如何使用numba在内核中分配数组？

Question

我试图在内核中的for循环中分配一些数组。内核如下所示：

@cuda.jit
def forcecudatiling(p_num,d_num,r,force):
    threadsInBlock=cuda.blockDim.x
    threadsInGrid=threadsInBlock*cuda.gridDim.x
    tid=cuda.threadIdx.x + cuda.blockIdx.x*cuda.blockDim.x
    tiles=p_num/cuda.blockDim.x + 1
    shared_p_mx = cuda.shared.array(0,dtype=np.float32)
    shared_p_my = cuda.shared.array(0,dtype=np.float32)
    alpha=(1.5)
    rho=(1.0)
    beta=(1.5*(1.0+alpha))
    for k in range(tid,p_num,threadsInGrid):
        r_k=cuda.device_array((d_num,p_num))
        forcetemp=cuda.device_array((d_num,p_num))
        r_k[0,k]=r[0,k]
        r_k[1,k]=r[1,k]
        forcetemp[0,k]=0.0
        forcetemp[1,k]=0.0

我试图分配的数组是r_k和forcetemp，但是使用上面的代码，我得到以下错误：

TypingError: Failed in nopython mode pipeline (step: nopython frontend) Unknown attribute 'device_array' of type Module()

File "", line 117: def forcecudatiling(p_num,d_num,r,force): for k in range(tid,p_num,threadsInGrid): r_k=cuda.device_array((d_num,p_num))

Answer 1

I am trying to allocate some arrays inside a for loop that is inside a Kernel

你不能那样做。Numba CUDA内核中没有内存分配或数组创建