Difference between revisions of "Instruction Set/fmafe"
From Mill Computing Wiki
m (Protected "Instruction Set/fmafe": generated ([Edit=<protect-level-bot>] (indefinite) [Move=<protect-level-bot>] (indefinite))) | |
(No difference) | |
Revision as of 01:35, 3 January 2015
realizing exu stream exu block compute phase operation in the binary floating point value domain and rounds to nearest, ties toward even adjacent value
Binary floating point fused multiply-add. As usual for those, it yields a higher precision than doing it separately, and is faster too. Rounds towards even.
operands: like Addf [ff:f]
Returns x*y+z on the belt.
encoding:
fmafe(f x)
,
exuArgs(op arg0, op arg1)
| Core | In Slots | Latencies |
|---|---|---|
| Silver | E0 E1 | w,w:w=6 wv,wv:wv=6 d,d:d=7 dv,dv:dv=7 q,q:q=8 qv,qv:qv=8 |
| Gold | E0 E1 E2 E3 | w,w:w=6 wv,wv:wv=6 d,d:d=7 dv,dv:dv=7 q,q:q=8 qv,qv:qv=8 |
fmafe(f x, f y, f z, f w) → f r0, f r1
operands: like Fmasf [ff:f]
This is a fused multiply-add-subtract. An excellent way to make full use of all Functional Units in the 2 Slots.
r0 is x*y+z*w
r1 is x*y-z*w
encoding:
fmafe(f x, f y)
,
exuArgs(op arg0, op arg1)
| Core | In Slots | Latencies |
|---|---|---|
| Silver | E0 | w,w:w,w=6,6 wv,wv:wv,wv=6,6 d,d:d,d=7,7 dv,dv:dv,dv=7,7 q,q:q,q=8,8 qv,qv:qv,qv=8,8 |
| Gold | E0 E2 | w,w:w,w=6,6 wv,wv:wv,wv=6,6 d,d:d,d=7,7 dv,dv:dv,dv=7,7 q,q:q,q=8,8 qv,qv:qv,qv=8,8 |
Instruction Set, alphabetical, Instruction Set by Category, Instruction Set, sortable, filterable
