Jump to content
  • Advertisement
Sign in to follow this  
nwdg

Vulkan Branching in compute kernels

This topic is 657 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

Hi,

I'm porting a D3D11 program I wrote to Vulkan, in this program I do volume ray casting.

It is a simple fog simulator I'm expering with. 
At each ray step I need to do some operations that requires a dynamic for loop, as the iterations number is evaluated on the fly at each step along the ray.
This means no loop unrolling or compiler optimizations.
On the other hand the hlsl code works just fine keeping a real-time speed. 
In my Vulkan implementation, it just crashes after the compute fence goes timeout because of resource locking.
What "fixed" the problem is using a constant value in the loop of course, but that kinda kills my algorithm.

Does anybody have tried anything like this? Have some infos on this matter? I would like to dig into this further.

Here's a pseudo-code of my ray marching algorithm

vec3 ForEachStep()
{
        vec3 retVal = 0, 0, 0;
        numIter = FunctionCallToDetermine();
        for i = 0 to numIter
       {
              retVal += FuncCall();
       }
       return retVal;
}
 
const uint numSteps = someValue;
 
void main()
{
       // do other unrelated stuff
       vec3 color = 0, 0, 0
       for i = 0 to numSteps
      {
           // do other unrelated stuff 
 
           color += ForEachStep();
 
           // do other unrelated stuff
      }
 
      // store color to texture target
}

Cheers!

Share this post


Link to post
Share on other sites
Advertisement

In my Vulkan implementation, it just crashes after the compute fence goes timeout because of resource locking.

So the shader finishes, and you get the crash afterwards?

Try a vkQueueWaitIdle() after vkQueueSubmit(), to quickly ensure all work has finished and see if it prevents the crash.

 

Or does the shader never finish? Usually this causes a bluescreen or a unstable system (Do NOT save your source files in this case - reboot instantly. I've lost some work with this)

Probably a infinite loop - implement an additional max counter to prevent this and see if the crash goes away.

 

 

Can you post some performance comparision if you get it to work?

Share this post


Link to post
Share on other sites

Fixed it! I overlooked an uint underflow when calculating the number of iterations! When the underflow occured it generated an infinite loop.

 

 

In my Vulkan implementation, it just crashes after the compute fence goes timeout because of resource locking.

So the shader finishes, and you get the crash afterwards?

Try a vkQueueWaitIdle() after vkQueueSubmit(), to quickly ensure all work has finished and see if it prevents the crash.

 

Or does the shader never finish? Usually this causes a bluescreen or a unstable system (Do NOT save your source files in this case - reboot instantly. I've lost some work with this)

Probably a infinite loop - implement an additional max counter to prevent this and see if the crash goes away.

 

 

Can you post some performance comparision if you get it to work?

 

 

 

 

I can only access my laptop in the weekend, I'll post some comparisons tomorrow!

Share this post


Link to post
Share on other sites

I can only access my laptop in the weekend, I'll post some comparisons tomorrow!

 

Hey, sorry for the late reply!

 

the GPU I've used for comparison is a GTX 970 3Gb

 

SPIR-V renders at 4/5 fps@720p, both ported to GLSL or directly compiled from HLSL.

HLSL renders at 25/27 fps@720p.

 

Did a second test with a bunch of spheres and simple lambertian reflectors/metal materials and the results are:

SPIR-V renders at ~2500 fps@720p

HLSL renders at ~400 fps@720p

 

Seems like that dynamic loop is killing SPIR-V!

Share this post


Link to post
Share on other sites

Thanks. Such varying results indicate they have some work to do at their drivers and it's still worth to try different APIs. :(

I may add a DX12 code path sometime...

 

Seems like that dynamic loop is killing SPIR-V!

Maybe in this case, but in general dynamic loops should be no problem really (I have them everywhere).

You would need to make much more test cases to verify an assumption like this.

If we could compare generated machine code, we could at least find out where the driver sucks.

 

I'm working on a large project and here i can compare Vulkan vs. OpenCL on AMD. Performance varies about 20-30%, mostly Vulkan is faster.

 

I've had only one extreme case where VK was 4 times faster. When adding fp16->fp32 conversion, CL compiler started wasting registers and occupancy went down.

Things like that are mostly the reason if 'good' code performs badly. Thus it's important to have a profiler tool showing register usage, occupancy etc.

Unfortunately AMDs CodeXL can't do this yet for VK (just DX, GL & CL), so coding feels like being blind folded at the moment.

Share this post


Link to post
Share on other sites

You would need to make much more test cases to verify an assumption like this.

 

Yes of course! Totally agreed.

 

If you are interested we can compare the machine code!

Share this post


Link to post
Share on other sites

NSight can show mapping between high level and machine code: https://devblogs.nvidia.com/parallelforall/cuda-pro-tip-view-assembly-code-correlation-nsight-visual-studio-edition/

Very nice. Try to see if this works for HLSL and Vulkan (although it's unlikely to work with GLSL->SPIR-V->PTX...)

 

Are you sure you upload the data properly to GPU memory using a staging buffer with Vulkan (data -> host visible buffer, host visible buffer -> device)?

Fetching the data from main memory would explain the bad performance.

Share this post


Link to post
Share on other sites

I'll try with NSight!

 

I've got a uniform buffer with a world matrix for the camera transform and a float4 filled with rendering mode data and volume info (this could probably be arranged better using push constants for the volume volume info and render mode).
Volume data is a 3D texture loaded with staging.
Raytracer output is a 2D texture that is not staged, created with STORAGE | SAMPLED usage flags, accessed as storage in compute, accessed with a sampler in fragment stage.

 

Would you do this different?

Share this post


Link to post
Share on other sites

Seems nothing wrong with that.

 

With NSight you should get at least exact timings for the various stages - maybe it's not the compute shader causing low fps.

(Also possible using vkCmdWriteTimestamp, but that's some extra work)

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement
×

Important Information

By using GameDev.net, you agree to our community Guidelines, Terms of Use, and Privacy Policy.

We are the game development community.

Whether you are an indie, hobbyist, AAA developer, or just trying to learn, GameDev.net is the place for you to learn, share, and connect with the games industry. Learn more About Us or sign up!

Sign me up!