当前位置：网站首页>Amway! How to provide high-quality issue? That's what Xueba wrote!

Amway! How to provide high-quality issue? That's what Xueba wrote!

2022-06-26 23:56:00 【Shengsi mindspire】

introduction

This tweet is for Shengsi MindSpore The quality of the community ISSUE, Amway gives you reference and learning , Click below “ Read the original ” You can jump gitee original text , more ISSUE Please see the link for details ：

https://gitee.com/mindspore/mindspore/issues

summary

UB The fusion process is as follows ：

adopt Pass Match the operator to be fused , take fusion id Set to node Properties of
Fuse the matched small operators into FusionOp

Initialize fusion information , Including determining fusion scope Included nodes 、 Determine the input and output nodes, etc
Check each fusion scope Whether it will form a ring （ call CheckCircle function ）
Compile the fusion operator , Compilation failed fusion scope Do not do UB The fusion
For each fusion scope, First, check whether the ring is formed （ call CheckCircle function ）, If it does not form a ring, create a fusion operator and replace it on the original graph

According to the actual measurement ,CheckCircle Function in UB The time-consuming proportion of fusion is 86%, So you need to CheckCircle To optimize .

UB There is no ring on the graph before merging , But after fusion, a ring may be formed . As shown in the figure below ：

hold A、B and C Nodes are merged into E after ,E and D Formed a ring .

Optimization point

Optimization point 1

From the above processing flow , There are two ring forming inspections ：

Check before compiling the operator for the first time , The goal is to prevent compilation from taking too long , Eliminate unnecessary operator compilation in advance through looping check , Lifting performance . From the test data , At this time CheckCircle() Function called 1719 Time , among 51 Secondary cyclization , The ring forming ratio is only 3%, So this check can delete .
The second time is at every FusionOp Before replacing , The check at this time cannot be deleted .

An optimization method ： Delete the first looping inspection .

Optimization point 2

Current looping inspection , from fusion scope Input node of （ Image below C and D） Start traversing its predecessor nodes （ The figure below C and D The precursors of all are B）.
As shown in the figure below , First input from C To traverse the C->B-A, And then input D To traverse the D->B->A, And that leads to B and A Repeatedly visited .

So I'm checking fusion scope when , You can record the visited nodes , Avoid repeated visits .

Optimization point 3

At present UB Integrated pattern Basically, they are single input and single output structures ：

According to this feature, some optimization can be done , Avoid unnecessary checks .

If all inputs are passed to fusion scope The first entry node in the （ The figure below 2 In a scene C node ）, Then the fusion will not form a ring , Looping check can be skipped .

If the input is passed to different entry nodes （ In the left C and D） Or not the first entry node （fusion scope The internal nodes are topologically ordered , Judging right graph from topological order C Not the first entry node ） It may form a ring .

in consideration of UB The fusion pattern Features and implementation complexity of , Make the following judgments ： If all inputs are passed to fusion scope The first node in the , Skip the looping inspection .

Because topology sorting has been done , So the first node must be the entry node .
This condition is too strict , Some other scenarios do not need to be checked , Consider other scenarios in UB The proportion of integration is very small , Temporary does not support .

It is found in the actual measurement that fusion scope Accounted for as ：1616/1719 = 94%.

Optimization point 4

If all outputs are generated by fusion scope The last exit node in the （ On the left of the figure below B And the one on the right C node ） produce , Then the fusion will not form a ring , Looping check can be skipped .

If all outputs are from different exit nodes （ In the left B and D） Produced or produced by intermediate nodes （ On the right B） It may form a ring .

in consideration of UB The fusion pattern Features and implementation complexity of , Make the following judgments ： If all outputs are generated by fusion scope The last exit node in the , Skip the looping inspection .

It is found in the actual measurement that fusion scope Accounted for as ：1617/1719 = 94%, And optimization points 3 The proportion is the same .

Verification effect

Time units in the following table ： second .