Semantics on top of profiles #115
Conversation
document/core/valid/instructions.rst
Outdated
| .. math:: | ||
| \frac{ | ||
| }{ | ||
| C \vdashinstr \shape\K{.}\vternop : [\V128~\V128~\V128] \to [\V128] |
There was a problem hiding this comment.
vternop -> RELAXEDLANESELECT.
Though, can this not actually be modelled as a ternop?
There was a problem hiding this comment.
It can for validation, but it doesn't match the usual ternop semantics, because it's not a lane-wise operation.
So if we group RELAXEDLANESELECT into vternop, it matches the vternop semantics, which is wrong.
|
Oh, another thing I forgot: please add a suitable subsection to the Changes appendix, probably as part of a new Version 2.1 section. |
ngzhian
changed the title
Done. |
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
relaxed dot allows the intermediate 16-bit to be saturated (if the underlying host uses signed * unsigned 8 bit, the dot product, 8-bit * 8-bit + 8-bit * -bit) overflows.
|
Ping |
|
The force-push screwed up the review history. :( |
Co-authored-by: Andreas Rossberg <[email protected]>
Co-authored-by: Andreas Rossberg <[email protected]>
|
Thanks for the review! |
|
Yes, makes sense. |
| \EXPROFDET & \relaxedmadd_N(z_1, z_2, z_3) &=& [ \fadd_N(\fmul_N(z_1, z_2), z_3), \fma_N(z_1, z_2, z_3) ] \\ | ||
| & \relaxedmadd_N(z_1, z_2, z_3) &=& \fma_N(z_1, z_2, z_3) \\ | ||
| \\ | ||
| & \fma_N(z_1, z_2, z_3) &=& \ieee_N(z_1 \cdot z_2 + z_3) \\ |
There was a problem hiding this comment.
Ah, sorry, I was a bit quick to approve. This of course needs to specify all the cases for NaNs and infinities, in the same manner as fadd etc.
(Also, since I didn't point it out explicitly before, please don't forget that all the definitions here still need a proper prose counterpart.)
There was a problem hiding this comment.
Yea, I missed it too, ptal #135.
No description provided.