Can anyone tell me how important state change management is, when the parameters actually do not change?

e.g. when several calls to "ID3D11DeviceContext::IASetPrimitiveTopology" are done with the same topology parameter.

Are such redundant state 'changes' handled quickly in D3D11 or should it be managed by the client program,

e.g. by just storing the parameter and then check if it must be changed or not:

StateManager::IASetPrimitiveTopology(D3D11_PRIMITIVE_TOPOLOGY topology)
{
    if (m_currentTopology != topology)
    {
        m_deviceContext->IASetPrimitiveTopology(topology);
        m_currentTopology = topology;
    }
}

That is actually what I currently do with OpenGL, to reduce the state changes. But I guess D3D11 is not such a heavy weighted state-machine as OpenGL is.

Thanks in advance.

State management isn’t as consequential as in OpenGL, but it still helps a bit or a lot depending on the state(s).
As mentioned above, set your states only when the draw call is issued. Otherwise it is easy to get redundant states even with your tracking method.

SetDepth( 1 );
Draw();
SetDepth( 0 ); // Just called after every draw call so that one draw call's state does not leak into another.
…
SetDepth( 1 );
Draw();


L. Spiro

I would recommend filtering out redundant states. Just do yourself a favor, and don't track the current state using global variables. Global variables won't work if you every switch to an API that supports multi-threaded command list generation.

To expand upon Spiro's point:

Don't use if-statements for state management. The if-statement alone is not free (and the D3D layer is already doing that for you, albeit behind a few more layers of abstraction/overhead). The checks also imply that your higher-level code doesn't know what state is being used and when.

For your example, if you have different topologies then that implies you're rendering rather vastly different things. Things should be rendered in like groups. e.g., render your meshes, then render your debug line overlays, etc. You never need an if-check then, because your draw code transforms like so:

// NOT GOOD
for (auto& object : objects) {
  render.set_state(object.material);
  render.draw(object);
}

// GOOD
for (auto& group : material_groups) {
  render.set_state(group.state);
  for (auto& object : group.objects)
    render.draw(object);
}

Of course, efficiently handling those groups takes some work. The gains can be quite worth it, though. Especially as it then enables automatic instancing/batching of your draw calls which keeps the hardware fed.

The hardware thing is key to remember. A lot of hardware can only run ~one program (with one set of state) at a time. If you're drawing e.g. quads, the vertex stage has all of 6 vertices to process, leaving potentially the rest of your 64-core wavegroup unused (if not the whole pipeline of hundreds or thousands of shader cores). You want all of the hardware working; you want to hand large chunks of data to the GPU and let it chew through the whole chunk in one go, rather than handing it many little chunks that it must inefficiently process sequentially. Basically, redundant state changes aren't your real problem; your renderer even making it possible to have redundant state changes in the first place is your problem. :)

I store a 64-bit integer with all the state ID's packed into it, and then XOR two of them to see what needs to change.
e.g.

raster_mask = 0xF;
raster_shift = 42;
u64 SetRaster( u64 packed, int rasterId ) { return (packed&~(raster_mask<<raster_shift)) | (rasterId << raster_shift); }
int GetRaster( u64 packed ) { return (packed>>raster_shift) & raster_mask; }
bool HasRaster( u64 packed ) { return 0!=(packed & (raster_mask<<raster_shift)); }

u64 oldState = ..., newState = ...;
u64 dirty = oldState ^ newState;
oldState = newState;
if( HasRaster(dirty) )
  SubmitRaster(device, GetRaster(newState));

Don't use if-statements for state management. The if-statement alone is not free (and the D3D layer is already doing that for you, albeit behind a few more layers of abstraction/overhead). The checks also imply that your higher-level code doesn't know what state is being used and when.

Eh, the overhead of those if statements is some number of nanoseconds per draw on a PC. In my D3D11 engine, a draw call (including state changes for it) takes somewhere on the order of a microsecond (mostly spent inside D3D), so a few nanosecond of extra if statements is pretty negligible :wink:
I'm also not sure if D3D will do the same check internally, or if it just forwards the redundant commands onto the driver anyway.

should be rendered in like groups. e.g., render your meshes, then render your debug line overlays, etc. You never need an if-check then, because your draw code transforms like so:

You can get that grouping automatically by sorting your draw calls before you submit them.

Also, while grouping to reduce state changes is good in general, it's not always the best strategy. e.g. for alpha-blended translucency it's not even a valid solution :) or for forward-rendered opaques without a depth-pre-pass, you often want to do a coarse front to back sort combined with a least-state-changes sort within each coarse depth range.

Individual states get merged into one PSO object, what's the strategy for cases where we have lots of objects with unique PSOs, does the switch cost related to how different two PSOs are?

https://developer.nvidia.com/dx12-dos-and-donts#pipeline

That was actually all I wanted to know:

The if-statement alone is not free (and the D3D layer is already doing that for you, albeit behind a few more layers of abstraction/overhead).

I know the basics of reasonable rendering state management, but I'm currently only writing a thin abstraction layer to have a unified interface for Direct3D and OpenGL.

So within this project, I have no render loops or material managment, just the basic render states.

Thanks and greetings,

Lukas