BLIS for RTL simulators and systemless gem5
39 views
Skip to first unread message
Stepan Nassyr
Nov 8, 2020, 4:28:52 PM11/8/20
to blis-devel
Hello,
In the context of implementing an SVE version of BLIS, I've been approached by multiple interested parties wanting to run BLIS on system-less gem5 and proprietary RTL simulators. In these cases there is no operating system or a threading environment, therefore pthread cannot be used. In case of gem5 there is also a full system mode that can support pthread - which is what we are currently using - this however comes with some overhead, so there is an incentive here as well. I am not quite sure pthread is the only hurdle to overcome in order to get it working, but it appears to be the first issue that needs to be taken care of.
I have made some naive attempts to disable pthread by hiding pthread* calls in some ifdefs, but ran into issues (I gave up after encountering segfaults, seeing bli_thrinfo_sup_grow in the call stack when calling gemm and deciding I'm in over my head).
How feasible would it be to decouple BLIS from pthread/any threading for this specific purpose? Or without modifying the code - would something like a "fake" pthread library that just serializes everything work?
In the context of implementing an SVE version of BLIS, I've been approached by multiple interested parties wanting to run BLIS on system-less gem5 and proprietary RTL simulators. In these cases there is no operating system or a threading environment, therefore pthread cannot be used. In case of gem5 there is also a full system mode that can support pthread - which is what we are currently using - this however comes with some overhead, so there is an incentive here as well. I am not quite sure pthread is the only hurdle to overcome in order to get it working, but it appears to be the first issue that needs to be taken care of.
I have made some naive attempts to disable pthread by hiding pthread* calls in some ifdefs, but ran into issues (I gave up after encountering segfaults, seeing bli_thrinfo_sup_grow in the call stack when calling gemm and deciding I'm in over my head).
How feasible would it be to decouple BLIS from pthread/any threading for this specific purpose? Or without modifying the code - would something like a "fake" pthread library that just serializes everything work?
Jeff Hammond
Nov 8, 2020, 4:44:31 PM11/8/20
to Stepan Nassyr, blis-devel
You should be able to build without pthreads with no code changes. There is a sequential build mode.
You can run a BLIS test program with strace to find the mandatory syscalls used during execution.
Jeff
Sent from my iPhone
On Nov 8, 2020, at 1:28 PM, Stepan Nassyr <s.na...@fz-juelich.de> wrote:
Hello,
In the context of implementing an SVE version of BLIS, I've been approached by multiple interested parties wanting to run BLIS on system-less gem5 and proprietary RTL simulators. In these cases there is no operating system or a threading environment, therefore pthread cannot be used. In case of gem5 there is also a full system mode that can support pthread - which is what we are currently using - this however comes with some overhead, so there is an incentive here as well. I am not quite sure pthread is the only hurdle to overcome in order to get it working, but it appears to be the first issue that needs to be taken care of.
I have made some naive attempts to disable pthread by hiding pthread* calls in some ifdefs, but ran into issues (I gave up after encountering segfaults, seeing bli_thrinfo_sup_grow in the call stack when calling gemm and deciding I'm in over my head).
How feasible would it be to decouple BLIS from pthread/any threading for this specific purpose? Or without modifying the code - would something like a "fake" pthread library that just serializes everything work?
--
You received this message because you are subscribed to the Google Groups "blis-devel" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blis-devel+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com.
Stepan Nassyr
Nov 8, 2020, 4:50:42 PM11/8/20
to blis-devel
I assumed that BLIS relies on pthread unconditionally for thread safety from reading https://github.com/flame/blis/issues/247 and looking through the source code. Have I misunderstood something?
Jeff Hammond
Nov 8, 2020, 4:54:48 PM11/8/20
to Stepan Nassyr, blis-devel
You are correct. I was lazy and responded without reading the code.
It should be possible to elide pthreads. I can try to help you with that after SC20.
Jeff
Sent from my iPhone
On Nov 8, 2020, at 1:50 PM, Stepan Nassyr <s.na...@fz-juelich.de> wrote:
I assumed that BLIS relies on pthread unconditionally for thread safety from reading https://github.com/flame/blis/issues/247 and looking through the source code. Have I misunderstood something?
To view this discussion on the web visit https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com.
Devin Matthews
Nov 8, 2020, 8:37:38 PM11/8/20
to blis-...@googlegroups.com
BLIS only uses pthreads in "single-threaded" mode for pthread_once.
This should be easy to insert a shim for on a truly single-threaded
system.
Devin
Devin
To view this discussion on the web visit https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com.
Field G. Van Zee
Nov 9, 2020, 12:14:03 PM11/9/20
to blis-...@googlegroups.com
Perhaps this should be turned into a configure-time option? Something
like "--thread-unsafe" with a description that explains why one would
want to give up thread safety voluntarily.
Shall we open an issue?
Field
>>>> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com?utm_medium=email&utm_source=footer>.
>>> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com?utm_medium=email&utm_source=footer>.
>> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com?utm_medium=email&utm_source=footer>.
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu?utm_medium=email&utm_source=footer>.
like "--thread-unsafe" with a description that explains why one would
want to give up thread safety voluntarily.
