Skip to main content
GameDev.net gamedev.net
🔒 Locked

Another multithreading problems with OpenGL

OpenGL
Started by MightyMartin Nov 27, 2006 at 12:48 AM 26 replies 9.3k views
Original Post
MightyMartin
MightyMartin
== short problem desc == - Can't display textures that loaded in another thread. == long problem desc == Hello. I've already search the GameDev and found many thread that discussed problems with multithreading with opengl, and most of answer suggest not to use multithreading with opengl. But it seems that I forced to use multithreading, based on my current condition: - I'm already in halfway through a game developing using OpenGL in Delphi. - The resource images that my program must preload before the game really is huge and the list keeps growing. That makes the waiting time became so long and long before I finally shown the opening screen. I'm also - Now I think about moving some of the loading parts into background threads that loading the leftover resources silently, when the foreground thread shows the opening screen. - The multithread is working nice. I just need to keep track of what texture that hasn't been loaded yet, and when that texture become available to the foreground thread system, I advance the opening screen to used the already loaded texture. For example: first I loaded the background screen in foreground thread. After that, I show the background screen moving, and I instruct the background thread to load the character face. So while the background screen is displayed, I silently load the character face. When the character's face is loaded completely, I slowly show the character's face to the screen in front of the background. And I instruct the background thread to load another images, and so on and on. - But the problem arise here. I'm definitely sure that the loading operation is working completely nice, because I'm only moved it and not making any modification, but it seems that *no* texture are being loaded to the openGL memory, even though the texture ID is valid (not zero). When I displayed one of the texture that already loaded by the background thread to the screen, it only shows a white rectangular, just like if the texture never been loaded to the memory. I checked out the texture ID once again, and yet it's still shows a non-zero ID (means that the texture already been loaded successfully into the memory). - I asked my friend that already has experienced the same sympton, and he said that it's really the limitation of OpenGL - that the OpenGL is not thread-safe. He said that once we leave the current thread which has the "true" OpenGL context, all the OpenGL calling method is not works, such as glBindTexture and such. But the one that keeps me wondering is that the OpenGL function that generates texture ID, the glGenTexture, is working! Why one function working, but another function is not working? I still doesn't understand until now. - I've read on another topic that the problem is located on the rendering context, that maybe I loaded the texture onto another context. - But, I've already passed the context that was created in the previous thread (foreground, renderer thread) onto the new thread (background, loader thread). And in the background thread, I used that context to load the texture, by using function such as the glBindTexture and gluBuild2DMipmaps, But still didn't change anything. But when I see some of the answers in this GameDev, even though it didn't answer my problems directly, I see that maybe there is a workaround solution over this limitation. Is there anyone can helps? Thanks.
If you can do it, so do I.
abdulla
abdulla
Are you calling glXMakeCurrent, aglSetCurrentContext or the Windows equivalent in a thread before you start issuing drawing commands? Also you can try having multiple contexts and sharing data, a good reference on this (if you're on Mac OS X) is this.
MightyMartin
MightyMartin
Update:

Uhhh, alright... when reading, it seems that the answer is the call to the wglMakeCurrent. But there is still one thing that bothers me. wglMakeCurrent needs hdc, in which hdc means the handle to the device context. Isn't the hdc is only one: the current and the only window's handle that display my game screen? If I need to make a call of wglMakeCurrent in the new thread, which new "handle" I supposed to pass? Is it the old handle? Or I must create a new window and get it's display context? I don't really understand either how was the wglMakeCurrent related to the Multithreading by reading the MSDN documentation about wglMakeCurrent.
If you can do it, so do I.
MightyMartin
MightyMartin
@Abdulla:

I don't really sure what you mean. Are you talking about wglMakeCurrent (glXMakeCurrent in windows) in the Foreground thread or wglMakeCurrent in Background thread?

If you are talking about wglMakeCurrent in foreground thread, absolutely I've already called that. But if it's in background thread, I haven't yet. Because I think that once the context has been already set, all the gl and glu successive calls will used the current context, regardless which thread calls. Is it true, or not?
If you can do it, so do I.
MightyMartin
MightyMartin
Another update:

I've read more thoroughly that to use multithread with wglMakeCurrent, I *must* have a rendering context for each thread. But a new rendering context means a new device context, and a new device context means a new window, is it right? I don't really get it...
If you can do it, so do I.
haegarr
haegarr
The following is written w/o knowledge especially of WGL, so there might be something better out there.

