Multi threading #5

Spacechild1 · 2022-02-22T12:09:59Z

dummy PR for development

configure with --enable-parallel!

check block~-help.pd and clone-help.pd for help and usage examples.

* public API to manage the DSP thread pool * private API to create DSP task queues and DSP tasks * PD_DSPTHREADS define to enable/disable the thread pool at compile time

* add global DSP task queue * reset and join DSP global task queue in dsp_tick() * add "-threads" command line option to set number of DSP threads

* add thread local variable for the current DSP thread index * if clocks are set/unset in the main thread (= 0), do as usual * if clocks are set/unset in a DSP helper thread (> 0), cache the desired time point and put the clock on a lock-free stack * at the end of dsp_tick() we take the list of deferred clocks and set/unset them for real.

* swapping/resizing an array doesn't require a DSP graph update * in control objects, garrayref can be faster because the garray is cached * in perform routines, garrayref allows thread-safe read or write access to garrays

protect st_soundout with spinlocks

expands to CLASS_THREADSAFE if PD_PARALLEL is 1

* add a_numthreads to t_audiosettings * add "threads" to preferences * add "Audio threads" option in audio settings dialog

by default, the DSP graph uses the global signal context. However, if a graph is processed in parallel, it needs its own signal context. For this purpose, we can also temporarily push/pop a new signal context with signalcontext_push() and signalcontext_pop().

* add global DSP task queue * parallel canvasses use their own signal context * vinlet makes a copy of the parent input signal (in the new context) * voutlet uses double buffering: voutlet_dspprolog() writes last buffer to parent output signal, voutlet_dsp() writes input signal to buffer.

* if "join" is true, block~ has its own DSP task queue * in ugen_donegraph, we push the DSP task queue, reset it and finally join it. this will synchronize all parallel DSP tasks in subpatches or child abstractions.

* each DSP task object registers itself with every outer switch~ object. If a switch~ object changes it state, it notifies all its DSP task children, so they can in turn notify their DSP task queue. * DSP tasks in switched-off parallel canvasses are simply not scheduled. * each DSP task queue maintains a count of switched-off DSP tasks, so that it won't lock up in case there is a switch~ *between* the queue and the task.

* if "parallel" is true, clone has its own signal context and DSP task queue. * child abstractions are scheduled as DSP tasks and finally joined. * the cloned abstractions don't need to know that they are being processed in parallel, we can carry out all necessary steps in clone_dsp().

dsptaskqueue_update() checks if the canvas sub tree contains non-thread-safe objects. Note that any non-thread-safe objects *outside* the tree are ignored. This function is called in ugen_done_graph() for canvases with block~ + "join" and also in ugen_start() for the toplevel queue. dsptaskqueue_check() checks if the task queue is thread-safe, and if false, posts the first N non-thread-safe object (only once per queue). This is called in ugen_done_graph() for canvases with block~ + "parallel". Each parallel canvas checks if the enclosing joining canvas is thread-safe and prints an error message if false. As an optimization, we traverse the complete object tree once in ugen_start() and mark every sub-tree (depth first) by setting gl_threadsafe. This speeds up subsequent calls to canvas_isthreadsafe() tremendously. clone is handled specially since it contains both the DSP task queue and the DSP tasks, so the check can be performed all inside clone_dsp(). Thread safety checks can be disabled with the "-nothreadsafe" command line option.

... and without -nothreadsafe

used in thread_physical_concurrency() and later for thread pinning

instead of going to sleep everytime the task queue is (temporarily) empty, the DSP helper threads only wake up once at the beginning of each DSP tick and then they spin-wait until all tasks have been finished. Pro: minimize wake-up latency = more stable performance Con: burning CPU cycles Can be enabled/disabled with the "-spinwait" resp. "-nospinwait" command line options.

allow to pin DSP threads to dedicated cores; useful for spinwaiting! Can be enabled/disabled with the "-affinity" resp. "-noaffinity" command line options

Spacechild1 · 2022-09-17T15:59:57Z

@umlaeute What does the dist-check job do? It seems like I have to add s_spinlock.h, but I don't know where... "Normal" autotools builds work fine.

umlaeute · 2022-09-17T16:45:47Z

You have to add all the headers to the src/Makefile.am

umlaeute

actually the headers are not installed (hence the noinst_ prefix), but that is just as it should be (it's just that the commit message is a bit misleading) :-P

Spacechild1 · 2022-09-17T22:51:40Z

(it's just that the commit message is a bit misleading)

It was just for testing :-)

actually the headers are not installed (hence the noinst_ prefix)

Actually, s_spinlock.h should be installed (it's a public header), but s_sync.h should not. I hope I got it right...

improvements to x1.htm

danomatika · 2022-11-07T22:44:15Z

Can you add screenshots of the block and clone help patches?

Spacechild1 · 2022-11-07T22:57:30Z

@danomatika done :-)

danomatika · 2022-11-07T23:01:19Z

src/m_pd.h

 #endif

+/* override for parallel processing support */
+#ifndef PD_PARALLEL


Personally, I would call this PDPARALLEL without the underscore to better match the existing PDINSTANCE and PDTHREADS defines.

Fair enough. On the other hand, there are quite a few constants/defines with a PD_ prefix:
PD_INTERNAL, PD_LONGINTTYPE, PD_FLOATSIZE, PD_BIGORSMALL, PD_<loglevel>, etc. Personally, I think PD_<name> is better style, but I don't care too much :-)

Generally, any suggestions for better naming are very much appreciated! (Naming things is hard...)

