https://community.intel.com/t5/Software-Tuning-Performance/choice-of-profiling-characteristic/m-p/1180044#M7412 <P>Sorry for the purely theoretical question.<BR /> I think there are great specialists.<BR /> I'm interested in your opinion.</P> <P>Can I find application characteristics that could describe the behavior<BR /> of the application without being bound to a particular processor architecture?<BR /> To understand what is the application needs.</P> <P>I understand that it's quite stupid to define this as the number of load / store per number<BR /> of floating point operations.</P> <P>Can I choose the optimal architecture for an application based on this characteristic ?</P> <P>Or is everything much more complicated?</P> <P>Cache, prefetch, different type of load instructions (load32bit load 64bit and etc),<BR /> the degree of parallelism and etc architecture optimizations do this characteristic<BR /> is meaningless?</P> <P>If so, is it possible to find some characteristics that can be used to determine the<BR /> optimal architecture for a given application?</P> <P>Is it possible to define these characteristics with the help of counters or should<BR /> we consider static analysis for example with the clang/llvm (the intermediate code<BR /> is much simpler. For example there is only one type of load / store memory access<BR /> instructions, and there are many other simplifications) ?</P> <P>Sorry for may be offtop questions and thanks for your time.</P> Sun, 06 May 2018 16:54:36 GMT SB17 2018-05-06T16:54:36Z https://community.intel.com/t5/Software-Tuning-Performance/choice-of-profiling-characteristic/m-p/1180044#M7412 <P>Sorry for the purely theoretical question.<BR /> I think there are great specialists.<BR /> I'm interested in your opinion.</P> <P>Can I find application characteristics that could describe the behavior<BR /> of the application without being bound to a particular processor architecture?<BR /> To understand what is the application needs.</P> <P>I understand that it's quite stupid to define this as the number of load / store per number<BR /> of floating point operations.</P> <P>Can I choose the optimal architecture for an application based on this characteristic ?</P> <P>Or is everything much more complicated?</P> <P>Cache, prefetch, different type of load instructions (load32bit load 64bit and etc),<BR /> the degree of parallelism and etc architecture optimizations do this characteristic<BR /> is meaningless?</P> <P>If so, is it possible to find some characteristics that can be used to determine the<BR /> optimal architecture for a given application?</P> <P>Is it possible to define these characteristics with the help of counters or should<BR /> we consider static analysis for example with the clang/llvm (the intermediate code<BR /> is much simpler. For example there is only one type of load / store memory access<BR /> instructions, and there are many other simplifications) ?</P> <P>Sorry for may be offtop questions and thanks for your time.</P> Sun, 06 May 2018 16:54:36 GMT https://community.intel.com/t5/Software-Tuning-Performance/choice-of-profiling-characteristic/m-p/1180044#M7412 SB17 2018-05-06T16:54:36Z