Sign in to follow this  

DX11 [DX11] Command Lists on a Single Threaded Renderer

This topic is 2220 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

One of the new major features of DirectX 11 is its support for multithreaded rendering using Immediate and Deferred Contexts, however it seems to me that the ability to create a Command List would potentially be beneficial even for a single threaded renderer. Is this correct?

Basically a Command List is a more efficient way of submitting a number of state and draw commands than calling each API separately, so even if you do all your rendering on one thread it would still seem more efficient to use Command Lists to perform repeat lists of actions.

If this is correct, then why isn't this mentioned more? Why are Deferred Contexts pretty much exclusively documented as a multi threading feature?

Thanks
ben

Share this post


Link to post
Share on other sites
I'm pretty sure I remembered reading somewhere that the runtime/drivers weren't optimized for this case. However I still think it would be worth trying out, especially as the drivers get better support for deferred command list generation.

Share this post


Link to post
Share on other sites
[quote name='MJP' timestamp='1321297420' post='4883879']
I'm pretty sure I remembered reading somewhere that the runtime/drivers weren't optimized for this case. However I still think it would be worth trying out, especially as the drivers get better support for deferred command list generation.
[/quote]
You are right, I just ran a test on 10,000 cubes, single threaded:
[list][*]Immediate: 200 fps[*]Deferred: 150 fps[/list]So Deferred for single threaded application is slower (at least on my machine). Note that checking the threading support for command list for my graphics card (AMD 6970M) is returning false, so I assume that It is not supported natively by the driver but "emulated" by DX11...

Share this post


Link to post
Share on other sites
I'm not sure if the AMD GPU's support it yet, but my GTX 470 on latest drivers says that it does. I'll have to check on my HD 6970 when I get home.

If your test is easily packagable, I would be happy to try it out on my machine to see how it performs.

Share this post


Link to post
Share on other sites
[quote name='xoofx' timestamp='1321370413' post='4884183']
[quote name='MJP' timestamp='1321297420' post='4883879']
I'm pretty sure I remembered reading somewhere that the runtime/drivers weren't optimized for this case. However I still think it would be worth trying out, especially as the drivers get better support for deferred command list generation.
[/quote]
You are right, I just ran a test on 100,000 cubes, single threaded:
[list][*]Immediate: 200 fps[*]Deferred: 150 fps[/list]So Deferred for single threaded application is slower (at least on my machine). Note that checking the threading support for command list for my graphics card (AMD 6970M) is returning false, so I assume that It is not supported natively by the driver but "emulated" by DX11...
[/quote]
I would caution against making blanket statements about the performance of single vs. deferred rendering - it totally depends on what your renderer does when it is submitting work to the API. For example, if your engine does lots of work in between the API calls that it makes, then it would likely be beneficial to utilize multiple threads which could reduce the total time needed to process a rendering path. On the other hand, if your submission routines are very bare bones and only submits API calls, there could be some benefit to compiling a long list of commands into a command list and then reusing it from frame to frame. This will depend on the hardware, the driver, your engine, and the application that is using your engine - you need to profile and see if it is worth it in a given context. You could even dynamically test it out on the first startup of your application and then choose the appropriate rendering method.

To that end - I would suggest setting up your rendering code to not know if it is using a deferred or immediate context so that you can delay the decision as long as possible as to which method you will use. That is just my own suggestion though - I have found it to be useful, but your mileage may vary!

Share this post


Link to post
Share on other sites
[quote name='xoofx' timestamp='1321370413' post='4884183']
[quote name='MJP' timestamp='1321297420' post='4883879']
I'm pretty sure I remembered reading somewhere that the runtime/drivers weren't optimized for this case. However I still think it would be worth trying out, especially as the drivers get better support for deferred command list generation.
[/quote]
You are right, I just ran a test on 100,000 cubes, single threaded:
[list][*]Immediate: 200 fps[*]Deferred: 150 fps[/list]So Deferred for single threaded application is slower (at least on my machine). Note that checking the threading support for command list for my graphics card (AMD 6970M) is returning false, so I assume that It is not supported natively by the driver but "emulated" by DX11...
[/quote]

Thanks all

xoofx, in your test do you re-create the command list every frame, or do you create the command list once at startup and then just re-execute it every frame?

I'm amazed that AMD cards don't support multithreading at the driver level yet!

I'd be interested in seeing the results on a NVidia card too.

Thanks
Ben

Share this post


Link to post
Share on other sites
If have released the executable test along some analysis about the results [url="http://code4k.blogspot.com/2011/11/direct3d11-multithreading-micro.html"]here[/url].

Of course, I agree with Jason.Z statements about taking carefully this kind of results, and the fact that a renderer can easily be built to switch transparently from a deferred context to an immediate context.

To respond to your initial question BenS1, It seems that hardware support for command list doesn't seems to change a lot (using a pre-prepared command list once and run it on an immediate context) compare to using the default Direct3D11 behavior.

Share this post


Link to post
Share on other sites
The problem with using a deferred context in a single threaded system is that you are doing more work per core in that situation; you have to prepare the CL, which takes some extra CPU overhead as the driver needs to do things and then you have to reaccess it again to send it to the card properly. Spread across multiple threads the cost-per-setup drops significantly and, if you batch them, your send arch will benifit greatly from code cache reuse (and depending on how it's stored maybe some data cache too).

Any speed gain also depends very much on what you are doing; in a test case at work which was setup to be heavily CPU bound, switching on mutli-threading CL support when NV's drivers were updated to fix it did give us a speed boost however it wasn't that much. I then spent some time playing with the test case and discovered that when we got over a certain threshold for data per CL we started spending more and more time in the buffer swapping function than anywhere else in the submission due to the driver having to do more. (I can't recall the specifics but from what I do recall drivers are limited memory wise or something like that... basically we blew a buffer right out).

However up until that point the MT CL rendering WAS making a significant difference with our CPU time usage and we had near perfect scaling [b]for the test case[/b].

The key point from all this; MT CL, if implimented by the drivers, will help but ONLY your CPU time.

I make a point of saying this because there is no 'hardware support' for CL; Command Lists are purely a CPU side thing, the difference is between letting the DX11 runtime cache the commands or letting the driver cache them and optimise them. (AMD still lacks support for this, although it is apprently 'coming soon')

(Also, as a side note, I do recall reading that 'create, store and reuse' isn't an optimal pattern for command lists. The runtime isn't really setup for this case and it assumes you'll be remaking them each frame, which is a fair assumption because as you can't chain them together to adjust each others state and most command lists will change each frame in a 'real world' situation it is best to test against this)

Share this post


Link to post
Share on other sites
[quote name='phantom' timestamp='1321785963' post='4885837']
I then spent some time playing with the test case and discovered that when we got over a certain threshold for data per CL we started spending more and more time in the buffer swapping function than anywhere else in the submission due to the driver having to do more. (I can't recall the specifics but from what I do recall drivers are limited memory wise or something like that... basically we blew a buffer right out).[/quote]
this could come from the Map/UnMap on command buffers, with an immediate context that is giving directly a kind of DMA to the GPU memory, but with a deferred context, it has to copy to a temporary buffer (which is probably on the RAM, but not sure it is on a shared memory on the GPU)...

[quote name='phantom' timestamp='1321785963' post='4885837']
I make a point of saying this because there is no 'hardware support' for CL; Command Lists are purely a CPU side thing, the difference is between letting the DX11 runtime cache the commands or letting the driver cache them and optimise them. (AMD still lacks support for this, although it is apprently 'coming soon')
[/quote]
Indeed, if it is natively supported by the driver, It can be optimized. A coworker found also on NVIDIA a performance boost when they introduced support for command list, though on AMD, It is already fine without the support from the driver... probably the command buffer on AMD is already layout in the same way DirectX11 command buffer is layout...

