| lxo | lkcl, is it the case that our processor will support neither altivec nor vsx? I'm looking at the remaining known problem with the little I've implemented of svp64 in gcc, and it follows from having both vsx and svp64 enabled. if that's not a relevant case, I shall not waste time on it, and arrange for them to be mutually exclusive instead | 00:47 |
|---|---|---|
| lxo | lkcl, I've implemented something to make them mutually exclusive, at least for now. if they aren't, it will take quite some work to combine the svp64 patterns with the corresponding vsx and altivec ones | 02:00 |
| *** wxie1 is now known as wxie | 03:38 | |
| lkcl | lxo: correct. SIMD is harmful, and significantly impairs and hinders design as well as costing us massive amounts of time and money. | 12:26 |
| lkcl | also it makes very little sense to have Cray-Vectorised-SIMD | 12:27 |
| lkcl | that said - *sigh* - there *may* be a requirement to *call* VSX EABI formatted functions. | 12:28 |
| lkcl | the rule of thumb is: VSX performance and support is absolute rock-bottom priority. if it can actually be made *worse* performance than "strictly necessary" that is an added bonus (but not if if takes more time and effort to do so) | 12:31 |
| lkcl | the more that people can be discouraged and driven away from VSX the better | 12:31 |
| lkcl | the SIMD simulator - which we're forced to do as a kernel-side module - will have something like a 200 to 1,000 instruction overhead or possibly even higher due to context-switching, and that's perfect. | 12:32 |
| lkcl | cesar[m]1, i moved the pc-setting (storing NIA) out of the execute FSM | 15:59 |
| lkcl | the idea is to have fetch communicate with issue, issue communicate with execute | 16:03 |
| lkcl | i just moved the setting of NIA into the fetch FSM | 16:21 |
| cesar[m]1 | Sure, makes sense. That way, you can keep fetching new instructions as long as VL=0 and it's a vector instruction. | 21:50 |
| cesar[m]1 | Likewise, srcstep should be updated in the execute FSM. That way, it can keep issuing and executing while it's a vector instruction and srcstep != VL-1. | 21:53 |
Generated by irclog2html.py 2.17.1 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!