commit | 17ea3bdb7fb090aa0b3487b78b94ed29aaa2a134 | [log] [tgz] |
---|---|---|
author | Victor Khimenko <khim@google.com> | Thu Mar 07 18:09:31 2024 +0000 |
committer | Victor Khimenko <khim@google.com> | Thu Mar 07 18:15:27 2024 +0000 |
tree | a3da06b8e8eac740314fc973374d55d7caf1add7 | |
parent | 1ebdc4bbedc0e26d3c986f9c9c63b97b330c5746 [diff] |
Remove DEFINE_2OP_ARITHMETIC_INTRINSIC_VS intrinsics. As a result of refactoring when masking was moved partialy to separate intrinsics and partially to the interpreter we ended up with pretty bizzare intrinsics that take two scalars and produce the desires result in all elements of vector. And then interpreter takes one element from that result. Such intrinsics technically work but are entirely useless and to design sensible handling of reduction instructions we first need to run real apps on the emulator and collect statistics: what kinds of reductions are in actual use, do they use masking or not, etc. Test: berberis_all Change-Id: I34929a3f38ca06e4dad3683e3052c236dc673616
Berberis: dynamic binary translator to run Android apps with riscv64 native code on x86_64 devices or emulators.