Fix the QuantizedAVX2 build issue

[lly-zero-one]

The QuantizedAVx2 does not support the int32 type. We switch to use at::quantize_vec function instead.

[jamesr66a]

Any perf checks?

[facebook-github-bot]

@llyfacebook has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[lly-zero-one]

Any perf checks?

Yes, I found it is even faster than the original one. (For bilinear2d interpolate case.)

**** torch.qint8 *****
/home/lingyiliu/local/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py:2494: UserWarning: Default upsampling behavior when mode=bilinear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.
  "See the documentation of nn.Upsample for details.".format(mode))
time/iter ms (float)    time/iter ms (quant)    quant/float
2.776181697845459       0.03426074981689453     0.012340960911702442
GB/s float      GB/s quant
1.165940976598206       23.619331285066114
**** torch.quint8 *****
time/iter ms (float)    time/iter ms (quant)    quant/float
2.8192615509033203      0.032978057861328125    0.011697409859245451
GB/s float      GB/s quant
1.1481247630121 24.538012620474262
**** torch.qint32 *****
time/iter ms (float)    time/iter ms (quant)    quant/float
2.0195364952087402      1.0663747787475586      0.5280294667996766
GB/s float      GB/s quant
1.602775690203824       3.0353906192358084

[hx89]

I got some undeclared identifier errors when build locally:
stderr: caffe2/aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp:654:523: error: use of undeclared identifier 'area_pixel_compute_s
cale'
caffe2/aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp:654:553: error: expected '(' for function-style cast or type construction
caffe2/aten/src/ATen/native/quantized/cpu/kernels/QuantizedOpKernels.cpp:654:621: error: use of undeclared identifier 'area_pixel_compute_scale'

[jamesr66a]

@llyfacebook yeah, QuantizeAVX2 operates on vectors of 32, but you were feeding it vectors of 8, so it was just running scalar code :p

[jamesr66a]

@llyfacebook I'd expect another 2x speedup if you switch to doing the float operations 4-wide and using QuantizeAVX2 again

[hx89]

LGTM! My local build pass after rebasing.

[facebook-github-bot]

@llyfacebook is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@llyfacebook merged this pull request in 428204d.