Sign in to follow this  

DX11 [Compute Shader] Groupshared memory as slow as VRAM

Recommended Posts

As far as I know, I have one of the earlier and cheaper mobile graphics cards that supported DX11. It's a AMD Radeon HD 5730M.

I've started optimizing graphics algorithms by porting them to compute shader and improving them by sharing memory and synchronizing the threads. This way I could improve the runtime of my Bloom from [eqn]O(n)[/eqn] to [eqn]O(log(n))[/eqn] per pixel.

But that was only the theoretical runtime. In reality, the algorithm performed so much worse than the original linear algorithm. I'm pretty sure I know the reason. Instead of let's say 32 read operations and 1 write operation, the algorithm now needs 1 read operation from VRAM, 5 read operations from groupshared memory, 5 write operations to groupshared memory and 1 write operation to VRAM.

Overall groupshared memory being L1 Cache should be way faster than 32 read operations from VRAM and it's even way less operations because of the algorithm having logarithmic runtime, but it's way slower (8ms instead of 0.5ms). The slowdown could be because of memory bank conflicts. But could they really cause such an enormous slowdown?

To me it looks like my graphics card might not even have an actual L1 cache residing on the Wavefront as groupshared memory at all. It performs just as bad as a UAV residing in VRAM would. So maybe they simply wrote a driver that uses 32kb of reserved memory in the VRAM as groupshared memory. Could that be the case or is it the bank conflicts?

I wish there were tools that could shine more light on such problems. Graphics cards and the tools should be more transparent in what's actually going on, so that the developers could improve the algorithms even further.

Update: After reading through NVidias CUDA documentation my shaders don't even cause any bank conflicts at all. Each half warp (16 threads) always accesses 16 different memory banks. Just a whole block (1024 threads) accesses them multiple times, which is normal and has nothing to do with bank conflicts. Edited by CryZe

Share this post

Link to post
Share on other sites
I suppose it's possible that your hardware doesn't actually have on-chip shared memory and just uses global memory instead, but I've not heard of that ever being the case. Although mobile hardware isn't usually well-documented, so who knows. You could try using GPU PerfStudio or AMD's APP profiling suite, but I'm not sure if either those will give you enough information to narrow down the problem. Perhaps you might want to try running some samples that make use of shared memory to see if they also perform poorly on your hardware.

Also just so you know, shared memory isn't L1. On AMD and Nvidia hardware It's its own special type of on-chip memory, and it's separate from the caches.

Share this post

Link to post
Share on other sites
[quote name='MJP' timestamp='1347309681' post='4978690']
Also just so you know, shared memory isn't L1. On AMD and Nvidia hardware It's its own special type of on-chip memory, and it's separate from the caches.
"As mentioned in Section F.4.1, for devices of compute capability 2.x and higher, the same on-chip memory is used for both L1 and shared memory, and how much of it is dedicated to L1 versus shared memory is configurable for each kernel call." Source: [url=""]CUDA Programming Guide[/url]

I've tried both the PerfStudio and AMD's APP Profiler. But they didn't work at all. PerfStudio wasn't able to catch a frame (endlessly trying to connect, even though it was already connected) and the APP Profiler showed me an error message in both of its modes. I'll probably try it again tomorrow.

[quote name='MJP' timestamp='1347309681' post='4978690']
Perhaps you might want to try running some samples that make use of shared memory to see if they also perform poorly on your hardware.
Oh, that's a good idea. I remember that the OIT11 Sample from the DirectX Sample Browser performs incredibly bad on my hardware (9FPS at 320x240). I don't know if it performs bad in comparison to the other samples on other hardware as well, though. I'll take a look into it's source to check out why it might perform that bad.

I'll also try to implement a bandwidth heavy compute shader that either performs an enormous amount of write operations to shared memory or to shared memory while causing as many bank conflicts as possible or to global memory. If the performance is the same the chances that my graphics card uses on chip memory as shared memory are pretty much zero. Edited by CryZe

