Just curious:
When would the number of instances in a gang equal to a small multiple of SIMD vector width?
huehue
So a call to an ISPC function results in 1 gang, which has multiple program instances to execute the function. But is it possible to have multiple gangs (each having multiple program instances) for a particular function call? If multiple tasks are created, then each would result in 1 gang. But is there any case where one task can have multiple gangs, or does one function call always correspond to exactly one gang?
cyl
The program counter (# of instances in a gang) of gates machine is 8. How do we determine that value?
vincom2
@cyl if you look at the next slide, it's programCount in ISPC code
Just curious: When would the number of instances in a gang equal to a small multiple of SIMD vector width?
So a call to an ISPC function results in 1 gang, which has multiple program instances to execute the function. But is it possible to have multiple gangs (each having multiple program instances) for a particular function call? If multiple tasks are created, then each would result in 1 gang. But is there any case where one task can have multiple gangs, or does one function call always correspond to exactly one gang?
The program counter (# of instances in a gang) of gates machine is 8. How do we determine that value?
@cyl if you look at the next slide, it's
programCount
in ISPC code@PandaX see my comment on slide
@huehue, each task maps to exactly one gang.
@huehue
Wouldn't multiple gangs mapped to a function be the exact same if not worse than just a singular larger gang?