floatexplorer

This is a little bit of code to dump the representation of some floating point types, including:

• e5m2,
• e4m3,
• fp16 (e5m10)
• bf16 (e8m7),
• ieee 32-bit (C float: e8m23),
• ieee 64-bit (C double, and long double on some platforms: e11m52),
• Intel 80-bit long double (e15m64: unlike most, this one doesn't use a IEEE like representation, and has an explicit leading mantissa bit)
• IEEE 128-bit (e15m112: Linux ARM long double, GCC libquadmath)

TODO

• See if the HIP API can also do the float type conversions as a cross check.

• Implement zArch mainframe "HEXFLOAT" types.

• Review outputs for E4M3 for:

  ./floatexplorer --e4m3 0x78 ./floatexplorer --e4m3 0xF8 ./floatexplorer --e4m3 0xF9 ./floatexplorer --e4m3 0xFA ./floatexplorer --e4m3 0xFB ./floatexplorer --e4m3 0xFC ./floatexplorer --e4m3 0xFD ./floatexplorer --e4m3 0xFE ./floatexplorer --e4m3 0xFF

(testbin/e4m3.sh: get different results for these in CUDA vs. non-CUDA configurations.)

• Same thing for:

  ./floatexplorer --e5m2 0xFD ./floatexplorer --e5m2 0xFE ./floatexplorer --e5m2 0xFF

(testbin/e5m2.sh)

• Review outputs from:

./testbin/bf16.sh | tee expected/bf16.all.txt ./testbin/fp16.sh | tee expected/fp16.all.txt

(those were generated with CUDA conversion api -- compare to non-CUDA)

Discussion

An early version of the code is described here:

https://peeterjoot.com/2025/04/23/a-little-float-explorer-program/

Build

Requires C++20 for <format>. Run:

make

Known to build on

• MacOS
• ARM Linux (fedora 42)
• x64 Linux (fedora 42)
• x64 Linux (fedora 41)
• x64 Linux (debian Ubuntu 24.04 -- WSL2)

Dependencies:

Linux x64:

sudo dnf -y install make
sudo dnf -y install g++
sudo dnf -y install gcc
sudo dnf -y install libquadmath
sudo dnf -y install libquadmath-devel

MacOs:

brew install gcc

CUDA dependencies

Support for GPU types (e5m2, e4m3, fp16, bf16) has been implemented. For string <> float conversions for these types, CUDA support is used if available (rudimentary auto-detection in the makefile.) CUDA is very specific about which fedora versions are required. CUDA 12.9 did not work on any version higher than fedora 41, and CUDA 13 is required for Fedora 42 (latest.)

The fedora42 installation sequence was:

sudo dnf config-manager addrepo --from-repofile=https://developer.download.nvidia.com/compute/cuda/repos/fedora42/$(uname -m)/cuda-fedora42.repo
sudo dnf module disable nvidia-driver
sudo dnf config-manager setopt cuda-fedora42-$(uname -m).exclude=nvidia-driver,nvidia-modprobe,nvidia-persistenced,nvidia-settings,nvidia-libXNVCtrl,nvidia-xconfig
sudo dnf -y install cuda-toolkit
sudo dnf config-manager setopt cuda-fedora42-x86_64.exclude=
sudo dnf install -y nvidia-driver-cuda --refresh

... with some reboots in the mix, and lots of uninstalls to remove artifacts of having attempted to installed cuda-12.9.

CUDA bugs:

Note that the CUDA 12.9 type conversion seems to misidentify some E4M3 NaNs, such as 0x7E. Using:

 __half half = __nv_cvt_fp8_to_halfraw( s, __NV_E4M3 );
 float output = __half2float( half );

we get 448 instead of NaN (and not even 240, which would be the big-value if NaNs weren't supported.)

i.e.: 240:

FromDigits["1111", 2]*2^(7 - 3)