@barneypitt That is just as an example that we may at some point want to generate code other languages rather than OpenCL only
*generate code to other languages...
instead of SYCL, could CUDA, or could be HSA
or nothing else for now
or OpenACC
who knows
@barneypitt if we have an AST like data than it is easier to support such use cases
if they arise
to have OpenCL 1.2 and 2.0
generators
as they are quite differente
and some devices are OpenCL 1.x only
and there will also be OpenCL Next
plus there is also Vulkan
and Metal on Mac
this vectorized programming technologies are getting quite fragmented
instead of joining forces to have a neat solution for everyone
Aparapi could be what Hibernate was for databases, or even better
Got to go soon. But to be clear, when you started talking about an AST I assumed you meant from source code. Which sounded great, but was not Aparapi. But I was unlear what "AST from bytecode" meant. To me, it means a hierarchical desciption of the bytecode, which is already what the Instruction class is.
But instruction is not a very friendly hierarchical description of the bytecode - doesn't really allow iteration/navigation. So if what we're talking about is a much-improved, more flexible, navigable successor to Instruction, then I wholeheartedly agree, bring it on!
Okay, cool, do you have links to those implementations? No point in us reinventing the wheel!
@barneypitt https://github.com/mirkosertic/Bytecoder
there are many others
I haven't had the time to play with them
they aren't true AST, but rather AST like data
"The JVM Bytecode is parsed and transformed into an intermediate representation. This intermediate representation is passed thru optimizer stages and sent to a backend implementation for target code generation."
Referring to Bytecoder
we would need to assess which are good and which aren't
time to rest
see you
So you can imagine I'm not diggin that somultion