Closed dzaima closed 1 month ago
@4vtomat could you take a look for this issue?
Somewhat-relatedly, all _vf_
widening float multiply-add intrinsics, e.g. __riscv_vfwmacc_vf_f32m1
(not just, but including, bfloat16 ones) name their scalar arguments vs1
where I think it should be rs1
. This appears to have been the case for a while.
Thanks @dzaima , let me take this!
Running
make bf16-all
I get it generating incorrect reinterpret instructions like__riscv_reinterpret_v_bf16m8_i16m8
.Note the lack of
v
prefix there, which matches nov
in["reinterpret"]
inbfloat16_inst.py
.