Multithreaded DirectX
Author
Hey,
I've implemented a resource loader that can create objects in a seperate thread, one being a CTexture object. I ran a test application that loads two textures back to back (so they both load in the seperate thread) and then renders them when they've been loaded.
Now, about every 10th time, D3DXCreateTextureFromFile(...) returns E_OUTOFMEMORY...
Is the video card being locked in the main thread as the loading thread tries to access it? What precausions do I need to take with DirectX and multithreading?
Thanks
Firstly, use D3DCREATE_MULTITHREADED flag when creating the device - this will instruct D3D to take more critical sections to protect your data.
Secondly, protect your own data (resource creation) by critical sections or other synchronization objects such as semaphores.
Thirdly, avoid using different thread from the one that created the device, when using D3D core calls. You can load all data from a file to your own buffer, and use the D3D thread to then create the actual resource instance by loading it from that buffer when the loader thread signals success.
Secondly, protect your own data (resource creation) by critical sections or other synchronization objects such as semaphores.
Thirdly, avoid using different thread from the one that created the device, when using D3D core calls. You can load all data from a file to your own buffer, and use the D3D thread to then create the actual resource instance by loading it from that buffer when the loader thread signals success.
yeah I think threading is more useful in terms of having a thread for your graphics and thread for other bits of your game (AI, business logic, network, whatever). I think if you start splitting your actual rendering over different threads you are probably asking for trouble.
Unless you are actually running on one of these fancy-pants duel core machines or more though, threading will add complication and randomness (eg "about every 10th time...") to your code without much gain - because on a single core machine you are just interrupting one task with another at the end of the day, to give the illusion of running more then one thing at once.
Unless you are actually running on one of these fancy-pants duel core machines or more though, threading will add complication and randomness (eg "about every 10th time...") to your code without much gain - because on a single core machine you are just interrupting one task with another at the end of the day, to give the illusion of running more then one thing at once.
He's talking about loading resources on a different thread, which I think does simplify it, since it means you can just read/write whatever without having to break it up into frame sized pieces.
As to the problem, I've never had an issue like that with a similar setup. Also, I do make D3D calls from my resource thread with do problems, which I assumed was the reason D3DCREATE_MULTITHREADED was required. If you never call D3D from another thread, you actually won't need the flag.
tj963
As to the problem, I've never had an issue like that with a similar setup. Also, I do make D3D calls from my resource thread with do problems, which I assumed was the reason D3DCREATE_MULTITHREADED was required. If you never call D3D from another thread, you actually won't need the flag.
tj963
Quote:Original post by tj963
As to the problem, I've never had an issue like that with a similar setup. Also, I do make D3D calls from my resource thread with do problems, which I assumed was the reason D3DCREATE_MULTITHREADED was required. If you never call D3D from another thread, you actually won't need the flag.
tj963
If you can absolutely guarantee that the device won't be accessed by more than one thread, you can choose not to use the flag. However, guaranteeing this is easier said than done, by far [smile]
Author
Actually, I didn't even know the D3DCREATE_MULTITHREADED flag existed =P
I added it, and the errors disappeared.
Thanks guys, it was an easy fix. =P
I added it, and the errors disappeared.
Thanks guys, it was an easy fix. =P
"If you can absolutely guarantee that the device won't be accessed by more than one thread"
Yea I'm still new to programming in this type of environment but I think that visual studio use threads for different form resources and you might not even know your calling from another thread. I use a seperate thread (I think) that is a backgroundtimer from the tool pallet when you create your windows form. I guess that is a thread (or am I mistaken) and I call my render method from that timer method which is described by the docs as an event.
Excuse me from running my sentences into a bluring mess, a bit tired :(
-Devin
Yea I'm still new to programming in this type of environment but I think that visual studio use threads for different form resources and you might not even know your calling from another thread. I use a seperate thread (I think) that is a backgroundtimer from the tool pallet when you create your windows form. I guess that is a thread (or am I mistaken) and I call my render method from that timer method which is described by the docs as an event.
Excuse me from running my sentences into a bluring mess, a bit tired :(
-Devin
Quote:Original post by DirectXFreak
Actually, I didn't even know the D3DCREATE_MULTITHREADED flag existed =P
I added it, and the errors disappeared.
Thanks guys, it was an easy fix. =P
Whether it bites you or not is hard to tell, but it is worth noting that specifying that flag can degrade performance. Sadly, easy fixes tend to bring other baggage along with them [smile]
Basically, that flag will force the D3D runtime to be much more careful (it'll take out CritSec's for example), all of which can add a slight overhead to the smooth running of your application.
Best performance, as was mentioned, is to put different "components" on different threads rather than split a single component across multiple threads.
One way around this would be to have a "File I/O" thread that manages a number of worker threads. Write yourself a good piece of code here and you can slot it into any other part of your application that requires reading/writing files.
hth
Jack
Do note that proper profiling will reveal the thread wait bottlenecks in your multi-threaded program. When you have the profiling data, it's easier to rearrange your worker thread logic than by guessing which operations take more time than others.
Also, check out the OpenMP pragmas. They offer semi-automatic thread synchronization and pooling for very little extra work over the original c++ code. I've experienced a little bit with it and D3D and this seems to work quite well.
Also, check out the OpenMP pragmas. They offer semi-automatic thread synchronization and pooling for very little extra work over the original c++ code. I've experienced a little bit with it and D3D and this seems to work quite well.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement