nvptx: '-mframe-malloc-threshold', '-Wframe-malloc-threshold' (was: Handling of large stack objects in GPU code generation -- maybe transform into heap allocation?)

Fri Dec 23 14:08:06 GMT 2022

Hi!

On 2022-11-11T15:35:44+0100, Richard Biener via Fortran <fortran@gcc.gnu.org> wrote:
> On Fri, Nov 11, 2022 at 3:13 PM Thomas Schwinge <thomas@codesourcery.com> wrote:
>> For example, for Fortran code like:
>>
>>     write (*,*) "Hello world"
>>
>> ..., 'gfortran' creates:
>>
>>     struct __st_parameter_dt dt_parm.0;
>>
>>     try
>>       {
>>         dt_parm.0.common.filename = &"source-gcc/libgomp/testsuite/libgomp.oacc-fortran/print-1_.f90"[1]{lb: 1 sz: 1};
>>         dt_parm.0.common.line = 29;
>>         dt_parm.0.common.flags = 128;
>>         dt_parm.0.common.unit = 6;
>>         _gfortran_st_write (&dt_parm.0);
>>         _gfortran_transfer_character_write (&dt_parm.0, &"Hello world"[1]{lb: 1 sz: 1}, 11);
>>         _gfortran_st_write_done (&dt_parm.0);
>>       }
>>     finally
>>       {
>>         dt_parm.0 = {CLOBBER(eol)};
>>       }
>>
>> The issue: the stack object 'dt_parm.0' is a half-KiB in size (yes,
>> really! -- there's a lot of state in Fortran I/O apparently).  That's a
>> problem for GPU execution -- here: OpenACC/nvptx -- where typically you
>> have small stacks.  (For example, GCC/OpenACC/nvptx: 1 KiB per thread;
>> GCC/OpenMP/nvptx is an exception, because of its use of '-msoft-stack'
>> "Use custom stacks instead of local memory for automatic storage".)
>>
>> Now, the Nvidia Driver tries to accomodate for such largish stack usage,
>> and dynamically increases the per-thread stack as necessary (thereby
>> potentially reducing parallelism) -- if it manages to understand the call
>> graph.  In case of libgfortran I/O, it evidently doesn't.  Not being able
>> to disprove existance of recursion is the common problem, as I've read.
>> At run time, via 'CU_JIT_INFO_LOG_BUFFER' you then get, for example:
>>
>>     warning : Stack size for entry function 'MAIN__$_omp_fn$0' cannot be statically determined
>>
>> That's still not an actual problem: if the GPU kernel's stack usage still
>> fits into 1 KiB.  Very often it does, but if, as happens in libgfortran
>> I/O handling, there is another such 'dt_parm' put onto the stack, the
>> stack then overflows; device-side SIGSEGV.
>>
>> (There is, by the way, some similar analysis by Tom de Vries in
>> <https://gcc.gnu.org/PR85519> "[nvptx, openacc, openmp, testsuite]
>> Recursive tests may fail due to thread stack limit".)
>>
>> Of course, you shouldn't really be doing I/O in GPU kernels, but people
>> do like their occasional "'printf' debugging", so we ought to make that
>> work (... without pessimizing any "normal" code).
>>
>> I assume that generally reducing the size of 'dt_parm' etc. is out of
>> scope.
>>
>> There is a way to manually set a per-thread stack size, but it's not
>> obvious which size to set: that sizes needs to work for the whole GPU
>> kernel, and should be as low as possible (to maximize parallelism).
>> I assume that even if GCC did an accurate call graph analysis of the GPU
>> kernel's maximum stack usage, that still wouldn't help: that's before the
>> PTX JIT does its own code transformations, including stack spilling.
>>
>> There exists a 'CU_JIT_LTO' flag to "Enable link-time optimization
>> (-dlto) for device code".  This might help, assuming that it manages to
>> simplify the libgfortran I/O code such that the PTX JIT then understands
>> the call graph.  But: that's available only starting with recent
>> CUDA 11.4, so not a general solution -- if it works at all, which I've
>> not tested.
>>
>> Similarly, we could enable GCC's LTO for device code generation -- but
>> that's a big project, out of scope at this time.  And again, we don't
>> know if that at all helps this case.
>>
>> I see a few options:
>>
>> (a) Figure out what it is in the libgfortran I/O implementation that
>> causes "Stack size [...] cannot be statically determined", and re-work
>> that code to avoid that, or even disable certain things for nvptx, if
>> feasible.

