Enable float requantization for avgpool/gavgpool ops. (#37037)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37037
For avgpool and gavgpool change requantization scheme:
Similar to conv and linear now convert the accumulated int32
values to float, apply requantization scale which includes the averaging
multiplier.
Conver the resulting float value to int32.
Add output_zero_point.
Benchmark numbers compared to baseline:
% speedup on pixel XL.
------------------------------
| | aarch32 | aarch64|
|avgpool | .4 | 13.6 |
|gavgpool | -2.6% | 3.5% |
-------------------------------
Test Plan:
Tested via q8avgpool-test, q8gavgpool-test, average-pooling-test and
global-average-pooling-test in PT QNNPACK.
Also via integated test_quantized.py.
python test/quantization/test_quantized.py
Imported from OSS
Differential Revision: D21168981
fbshipit-source-id: 9060324304603ca7fd380c788a87b01a6d586c5c
