devicearray.py:139(copy_to_device)

24938 0.080 0.000 2.157 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\driver.py:1573(host_to_device)

24938 0.261 0.000 1.848 0.000 C:\Users\na\AppData\Local\Continuum\Anaconda3_2\lib\site-packages\numba\cuda\cudadrv\driver.py:667(memalloc)

pyculib的fft功能大约需要20秒，而numpy fft需要大约0.6秒 . 为什么pyculib的功能需要这么长时间？有没有办法改进代码以缩短这个时间？或者使用不同的库更好吗？