Implementation of numpy function .view(uint8)

1

I’m looking to get the byte-decomposed values of Float32 tensor on GPU. Would anyone know how to do this? In numpy, this would be:

float_nums = [0.16474093, -0.06143471,  0.09829687]
float_arr = np.array(float_nums, dtype=np.float32)
uint8_arr = float_nums.view(np.uint8)
# uint8_arr is now 4 times the length float_arr 
#   and I can perform various bit operations, 
#   like bit masking the float mantissa, etc.

PyTorch’s bit operators (^/&/etc.) require ByteTensors - this is my underlying use-case.

Unfortunately, the .byte() function doesn’t re-interpret the same binary bytes, instead uses the floating point value to round and wraps around to nearest byte value. For example, float_arr.byte(), in the above snippet, would result in an all-zeroes tensor.

I could interconvert to numpy and back, but these Tensors are on GPU and shuffling all these tensors back and forth to main-memory is wasteful and begins to dominate my inference and training time.

Any suggestions on how to achieve this same functionality would be appreciated.

Thanks.

2

Hi,

I’m afraid there is no way to do this at the moment.
Would a reinterpret function would be an interesting feature @smth ?

3

Thanks for the quick response. If this is the case, I might look into writing a CUDA extension for this little bit manipulation and the feasibility of doing the re-interpreting inside a custom kernel.

If so desired, I could look into extending this to a more general reinterpret bytes to submit upstream.

4

@albanD yes a reinterpret function would be nice, though not sure how far the rabbit-hole goes in implementing it :slight_smile:

5

Hi,

Yes a simple cuda extension with the new cpp extension should be very simple to do. And the easiest way to do this.

@smth There is definitly no api that lets you do that easily. I’ll take a look when I have a bit of free time and I’m done with the hook thing.