Declarative abstraction sometimes is faster than imperative. When computer decides some operations (read and write) within "atomic" can be performed in parallel. This is explained in detail in lecture 15.
This comment was marked helpful 0 times.
retterermoore
How would you determine which to use in practice without actively writing both and testing them? You could test out both, but if imperative involves writing a large codebase with a complex series of locks it could be easier to know how much faster it would be than declarative without having to write it out.
This comment was marked helpful 0 times.
jinsikl
@retterermoore I'm not sure there is a better way to find out which is faster. Determining that without actually trying it would involve intricate knowledge of how the compiler translates the declarative code, target architecture, etc.
This comment was marked helpful 0 times.
idl
I believe that when stated in a declarative way, this gives the compiler/runtime the chance to heavily optimize how it should atomize the code. This could potentially result in a much faster implementation than if the developer worries about it.
This comment was marked helpful 0 times.
ruoyul
ISPC forall loop is declarative, it let the compiler decide how to run the gangs and assign work. While the for loops is imperative, the programmer states how assignment should be done.
This comment was marked helpful 0 times.
bxb
@ruoyul I was having the same thoughts about explicit vs. implicit parallelism. Declarative is the same as explicit, in that the programmer has to specify the work division and how it should be done; Imperative is like implicit in which there is usually some construct that says how something will be done to an extent, but the actual implementations are usually hidden from the programmer.
This comment was marked helpful 0 times.
drayson
@bxb Shouldn't those be the other way around?
This comment was marked helpful 0 times.
tcz
I think declarative code would be more portable and much easier to write. SQL is an example of a declarative language.
Declarative abstraction sometimes is faster than imperative. When computer decides some operations (read and write) within "atomic" can be performed in parallel. This is explained in detail in lecture 15.
This comment was marked helpful 0 times.
How would you determine which to use in practice without actively writing both and testing them? You could test out both, but if imperative involves writing a large codebase with a complex series of locks it could be easier to know how much faster it would be than declarative without having to write it out.
This comment was marked helpful 0 times.
@retterermoore I'm not sure there is a better way to find out which is faster. Determining that without actually trying it would involve intricate knowledge of how the compiler translates the declarative code, target architecture, etc.
This comment was marked helpful 0 times.
I believe that when stated in a declarative way, this gives the compiler/runtime the chance to heavily optimize how it should atomize the code. This could potentially result in a much faster implementation than if the developer worries about it.
This comment was marked helpful 0 times.
ISPC forall loop is declarative, it let the compiler decide how to run the gangs and assign work. While the for loops is imperative, the programmer states how assignment should be done.
This comment was marked helpful 0 times.
@ruoyul I was having the same thoughts about explicit vs. implicit parallelism. Declarative is the same as explicit, in that the programmer has to specify the work division and how it should be done; Imperative is like implicit in which there is usually some construct that says how something will be done to an extent, but the actual implementations are usually hidden from the programmer.
This comment was marked helpful 0 times.
@bxb Shouldn't those be the other way around?
This comment was marked helpful 0 times.
I think declarative code would be more portable and much easier to write. SQL is an example of a declarative language.
This comment was marked helpful 0 times.