Share this post


Link to post
Share on other sites
[quote name='xoofx' timestamp='1321774033' post='4885817']
If have released the executable test along some analysis about the results [url="http://code4k.blogspot.com/2011/11/direct3d11-multithreading-micro.html"]here[/url].

Of course, I agree with Jason.Z statements about taking carefully this kind of results, and the fact that a renderer can easily be built to switch transparently from a deferred context to an immediate context.

To respond to your initial question BenS1, It seems that hardware support for command list doesn't seems to change a lot (using a pre-prepared command list once and run it on an immediate context) compare to using the default Direct3D11 behavior.
[/quote]

Wow, great article! Thanks.

Its a shame that the results show that command lists aren't really a faster way of repeating the same drawing commands over and over for a single threaded renderer.

I suspect they have the potential to be faster if the driver developers had sufficient motivation to optimise this area of their code, especially if they optimised the command list when you call FinishCommandList. I guess the problem is that the driver has no idea if you're only going to use the command list once and throw it away (In which case the act of optimising the command list may cost more than the potential gains), if if you're going to create the comand list once and execute it many times (In which case optimisng the list may be beneficial).

I guess we'd need a tweak to the API so that you can either pass in a boolean to FinishCommandList to tell the driver whether the command list should be optimised or not, or maybe there could be a separate explicit OptimizeCommandList method.

Thanks again for your detailed analysis.

Thanks
Ben

Share this post


Link to post
Share on other sites
[quote name='phantom' timestamp='1321785963' post='4885837']
The problem with using a deferred context in a single threaded system is that you are doing more work per core in that situation; you have to prepare the CL, which takes some extra CPU overhead as the driver needs to do things and then you have to reaccess it again to send it to the card properly. Spread across multiple threads the cost-per-setup drops significantly and, if you batch them, your send arch will benifit greatly from code cache reuse (and depending on how it's stored maybe some data cache too).

<snip>

(Also, as a side note, I do recall reading that 'create, store and reuse' isn't an optimal pattern for command lists. The runtime isn't really setup for this case and it assumes you'll be remaking them each frame, which is a fair assumption because as you can't chain them together to adjust each others state and most command lists will change each frame in a 'real world' situation it is best to test against this)
[/quote]

Thanks Phantom, but in my case I was thinking of creating the command list once and then executing it for each frame.

As I'm sure you know, a command list containing a constant buffer will only contain references (Or pointers) to the constant buffer and not the actual data containined int he buffer itself, so an app can still change the data in the constant buffer from frame to frame without having to create a new command list.

So for example I was thinking:
1. At startup create a command list (DrawTankCL) that draws a Tank at a position defined in a Contant Buffer ("TankCB")
2. Update TankCB.position on the CPU based on user input, physics etc
3. ExecuteCommandList(DrawTankCL)
4. Repeat from step 2.

As you can see the command list is created once and executed over and over, and yet the tanks posiiton is still dynamic.

Its a shame that this "create, store, reuse" pattern is not optimised int he drivers.

Anyway, at least now I know the answer so I code my game accordingly.

Thanks for your help
Ben

Share this post


Link to post
Share on other sites
There are two problems with your idea.

Firstly, you are being too fine grain with your CL for it to really be useful. There is a good PDF from GDC2011 which covers some of this (google: Jon Jansen DX11 Performance Gems, that should get you it). The main thing is that a CL has overhead, apprently a few dozen API calls so doing too little work in one is going to be a problem as it will just get swamped with overhead. Depending on your setup scenes or material groups are better fits for CL building and execution.

Secondly; you run the risk of suffering a stall at step 2. The driver buffers commands and the GPU should be working at the same time as you execute other work, so there is a chance that when you come to update in step 2 you could be waiting a 'significant' amount of time for the GPU to be done with your buffer and release it so that you can update it again. Discard/lock or other update [i]might[/i] avoid the problem, I've not tried it myself, but it still presents an issue.