The GL context stores the state of GL. Each thread can have its own GL context, allowing the use of GL also in multi-threaded environments.

Making a GL context current defines its use by the calling thread! AFAIK it should be possible to use the same GL context in several threads but it can be made current for at most one thread at a time. So if you want to "share" the context you have to ensure exclusive use, e.g. by protecting the context with a semaphore. This also requires the owning thread to release the context as soon as possible to make it available to other threads.

On invocations of wglMakeCurrent the device context and the render context need not always be the same. However, there are constraints like that the device context supports GL (of course) and has a matching pixel format.

Normally a context sharing is provided for display lists, textures, and buffer objects, but I don't found such a thing in WGL (what doesn't mean there isn't one).
DaBono
DaBono
As previous posters said: the OpenGL context is only current for one thread at the time.

The most sensible solution to your problem (I think, since I'm using it [wink]) is to load and decode the texture in the resource thread. Then, give the render thread a signal that there is an image waiting to be uploaded to video memory. The render thread can then look at the list of waiting textures and upload them at the start of each frame.
That way you have the CPU and I/O-intensive stuff in a separate thread, and on-screen textures.
MightyMartin
MightyMartin
Thanks for your reply! So much mystery is answered. It still has some question left though.


@Haegarr:

- You said that a rendering context can be used by many thread, but only *one* thread can make the rendering context current. Sorry, maybe my knowledge is still low, but isn't that rendering context is made current to the "context device", not to a "specific thread" -- and a thread is not necessarily must own a unique DC, isn't it? So, which device context I must use? Or, if I try to make assumption, I just re-call the wglMakeCurrent to the foreground thread's device context (that way, they used the same device context), but the call must be made within the new thread? -- I mean, it's not about the "device context" itself, it's about what device context I must supplied to the wglMakeCurrent in the new thread?

but I'll try your valuable advice that stated that the previous thread must release the context before the new thread use it. I didn't know that!


@DaBono:

- If I must call wglMakeCurrent for every frame, doesn't it means that the animation will not run smoothly when the signal is set? because what I render in foreground is not merely a "loading animation", but a full game opening.

but your advice is somewhat needs to be considered though... Thanks.
If you can do it, so do I.
haegarr
haegarr
In general, the device context defines where the output will be written to, e.g. which window will show the graphics. The render context is OpenGL's collection of all made textures, display lists, VBOs, and so on. It also stores the current depth comparison function, background clear color, whether or not you've currently invoked a glBegin but not a glEnd yet, and much more of those things. In short: It stores the OpenGL rendering state.

Each thread has the possibility to have at most 1 OpenGL context current. When a thread invokes wglMakeCurrent, it binds the overhanded rendering context to the device context, what means that output generated by using the rendering context has to go to the device context. Moreover, the context is set for the thread from which wglMakeCurrent was called. That means that exactly that thread is now using that rendering context. Until the thread releases the context no other thread is allowed to invoke wglMakeCurrent with just that context.

In fact, wglMakeCurrent binds 3 parts together: The device context, the rendering context, and the invoking thread. The latter dependency is of interest in multi-threaded environments only, and is hence often mentioned with less emphasis.

A thread can deal with multiple device contexts. It is also able to deal with several rendering contexts, but only 1 can be made current at a time. And that rendering context can only be current to 1 thread at a time.


EDIT: I bet that performance will suffer from often rebinding contexts. Since WGL seems to lack support of native context sharing, I think an approach as supposed by DaBono would be best if the desire is just to decouple loading and using of a texture, so at most the texture resource may be needed to be protected by a semaphore.
DaBono
DaBono
Quote:
Original post by MightyMartin
@DaBono:
- If I must call wglMakeCurrent for every frame, doesn't it means that the animation will not run smoothly when the signal is set? because what I render in foreground is not merely a "loading animation", but a full game opening.


You don't need to call wglMakeCurrent each frame, but just once. I'll try to make myself a bit clearer. Your resource thread in pseudo-code will do something like:
while( resourcesToLoad ) {   FileHandle fp = filesystem.open( textureName );   Image *image  = JPGLoader::load(fp);   textureQueue.push( image );}
Then, your render thread is something like:
createWindow();wglMakeCurrent( hwnd, ... );InitGL();while( gameRunning ) {   while( textureQueue.size() > 0 )      glTexImage2D( textureQueue.front(), .... );   renderFrame();}

So you see, only one call for wglMakeCurrent is needed.
Unfortunately, the code above is a bit too simple (e.g. you need to do proper locking around the queue), but I think you can get started with this.
MightyMartin
MightyMartin
@Haegarr:

I understand now! It's really depicted something that I feared most, that I must re-make current the rendering context everytime the control changes from drawing next frame with loading resources. So there is no escape. This means that I must tricking by inserting the loading process to be happen while the screen is showing a more static screen than the opening screen (eg: the game logo). Thanks for your help! Your information about wglMakeCurrent actually binds 3 things really strikes the point!


