Verify MKLDNN benchmark using 0.11.0 image

[luotao1]

IntelOptimizedPaddle.md use paddle:latest and paddle:latest-openblas, we should verify it using paddle:0.11.0 and paddle:0.11.0-openblas.

[luotao1]

Training: before->0.11.0

• MKLML and MKL-DNN, the benchmark is almost the same.
• OpenBlas: Not finished yet.

Compared Result:

• VGG-19

BatchSize	64	128	256
OpenBLAS	7.80->7.57	9.00->8.93	10.80->10.60
MKLML	12.12->12.59	13.70->14.36	16.18->17.11
MKL-DNN	28.46->28.54	29.83->29.85	30.44->30.62

• ResNet-50

BatchSize	64	128	256
OpenBLAS	25.22->22.81	25.68->23.83	27.12->25.85
MKLML	32.52->32.27	31.89->32.14	33.12->33.28
MKL-DNN	81.69->81.16	82.35->83.84	84.08->85.28

• GoogLeNet

BatchSize	64	128	256
OpenBLAS	89.52->88.49	96.97->92.99	108.25->107.22
MKLML	128.46->129.37	137.89->136.49	158.63->157.50
MKL-DNN	250.46->253.73	264.83->254.01	269.50->270.79

[luotao1]

Inference: before->0.11.0

• MKLML and MKL-DNN: the benchmark is almost the same except GoogleNet-MKLDNN.
• OpenBlas: Not finished yet.

Compared Result:

• VGG-19

BatchSize	1	2	4	8	16
OpenBLAS	1.07->1.10	1.08->1.96	1.06->3.62	0.88->3.63	0.65->2.25
MKLML	5.58->5.51	9.80->9.52	15.15->15.81	21.21->21.38	28.67->32.56
MKL-DNN	75.07->71.23	88.64->88.52	82.58->89.76	92.29-> 92.15	96.75->97.04

• ResNet-50

BatchSize	1	2	4	8	16
OpenBLAS	3.35->3.31	3.19->6.72	3.09->11.59	2.55->13.17	1.96->9.27
MKLML	6.33->6.28	12.02->11.85	22.88->21.67	40.53->40.23	63.09->62.47
MKL-DNN	107.83->108.94	148.84->151.12	177.78->188.24	189.35->181.82	217.69->224.56

• GoogLeNet

BatchSize	1	2	4	8	16
OpenBLAS	12.04->12.06	11.31->23.56	10.00->34.48	9.07->36.45	4.34->23.12
MKLML	22.74->22.04	41.56->40.82	81.22->81.89	133.47->133.75	210.53->190.48
MKL-DNN	175.10->238.81	272.92->309.93	450.70->269.47	512.00->341.33	600.94->355.56

• AlexNet

BatchSize	1	2	4	8	16
OpenBLAS	3.53	6.23	15.04	26.06	31.62
MKLML	21.32	36.55	73.06	131.15	192.77
MKL-DNN	442.91	656.41	719.10	847.68	850.51

[tensor-tang]

I've tested on 0.11.0 and I think below data need double check:

• Training GoogLeNet

BatchSize	128
MKL-DNN	264.83->254.01

• Inference GoogLeNet

BatchSize	8	16
MKL-DNN	512.00->341.33	600.94->355.56

My data did not reduce that much.

[luotao1]

My data did not reduce that much.

Yes, The double check is ok. Thus, both results of mklml and mkl-dnn are ok.

• Training GoogLeNet

BatchSize	128
MKL-DNN	264.83->265.41

• Inference GoogLeNet

BatchSize	4	8	16
MKL-DNN	450.70->415.58	512.00->524.59	600.94->603.77

BatchSize	16
MKL-DNN	210.53->204.47

[luotao1]

openblas results are all error due to not set env.

BatchSize	64
VGG-19	7.80->0.23
ResNet-50	25.22->0.69
GoogleNet-v1	89.52->1.53

[luotao1]

openblas (Training) results are almost same after #7104

• AlexNet (Training) : need double check

BatchSize	64	128	256
OpenBLAS	45.62->50.80	72.79->73.38	107.22->110.81

[tensor-tang]

I checked these data, LGTM.