c_api内存池的几个bug

[qiqikit]

1. 修了一下c_api allocator参数的几个bug
2. 做了个内存池最大使用内存限制功能，功能比较粗糙
3. 提升了 allocator 的 emptyCache() 方法，小内存的设备需要

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov-commenter]

Codecov Report

Merging #3817 (9b917d5) into master (f79073c) will increase coverage by 1.53%. The diff coverage is 17.36%.

@@            Coverage Diff             @@
##           master    #3817      +/-   ##
==========================================
+ Coverage   91.75%   93.29%   +1.53%     
==========================================
  Files         674      674              
  Lines      171538   172106     +568     
==========================================
+ Hits       157402   160564    +3162     
+ Misses      14136    11542    -2594     

Impacted Files	Coverage Δ
src/allocator.h	87.50% <ø> (ø)
src/c_api.cpp	31.72% <2.87%> (-5.19%)	:arrow_down:
src/allocator.cpp	75.75% <56.86%> (-0.13%)	:arrow_down:
src/cpu.cpp	63.81% <0.00%> (-1.71%)	:arrow_down:
src/net.cpp	65.37% <0.00%> (-0.12%)	:arrow_down:
src/layer/vulkan/pooling_vulkan.cpp	98.42% <0.00%> (+<0.01%)	:arrow_up:
src/mat.h	89.87% <0.00%> (+0.05%)	:arrow_up:
src/layer/x86/convolution_3x3_pack8to4_int8.h	99.43% <0.00%> (+0.09%)	:arrow_up:
src/layer/x86/convolution_sgemm_pack8to1_int8.h	99.10% <0.00%> (+0.09%)	:arrow_up:
src/layer/x86/convolution_3x3_pack8to1_int8.h	99.06% <0.00%> (+0.12%)	:arrow_up:
... and 44 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update c2fb93b...9b917d5. Read the comment docs.

[lgtm-com[bot]]

This pull request introduces 2 alerts when merging 9b917d55c1c5458f2e08a79889e9c078f5c93deb into c2fb93b6ff99045dd76aae2d41218a15df189247 - view on LGTM.com

new alerts:

• 2 for Virtual call from constructor or destructor

[nihui]

emptyCache 看起来就是原来的 clear 嘛？所以直接把 clear 提升到基类也ok？ gpu 那边的 VkAllocator 有 clear 方法

[qiqikit]

对的，一些内存比较少的国产环境中顺序推理多张图片时，识别完一张图片时，调用一下emptyCache() 可以为后续图像算法之类争取到更多一点儿内存…… 

叶子

 

------------------ 原始邮件 ------------------ 发件人: @.>; 发送时间: 2022年5月21日(星期六) 晚上9:36 收件人: @.>; 抄送: @.>; @.>; 主题: Re: [Tencent/ncnn] c_api内存池的几个bug (PR #3817)

emptyCache 看起来就是原来的 clear 嘛？所以直接把 clear 提升到基类也ok？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[qiqikit]

我试了直接提升clear() 好像改动会更大一些，vk也有重名的，就单独加了名字。要是 cpu 和 vk 能统一内存池接口应该是最好的方案。没有写内存池的经验，逻辑可能欠妥，看着改就行。另外 clear() 不如 emptyCache()描述形象， pytorch里用的是后者。  

叶子

 

[nihui]

move part of this pr to https://github.com/Tencent/ncnn/pull/4360