> Shrink st_parameter_dt (it's part of the ABI though, kind of).  Lots of the
> bloat is from things that are unused for simpler I/O cases (so some
> "inheritance" could help), and lots of the bloat is from using
> string/length pairs using char * + size_t for what looks like could be
> encoded a lot more efficiently.
>
> There's probably not much low-hanging fruit.

(Similarly comments in Janne's email.)

Well, as had to be expected, libgfortran I/O is really just one example,
but the underlying problem may also be triggered in other ways (via other
newlib/libc functions, for example).

So, really a generic solution seems to be called for.

>> (b) Also for GCC/OpenACC/nvptx use the GCC/OpenMP/nvptx '-msoft-stack'.
>> I don't really want to do that however: it does introduce a bit of
>> complexity in all the generated device code and run-time overhead that we
>> generally would like to avoid.

Directly using '-msoft-stack' isn't actually possible: it does implement
"one stack per 32-threads warp", but for OpenACC we need "one stack per
thread of a warp" (that is, each OpenACC 'vector' independently), and
pre-allocating from device memory all those stacks (which may be a lot!)
I foresee to really negatively impact overall performance?

>> (c) I'm contemplating a tweak/compiler pass for transforming such large
>> stack objects into heap allocation (during nvptx offloading compilation).
>> 'malloc'/'free' do exist; they're slow, but that's not a problem for the
>> code paths this is to affect.  (Might also add some compile-time
>> diagnostic, of course.)  Could maybe even limit this to only be used
>> during libgfortran compilation?  This is then conceptually a bit similar
>> to (b), but localized to relevant parts only.  Has such a thing been done
>> before in GCC, that I could build upon?
>>
>> Any other clever ideas?

> Converting to heap allocation is difficult outside of the frontend and you
> have to be very careful with memleaks.

Heh, in fact it seems to be pretty simple!  (Famous last words?)  See
"[WIP] nvptx: '-mframe-malloc-threshold', '-Wframe-malloc-threshold'"
attached.  What do people think about such a thing?

Still to be discussed are '-Wframe-malloc-threshold' (default-on vs.
'-Wextra'; or '-fopt-info' 'missed: [...]' or 'note: [...]' instead?),
default value for '-mframe-malloc-threshold=[...]' (potentially different
for GCC/nvptx target libraries build vs. user-compiled code?), etc.

> The library is written in C and
> I see heap allocated temporaries there but in at least one
> place a stack one is used:
>
> void
> st_endfile (st_parameter_filepos *fpp)
> {
> ...
>       if (u->current_record)
>         {
>           st_parameter_dt dtp;
>           dtp.common = fpp->common;
>           memset (&dtp.u.p, 0, sizeof (dtp.u.p));
>           dtp.u.p.current_unit = u;
>           next_record (&dtp, 1);
>
> that might be a mistake though - maybe it's enough to change that
> to a heap allocation?  It might be also totally superfluous since
> only 'u' should matter here ... (not sure if the above is the case
> you are running into).

(Have not yet looked into that; won't solve the general issue.)

Grüße
 Thomas

-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-WIP-nvptx-mframe-malloc-threshold-Wframe-malloc-thre.patch
Type: text/x-diff
Size: 16586 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc-patches/attachments/20221223/d558af9d/attachment-0001.bin>