3 ms·

>However, unknown to any of us, it turns out that PEXT is extremely slow on AMD Zen and Zen 2 architectures. Is it because of some patents? Anyone knows?

craftyguy 7y ago
I like how an honest, on-topic question like this on 'hacker' 'news' is downvoted into oblivion. Nice quality site this is.
- dang 7y ago
  Your comment breaks the site guidelines. Since you've quoted them recently in other contexts, I assume you know this. One reason we have that rule is that once users come along and (as they typically do) give corrective upvotes to the unfairly-downvoted comment, a comment like yours becomes false. Unfortunately, such posts don't then garbage-collect themselves—they stick around, adding noise to the thread. https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html
BeeOnRope 7y ago
Yeah, it could be patents. The efficient hardware implementation of these instructions was patented by the Princeton researchers that developed it, and this work was sponsored by Intel (and the Department of Defense). One could reasonably assume that the sponsor of the work would have access to the patent on very favorable terms. Some details at [1]. [1] https://twitter.com/trav_downs/status/1202863930210734080 https://twitter.com/trav_downs/status/1202863930210734080
gioele 7y ago
While on the topic, I always hope that HW manufacturers could start being explicit about what is _supported_ and what is "supported". We have this problem with CPU manufacturers stating that something is supported, when it is instead emulated (hello checks on CPUID). Disk drives acknowledge writes before they actually write the data to the disk (hello layers of unreliable fsyncs). Network cards claim that packet have been shipped while they never left the circular buffer (hello bufferbloat). Please HW manufacturers, stop lying to us.

craftyguy 7y ago

I like how an honest, on-topic question like this on 'hacker' 'news' is downvoted into oblivion. Nice quality site this is.

dang 7y ago
Your comment breaks the site guidelines. Since you've quoted them recently in other contexts, I assume you know this. One reason we have that rule is that once users come along and (as they typically do) give corrective upvotes to the unfairly-downvoted comment, a comment like yours becomes false. Unfortunately, such posts don't then garbage-collect themselves—they stick around, adding noise to the thread. https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html

BeeOnRope 7y ago

Yeah, it could be patents. The efficient hardware implementation of these instructions was patented by the Princeton researchers that developed it, and this work was sponsored by Intel (and the Department of Defense). One could reasonably assume that the sponsor of the work would have access to the patent on very favorable terms. Some details at [1]. [1] https://twitter.com/trav_downs/status/1202863930210734080 https://twitter.com/trav_downs/status/1202863930210734080

gioele 7y ago

While on the topic, I always hope that HW manufacturers could start being explicit about what is _supported_ and what is "supported". We have this problem with CPU manufacturers stating that something is supported, when it is instead emulated (hello checks on CPUID). Disk drives acknowledge writes before they actually write the data to the disk (hello layers of unreliable fsyncs). Network cards claim that packet have been shipped while they never left the circular buffer (hello bufferbloat). Please HW manufacturers, stop lying to us.