Share this post

Link to post
Share on other sites

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  

  • Announcements

  • Forum Statistics

    • Total Topics
    • Total Posts
  • Similar Content

    • By KarimIO
      Hey guys,
      I'm trying to work on adding transparent objects to my deferred-rendered scene. The only issue is the z-buffer. As far as I know, the standard way to handle this is copying the buffer. In OpenGL, I can just blit it. What's the alternative for DirectX? And are there any alternatives to copying the buffer?
      Thanks in advance!
    • By joeblack
      im reading about specular aliasing because of mip maps, as far as i understood it, you need to compute fetched normal lenght and detect now its changed from unit length. I’m currently using BC5 normal maps, so i reconstruct z in shader and therefore my normals are normalized. Can i still somehow use antialiasing or its not needed? Thanks.
    • By 51mon
      I want to change the sampling behaviour to SampleLevel(coord, ddx(coord.y).xx, ddy(coord.y).xx). I was just wondering if it's possible without explicit shader code, e.g. with some flags or so?
    • By GalacticCrew
      I want to improve the performance of my game (engine) and some of your helped me to make a GPU Profiler. After creating the GPU Profiler, I started to measure the time my GPU needs per frame. I refined my GPU time measurements to find my bottleneck.
      Searching the bottleneck
      Rendering a small scene in an Idle state takes around 15.38 ms per frame. 13.54 ms (88.04%) are spent while rendering the scene, 1.57 ms (10.22%) are spent during the SwapChain.Present call (no VSync!) and the rest is spent on other tasks like rendering the UI. I further investigated the scene rendering, since it takes über 88% of my GPU frame rendering time.
      When rendering my scene, most of the time (80.97%) is spent rendering my models. The rest is spent to render the background/skybox, updating animation data, updating pixel shader constant buffer, etc. It wasn't really suprising that most of the time is spent for my models, so I further refined my measurements to find the actual bottleneck.
      In my example scene, I have five animated NPCs. When rendering these NPCs, most actions are almost for free. Setting the proper shaders in the input layout (0.11%), updating vertex shader constant buffers (0.32%), setting textures (0.24%) and setting vertex and index buffers (0.28%). However, the rest of the GPU time (99.05% !!) is spent in two function calls: DrawIndexed and DrawIndexedInstance.
      I searched this forum and the web for other articles and threads about these functions, but I haven't found a lot of useful information. I use SharpDX and .NET Framework 4.5 to develop my game (engine). The developer of SharpDX said, that "The method DrawIndexed in SharpDX is a direct call to DirectX" (Source). DirectX 11 is widely used and SharpDX is "only" a wrapper for DirectX functions, I assume the problem is in my code.
      How I render my scene
      When rendering my scene, I render one model after another. Each model has one or more parts and one or more positions. For example, a human model has parts like head, hands, legs, torso, etc. and may be placed in different locations (on the couch, on a street, ...). For static elements like furniture, houses, etc. I use instancing, because the positions never change at run-time. Dynamic models like humans and monster don't use instancing, because positions change over time.
      When rendering a model, I use this work-flow:
      Set vertex and pixel shaders, if they need to be updated (e.g. PBR shaders, simple shader, depth info shaders, ...) Set animation data as constant buffer in the vertex shader, if the model is animated Set generic vertex shader constant buffer (world matrix, etc.) Render all parts of the model. For each part: Set diffuse, normal, specular and emissive texture shader views Set vertex buffer Set index buffer Call DrawIndexedInstanced for instanced models and DrawIndexed models What's the problem
      After my GPU profiling, I know that over 99% of the rendering time for a single model is spent in the DrawIndexedInstanced and DrawIndexed function calls. But why do they take so long? Do I have to try to optimize my vertex or pixel shaders? I do not use other types of shaders at the moment. "Le Comte du Merde-fou" suggested in this post to merge regions of vertices to larger vertex buffers to reduce the number of Draw calls. While this makes sense to me, it does not explain why rendering my five (!) animated models takes that much GPU time. To make sure I don't analyse something I wrong, I made sure to not use the D3D11_CREATE_DEVICE_DEBUG flag and to run as Release version in Visual Studio as suggested by Hodgman in this forum thread.
      My engine does its job. Multi-texturing, animation, soft shadowing, instancing, etc. are all implemented, but I need to reduce the GPU load for performance reasons. Each frame takes less than 3ms CPU time by the way. So the problem is on the GPU side, I believe.
    • By noodleBowl
      I was wondering if someone could explain this to me
      I'm working on using the windows WIC apis to load in textures for DirectX 11. I see that sometimes the WIC Pixel Formats do not directly match a DXGI Format that is used in DirectX. I see that in cases like this the original WIC Pixel Format is converted into a WIC Pixel Format that does directly match a DXGI Format. And doing this conversion is easy, but I do not understand the reason behind 2 of the WIC Pixel Formats that are converted based on Microsoft's guide
      I was wondering if someone could tell me why Microsoft's guide on this topic says that GUID_WICPixelFormat40bppCMYKAlpha should be converted into GUID_WICPixelFormat64bppRGBA and why GUID_WICPixelFormat80bppCMYKAlpha should be converted into GUID_WICPixelFormat64bppRGBA
      In one case I would think that: 
      GUID_WICPixelFormat40bppCMYKAlpha would convert to GUID_WICPixelFormat32bppRGBA and that GUID_WICPixelFormat80bppCMYKAlpha would convert to GUID_WICPixelFormat64bppRGBA, because the black channel (k) values would get readded / "swallowed" into into the CMY channels
      In the second case I would think that:
      GUID_WICPixelFormat40bppCMYKAlpha would convert to GUID_WICPixelFormat64bppRGBA and that GUID_WICPixelFormat80bppCMYKAlpha would convert to GUID_WICPixelFormat128bppRGBA, because the black channel (k) bits would get redistributed amongst the remaining 4 channels (CYMA) and those "new bits" added to those channels would fit in the GUID_WICPixelFormat64bppRGBA and GUID_WICPixelFormat128bppRGBA formats. But also seeing as there is no GUID_WICPixelFormat128bppRGBA format this case is kind of null and void
      I basically do not understand why Microsoft says GUID_WICPixelFormat40bppCMYKAlpha and GUID_WICPixelFormat80bppCMYKAlpha should convert to GUID_WICPixelFormat64bppRGBA in the end
  • Popular Now