将非常大的空集分配给 CUDA 时，Numpy/Numba 引发错误

Question

我正在使用 Numba/Numpy 编写 Mandelbrot 集合生成器。其中一项优化是使用 cudatoolkit 通过 Numba 将计算推送到 CUDA。该脚本适用于低分辨率集，但在尝试计算大集时会出错。

import numpy as np
from pylab import imshow, show
import time
from numba import cuda
from numba import *
import matplotlib

def mandel(x, y, max_iters):

  c = complex(x, y)
  z = 0.0j
  for i in range(max_iters):
    z = z*z + c
    if (z.real*z.real + z.imag*z.imag) >= 4:
      return i

  return max_iters

mandel_gpu = cuda.jit(device=True)(mandel)

@cuda.jit
def mandel_kernel(min_x, max_x, min_y, max_y, image, iters):
  height = image.shape[0]
  width = image.shape[1]

  pixel_size_x = (max_x - min_x) / width
  pixel_size_y = (max_y - min_y) / height

  startX, startY = cuda.grid(2)
  gridX = cuda.gridDim.x * cuda.blockDim.x;
  gridY = cuda.gridDim.y * cuda.blockDim.y;

  for x in range(startX, width, gridX):
    real = min_x + x * pixel_size_x
    for y in range(startY, height, gridY):
      imag = min_y + y * pixel_size_y 
      image[y, x] = mandel_gpu(real, imag, iters) / iters


gimage = np.zeros((65536, 65536), dtype = np.uint8)
#gimage = np.zeros((1024, 1024), dtype = np.uint8)
blockdim = (32, 8)
griddim = (32,16)

start = time.time()
d_image = cuda.to_device(gimage)
mandel_kernel[griddim, blockdim](-2.0, 2.0, -2.0, 2.0, d_image, 10000) 
d_image.to_host()

dt = time.time() - start

print ("Mandelbrot created in " + str(dt) + " seconds")

imshow(gimage, 'gray')
show()

#matplotlib.image.imsave("mandel.png", gimage)

超过 46000 x 46000 像素，python 会引发以下错误：

Traceback (most recent call last):
  File "C:\_main\Files\Mandel\mandel_cuda.py", line 46, in <module>
    d_image = cuda.to_device(gimage)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devices.py", line 212, in _require_cuda_context
    return fn(*args, **kws)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\api.py", line 103, in to_device
    to, new = devicearray.auto_device(obj, stream=stream, copy=copy)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 688, in auto_device
    devobj.copy_to_device(obj, stream=stream)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devices.py", line 212, in _require_cuda_context
    return fn(*args, **kws)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 181, in copy_to_device
    sentry_contiguous(self)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 657, in sentry_contiguous
    core = array_core(ary)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 647, in array_core
    return ary[tuple(core_index)]
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devices.py", line 212, in _require_cuda_context
    return fn(*args, **kws)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 479, in __getitem__
    return self._do_getitem(item)
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 493, in _do_getitem
    newdata = self.gpu_data.view(*extents[0])
  File "C:\ProgramData\Anaconda3\lib\site-packages\numba\cuda\cudadrv\driver.py", line 1227, in view
    raise RuntimeError("non-empty slice into empty slice")
RuntimeError: non-empty slice into empty slice

脚本在具有 4GB VRAM 的 1050ti 上运行。对于 46000 x 46000 像素，显存使用量仅为 2.1GB。应该有足够的 VRAM 用于 46000 x 46000 以上的渲染。

Answer 1

似乎是 VRAM 溢出问题。对于渲染的前 30 秒，更多的 VRAM 用于存储空集。初始化时，很快就达到了 4GB 的限制，导致脚本崩溃。