[MKL-DNN] Extended LRN with reusing via Acquire API

[jczaja]

Changes discussed here are reimplementing LRN mkl-dnn ops to use common Acquire API .

Benefits:

• Easier to modify and maintain code e.g. further modifications for Multi-threading and mkl-dnn 1.0 will be easier
• Performance improvement due to Reusing introduced: Google Net v1 (model with LRN) inference via CAPI with bs=1 is ~9% faster on AVX512 platform e.g. SKX (8180)

[Sand3r-]

@jczaja Thank you so much for the contribution and the speedup it brings on GoogleNet v1!

[jczaja]

@luotao1 Could you please review this PR? Again your approval is needed to have CI passed

[luotao1]

LGTM

[luotao1]

I test Google Net v1 (model with LRN) inference via CAPI with bs=1 on AVX2 platform e.g. E5-2650 v4.

./test_analyzer_image_classification --gtest_filter=Analyzer_resnet50.profile_mkldnn --batch_size=1 --warmup --repeat=100 --paddle_num_threads=4 --infer_model=googlenetfluid/ --profile

lrn from 241ms to 208ms, speedup 13%

• before

Event              Calls       Total       Min.        Max.        Ave.        Ratio.
thread0::conv2d    5700        1881.6      0.085055    3.01909     0.330106    0.787164
**thread0::lrn       200         241.991     0.716039    2.08498     1.20995     0.101236**
thread0::pool2d    1400        178.921     0.0452      0.266267    0.127801    0.0748511
thread0::concat    900         81.7893     0.04769     0.496502    0.090877    0.0342163
thread0::fc        100         4.40745     0.042422    0.046222    0.0440745   0.00184385
thread0::fetch     100         1.21941     0.010564    0.013756    0.0121941   0.000510137
thread0::feed      100         0.427123    0.003667    0.005537    0.00427123  0.000178686

• after

Event              Calls       Total       Min.        Max.        Ave.        Ratio.
thread0::conv2d    5700        1872.48     0.086678    3.06727     0.328505    0.799478
**thread0::lrn       200         208.1       0.540581    1.58649     1.0405      0.0888512**
thread0::pool2d    1400        178.047     0.046527    0.557823    0.127176    0.0760194
thread0::concat    900         77.3396     0.047206    0.470094    0.0859329   0.0330212
thread0::fc        100         4.48958     0.042211    0.063319    0.0448958   0.00191689
thread0::fetch     100         1.24835     0.011442    0.016381    0.0124835   0.000532998
thread0::feed      100         0.421286    0.003702    0.006086    0.00421286  0.000179874

[luotao1]

~9% faster

Is this Op-level speedup or Model-level speedup? @jczaja

[jczaja]

@luotao1 Reported speedup is on model level (SKX 8180 AVX512). I stored only partial log:

before:
sample latency: 26.7129 fps: 37.43

...
thread0::lrn: 2524.55
...

After:
sample latency: 24.4794 fps: 40.8506

....
thread0::lrn: 1194.76
......