Make USE_PTHREADS work with MODULARIZE mode. #5016
Conversation
bvibber
force-pushed
the
modular-pthread
branch
from
d666646
to
91b054b
Compare
Works, but currently halts the main thread :D Gives a roughly 20-40% improvement on VP9 decode at 480p and 720p on MacBook Pro early 2015 (2 cores plus hyperthreading for 4 worker threads). Requires patched emscripten for modularize stuff emscripten-core/emscripten#5016 todo: * get the emscripten fixes upstream * don't block main thread * see if helps VP8 too * do benchmarks * autodetect etc
bvibber
force-pushed
the
modular-pthread
branch
from
91b054b
to
307235e
Compare
|
Updated against current incoming branch. |
|
Updated against current incoming branch. |
pthread-main.js was touching globals directly, which aren't exposed as globals when the Module object is encapsulated into a function constructor with MODULARIZE and EXPORT_NAME options. Modified pthread-main.js and library_pthread.js (and a little modification to shell.js and jsifier.js) to use accessors for a couple things, and to inject the initialized stuff for pthread workers via options on the Module object instead of directly. emcc.py is changed to support checking document.currentScript.src at load time, outside the module wrapper constructor, and passes the value in to the module via Module.currentScriptUrl; this is overriding the document.currentScript.src check in shell.js. Shouldn't break existing non-modular code, but does require that pthread-main.js be updated when recompiling (as should happen normally). Used in ogv.js 1.4.x for experimental multithread builds. Fixes emscripten-core#5009
|
(and one last tweak for the latest merges) |
|
Thanks for this, @Brion. I have it working now but when I compiled with the default EXPORT_NAME, it looked like the worker threads were trying to spawn their own workers and were breaking. It might be nice to guard against that. |
|
ping @juj |
|
re-ping @juj |
|
Hey, sorry for the delay on this one. There's quite a few changes coming in to this, so I'll need to revise this somewhat to land. @Brion : what's your method of testing/using this? Do you have a specific setup, or just enabling the |
|
Yeah, it's a little messy and may need cleanup, I kind of hacked it until it worked. :) I was doing ad-hoc testing with ogv.js's libvpx wrapper, which unfortunately has a lot of moving parts. I've disabled pthreads from the mainline ogv.js build process until wasm threading arrives but if you want to try it, the last version with the pthreads build included is commit 4b46eb44cf490fe561402d251595d2b8dd273e43. Module source (linked to libvpx): https://github.com/brion/ogv.js/blob/master/src/c/ogv-decoder-video-vpx.c |
|
Obsoleted by #7667 which does it better :) |
pthread-main.js was touching globals directly, which aren't
exposed as globals when the Module object is encapsulated
into a function constructor with MODULARIZE and EXPORT_NAME
options.
Modified pthread-main.js and library_pthread.js (and a little
modification to shell.js and jsifier.js) to use accessors for
a couple things, and to inject the initialized stuff for pthread
workers via options on the Module object instead of directly.
emcc.py is changed to support checking document.currentScript.src
at load time, outside the module wrapper constructor, and passes
the value in to the module via Module.currentScriptUrl; this is
overriding the document.currentScript.src check in shell.js.
Shouldn't break existing non-modular code, but does require that
pthread-main.js be updated when recompiling (as should happen
normally).
Fixes #5009