Ben Vanik
Ben Vanik
Good find! This looks like a case that specialization should be able to handle - the analysis information derived from the dispatch sites should be present during executable configuration, but...
(also, great triage! thanks for digging in!)
It's also a compile time tradeoff - compiling 100 executables is way faster than compiling 1500. When we specialize we have to be certain we're doing it for meaningful reasons....
The assume ops in the executable are derived from the host information, so by definition it has the information. Whether we have the right folders or not is another question...
My hope (having looked at your PR) is that we can fold the arith AndI/etc ops - if not, we'll need to have some ops we can fold (`util.int.whatever_the_arithmetic_is_doing`) emitted...
(closing as stale)
(closing as stale)
(closing as stale)
(we _should_ be able to handle it but we may not today and upstream stuff like bufferization may not either - I believe scf does work and on the TODO...
yep! rocminfo will show you for your particular device, but nowadays the norm is: ``` Grid Max Size: 4294967295(0xffffffff) Grid Max Size per Dimension: x 2147483647(0x7fffffff) y 65535(0xffff) z 65535(0xffff)...