@DaBono:

Hmmm... I think I understand the concept. So you mean to put only the "non-openGL image processing" parts in the thread, and when the image data is ready, the only process that matters is only to generate the mipmaps that shares the same thread process.

That's a brilliant idea, ummm... I don't know whether the build mipmap process is run fast enough or not to chase "a smooth framerate while loading", but it worth a try. Thanks. :)


actually... both of your ideas can be combined, Haegarr and DaBono... :)
If you can do it, so do I.
MightyMartin
MightyMartin
@Rick Appleton:

I have once again read thoroughly the links that you suggest, and found out that I still can done my previous way, but by using wglShareLists. Unfortunately I've neither ever using list nor learnt something about list. I always used the traditional method, by acquired new TextureID to a new variable using glGenTextures. Right now, I'm doing some research about the wgl list, but if someone willing to share a general or just a very basic starting information about how to using list in wgl, like what function is commonly used when working with wgl list, I will appreciate it. Thanks! :)
If you can do it, so do I.
rick_appleton
rick_appleton
wglShareList is a bit badly named. It's a holdover from the old days when there weren't any objects except for display lists.

Nowadays we have all kinds of objects in OpenGL (texture, VBO, GLSL shader, ASM shaders, etc). All of these are shared with wglShareList.

So if you call wglShareList at the correct time in the correct thread, you can just keep using glGenTextures and family and the texture will be shared between the different contexts.
haegarr
haegarr
Quote:
Original post by rick_appleton
wglShareList is a bit badly named. It's a holdover from the old days when there weren't any objects except for display lists.

Nowadays we have all kinds of objects in OpenGL (texture, VBO, GLSL shader, ASM shaders, etc). All of these are shared with wglShareList.

Ah, that's exactly what I had in mind at the bottom of my 1st reply. Good luck that there actually _is_ such context sharing also in WGL :)
GamerSg
GamerSg
Quote:
Original post by MightyMartin

That's a brilliant idea, ummm... I don't know whether the build mipmap process is run fast enough or not to chase "a smooth framerate while loading", but it worth a try. Thanks. :)


actually... both of your ideas can be combined, Haegarr and DaBono... :)


This is also the way i do it. You will be pleased to know that by using this before glTexImage2D, you can let the gfx card build mipmaps for you. This is much much faster than using gluBuild2DMipMaps.

glTexParameteri(GL_TEXTURE_2D,GL_GENERATE_MIPMAP_SGIS,GL_TRUE);
MightyMartin
MightyMartin
@Rick_appleton, Haegarr:


I've read this tutorial:

http://www.lighthouse3d.com/opengl/displaylists/

and I think I've gain an insight about the display list and wglShareLists. The display list itself is part of openGL programming. The only thing that really belongs to the wgl is the wglShareLists. But I have some common questions:

- from what I learnt, the display list is some kind of macro storing commands to be used later. Is it right?

- I don't really understood that the wglShareLists really "share" textures (and VBO, GLSL shader, ASM shaders, etc). Isn't that what wglShareLists really share is a list of ... command list? I mean, if you stated that wglShareLists share almost "anything", including textures, then I don't even need the use of display list -- I just called wglShareLists at the beginning of the thread call, and then, *poof*, all the successive openGL commands like glBindTexture, glGenTexture, and gluBuild2DMipMaps in the background loader thread will be valid. Is it right, or ... else?

- btw, I have read the zppz's tutorial, and searching again through the MSDN, and get conclusion that wglShareLists (hglrc1, hglrc2) must be called *before* the rendering context 2 (hglrc2) load any resources, such as textures. So, like you said, it's alright if the hglrc1 already has some resource loaded. That resource is also shared to hglrc2 when the wglSharelists is called. But from what I read, it's not right if the hglrc2 also has already some resource loaded too. So, in hglrc2, there must be no resource preloaded before the share attempt. I don't know what will happen if this is done. I tried to search from where I read this, but I can't found it again.

