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)

Hi there,

I've been following your issue on pytorch and when I test your gist, after modifying it slightly I get a CUDA_ERROR_INVALID_VALUE
when it calls .copy_to_device().

Here's the modified code running on PyTorch 1.2 (it breaks otherwise, some things are missing and get_context is only available oncudadrv.devices)

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

def devndarray2torch(dev_arr):
    t = torch.empty(size=dev_arr.shape, dtype=torch.float32).cuda()
    ctx = cuda.cudadrv.devices.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 = devndarray2torch(d_arr)

print(tensor)

Would you mind pointing me in the right direction as to where this error might come from?

sure, can you provide the exact stack trace? that may be helpful to debug.

Thanks! Here it is:

Traceback (most recent call last):
  File "tensor_copy.py", line 29, in <module>
    tensor = devndarray2torch(d_arr)
  File "tensor_copy.py", line 24, in devndarray2torch
    tmp_arr.copy_to_device(dev_arr)
  File "C:\Users\Barak\AppData\Local\Programs\Python\Python37\lib\site-packages\numba\cuda\cudadrv\devices.py", line 225, in _require_cuda_context
    return fn(*args, **kws)
  File "C:\Users\Barak\AppData\Local\Programs\Python\Python37\lib\site-packages\numba\cuda\cudadrv\devicearray.py", line 188, in copy_to_device
    _driver.device_to_device(self, ary, self.alloc_size, stream=stream)
  File "C:\Users\Barak\AppData\Local\Programs\Python\Python37\lib\site-packages\numba\cuda\cudadrv\driver.py", line 1932, in device_to_device
    fn(device_pointer(dst), device_pointer(src), size, *varargs)
  File "C:\Users\Barak\AppData\Local\Programs\Python\Python37\lib\site-packages\numba\cuda\cudadrv\driver.py", line 294, in safe_cuda_api_call
    self._check_error(fname, retcode)
  File "C:\Users\Barak\AppData\Local\Programs\Python\Python37\lib\site-packages\numba\cuda\cudadrv\driver.py", line 329, in _check_error
    raise CudaAPIError(retcode, msg)
numba.cuda.cudadrv.driver.CudaAPIError: [1] Call to cuMemcpyDtoD results in CUDA_ERROR_INVALID_VALUE

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?