MKLDNN inference result on Mac

[luotao1]

Use run_mkldnn_infer.sh to test MKLDNN/MKLML inference of VGG19/ResNet50/GoogleNet on Mac machine.

• docker environment: paddlepaddle/paddle:latest
• machine: Mac Pro A1502, CPU: 2.7 GHz Intel Core i5, Memory: 8 GB 1867 MHz DDR3

[kuke]

GoogleNet (images/sec)

BatchSize	1	2	4	8
MKLML	11.33	17.39	25.84	23.59
MKL-DNN	37.43	39.42	38.50	29.10

ResNet-50 (images/sec)

BatchSize	1	2	4	8
MKLML	3.82	5.33	6.43	5.27
MKL-DNN	13.70	13.91	14.61	14.40

VGG-19 (images/sec)

BatchSize	1	2	4	8
MKLML	1.75	2.37	2.62	2.37
MKL-DNN	3.23	3.20	3.75	3.81

[luotao1]

It seems that MKLDNN has speedup on Mac:

• GoogleNet: 1.2x~3.3x
• ResNet-50: 2.3x~3.6x
• VGG-19: 1.3x~1.8x

[NHZlX]

Is there any speed test without mklml and mklDNN.

[luotao1]

Do you mean openblas? You can see https://github.com/tensor-tang/Paddle/blob/ee0d365a3a21ab8881ad88b252a5d117b87bc726/benchmark/IntelOptimizedPaddle.md . It is tested on server.