Re: [PATCH 3/8] nvptx -muniform-simt

[Alexander Monakov <amonakov@ispras.ru>]

On Sun, 12 Jun 2016, Sandra Loosemore wrote:
> On 06/09/2016 10:53 AM, Alexander Monakov wrote:
> > +@item -muniform-simt
> > +@opindex muniform-simt
> > +Generate code that allows to keep all lanes in each warp active, even when
> 
> Allows *what* to keep?  E.g. what is doing the keeping here?  If it is the
> generated code itself, please rephrase as
> 
> Generate code that keeps....

Let me try to expand and rephrase what I meant:

Allows the compiler to emit code that, at run time, may have all lanes active,
particularly in those regions of the program where observable effects from
execution must happen as if one lane is active (outside of SIMD loops).

But nevertheless generated code can run just like conventionally generated
code does: with each lane being active/inactive independently, and side
effects happening from each active lane (inside of SIMD loops).

Whether it actually runs in the former (let's call it "uniform") or the latter
("conventional") way is switchable at run time. The compiler itself is
responsible for emitting mode changes at SIMD region boundaries.

Does this help? Below I went with your suggestion, but changed "keeps" to "may
keep" because that's generally true only outside of SIMD regions.

> > +observable effects from execution should appear as if only one lane was
> 
> s/was/is/
> 
> > +active. This is achieved by instrumenting syscalls and atomic instructions
> > in
> > +a lightweight way that allows to switch behavior at runtime. This code
> 
> Same issue here....  allows *what* to switch behavior?  (And how would you
> select which run-time behavior you want?)

Sorry. This gives compiler itself a way to emit code that will switch behavior
of the subsequently running code.

> Also, in the snippet above where it is used as a noun, please
> s/runtime/run time/

Thanks. Does the following look better?

@item -muniform-simt
@opindex muniform-simt
Generate code that may keep all lanes in each warp active, even when
observable effects from execution must appear as if only one lane is active.
This is achieved by instrumenting syscalls and atomic instructions in a
lightweight way, allowing the compiler to emit code that can switch at run
time between this and conventional execution modes. This code generation
variant is used for OpenMP offloading, but the option is exposed on its own
for the purpose of testing the compiler; to generate code suitable for linking
into programs using OpenMP offloading, use option @option{-mgomp}.

Alexander