From: lkcl <lkcl@web>
Date: Tue, 21 Jun 2022 18:29:07 +0000 (+0100)
Subject: (no commit message)
X-Git-Tag: opf_rfc_ls005_v1~1619
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=24d274b8edda5fa4bbc75f019ba9eea04a4da7a0;p=libreriscv.git

---

diff --git a/openpower/sv/svp64/appendix.mdwn b/openpower/sv/svp64/appendix.mdwn
index 6a4b3e194..4af881682 100644
--- a/openpower/sv/svp64/appendix.mdwn
+++ b/openpower/sv/svp64/appendix.mdwn
@@ -497,6 +497,15 @@ When Rc=1 a corresponding Vector of co-resultant CRs is also
 created.  No special action is taken: the result and its CR Field
 are stored "as usual" exactly as all other SVP64 Rc=1 operations.
 
+Implementors must keep in mind that Parallel Reduction is an abstracted
+concept of a "Schedule" that issues Scalar Operations.  It is
+**not permitted** to assume that the accuracy can be made higher than that
+of performing the exact same Scalar instructions just because a Parallel
+Reduction Schedule is being used.
+
+Also note that the Schedule only makes sense on top of certain instructions:
+X-Form with a Register Profile of `RT,RA,RB` is fine.
+
 # Fail-on-first
 
 Data-dependent fail-on-first has two distinct variants: one for LD/ST