```cupyx.scipy.ndimage.zoom``` returns incorrect values

Description

Hi, I am currently working on a project that I need to scale a 4-d uint8 array by a factor in each axis. I first tried the scipy.ndimage.zoom, but it was slow. I later found out that cupyx supports this function so I changed to cupyx.

However, the array returned is not the same as the scipy one. Around 0.0044% of the values are not the same. My guess is that cupyx set some of the values to 0 directly (maybe due to error, I am not sure).

This is the same cross-section of the two arrays: Cupyx

Scipy

You can see there are differences on the top.

To Reproduce

import numpy as np
import cupy as cp
import cupyx.scipy.ndimage
import scipy.ndimage

# Inputs:
# orginal_array: 4d uint8 numpy shape: (208, 208, 224, 21) 
# space_directions: [0.97212501, 1.37501857, 1.15326139, 1]

data_array_cupyx = cupyx.scipy.ndimage.zoom(cp.array(original_array, dtype=cp.uint8), space_directions, output=cp.uint8).get().astype(np.uint8)
data_array_scipy = scipy.ndimage.zoom(original_array, space_directions, output=np.uint8)

diff_array = data_array_cupyx != data_array_scipy
print(np.prod(data_array_cupyx.shape))
# >>> 313008696
print(np.sum(diff_array))
# >>> 13921

Installation

Source (pip install cupy)

Environment

OS                           : Windows-10-10.0.19041-SP0
Python Version               : 3.8.3
CuPy Version                 : 10.2.0
CuPy Platform                : NVIDIA CUDA
NumPy Version                : 1.22.3
SciPy Version                : 1.8.0
Cython Build Version         : 0.29.28
Cython Runtime Version       : None
CUDA Root                    : C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2
nvcc PATH                    : C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\bin\nvcc.EXE
CUDA Build Version           : 11020
CUDA Driver Version          : 11060
CUDA Runtime Version         : 11020
cuBLAS Version               : (available)
cuFFT Version                : 10401
cuRAND Version               : 10203
cuSOLVER Version             : (11, 1, 0)
cuSPARSE Version             : (available)
NVRTC Version                : (11, 2)
Thrust Version               : 101000
CUB Build Version            : 101000
Jitify Build Version         : <unknown>
cuDNN Build Version          : 8101
cuDNN Version                : 8101
NCCL Build Version           : None
NCCL Runtime Version         : None
cuTENSOR Version             : None
cuSPARSELt Build Version     : None
Device 0 Name                : NVIDIA GeForce GTX 1060 3GB
Device 0 Compute Capability  : 61
Device 0 PCI Bus ID          : 0000:01:00.0

Additional Information

No response

[PCRedHot]

Sure, here is a subset of the data. part.zip

Script

import numpy as np
import cupy as cp
import cupyx.scipy.ndimage
import scipy.ndimage

numpy_file_path = 'path_to/part.npy'

space_directions = [0.97212501, 1.37501857, 1.15326139]
original_array = np.load(numpy_file_path)

# original_array shape: (208, 208, 224) uint8

data_array_cupyx = cupyx.scipy.ndimage.zoom(cp.array(original_array, dtype=cp.uint8), space_directions, output=cp.uint8).get()
data_array_scipy = scipy.ndimage.zoom(np.array(original_array, dtype=np.uint8), space_directions, output=np.uint8)

# new array shape: (202, 286, 258)

# Differences
diff_array = data_array_cupyx != data_array_scipy
print('Different Values:', np.sum(diff_array), 'out of', np.prod(data_array_cupyx.shape))
# >>> Different Values: 568 out of 14905176

[grlee77]

Okay, that is good that the floating point case followed by casting is consistent. That shouldn’t be necessary as ideally output=cp.uint8 should work the same way as in scipy. If you are able to make a small subset of the data where you see the issue available for us to test with, that could help look into it a bit more.