Created by: peterzhang2029

text_classification的Profiling结果，源码https://github.com/PaddlePaddle/models/tree/develop/fluid/text_classification：

训练配置：

batch_size: 4
num_of_batch: 500
dict_size: 89528

Profiling Report 指标：

Calls: 表示总共调用的数量
Total：表示总共运行的时长
Min：表示所有的调用中最短的一次调用时长
Max：表示所有调用中最长的一次调用时长
Ave：表示所有调用的平均调用时长

训练500个batch的Profiling Report 结果：

The dictionary size is : 89528
Total time: 16.033737

------------------------->     Profiling Report     <-------------------------

Place: CPU
Time unit: ms
Sorted by total time in descending order in the same thread

Event                            Calls       Total       Min.        Max.        Ave.
thread0::sgd                     3493        4830.48     0.009502    15.5302     1.3829
thread0::lookup_table_grad       499         4775.31     9.31438     12.5104     9.56976
thread0::sequence_conv_grad      499         1683.32     1.04954     15.5015     3.3734
thread0::sequence_conv           499         1048.58     0.698105    14.0347     2.10136
thread0::elementwise_add_grad    1497        210.856     0.013004    1.14577     0.140852
thread0::elementwise_add         1497        122.64      0.00811     0.577024    0.0819238
thread0::tanh                    499         117.027     0.090373    0.623305    0.234523
thread0::sequence_pool           499         104.869     0.100548    0.479671    0.210159
thread0::lookup_table            499         67.5452     0.059546    0.555258    0.135361
thread0::tanh_grad               499         54.0636     0.044505    0.316446    0.108344
thread0::sequence_pool_grad      499         36.3875     0.039143    0.18097     0.0729209
thread0::mul_grad                998         35.8055     0.023166    0.237668    0.0358772
thread0::mul                     998         26.0178     0.011744    0.074651    0.02607
thread0::cast                    1996        19.4615     0.0061      0.158941    0.00975027
thread0::sum                     998         15.1658     0.009326    0.09614     0.0151962
thread0::top_k                   499         14.7059     0.025781    0.0949      0.0294708
thread0::accuracy                499         10.6152     0.018964    0.059349    0.021273
thread0::softmax                 499         8.14185     0.014948    0.044151    0.0163163
thread0::cross_entropy_grad      499         7.07577     0.0132      0.023645    0.0141799
thread0::cross_entropy           499         6.96024     0.012991    0.02396     0.0139484
thread0::feed                    998         6.73899     0.004097    0.035801    0.00675249
thread0::softmax_grad            499         6.43104     0.011922    0.032686    0.0128879
thread0::fetch                   1497        5.35905     0.001853    0.019403    0.00357986
thread0::elementwise_div         499         5.16685     0.008804    0.073607    0.0103544
thread0::mean_grad               499         4.91204     0.009089    0.044894    0.00984377
thread0::mean                    499         3.64095     0.006551    0.041162    0.00729649
thread0::fill_constant           499         3.02958     0.005365    0.031708    0.00607131

训练1个pass的Profiling Report 结果：

The dictionary size is : 89528
Total time: 246.533985

------------------------->     Profiling Report     <-------------------------

Place: CPU
Time unit: ms
Sorted by total time in descending order in the same thread

Event                            Calls       Total       Min.        Max.        Ave.
thread0::lookup_table_grad       6248        85039.4     12.9076     32.3335     13.6107
thread0::sgd                     43736       68442.7     0.009991    35.3864     1.5649
thread0::sequence_conv_grad      6248        24102.5     0.776669    45.2805     3.85763
thread0::sequence_conv           6248        16146       0.47397     29.5967     2.58418
thread0::elementwise_add_grad    18744       2645.15     0.013075    1.76686     0.14112
thread0::elementwise_add         18744       1515.66     0.007877    1.35418     0.0808611
thread0::tanh                    6248        1412.74     0.078598    2.58319     0.226111
thread0::sequence_pool           6248        1296.17     0.084166    1.01645     0.207454
thread0::lookup_table            6248        996.414     0.061383    2.69139     0.159477
thread0::tanh_grad               6248        686.66      0.040493    3.34762     0.109901
thread0::sequence_pool_grad      6248        460.941     0.03908     0.767207    0.0737741
thread0::mul_grad                12496       455.801     0.023004    1.12555     0.0364757
thread0::mul                     12496       323.891     0.011963    0.718241    0.0259196
thread0::cast                    24992       319.577     0.006238    24.1128     0.0127872
thread0::top_k                   6248        238.347     0.026201    12.0761     0.0381477
thread0::sum                     12496       227.388     0.009395    16.0632     0.0181969
thread0::accuracy                6248        171.487     0.020024    15.2487     0.0274467
thread0::softmax                 6248        105.258     0.015404    0.702274    0.0168467
thread0::feed                    12496       94.2271     0.003959    4.75252     0.00754058
thread0::cross_entropy_grad      6248        90.4667     0.013344    0.148185    0.0144793
thread0::cross_entropy           6248        88.0551     0.012642    0.698313    0.0140933
thread0::softmax_grad            6248        84.509      0.012117    0.699262    0.0135258
thread0::fetch                   18744       76.1771     0.001746    0.098168    0.00406408
thread0::elementwise_div         6248        69.8883     0.009183    1.36969     0.0111857
thread0::mean_grad               6248        62.9747     0.008981    0.056408    0.0100792
thread0::mean                    6248        45.4185     0.006505    0.045735    0.00726928
thread0::fill_constant           6248        38.8432     0.005422    0.637213    0.0062169

结论： 其中 lookup_table_grad 时间占用明显过长