Shall we open an issue?
Field
>>>
>>> --
>>> You received this message because you are subscribed to the Google
>>> Groups "blis-devel" group.
>>> To unsubscribe from this group and stop receiving emails from it,
>>> send an email to blis-devel+...@googlegroups.com
>>> <mailto:blis-devel+...@googlegroups.com>.
>>> --
>>> You received this message because you are subscribed to the Google
>>> Groups "blis-devel" group.
>>> To unsubscribe from this group and stop receiving emails from it,
>>> send an email to blis-devel+...@googlegroups.com
>>> To view this discussion on the web visit
>>> https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com
>>> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com?utm_medium=email&utm_source=footer>.
>> --
>> You received this message because you are subscribed to the Google
>> Groups "blis-devel" group.
>> To unsubscribe from this group and stop receiving emails from it, send
>> an email to blis-devel+...@googlegroups.com
>> <mailto:blis-devel+...@googlegroups.com>.
>> You received this message because you are subscribed to the Google
>> Groups "blis-devel" group.
>> To unsubscribe from this group and stop receiving emails from it, send
>> an email to blis-devel+...@googlegroups.com
>> To view this discussion on the web visit
>> https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com
>> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com?utm_medium=email&utm_source=footer>.
>
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel+...@googlegroups.com>.
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to blis-devel+...@googlegroups.com
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu?utm_medium=email&utm_source=footer>.
Devin Matthews
Nov 9, 2020, 12:23:31 PM11/9/20
to blis-...@googlegroups.com
I would suggest something like "--without-system" as there may be other
non-threading changes that would fall under that umbrella.
Devin
non-threading changes that would fall under that umbrella.
Devin
Minh Quan HO
Nov 9, 2020, 12:23:31 PM11/9/20
to blis-devel
Maybe we can implement a custom pthread_once() based on a lock-free mutex by atomic builtins in GCC ? https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html
Quan
Minh Quan HO
Nov 9, 2020, 12:27:23 PM11/9/20
to blis-devel
LLVM support them, too : https://clang.llvm.org/docs/LanguageExtensions.html#builtin-functions
Jeff Hammond
Nov 9, 2020, 12:30:49 PM11/9/20
to Devin Matthews, blis-discuss
This is the right way to go. Unless BLIS plans to support real user on systems that do not have Pthreads, it makes sense to target the use case requested by Stepan.
Jeff
To view this discussion on the web visit https://groups.google.com/d/msgid/blis-devel/9ab68a9d-28fa-cc99-75e8-9314cd9476c7%40smu.edu.
Field G. Van Zee
Nov 10, 2020, 2:00:15 PM11/10/20
to blis-...@googlegroups.com
I'll start work on this shortly. Once I have something to present, I'll
push it to 'dev', and from there we can make further tweaks if necessary.
Field
> <s.na...@fz-juelich.de <mailto:s.na...@fz-juelich.de>>
> <mailto:blis-devel%2B...@googlegroups.com>.
> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com>
> >>>>>
> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com?utm_medium=email&utm_source=footer
> >>>> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> >>> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> >> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> <https://groups.google.com/d/msgid/blis-devel/9ab68a9d-28fa-cc99-75e8-9314cd9476c7%40smu.edu>.
>
>
>
> --
> Jeff Hammond
> jeff.s...@gmail.com <mailto:jeff.s...@gmail.com>
> http://jeffhammond.github.io/ <http://jeffhammond.github.io/>
> <https://groups.google.com/d/msgid/blis-devel/CAGKz%3DuJp5fAwsqLPB6tpS9K1pcc60WOLO_ceqfoY4-QxOYTFTQ%40mail.gmail.com?utm_medium=email&utm_source=footer>.
push it to 'dev', and from there we can make further tweaks if necessary.
Field
> >>>> wrote:
> >>>>
> >>>> I assumed that BLIS relies on pthread unconditionally for thread
> >>>> safety from reading https://github.com/flame/blis/issues/247
> <https://github.com/flame/blis/issues/247> and
> >>>> looking through the source code. Have I misunderstood something?
> >>>>
> >>>> On Sunday, November 8, 2020 at 10:44:31 PM UTC+1 Jeff Hammond
> wrote:
> >>>>
> >>>> You should be able to build without pthreads with no code
> >>>> changes. There is a sequential build mode.
> >>>>
> >>>> You can run a BLIS test program with strace to find the
> mandatory
> >>>> syscalls used during execution.
> >>>>
> >>>> Jeff
> >>>>
> >>>> Sent from my iPhone
> >>>>
> >>>>> On Nov 8, 2020, at 1:28 PM, Stepan Nassyr
> >>>>
> >>>> I assumed that BLIS relies on pthread unconditionally for thread
> >>>> safety from reading https://github.com/flame/blis/issues/247
> <https://github.com/flame/blis/issues/247> and
> >>>> looking through the source code. Have I misunderstood something?
> >>>>
> >>>> On Sunday, November 8, 2020 at 10:44:31 PM UTC+1 Jeff Hammond
> wrote:
> >>>>
> >>>> You should be able to build without pthreads with no code
> >>>> changes. There is a sequential build mode.