- and furthermore, by the existence of wglShareLists (which from what you both depicted, should be called wglShareResources... =] ), isn't this really a simple one-for-all solution to OpenGL Multithreading -- I mean, just call wglShareLists, and then all the multithreading concept that reside inside most of the newbies will be work? (oh! I forgot presence of the wglMakeCurrent! is it still have to be taken into account to?)

Thanks.
If you can do it, so do I.
haegarr
haegarr
Quote:
Original post by MightyMartin
- from what I learnt, the display list is some kind of macro storing commands to be used later. Is it right?

Yep. There is some kind of internal compilation that allows execution of GL routines (what you've named "commands") being more effective than passing each one for its own. Not all GL routines are allowed to be invoked in a display list, and some other caveats may appear, too.

Display lists are nice for things like bitmap based fonts (e.g. 1 display list for each glyph), macros for primitive shapes, and similar things.

Quote:
Original post by MightyMartin
- I don't really understood that the wglShareLists really "share" textures (and VBO, GLSL shader, ASM shaders, etc). Isn't that what wglShareLists really share is a list of ... command list? I mean, if you stated that wglShareLists share almost "anything", including textures, then I don't even need the use of display list -- I just called wglShareLists at the beginning of the thread call, and then, *poof*, all the successive openGL commands like glBindTexture, glGenTexture, and gluBuild2DMipMaps in the background loader thread will be valid. Is it right, or ... else?

Notice that a context share does not share everything. A GL context, as said somewhere above, contains "resources" like display lists, textures, VBOs, ... but also the "render state", like the clear color, whether depth comparison is on and which functions it uses, and bah, everything else. Context sharing means to share the "resources" but not the "render state"!

In the very early days there were no VBOs, no texture objects (i.e. only 1 texture at a time), but there were already display lists. The routine wglShareLists originates from that time. The shared resources got more and more since then, but the name of the routines was never changed. (Notice that I repeat rick_appleton's statement here.)

glBindTexture is executed in a thread and alters the "render state" but not yet the "resources", so it has no shared effect. _But_, if the thread executes a glGenTextures, then the texture object is allocated so that is available across the shared contexts. If then a thread executes glTexImage then the texture object is filled with data and hence pixel data becomes available across the shared contexts. That's the way it works.

Quote:
Original post by MightyMartin
- btw, I have read the zppz's tutorial, and searching again through the MSDN, and get conclusion that wglShareLists (hglrc1, hglrc2) must be called *before* the rendering context 2 (hglrc2) load any resources, such as textures. So, like you said, it's alright if the hglrc1 already has some resource loaded. That resource is also shared to hglrc2 when the wglSharelists is called. But from what I read, it's not right if the hglrc2 also has already some resource loaded too. So, in hglrc2, there must be no resource preloaded before the share attempt. I don't know what will happen if this is done. I tried to search from where I read this, but I can't found it again.

In all other windowing systems I know (X Window's GLX, MacOS X's AGL, and AFAIK PS/2's PGL), you must specify the shared context already when _creating_ another context. Only in WGL there is such a problem of "when to do the sharing". Well, I suggest you definitely to execute the sharing very close after creating the 2nd context! To make it clear: It better not to relate sharing to the beginning of the 2nd thread, but to the creation of the 2nd context.

Why to share _before_ allocating local resources: E.g. if both belonging contexts have already allocated a (different) texture by glGenTextures, how should they share them if both have the texture name 1? It couldn't work.

Quote:
Original post by MightyMartin
- and furthermore, by the existence of wglShareLists (which from what you both depicted, should be called wglShareResources... =] ), isn't this really a simple one-for-all solution to OpenGL Multithreading -- I mean, just call wglShareLists, and then all the multithreading concept that reside inside most of the newbies will be work?

Sorry, but I don't understand what you mean.

Quote:
Original post by MightyMartin
(oh! I forgot presence of the wglMakeCurrent! is it still have to be taken into account to?)

wglMakeCurrent must still be executed, right you are. Most GL functions does not work otherwise. However, you have to invoke it normally only once per thread.
rick_appleton
rick_appleton
First a thanks to Haegarr for the explanations. They are more detailed than I could be bothered to do, and are pretty much spot on.

I'll quickly go through your list of questions.
Quote:
Original post by MightyMartin
- from what I learnt, the display list is some kind of macro storing commands to be used later. Is it right?

That's pretty much correct yes.

