Speed up equalize transform: use bincount instead of histc

[NicolasHug]

bincount is faster than histc on CPU, so this should speed up (~2X) the equalize transforms.

This is sort of a follow-up to #3334 and #3173

On GPU:

In [16]: img = torch.randint(0, 256, size=(460 * 680,)).to('cuda')

In [17]: %timeit torch.bincount(img, minlength=256)
1.52 ms ± 25.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [18]: %timeit torch.histc(img.to(torch.float32), bins=256, min=0, max=255)
127 µs ± 1.07 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

on CPU (different machine):

img = torch.randint(0, 256, size=(460 * 680,))

%timeit torch.histc(img.to(torch.float32), bins=256, min=0, max=255)
696 µs ± 10.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%timeit torch.bincount(img, minlength=256)
289 µs ± 1.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

[codecov]

Codecov Report

Merging #3493 (c294ec0) into master (414427d) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #3493   +/-   ##
=======================================
  Coverage   78.75%   78.75%           
=======================================
  Files         105      105           
  Lines        9748     9750    +2     
  Branches     1565     1566    +1     
=======================================
+ Hits         7677     7679    +2     
  Misses       1581     1581           
  Partials      490      490           

Impacted Files	Coverage Δ
torchvision/transforms/functional_tensor.py	79.92% <100.00%> (+0.07%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 414427d...c294ec0. Read the comment docs.

[fmassa]

There is a huge difference on CUDA. I think it would be worth opening an issue in PyTorch to point out that the implementation of bincount is suboptimal?

Also, it would be good to check the number of CUDA again, but this time using torch.cuda.synchronize() to get more accurate timings

[datumbox]

Do you need to clip values between 0 and 255 to keep things consistent between cuda/cpu?

[NicolasHug]

Passing minlength=256 ensures that bincount will count the number of values for all values in [0, 255], so histc and bincount would be equivalent.

In the histc call we have bins=256, min=0, max=255, which IIUC assumes that the image is within [0, 255] already, is this not the case?

[NicolasHug]

I created a new "Improvement" label, in accordance with #3351 (comment)

[fmassa]

Thanks!