> >>>>
> >>>> You can run a BLIS test program with strace to find the
> mandatory
> >>>> syscalls used during execution.
> >>>>
> >>>> Jeff
> >>>>
> >>>> Sent from my iPhone
> >>>>
> >>>>> On Nov 8, 2020, at 1:28 PM, Stepan Nassyr
> >>>>> To view this discussion on the web visit
> >>>>>
> https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com
> >>>>>
> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com>
> >>>>>
> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com?utm_medium=email&utm_source=footer
> <https://groups.google.com/d/msgid/blis-devel/0742385c-6c65-45ae-bc0e-f45b0e97607en%40googlegroups.com?utm_medium=email&utm_source=footer>>.
> >>>>
> >>>> --
> >>>> You received this message because you are subscribed to the
> Google
> >>>> Groups "blis-devel" group.
> >>>> To unsubscribe from this group and stop receiving emails from it,
> >>>> send an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>
> >>>>
> >>>> --
> >>>> You received this message because you are subscribed to the
> >>>> Groups "blis-devel" group.
> >>>> To unsubscribe from this group and stop receiving emails from it,
> >>>> send an email to blis-devel+...@googlegroups.com
> >>>> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> >>>> To view this discussion on the web visit
> >>>>
> https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com
> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com>
>
> >>>>
> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com?utm_medium=email&utm_source=footer
> >>>>
> https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com
> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com>
>
> >>>>
> <https://groups.google.com/d/msgid/blis-devel/3f2b3b1e-44e2-4889-a423-d609cabecd8fn%40googlegroups.com?utm_medium=email&utm_source=footer>>.
>
> >>>>
> >>> --
> >>> You received this message because you are subscribed to the Google
> >>> Groups "blis-devel" group.
> >>> To unsubscribe from this group and stop receiving emails from it,
> >>> send an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>
>
> >>>>
> >>> --
> >>> You received this message because you are subscribed to the Google
> >>> Groups "blis-devel" group.
> >>> To unsubscribe from this group and stop receiving emails from it,
> >>> send an email to blis-devel+...@googlegroups.com
> >>> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> >>> To view this discussion on the web visit
> >>>
> https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com
> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com>
>
> >>>
> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com?utm_medium=email&utm_source=footer
> >>>
> https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com
> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com>
>
> >>>
> <https://groups.google.com/d/msgid/blis-devel/D407EA4E-0DEF-466C-A886-FE3980F48CA1%40gmail.com?utm_medium=email&utm_source=footer>>.
>
> >>>
> >>
> >> --
> >> You received this message because you are subscribed to the Google
> >> Groups "blis-devel" group.
> >> To unsubscribe from this group and stop receiving emails from it,
> >> send an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>
>
> >>>
> >>
> >> --
> >> You received this message because you are subscribed to the Google
> >> Groups "blis-devel" group.
> >> To unsubscribe from this group and stop receiving emails from it,
> >> send an email to blis-devel+...@googlegroups.com
> >> <mailto:blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>>.
> >> To view this discussion on the web visit
> >>
> https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu>
>
> >>
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu?utm_medium=email&utm_source=footer
> >>
> https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu>
>
> >>
> <https://groups.google.com/d/msgid/blis-devel/848c0286-fd6a-ac6a-816d-164749018a4a%40smu.edu?utm_medium=email&utm_source=footer>>.
>
> >>
> >
>
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it,
> send an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel%2Bunsu...@googlegroups.com>.
>
> >>
> >
>
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it,
> send an email to blis-devel+...@googlegroups.com
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/blis-devel/9ab68a9d-28fa-cc99-75e8-9314cd9476c7%40smu.edu
> <https://groups.google.com/d/msgid/blis-devel/9ab68a9d-28fa-cc99-75e8-9314cd9476c7%40smu.edu>.
>
>
>
> --
> Jeff Hammond
> jeff.s...@gmail.com <mailto:jeff.s...@gmail.com>
> http://jeffhammond.github.io/ <http://jeffhammond.github.io/>
>
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel+...@googlegroups.com>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/blis-devel/CAGKz%3DuJp5fAwsqLPB6tpS9K1pcc60WOLO_ceqfoY4-QxOYTFTQ%40mail.gmail.com
> --
> You received this message because you are subscribed to the Google
> Groups "blis-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to blis-devel+...@googlegroups.com
> <mailto:blis-devel+...@googlegroups.com>.
> To view this discussion on the web visit
> <https://groups.google.com/d/msgid/blis-devel/CAGKz%3DuJp5fAwsqLPB6tpS9K1pcc60WOLO_ceqfoY4-QxOYTFTQ%40mail.gmail.com?utm_medium=email&utm_source=footer>.
Field G. Van Zee
Nov 10, 2020, 2:07:53 PM11/10/20
to blis-...@googlegroups.com
Let's take this discussion to [1] henceforth. Thanks everyone.
[1] https://github.com/flame/blis/issues/454
[1] https://github.com/flame/blis/issues/454
Reply all
Reply to author
Forward
0 new messages