[MKL-DNN] Fully Connected

[Sand3r-]

Introduces Fully Connected operator. It has a comparable performance to the reference version, although it is necessary to be merged, so that the base for the int8 implementation and its performance gains is established.

Update 05.04.2019
Compared to a reference FC, the MKL-DNN version provides:

• 51% speedup on SKX 6148
• 35% on CLX 8280

[tensor-tang]

Could you tell us the difference or improvment over the current fc implementation.

Does it just for infer?

If it is performance improvement, could you give some brief data of before and after?

[hshen14]

@Sand3r- any update on this?

[Sand3r-]

@Sand3r- any update on this?

@hshen14 As mentioned in an internal mail, Paddle's CI VM yields flawed output for some of the convolutions and fc. This results in insufficient accuracy to have test_analyzer_resnet50 passed. This isn't the case for all of our local machines, where everything works as intended.

[CLAassistant]

All committers have signed the CLA.

[kbinias]

The problem with the FC operator is that on CI MKL library is using AVX2 instruction set, while it should use AVX512. We reproduced the problem locally by enforcing AVX2 in MKL using MKL_CBWR=AVX2 environment setting.
@luotao1 We need more information about CI running environment. Could you send us info about current CI configuration (results of commands: set and env) ?

[luotao1]

We need more information about CI running environment.

All the CI machines have AVX512, see #15032 (comment) and #15032 (comment)

Could you send us info about current CI configuration (results of commands: set and env)

Do you mean MKL_CBWR? We don't set this environment. And you can echo $MKL_CBWR yourself. Maybe you can do echo in cmake_gen

Paddle/paddle/scripts/paddle_build.sh

Line 55 in 41b8cf0

function cmake_gen() {

[kbinias]

@luotao1 I noticed differences between the downloaded versions of the MKLML libraries.
CI download: Glibc225_vsErf_mklml_lnx_2019.0.1.20181227
Public PaddlePaddle: mklml_lnx_2019.0.1.20181227
These libraries do not match at the binary level.
I am not sure if MKLML from CI support AVX512.

[Sand3r-]

Done.

[Sand3r-]

Didn't expect the happiest day of my life would come today.