Releases: tenstorrent/sfpi
test
7.25.0-enable-36930 wip
test
7.25.0-rename-36930 Unify builtin fn names
Compiler update
Compiler Update
- The
__builtin_sfplutfp32_3rbuiltin has been corrected - Other stability fixes
Compiler update
- Correct the compiler's representation of
sfpshft2instruction. - Other cleanups and simplifications
Compiler update
The __builtin_rvtt_sfptransp builtin has been corrected in the same way sfpswap has been, to represent data flow correctly.
auto r = __builtin_rvtt_sfptransp(a, b, c, d);
auto a = __builtin_rvtt_sfpselect4 (r, 0);
auto b = __builtin_rvtt_sfpselect4 (r, 1);
auto c = __builtin_rvtt_sfpselect4 (r, 2);
auto d = __builtin_rvtt_sfpselect4 (r, 3);
The sfpi library's subvec_transp function has been adjusted.
Compiler update
The __builtin_rvtt_sfpswap builtin has been corrected. It now returns a longer vector type that may be inspected with a new __builtin_rvtt_sfpselect2 builtin. For instance:
auto r = __builtin_rvtt_sfpswap (a, b, mod);
auto ra = __builtin_rvtt_select2 (r, 0);
auto rb = __builtin_rvtt_select2 (r, 1);
This change is necessary to properly represent the data flow within the compiler.
The sfpi library's vec_swap and vec_min_max functions have been adjusted.
Improved replay optimization
The replay optimization has been extended to share the replay buffer with user's explicit use. Previously the replay optimization was disabled in functions that had explicit uses.
Compiler update
- Fixes for incorrectly implemented quasar
roccextensions - Improved replay optimization
Compiler update
- Update to improve compiler stability
- Reparent newlib repo's origin
The GitHub-hosted newlib mirror went away. This uses a new mirror under Tenstorrent control. There are no SFPI-related changes to newlib, but relying on sourceware was proving unreliable. Releases between 7.14.0 and 7.17.0 used the now non-existent bminor mirror. Git hashes between all three repos are the same, and those releases may use either sourceware or tenstorrent-forks repos.