Share this post


Link to post
Share on other sites
[quote name='xoofx' timestamp='1321790455' post='4885852']
Indeed, if it is natively supported by the driver, It can be optimized. A coworker found also on NVIDIA a performance boost when they introduced support for command list, though on AMD, It is already fine without the support from the driver... probably the command buffer on AMD is already layout in the same way DirectX11 command buffer is layout...
[/quote]

NV is a strange beast; before they had 'proper' support they kinda emulated it by spinning up a 'server' thread and serialising the CL creation via that. Amusingly if any of your active threads ended up on the same core as the server thread it tended to murder performance but by staying clear you could get a small improvement. Once the drivers came out which did the work correctly this problem went away.

In our test NV with proper support soundly beat AMD without it; this was a 470GTX vs 5870 on otherwise basically identical hardware (i7 CPUs, the NV one had a few hundred Mhz over the AMD one, but not enough for the performance delta seen). AMD's performance was more in line with the single thread version. However our test was a very heavy CPU bound one; 15,000 draw calls spread over 6 cores each one drawing a single flat shaded cube. Basically an API worse nightmare ;)

(Amusing side note; the same test/code on an X360 @ 720p could render at a solid 60fps with a solid 16.6ms frame time. That's command lists being generated each frame over 6 cores; shows just how much CPU overhead/performance loss you take when running on Windows :( )

Share this post


Link to post
Share on other sites

This topic is 2220 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

Sign in to follow this  

  • Forum Statistics

    • Total Topics
      628766
    • Total Posts
      2984583
  • Similar Content

    • By GreenGodDiary
      Having some issues with a geometry shader in a very basic DX app.
      We have an assignment where we are supposed to render a rotating textured quad, and in the geometry shader duplicate this quad and offset it by its normal. Very basic stuff essentially.
      My issue is that the duplicated quad, when rendered in front of the original quad, seems to fail the Z test and thus the original quad is rendered on top of it.
      Whats even weirder is that this only happens for one of the triangles in the duplicated quad, against one of the original quads triangles.

      Here's a video to show you what happens: Video (ignore the stretched textures)

      Here's my GS: (VS is simple passthrough shader and PS is just as basic)
      struct VS_OUT { float4 Pos : SV_POSITION; float2 UV : TEXCOORD; }; struct VS_IN { float4 Pos : POSITION; float2 UV : TEXCOORD; }; cbuffer cbPerObject : register(b0) { float4x4 WVP; }; [maxvertexcount(6)] void main( triangle VS_IN input[3], inout TriangleStream< VS_OUT > output ) { //Calculate normal float4 faceEdgeA = input[1].Pos - input[0].Pos; float4 faceEdgeB = input[2].Pos - input[0].Pos; float3 faceNormal = normalize(cross(faceEdgeA.xyz, faceEdgeB.xyz)); //Input triangle, transformed for (uint i = 0; i < 3; i++) { VS_OUT element; VS_IN vert = input[i]; element.Pos = mul(vert.Pos, WVP); element.UV = vert.UV; output.Append(element); } output.RestartStrip(); for (uint j = 0; j < 3; j++) { VS_OUT element; VS_IN vert = input[j]; element.Pos = mul(vert.Pos + float4(faceNormal, 0.0f), WVP); element.Pos.xyz; element.UV = vert.UV; output.Append(element); } }  
      I havent used geometry shaders much so im not 100% on what happens behind the scenes.
      Any tips appreciated! 
    • By mister345
      Hi, I'm building a game engine using DirectX11 in c++.
      I need a basic physics engine to handle collisions and motion, and no time to write my own.
      What is the easiest solution for this? Bullet and PhysX both seem too complicated and would still require writing my own wrapper classes, it seems. 
      I found this thing called PAL - physics abstraction layer that can support bullet, physx, etc, but it's so old and no info on how to download or install it.
      The simpler the better. Please let me know, thanks!
    • By Hexaa
      I try to draw lines with different thicknesses using the geometry shader approach from here:
      https://forum.libcinder.org/topic/smooth-thick-lines-using-geometry-shader
      It seems to work great on my development machine (some Intel HD). However, if I try it on my target (Nvidia NVS 300, yes it's old) I get different results. See the attached images. There
      seem to be gaps in my sine signal that the NVS 300 device creates, the intel does what I want and expect in the other picture.
      It's a shame, because I just can't figure out why. I expect it to be the same. I get no Error in the debug output, with enabled native debugging. I disabled culling with CullMode.None. Could it be some z-fighting? I have little clue about it but I tested to play around with the RasterizerStateDescription and DepthBias properties with no success, no change at all. Maybe I miss something there?
      I develop the application with SharpDX btw.
      Any clues or help is very welcome
       


