Attempt to Add float16_t support

[spinicist]

Hello,

As discussed here fmtlib/fmt#4725 I need to format float16_ts. I started trying to work on this today, but have failed early on.

I am working with GCC 15.2 (Clang does not have stdfloat support yet). I have changed CMakeLists.txt to compile in C++23 mode. I then started to add float16_t overloads. However, if I compile with the set of changes in this MR I get:

/software/system/gcc/15.2/bin/g++  -I/home/k1078535/Code/zmij/. -std=gnu++23 -pthread -MD -MT CMakeFiles/zmij.dir/zmij.cc.o -MF CMakeFiles/zmij.dir/zmij.cc.o.d -o CMakeFiles/zmij.dir/zmij.cc.o -c /home/k1078535/Code/zmij/zmij.cc
In file included from /home/k1078535/Code/zmij/zmij.cc:25:
/nan/pacific/network/system/el8/gcc/15.2/include/c++/15.2.0/limits:2015:27: error: ‘__gnu_cxx’ was not declared in this scope; did you mean ‘zmij::__gnu_cxx’?
 2015 |     struct numeric_limits<__gnu_cxx::__bfloat16_t>
      |                           ^~~~~~~~~
      |                           zmij::__gnu_cxx
In file included from /nan/pacific/network/system/el8/gcc/15.2/include/c++/15.2.0/stdfloat:33,
                 from /home/k1078535/Code/zmij/zmij.h:66,
                 from /home/k1078535/Code/zmij/zmij.cc:8:
/nan/pacific/network/system/el8/gcc/15.2/include/c++/15.2.0/x86_64-pc-linux-gnu/bits/c++config.h:373:11: note: ‘zmij::__gnu_cxx’ declared here
  373 | namespace __gnu_cxx
      |           ^~~~~~~~~

This doesn't make any sense to me - any ideas are most welcome.

(I have checked that a simple test float16_t program not using zmij compiles correctly on my setup).

[Alcaro]

This doesn't make any sense to me - any ideas are most welcome.

Your #include is inside a namespace zmij {.

#include in namespaces is not allowed, move it outside.

You'll most likely need to duplicate that #ifdef, a namespace zmij {, or something.

[spinicist]

I swear I had tried that initially and seen the same error. Apologies for wasting your time.

[spinicist]

I have made some progress now. I have added enough #ifdef __STDCPP_FLOAT16_T__ that I can compile a modified example.cc to output a float16_t. However, the output is incorrect. It prints 0.0001 instead of 1. I suspect that I have the traits set up wrong, but I can't see where. I'll continue to hack on this, but any suggestions are very welcome.

P.S. One definite issue is that I'm getting bad information from dec_exp_formats.get<>. This gives a start_pos of 2 for float16_t when by comparison to float I think it should be 0.

[vitaut]

Thanks for the PR. Please revert any unrelated changes such as commenting out sse tests.

[spinicist]

Sorry - I should have made it clearer that:

1. This PR isn't ready to merge. If there's a better forum to ask for help with getting it ready, let me know.
2. I forgot that I disabled SIMD entirely for now. I thought it best to get it correct first and then re-enable and work out the correct assembly for that section.

[Antares0982]

I'm also very interested in float16-to-string conversion. But I'm wondering whether we can avoid using std::float16_t and instead use a uint16_t or another type of the same width as a substitute, since std::float16_t requires additional compiler support and checks.

[vitaut]

I'm wondering whether we can avoid using std::float16_t and instead use a uint16_t or another type of the same width as a substitute, since std::float16_t requires additional compiler support and checks.

Let's first make std::float16_t work without these complications.

[spinicist]

@vitaut A better explanation: The SSE tests are commented out because:

1. I was getting SIMD issues in zmij.cc
2. Hence I set # define ZMIJ_USE_SIMD 0 for now
3. This then made the SSE tests fail to compile

What is the correct way to disable SIMD for both zmij and the tests?

[vitaut]

You shouldn't be adding a new trait but update the main trait to handle 16-bit float correctly (mostly num_bits).

[vitaut]

What is the correct way to disable SIMD for both zmij and the tests?

In your local build you can set ZMIJ_USE_SIMD macro to 0. In zmij itself you shouldn't be disabling anything - if there is an issue it should be investigated and fixed.

[vitaut]

half -> float16

[spinicist]

I have rebased and rewritten this incorporating your requests. It still gives incorrect output.

I think the key issue is here: https://github.com/vitaut/zmij/blob/main/zmij.cc#L1063

I assume this needs an extra special case for 16 bits, but I don't know which constants then need changing. Any pointers much appreciated.

[vitaut]

I think the key issue is here: https://github.com/vitaut/zmij/blob/main/zmij.cc#L1063

I think the code you were referring to has moved. Use a permalink to prevent this.

[spinicist]

I think the code you were referring to has moved. Use a permalink to prevent this.

Yes, it has moved up slightly:

zmij/zmij.cc

Line 975 in f1126cd

if (num_bits == 32) {

(This is a block that checks for 32 bits)

[vitaut]

That particular check could be changed to <= 32 although we'll probably need more adjustments elsewhere.