Hi NVCOMP Team,
I’d like to report a bug that seems straightforward yet unexpected. Since the NVCOMP GitHub repository appears less active, I’m reaching out here instead.
I am running the exact code provided in the NVCOMP Python API documentation. However, I’ve noticed that the datatype in nvcomp.Array is always set to uint8, which is incorrect for this use case. The issue arises directly from the nvcomp.Array, and it disrupts subsequent operations.
Code:
import numpy as np
import cupy as cp
from nvidia import nvcomp
print("nvcomp version:", nvcomp.__version__)
print("nvcomp cuda version:", nvcomp.__cuda_version__)
ascending = np.arange(0, 4096, dtype=np.int32)
nvarr_h = nvcomp.as_array(ascending)
print(ascending.__array_interface__)
print(nvarr_h.__array_interface__)
print(nvarr_h.__cuda_array_interface__)
print(nvarr_h.buffer_size)
print(nvarr_h.buffer_kind)
print(nvarr_h.ndim)
print(nvarr_h.dtype)
print(nvarr_h.shape)
print(nvarr_h.strides)
print(nvarr_h.item_size)
print(nvarr_h.size)
Output:
nvcomp version: 4.1.0
nvcomp cuda version: 12050
{'data': (94855846852368, False), 'strides': None, 'descr': [('', '<i4')], 'typestr': '<i4', 'shape': (4096,), 'version': 3}
{'shape': (4096,), 'strides': None, 'typestr': '|u1', 'data': (94855846852368, False), 'version': 3}
{'shape': (4096,), 'strides': None, 'typestr': '|u1', 'data': (94855846852368, False), 'version': 3, 'stream': 1}
4096
ArrayBufferKind.STRIDED_HOST
1
uint8
(4096,)
(1,)
1
4096
This behavior seems too basic to be a widespread bug, so I’m wondering if it’s specific to my environment. How can I ensure the correct datatype is used?