我有一个代码,可以用python3.5和numba和CUDA8.0在GPU中进行一些计算。当输入一个大小为(50,27)的数组时,它成功运行并得到正确的结果。我把输入的数据改为size(200340),它有一个错误。在
我在代码中使用共享内存。共享内存不足吗?或者网格大小和块大小不好?我不知道如何识别它并为网格和块选择合适的大小。在
我设置了小网格大小和块大小,误差是一样的。在
我该怎么解决这个问题?谢谢你的建议。在
我简化了我的代码,它也有同样的错误。这里可以方便地设置输入数据的大小:df = np.random.random_sample((300, 200)) + 10
。在
代码:
import os,sys,time,math
import pandas as pd
import numpy as np
from numba import cuda, float32
os.environ['NUMBAPRO_NVVM']=r'D:\NVIDIA GPU Computing Toolkit\CUDA\v8.0\nvvm\bin\nvvm64_31_0.dll'
os.environ['NUMBAPRO_LIBDEVICE']=r'D:\NVIDIA GPU Computing Toolkit\CUDA\v8.0\nvvm\libdevice'
bpg = 8
tpb = (4,32)
tsize = (3,4)
hsize = (1,4)
@cuda.jit
def calcu_T(D, T):
gw = cuda.gridDim.x
bx = cuda.blockIdx.x
tx = cuda.threadIdx.x
bw = cuda.blockDim.x
ty = cuda.threadIdx.y
bh = cuda.blockDim.y
c_num = D.shape[1]
c_index = bx
while c_index<c_num*c_num:
c_x = int(c_index/c_num)
c_y = c_index%c_num
if c_x==c_y:
T[c_x,c_y] = 0.0
else:
X = D[:,c_x]
Y = D[:,c_y]
hbuf = cuda.shared.array(hsize, float32)
h = tx
Xi = X[h:]
Xi1 = X[:-h]
Yih = Y[:-h]
sbuf = cuda.shared.array(tsize, float32)
L = len(Xi)
#mean
if ty==0:
Xi_m = 0.0
Xi1_m = 0.0
Yih_m = 0.0
for i in range(L):
Xi_m += Xi[i]
Xi1_m += Xi1[i]
Yih_m += Yih[i]
Xi_m = Xi_m/L
Xi1_m = Xi1_m/L
Yih_m = Yih_m/L
sbuf[0,tx] = Xi_m
sbuf[1,tx] = Xi1_m
sbuf[2,tx] = Yih_m
cuda.syncthreads()
sl = cuda.shared.array(tpb, float32)
r_index = ty
s_l = 0.0
while r_index<L:
s1 = 0.0
for i in range(L):
s1 += (Xi[r_index]+Xi1[i])/sbuf[0,tx]
s_l += s1
r_index +=bh
sl[tx,ty] = s_l
cuda.syncthreads()
#
if ty==0:
ht = 0.0
for i in range(bh):
ht += sl[tx,i]
hbuf[0,tx] = ht/L
cuda.syncthreads()
#max
if tx==0 and ty==0:
m_t = 0.0
for index,ele in enumerate(hbuf[0]):
if index==0:
m_t = ele
elif ele>m_t:
m_t = ele
T[c_x,c_y] = m_t
c_index +=gw
df = np.random.random_sample((300, 200)) + 10
D = np.array(df, dtype=np.float32)
r,c = D.shape
T = np.empty([c,c])
dD = cuda.to_device(D)
dT = cuda.device_array_like(T)
calcu_T[bpg, tpb](dD,dT)
dT.copy_to_host(T)
错误:
^{pr2}$我的设备信息:
Device 0:
CUDA Driver Version / Runtime Version 8.0 / 8.0
CUDA Capability Major/Minor version number: 5.0
Total amount of global memory: 2048 MBytes (2147483648 bytes)
( 5) Multiprocessors, (128) CUDA Cores/MP: 640 CUDA Cores
Maximum Texture Dimension Size (x,y,z) 1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
Maximum Layered 1D Texture Size, (num) layers 1D=(16384), 2048 layers
Maximum Layered 2D Texture Size, (num) layers 2D=(16384, 16384), 2048 layers
Total amount of constant memory: 65536 bytes
Total amount of shared memory per block: 49152 bytes
Total number of registers available per block: 65536
Warp size: 32
Maximum number of threads per multiprocessor: 2048
Maximum number of threads per block: 1024
Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
Max dimension size of a grid size (x,y,z): (2147483647, 65535, 65535)
Maximum memory pitch: 2147483647 bytes
Texture alignment: 512 bytes
你的代码没有问题。如果我在GTX970上运行你的代码,我会得到:
也就是说,没有运行时错误,但是包含内核的python代码需要6.6秒才能运行。如果我用CUDA分析器分析代码:
^{pr2}$您可以看到您发布的内核需要6.5秒才能运行。在
您没有提供详细信息,但我猜您是在Windows上运行的,您的GPU是一个显示GPU,并且您的代码运行速度非常慢,以至于它达到了WDDM显示管理器看门狗超时限制。这是一个非常好的文档,并且已经被问过几百次了,例如here。在
你所选择的搜索引擎和CUDA Windows入门指南将为你提供从操作系统和硬件角度改善情况的替代方案的信息。然而,最明显的是改进代码,使其运行更快。在
相关问题 更多 >
编程相关推荐