search

riscv-non-isa / rvv-intrinsic-doc

https://jira.riscv.org/browse/RVG-153
BSD 3-Clause "New" or "Revised" License
287 stars 90 forks source link

Fix and cleanup examples #254

Closed camel-cdr closed 1 year ago

camel-cdr commented 1 year ago

This resolves https://github.com/riscv-non-isa/rvv-intrinsic-doc/issues/252, https://github.com/riscv-non-isa/rvv-intrinsic-doc/issues/244, unsigned char instead of char, and cleans up the code a bit.

The problem with using char in combination with vint8mX_t is that char can be either signed or unsigned, so to remedy this problem, I've changed the examples to use unsigned char explicitly (and vuint8m8_t).

rvv_reduce.c had an error, where a agnostic tail policy could make it produce the wrong results, this has been fixed by using the _tu intrinsic variant.

dzaima commented 1 year ago

__riscv_vmfne_vv_f64m1_b64(vec_a, vec_zero, vl); in rvv_reduce.c could also be replaced with __riscv_vmfne_vf_f64m1_b64

eopXD commented 1 year ago

Nit: May you add "Resolve #XXX" into the commit messages so they can be interlinked with the issue?

camel-cdr commented 1 year ago

@eopXD Is it important that it's in the correct commit, or can just add that in a commit renaming branch to branch_vec? I haven't done a rebase & amend & force push with an existing PR, so idk how that would behave.

eopXD commented 1 year ago

@eopXD Is it important that it's in the correct commit, or can just add that in a commit renaming branch to branch_vec? I haven't done a rebase & amend & force push with an existing PR, so idk how that would behave.

Yes, please rebase and amend it then force push again. Thank you.

camel-cdr commented 1 year ago

The only objection I have is that I don't find abbreviating src/dst to s/d very informative.

I haven't addressed this yet, what name would you suggest?

I've also wondered about the choices of LMUL, was the idea to show a few different ones? Because usually maximizing LMUL given the number of required registers seems to be the best option.

eopXD commented 1 year ago

The only objection I have is that I don't find abbreviating src/dst to s/d very informative.

I haven't addressed this yet, what name would you suggest?

Welp, maybe input_src/input_dst for parameter and src / dst for iteration?

I've also wondered about the choices of LMUL, was the idea to show a few different ones? Because usually maximizing LMUL given the number of required registers seems to be the best option.

Yes the intent was to show different LMUL's.

camel-cdr commented 1 year ago

Ok, I've now used source, destination, src, dst.