Summary of performance optimization：02.Recognize_digits (WIP)

[chengduoZH]

Config and Env

• export OMP_NUM_THREADS=1
• export MKL_NUM_THREADS=1
• Env: docker, CPU, profiling machine.
• Compling mode: Release
• By code: https://github.com/chengduoZH/benchmark/tree/profile/02_refine
  • commit: b594b0b

CPU

Performance analysis tools: cpu_profiling.md

optimized place:

• Refine enforce.h
• Refine enforce.h
• elementwise_add_op.h
• elementwise_add_op.h and im2col.cc

Current performance analysis results (time: s )

pass	v2	fluid
0	34.803	30.529
1	34.8	29.695
2	34.522	31.125
3	34.981	30.852
4	34.755	30.229
avg	34.7722	30.486

GPU

optimized place:

• accuracy_op.cu
• feed_data(Sync->Async)
• fetch_data

Current performance analysis results (time: s )

pass	v2	fluid
0	2.365	2.967
1	2.333	2.929
2	2.266	2.98
3	2.217	2.93
4	2.255	2.963
avg	2.2872	2.9538