compilation problem

[jtlap]

Hello, I'm one of the EVE developer. Thank you for using our library.

As I have seen vstat, I wanted to try it and I followed your instructions, but I do not see how to compile the tests
and it seems to me that some include are not here (#include <doctest/doctest.h> in vstat_test.cpp for instance)
and I don't know how to generate them.
I will be glad to know how to do that.

However, I just had a glance to the code, and I have a question and (perhaps) an advice but not supported by trials yet.

The question is : I have seem that internally in templated functions some local variables and/or results are always taken as double.
Pure float computation is not possible?

The advice (that can be related to the question and to accuracy/speed) is that in EVE we have a quite well populated family of fma function
and also diff_of_prod and sum_of_prod that can improve the numeric computations
(for instance 'sum_w * x - sum_x * w' can be a diff_of_prod , 'T dx = x * sum_w - sum_x;' can be a fms)

I don't know if using fma where possible, will be better but at least it cannot degrade the performances
For sum_of_prod it depends of a pedantic option. With no pedantic just one fma is used. With pedantic it is more costly but
with an error correction and more fmas that could improve accuracy.

P.S.: just a remark about the documentation. As you use EVE it is a C++-20 project not 17!

[foolnotion]

Hi,

Thanks for reaching out!

As I have seen vstat, I wanted to try it and I followed your instructions, but I do not see how to compile the tests
and it seems to me that some include are not here (#include <doctest/doctest.h> in vstat_test.cpp for instance)
and I don't know how to generate them.
I will be glad to know how to do that.

I use nix to manage dependencies, if you want to build the tests you must provide the following dependencies: boost doctest gsl pkg-config linasm.

The question is : I have seem that internally in templated functions some local variables and/or results are always taken as double.
Pure float computation is not possible?

Pure float computation is possible (unless you have encountered a bug), we only use double for the final aggregated statistics; this should not have any impact on performance.

The advice (that can be related to the question and to accuracy/speed) is that in EVE we have a quite well populated family of fma function and also diff_of_prod and sum_of_prod that can improve the numeric computations (for instance 'sum_w * x - sum_x * w' can be a diff_of_prod , 'T dx = x * sum_w - sum_x;' can be a fms). I don't know if using fma where possible, will be better but at least it cannot degrade the performances
For sum_of_prod it depends of a pedantic option. With no pedantic just one fma is used. With pedantic it is more costly but with an error correction and more fmas that could improve accuracy.

Thanks for the suggestions, I will implement them and test performance.

P.S.: just a remark about the documentation. As you use EVE it is a C++-20 project not 17!

Thanks, this is probably an oversight in the documentation, due to the fact that in the beginning I was using vectorclass2 (C++17) as a SIMD backend. Later I moved to EVE.