    • By Beny Benz
      Hi,
      I'm currently trying to write a shader which shoud compute a fast fourier transform of some data, manipulating the transformed data, do an inverse FFT an then displaying the result as vertex offset and color. I use Unity3d and HLSL as shader language. One of the main problems is that the data should not be passed from CPU to GPU for every frame if possible. My original plan was to use a vertex shader and do the fft there, but I fail to find out how to store changing data betwen shader calls/passes. I found a technique called ping-ponging which seems to be based on writing and exchangeing render targets, but I couldn't find an example for HLSL as a vertex shader yet.
      I found https://social.msdn.microsoft.com/Forums/en-US/c79a3701-d028-41d9-ad74-a2b3b3958383/how-to-render-to-multiple-render-targets-in-hlsl?forum=xnaframework
      which seem to use COLOR0 and COLOR1 as such render targets.
      Is it even possible to do such calculations on the gpu only? (/in this shader stage?, because I need the result of the calculation to modify the vertex offsets there)
      I also saw the use of compute shaders in simmilar projects (ocean wave simulation), do they realy copy data between CPU / GPU for every frame?
      How does this ping-ponging / rendertarget switching technique work in HLSL?
      Have you seen an example of usage?
      Any answer would be helpfull.
      Thank you
      appswert
    • By ADDMX
      Hi
      Just a simple question about compute shaders (CS5, DX11).
      Do the atomic operations (InterlockedAdd in my case) should work without any issues on RWByteAddressBuffer and be globaly coherent ?
      I'v come back from CUDA world and commited fairly simple kernel that does some job, the pseudo-code is as follows:
      (both kernels use that same RWByteAddressBuffer)
      first kernel does some job and sets Result[0] = 0;
      (using Result.Store(0, 0))
      I'v checked with debugger, and indeed the value stored at dword 0 is 0
      now my second kernel
      RWByteAddressBuffer Result;  [numthreads(8, 8, 8)] void main() {     for (int i = 0; i < 5; i++)     {         uint4 v0 = DoSomeCalculations1();         uint4 v1 = DoSomeCalculations2();         uint4 v2 = DoSomeCalculations3();                  if (v0.w == 0 && v1.w == 0 && v2.w)             continue;         //    increment counter by 3, and get it previous value         // this should basically allocate space for 3 uint4 values in buffer         uint prev;         Result.InterlockedAdd(0, 3, prev);                  // this fills the buffer with 3 uint4 values (+1 is here as the first 16 bytes is occupied by DrawInstancedIndirect data)         Result.Store4((prev+0+1)*16, v0);         Result.Store4((prev+1+1)*16, v1);         Result.Store4((prev+2+1)*16, v2);     } } Now I invoke it with Dispatch(4,4,4)
      Now I use DrawInstancedIndirect to draw the buffer, but ocassionaly there is missed triangle here and there for a frame, as if the atomic counter does not work as expected
      do I need any additional synchronization there ?
      I'v tried 'AllMemoryBarrierWithGroupSync' at the end of kernel, but without effect.
      If I do not use atomic counter, and istead just output empty vertices (that will transform into degenerated triangles) the all is OK - as if I'm missing some form of synchronization, but I do not see such a thing in DX11.
      I'v tested on both old and new nvidia hardware (680M and 1080, the behaviour is that same).
       
  • Popular Now