无法正确打印两个值（python3.5+numba+CUDA8.0）

Question

有一个数组，我会在 GPU 中用它做一些计算。

在计算之前，我应该得到这个数组的子集。

当我打印子集时，发现两个值都不对。

代码如下：

import os,sys,time
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 = (3,1)  
tpb = (2,2)  

@cuda.jit
def calcu_TE(D,TE):
    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[0]
    #print(c_num)
    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:
            TE[0] = 0.0
        else:
            X = D[c_x,:]
            Y = D[c_y,:]
            if bx==1 :
                print('c_index,bx,tx,ty,X: ',c_index,bx,tx,ty,'  ',X[0],X[1],X[2],X[3],X[4],X[5],X[6],X[7],X[8],X[9])
                print('c_index,bx,tx,ty,Y: ',c_index,bx,tx,ty,'  ',Y[0],Y[1],Y[2],Y[3],Y[4],Y[5],Y[6],Y[7],Y[8],Y[9])
            #print('c_index,bx,tx,ty,Y: ',c_index,bx,tx,ty,Y[0],Y[1],Y[2],Y[3],Y[4],Y[5],Y[6],Y[7],Y[8],Y[9])
            h = tx
            if h==0:
                Xi = X[1:]
                Xi1 = X[:-1]
                Yi = Y[1:]
                if bx==1 :
                    print('bx,tx,ty: ',bx,tx,ty,'\n Xi',Xi[0],Xi[1],Xi[2],Xi[3],Xi[4],Xi[5],Xi[6],Xi[7],Xi[8],
                          '\n Xi1',Xi1[0],Xi1[1],Xi1[2],Xi1[3],Xi1[4],Xi1[5],Xi1[6],Xi1[7],Xi1[8],
                          '\n Yi',Yi[0],Yi[1],Yi[2],Yi[3],Yi[4],Yi[5],Yi[6],Yi[7],Yi[8])
        c_index +=gw


D = np.array([[ 0.42487645,0.41607881,0.42027071,0.43751907,0.43512794,0.43656972,0.43940639,0.43864551,0.43447691,0.43120232],
              [2.989578,2.834707,2.942902,3.294948,2.868170,2.975180,3.066900,2.712719,2.835360,2.607334]], dtype=np.float32)
TE = np.empty([1,1])
print('D: ',D)

stream = cuda.stream()
with stream.auto_synchronize():
    dD = cuda.to_device(D, stream)
    dTE = cuda.to_device(TE, stream)
    calcu_TE[bpg, tpb, stream](dD,dTE)

输出是：

D:  [[ 0.42487645  0.41607881  0.42027071  0.43751907  0.43512794  0.43656972
   0.43940639  0.43864551  0.43447691  0.43120232]
 [ 2.98957801  2.83470702  2.94290209  3.2949481   2.86817002  2.97517991
   3.06690001  2.71271896  2.83536005  2.6073339 ]]
c_index,bx,tx,ty,X:  1 1 0 0    0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
c_index,bx,tx,ty,X:  1 1 1 0    0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
c_index,bx,tx,ty,X:  1 1 0 1    0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
c_index,bx,tx,ty,X:  1 1 1 1    0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
c_index,bx,tx,ty,Y:  1 1 0 0    2.989578 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 2.835360 2.607334
c_index,bx,tx,ty,Y:  1 1 1 0    2.989578 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 2.835360 2.607334
c_index,bx,tx,ty,Y:  1 1 0 1    2.989578 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 2.835360 2.607334
c_index,bx,tx,ty,Y:  1 1 1 1    2.989578 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 2.835360 2.607334

bx,tx,ty:  1 0 0
 Xi 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
 Xi1 0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477
 Yi 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 0.000000 18949972373983835000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000
bx,tx,ty:  1 0 1
 Xi 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477 0.431202
 Xi1 0.424876 0.416079 0.420271 0.437519 0.435128 0.436570 0.439406 0.438646 0.434477
 Yi 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 0.000000 18949972373983835000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000

这太奇怪了。

Yi 应该是Yi 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 2.835360 2.607334。

但是打印的是Yi 2.834707 2.942902 3.294948 2.868170 2.975180 3.066900 2.712719 0.000000 18949972373983835000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000.000000。

有两个值错误。

我不知道为什么会这样。有什么我忽略的吗？

Answer 1

这似乎是 Numba 编译器为内核中很长的打印语句生成代码的方式存在问题，与内核的正确性无关。如果您像这样更改代码（即缩短打印语句）：

@cuda.jit
def calcu_TE(D,TE):
    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[0]
    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:
            TE[0] = 0.0
        else:
            X = D[c_x,:]
            Y = D[c_y,:]
            if bx==1 :
                print('c_index,bx,tx,ty,X: ',c_index,bx,tx,ty,'  ',X[0],X[1],X[2],X[3],X[4],X[5],X[6],X[7],X[8],X[9])
                print('c_index,bx,tx,ty,Y: ',c_index,bx,tx,ty,'  ',Y[0],Y[1],Y[2],Y[3],Y[4],Y[5],Y[6],Y[7],Y[8],Y[9])
            h = tx
            if h==0:
                Xi = X[1:]
                Xi1 = X[:-1]
                Yi = Y[1:]
                if bx==1 :
                    print('bx,tx,ty,Yi:',bx,tx,ty,'  ',Yi[0],Yi[1],Yi[2],Yi[3],Yi[4],Yi[5],Yi[6],Yi[7],Yi[8])
        c_index +=gw

您应该会发现Yi 打印正确。一般来说，在 CUDA 中依靠 print 语句来检测内核是一个相当糟糕的主意，而且通常你这样做只会让自己感到困惑，就像在这种情况下一样。