| commit | 5632facc4a2d6892a35968e0a0e5c38c554079ab | [log] [tgz] |
|---|---|---|
| author | Victor Khimenko <khim@google.com> | Sun Feb 25 20:51:10 2024 +0000 |
| committer | Victor Khimenko <khim@google.com> | Mon Feb 26 11:48:30 2024 +0000 |
| tree | 805b7a14d9c8db022c46c1a222cd069f4fd24af0 | |
| parent | 8f6963a5b7c1b8a5d152c37bb5dcaca7f9aa4649 [diff] |
Add Float8 (phony) and Float16 (real) types. Certain vector floating point operations deal with double-width floats. For actual type Float32 we need to specify width of Float16 and for Float16 we would need Float8. For Float16 we may use one of three base types supported: __fp16, _Float16, or __bf16. If we would decide to support half-precision floats _Float16 is better than __fp16. Type Float8 can be specified as a template argument but couldn't be actually instantiated and used. Test: m berberis_all Change-Id: Ia8869613980eb324d7dd496935e7f0f8248add2b
Berberis: dynamic binary translator to run Android apps with riscv64 native code on x86_64 devices or emulators.