Spacechild1 changed the base branch from master to develop February 22, 2022 12:10

Spacechild1 force-pushed the multi-threading branch 3 times, most recently from ce378da to c6f64bd Compare February 23, 2022 00:44

Spacechild1 force-pushed the multi-threading branch 3 times, most recently from 5f513e0 to f55c4b3 Compare March 31, 2022 11:18

Spacechild1 changed the base branch from develop to master March 31, 2022 11:37

Spacechild1 changed the base branch from master to develop March 31, 2022 11:37

Spacechild1 force-pushed the multi-threading branch from f55c4b3 to 8b8d059 Compare August 29, 2022 21:34

Spacechild1 force-pushed the multi-threading branch 3 times, most recently from 3550101 to 73979e9 Compare September 17, 2022 13:16

Spacechild1 added 17 commits September 17, 2022 15:39

add "s_spinlock.h"

b945cd2

add s_sync.h / s_sync.c

acc1068

add DSP thread pool API

9f4db0c

* public API to manage the DSP thread pool * private API to create DSP task queues and DSP tasks * PD_DSPTHREADS define to enable/disable the thread pool at compile time

implement internal thread pool

6af56fc

* add global DSP task queue * reset and join DSP global task queue in dsp_tick() * add "-threads" command line option to set number of DSP threads

add t_garrayref API

d8e9060

* swapping/resizing an array doesn't require a DSP graph update * in control objects, garrayref can be faster because the garray is cached * in perform routines, garrayref allows thread-safe read or write access to garrays

make table DSP objects thread-safe

afe8644

make delay DSP objects thread-safe

3a735f6

make send~/receive~ and throw~/catch~ thread-safe

60ce2af

make dac~ thread-safe

9d27289

protect st_soundout with spinlocks

make FFT thread-safe

7767d49

make sys_vgui() thread-safe (mainly for posting)

3be5d4a

mark API functions as threadsafe

f800d38

add PD_PARALLEL define to m_pd.h and implement it in class_new()

ac41950

use CLASS_DEFAULT for all DSP objects

30846ac

expands to CLASS_THREADSAFE if PD_PARALLEL is 1

configure add --enable-parallel option (disabled by default)

3737cae

DSP threads are now part of audio settings

83c1174

* add a_numthreads to t_audiosettings * add "threads" to preferences * add "Audio threads" option in audio settings dialog

Spacechild1 added 13 commits September 17, 2022 15:39

add t_signalcontext

96aa0fd

by default, the DSP graph uses the global signal context. However, if a graph is processed in parallel, it needs its own signal context. For this purpose, we can also temporarily push/pop a new signal context with signalcontext_push() and signalcontext_pop().

fix dsp_done() signature

a584da3

block~: implement "join" method

78bcdf2

* if "join" is true, block~ has its own DSP task queue * in ugen_donegraph, we push the DSP task queue, reset it and finally join it. this will synchronize all parallel DSP tasks in subpatches or child abstractions.

post all non-thread-safe DSP objects if Pd is started with -verbose

d40c6eb

... and without -nothreadsafe

parse hardware topology

260ea0a

used in thread_physical_concurrency() and later for thread pinning

add thread pinning option

5b9cfee

allow to pin DSP threads to dedicated cores; useful for spinwaiting! Can be enabled/disabled with the "-affinity" resp. "-noaffinity" command line options

update block~-help.pd and clone-help.pd

bfde087

improve garrayref comments

fa51f3e

Spacechild1 force-pushed the multi-threading branch 3 times, most recently from 7f78314 to d35002f Compare September 17, 2022 15:48

enable parallel processing in makefiles

29ced3f

Spacechild1 force-pushed the multi-threading branch from d35002f to 29ced3f Compare September 17, 2022 15:54

install headers

0cc438e

umlaeute reviewed Sep 17, 2022

View reviewed changes

Spacechild1 pushed a commit that referenced this pull request Oct 17, 2022

Merge pull request #5 from porres/patch-7

c21db24

improvements to x1.htm

danomatika reviewed Nov 7, 2022

View reviewed changes

Spacechild1 changed the base branch from develop to master March 23, 2023 12:58

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Multi threading #5

Multi threading #5

Uh oh!

Spacechild1 commented Feb 22, 2022 •

edited

Loading

Uh oh!

Spacechild1 commented Sep 17, 2022

Uh oh!

umlaeute commented Sep 17, 2022 via email •

edited

Loading

Uh oh!

umlaeute left a comment

Uh oh!

Spacechild1 commented Sep 17, 2022

Uh oh!

danomatika commented Nov 7, 2022

Uh oh!

Spacechild1 commented Nov 7, 2022

Uh oh!

danomatika Nov 7, 2022 •

edited

Loading

Uh oh!

Spacechild1 Nov 7, 2022 •

edited

Loading

Uh oh!

Spacechild1 Nov 7, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Multi threading #5

Are you sure you want to change the base?

Multi threading #5

Uh oh!

Conversation

Spacechild1 commented Feb 22, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Spacechild1 commented Sep 17, 2022

Uh oh!

umlaeute commented Sep 17, 2022 via email • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

umlaeute left a comment

Choose a reason for hiding this comment

Uh oh!

Spacechild1 commented Sep 17, 2022

Uh oh!

danomatika commented Nov 7, 2022

Uh oh!

Spacechild1 commented Nov 7, 2022

Uh oh!

danomatika Nov 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Spacechild1 Nov 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Spacechild1 Nov 7, 2022

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Spacechild1 commented Feb 22, 2022 •

edited

Loading

umlaeute commented Sep 17, 2022 via email •

edited

Loading

danomatika Nov 7, 2022 •

edited

Loading

Spacechild1 Nov 7, 2022 •

edited

Loading