Quote:
Original post by MightyMartin
- I don't really understood that the wglShareLists really "share" textures (and VBO, GLSL shader, ASM shaders, etc). Isn't that what wglShareLists really share is a list of ... command list? I mean, if you stated that wglShareLists share almost "anything", including textures, then I don't even need the use of display list -- I just called wglShareLists at the beginning of the thread call, and then, *poof*, all the successive openGL commands like glBindTexture, glGenTexture, and gluBuild2DMipMaps in the background loader thread will be valid. Is it right, or ... else?

Like Haegarr said, resources will be shared. OpenGL has state, and that is not shared.

Quote:
Original post by MightyMartin
- btw, I have read the zppz's tutorial, and searching again through the MSDN, and get conclusion that wglShareLists (hglrc1, hglrc2) must be called *before* the rendering context 2 (hglrc2) load any resources, such as textures. So, like you said, it's alright if the hglrc1 already has some resource loaded. That resource is also shared to hglrc2 when the wglSharelists is called. But from what I read, it's not right if the hglrc2 also has already some resource loaded too. So, in hglrc2, there must be no resource preloaded before the share attempt. I don't know what will happen if this is done. I tried to search from where I read this, but I can't found it again.

This is indeed what most sources of information say. From my tests however, it appears that all resources loaded in either thread at any time are shared across all contexts. It is probably safest to not rely on this fact though, so you should just code as if only resources loaded after wglShareLists are shared. The issue mentioned by Haegarr could theoretically happen, but I suspect that the drivers aren't aware of this up to that point. I haven't tested this, but I wouldn't be surprised if you create two textures in two totally unrelated apps, you actually get texture IDs 1 and 2 for example, and not 1 and 1.

Quote:
Original post by MightyMartin
- and furthermore, by the existence of wglShareLists (which from what you both depicted, should be called wglShareResources... =] ), isn't this really a simple one-for-all solution to OpenGL Multithreading -- I mean, just call wglShareLists, and then all the multithreading concept that reside inside most of the newbies will be work?

It does indeed seem to work that way. Note that you still have to be careful that you don't use a texture at the same time you are uploading it's data in another thread. Something like that will likely lead to 'undefined' behaviour, which basically means that anything can happen.

Quote:
Original post by MightyMartin
(oh! I forgot presence of the wglMakeCurrent! is it still have to be taken into account to?)

As Haegarr already mentioned, you still have to call wglMakeCurrent once from inside each thread.
MightyMartin
MightyMartin
About the usage of wglMakeCurrent:

can I do something like these:

In the global variable:
int protection = 0;


In foreground, renderer thread:
void mainloop (){  if (protection != 0) return;  // beginning entering mainloop, set the flag.  protection = 1;  wglMakeCurrent (hDC1, hglrc1);  // do the drawing things here;  ...  ...  wglMakeCurrent (NULL, NULL);  // release the protection  protection = 0;}


and in the background, loader thread:

// this is used to check the availability to "make current"int protector (){  if (protection != 0) return 0;  protection = 2;  wglMakeCurrent (hDC1, hglrc1);  return 1;}void releaseprotection (){  protection = 0;  wglMakeCurrent (NULL, NULL);}void backgroundthread (){  // the background thread will stuck in here until foreground thread has  // already finished it's circle  while (!protector()) {}  LoadTexture1 ();  releaseprotection ();    // the background thread will stuck in here until foreground thread has  // already run once.  while (protection != 1) {}  while (!protector()) {}  LoadTexture2 ();  releaseprotection ();  while (protection != 1) {}  while (!protector()) {}  LoadTexture3 ();  releaseprotection ();  ...  ...  ...}


so this way, the foreground thread still has some time left to render next frame between each texture loading. but, raegarr, you said that this solution will suffer the performance because of the intensive use of wglMakeCurrent, right?


Quote:
A thread can deal with multiple device contexts. It is also able to deal with several rendering contexts, but only 1 can be made current at a time. And that rendering context can only be current to 1 thread at a time.


by the way, from your sentence, is it possible to wglMakeCurrent (hDC1, hglrc1) in the first thread and wglMakeCurrent (hDC2, hglrc2) in the second thread at one simultaneous time? (where hDC1 != hDC2, and hglrc1 != hglrc2)


Quote:
As Haegarr already mentioned, you still have to call wglMakeCurrent once from inside each thread.


only once inside each thread? do you mean only need for the beginning of the thread only? don't we need to call wglMakeCurrent each time we moved across the thread?
If you can do it, so do I.

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.

Sign in to reply to this topic.