convert numba cuda array to pytorch tensor

numba_to_pytorch.py

from numba import cuda
import ctypes
import numpy as np
import torch

def devndarray2torch(dev_arr):
    t = torch.empty(size=dev_arr.shape, dtype=dtyp).cuda()
    ctx = cuda.cudadrv.driver.driver.get_context()
	
    # constant value of #bytes in case of float32 = 4
    mp = cuda.cudadrv.driver.MemoryPointer(ctx, ctypes.c_ulong(t.data_ptr()), t.numel()*4)
    tmp_arr = cuda.cudadrv.devicearray.DeviceNDArray(t.size(), [i*4 for i in t.stride()], np.dtype('float32'), 
                                            gpu_data=mp, stream=torch.cuda.current_stream().cuda_stream)
											
    # To verify whether the data pointer is same or not.
    # print(tmp_arr.__cuda_array_interface__)
    # print(dev_arr.__cuda_array_interface__)
    tmp_arr.copy_to_device(dev_arr)
    return t

d_arr = cuda.to_device(np.array([[10,20,30],[40,50,60.0]], dtype=np.float32))
tensor = devndarray2tensor(d_arr)
print(tensor)

your code is running as it is, in colab. here is the running notebook: https://colab.research.google.com/drive/1R9V8qNo2qj-yUbuNnq8IaEkoB1pw-oUP

you are running on windows, I haven't tested on windows machines.

what is the output of these lines??

print(tmp_arr.__cuda_array_interface__)
print(dev_arr.__cuda_array_interface__)

Huh interesting, I'm running on Windows with PyTorch 1.2.0, numba 0.45.1, Python 3.7.4

The output is:

{'shape': (2, 3), 'strides': (12, 4), 'data': (37748736, False), 'typestr': '<f4', 'version': 1}
{'shape': (2, 3), 'strides': (12, 4), 'data': (21510488064, False), 'typestr': '<f4', 'version': 1}

I also encountered the same problem running on Windows.

@AK-ayush, @BarakChamo
any ideas?