Milestones

[John Rose]

Ah, please ignore the reindeer. That's a corporate trade secret weapon. 

Pretend I said "rendered".

-- John  (on my iPhone)

On Nov 13, 2012, at 4:53 PM, John Rose <john.r.rose at oracle.com> wrote:

> On Nov 8, 2012, at 6:18 AM, "Harle, Christophe" <christophe.harle at amd.com> wrote:
> 
>> This proposed milestone involves moving enough of Aparapi's implementation down into the Project Lambda enabled JVM to allow Lambda enabled Aparapi demos/workloads to execute directly (no Aparapi.dll). Whilst this effort would constrain demos/examples to Aparapi's existing programming model constraints (parallel primitive arrays only) it would act as a 'tracer bullet' through the code base, allowing us to familiarize ourselves with the various major JVM components that we think will to impacted via a Sumatra implementation.
> 
> The "tracer bullet" is good.
> 
> Do we really need lambda for a first victory?  I think online code generation, just by itself, even without execution, would be huge.
> 
> Other milestones:
> • First use of CompileBroker to trigger GPU codegen. 
> • First use of C2 IR rendered to GPU code. 
> (Note: Codegen does not imply execution. Could be OptoNoExecute mode.)
> • First run of GPU code from online generation. 
> • First run of GPU code from (one-off) lambda collection library. 
> • First run of GPU code from (enhanced) standard parallel collection. 
> • First profitable run of ditto. 
> 
> On the data front: 
> • First use of pointer-chained Java data structure from GPU. 
> • First GC tracing through GPU task data. 
> • First call and return of GPU method with zero copying. (Already done?)  Here, "zero" is any O(#GPUs), which is less than the bulk data size. 
> 
> Not sure exactly how to order these. 
> 
> I am convince that there are about 3 fronts that need roghly simultaneous advance:
> • code: how to get code into the GPUs. Should be an online hot-spot reindeer by a JIT. 
> • data: how to get data in and out. Should contain managed pointers, but not as many as classic Java data. I.e. flatter with more "little structs".  See Arrays 2.0 and value types. 
> • Libraries: teach parallel collections to generate GPU commands on enabled systems. Translate lambda bodies and loop kernels to GPU code. 
> 
> Best wishes,
> – John