topic in Intel® Integrated Performance Primitives

Building a custom DLL from the command line for an old IPP distribution

georgiswan — Fri, 09 Mar 2018 17:41:39 GMT

Hello,

I am trying to build a custom IPP DLL with the command line tool ipp_custom_library_tool .

I have compilers_and_libraries_2018 installed and I find the custom library tool in /opt/intel/ipp/tools/custom_library_tool .

I have created a file with the list of needed symbols and they are recognized by the tool.

My problem is that we actually link with and use and old IPP ( composer_xe_2015.7.234 ) but we compile with 2018. The custom_library_tool does not exist (to my knowledge) in that old 2015 distribution and when I use the -c option of custom_library_tool to point to the old distribution, I get an error:

Error: Compilers and Libraries files aren't found

If I dont use -c, the tool completes but warns about many not found symbols that exist in 2015.7 but do not anymore in 2018 (ex: ippiErode3x3_32f_C3IR ).

Is it possible to build a custom library from an old IPP distribution ?

Thanks

Gilbert

georgiswan — Wed, 14 Mar 2018 22:25:46 GMT

I thought I would add some more precise details in hope to get some reply.

I created the list of ipp symbols used in my dylib (Mac OS) with:

nm libMylib.dylib |grep 'U _ipp'| sed -E 's/^[ ]+U _(.+)$/\1/' > /tmp/ippFuncs.txt

Then I went to the custom_library_tool :

cd /opt/intel/compilers_and_libraries_2018.1.126/mac/ipp/tools/custom_library_tool

and I tried to create a custom library with:

./ipp_custom_library_tool  -n libMyCustLib -l /tmp/ippFuncs.txt  -o /tmp/ipp_customs/ -intel64  -macos

I got many warnings for missing symbols:

Warning: ippiCrossCorrSame_NormLevel_32f_C1R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_32f_C3R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_32f_C4R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u32f_C1R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u32f_C3R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u32f_C4R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u_C1RSfs doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u_C3RSfs doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_NormLevel_8u_C4RSfs doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_Norm_32f_C1R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_Norm_32f_C3R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_Norm_32f_C4R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_Norm_8u32f_C1R doesn't match with any Intel(R) IPP function
Warning: ippiCrossCorrSame_Norm_8u32f_C3R doesn't match with any Intel(R) IPP function

This is because my libMylib.dylib is linked against ipp2015.7 but the ipp_custom_library_tool seems to want to use the 2018 IPP libraries to build the custom library.

Any hint on possibly building a custom library for the 2015.7 (or any older IPP libraries) ?

georgiswan — Wed, 14 Mar 2018 22:31:06 GMT

Another question regarding the ipp_custom_library_tool, will such a custom library contain the symbols for all architectures ?

For example, if my library needs ippsSet_16s will I get it for e9, g9, m7, mx, p8, . So that my custom library can be used to build an application that could run on any architecture ?

These two functions was

Chao_Y_Intel — Mon, 26 Mar 2018 03:31:51 GMT

These two functions was deprecated in the old release, and not included in the latest pacakge:

Could you use the following new functions to replace the old one?

ippiCrossCorrSame_NormLevel_32f_C1R ippiCrossCorrNorm_32f_C1R+algType=ippiROISame|ippiNormCoefficient

ippiCrossCorrSame_Norm_32f_C1R ippiCrossCorrNorm_32f_C1R+algType=ippiROISame|ippiNorm

yes, it is right. ipp_custom_library_tool build the library for multiple processors, and it will do the runtime dispatch for your code.

Thanks,

Chao