What is sideeffect

[FuXiii]

When I read test in test.cpp

builder.setSideEffect();

FrameGraph::Builder &FrameGraph::Builder::setSideEffect() {
  m_passNode.m_hasSideEffect = true;
  return *this;
}

and

FrameGraphResource FrameGraph::Builder::write(FrameGraphResource id) {
  assert(m_frameGraph.isValid(id));
  if (m_frameGraph._getResourceEntry(id).isImported()) setSideEffect();
...
}

What is the m_hasSideEffect use for?

And the FrameGraph-Example just a little complex, I am just a fresh hand. Can you update an example about how to draw a box with index buffer? :D Thanks~

[skaarj1989]

setSideEffect ensures that a pass is not culled during the compilation phase. (writing to imported resource counts as side-effect).

Moving to FrameGraph is a big paradigm shift. It might be overwhelming at first but once you understand it you won't look back ;)

I encourage you to watch/read:

• https://www.gdcvault.com/play/1024045/FrameGraph-Extensible-Rendering-Architecture-in
• https://www.gdcvault.com/play/1024612/FrameGraph-Extensible-Rendering-Architecture-in

[FuXiii]

Oh! I have generally finished reading your project and it wrote very well and clearly, Thank you very much~ :D.

---

When I try to read FrameGraph in Filament. FrameGraph looks like rebuild every frame? If I have many many Node and Resource in one rendering frame, if some Node not change but I must rebuild it at next new frame? Is that will waste performance？ Is that really necessary to rebuild FrameGraph every frame?

---

If I create a PassNode and create an TextureNode for read, there have assert code in FrameGraph::Builder::read(...):

FrameGraphResource FrameGraph::Builder::read(FrameGraphResource id) {
  assert(m_frameGraph.isValid(id));
  return m_passNode._read(id);
                      |_ assert(!creates(id) && !writes(id));
}

It looks like FrameGraph not support PassNode to create ResourceNode for read.

---

If I want create an TextureNode for ColorDrawPass and it will sampled in fragment shader.

1. How I binding the shaders in PassNode
2. Resource are create and destroy by ResourceEntry::Concept or Model . How to initial the texture's color pixel? If I import this texture into FrameGraph how to do it? the test in TEST_CASE("Imported resource", "[FrameGraph]") the id for the import already existed

 constexpr auto kBackbufferId = 777;
 FrameGraph fg;
 const auto backbuffer = fg.import("Backbuffer", {1280, 720}, FrameGraphTexture{kBackbufferId});

In really case the code maybe looks like:

Texture texture = ImageLoader::Load("./SomePicture.png");
FrameGraph fg;
fg.import(..., texture );

Is that right?

or just create an HostWritePass for host texture memory write? but I think this just a little odd.

[skaarj1989]

Is that really necessary to rebuild FrameGraph every frame?

Yes, you need to rebuild the FrameGraph every frame.

(...) Is that will waste performance?

Here I have a FrameGraph with 38 passes and 51 resources:

(measured with Tracy Profiler).
In my opinion, the cost is negligible.

It looks like FrameGraph not support PassNode to create ResourceNode for read.

Do you want to create a resource in a pass and read it immediately in the same pass? That will give you a garbage data.
Usually, a pass creates a resource (writes to it) and another pass reads from it.

How I binding the shaders in PassNode

It's up to you. Pass whatever you need to ctx in an execution callback.

struct PassData {};
fg.addCallbackPass<PassData>("SimplePass",
  [&](FrameGraph::Builder &builder, PassData &data) {
    // ...
  },
  [=](const PassData &data, FrameGraphPassResources &resources, void *ctx) {
    auto &rc = *static_cast<RenderContext *>(ctx);
    rc.bindPipeline(/* whatever, could be a member variable */);
    // ...
  }
);

Here you have an example:
https://github.com/skaarj1989/FrameGraph-Example/blob/7867df8e3146acfffaf41173d5776b6ac265fd50/src/WorldRenderer.cpp#L415

---

1. Don't use a FrameGraph to create a resource that is initialized once (like mesh vertex buffer or a material texture).
2. Don't import everyting into a FrameGraph.

Here, the SSAO pass is the owner of a noise texture and a kernel uniform buffer. Both are initialized by the constructor and then captured by the execution callback.

---

To create a resource, the FrameGraph class uses given allocator (transient resource manager/system).

void execute(void *context = nullptr, void *allocator = nullptr);

Take a look at here, for an example of a transient resource system.

[FuXiii]

Oh~ Thank you very very much~ :D
I think I need to read your FrameGraph-Example project .♪(^∇^*)

[FuXiii]

When to judge ID Valid

bool FrameGraph::isValid(FrameGraphResource id) const {
  const auto &node = _getResourceNode(id);
  auto &resource = m_resourceRegistry[node.m_resourceId];
  return node.m_version == resource.m_version;
}

the code compare the version between ResourceNode and ResourceRegistry. Why? Why to compare the version?

I know if different pass write to the same ResourceEntry ,the version of ResourceEntry will plus one for each clone and set the new ResourceEntry version into the ResourceNode version, that will let track resource very easy, but I don't understand why use the node.m_version == resource.m_version to judge whether it is valid？

[skaarj1989]

It enforces a specific execution order of the render passes (prevents mistakes).

Imagine that you have a pass P0, that writes to a resource R.
Followed by a pass P1, that writes to R (version is bumped)
Then you want to read R in yet another pass P2
you could by mistake give the handle of the first write operation to P2 (you should use the most recent one)

I'm gonna convert this issue to a discussion (since it's not really an issue).

The original question is about the sideEffect. If you have more questions just create a new discussion.

[FuXiii]

https://themaister.net/blog/2017/08/15/render-graphs-and-vulkan-a-deep-dive/

In this blog I remember it has cover some topic about reorder the PassNode.

please check the Traverse dependency graph section

[skaarj1989]

I am not going to implement pass reordering. I don't need that feature.

[FuXiii]

And in ResourceEntry::ResourceEntry(...)

template <typename T>
inline ResourceEntry::ResourceEntry(uint32_t id, typename T::Desc &&desc,
                                    T &&obj, uint32_t version, bool imported)
    : m_id{id}, 
     m_concept{std::make_unique<Model<T>>(
                  std::forward<typename T::Desc>(desc), std::forward<T>(obj))}

the

T &&obj

It looks like not had been used.
It just use T to define a T resource in template <typename T> struct Model

[skaarj1989]

It's part of type erasure. I don't understand your point here.

[FuXiii]

Oh, the type erasure looks like a feature of Templete Program, I am not very familiar with type erasure

[FuXiii]

Hello I am coming again.
In FrameGraphPassConcept this struct has declare virtual void operator()(FrameGraphPassResources &, void *) = 0; , and the child class FrameGraphPass use this to reload operate() member function to invoke the execFunction function.
I have not find some code like:

//FrameGraphPass fgp = ...;
//FrameGraphPassResources fgpr = ...;
//void* custom_data = ...;
fgp(fgpr, custom_data);

I only find the std::invoke(...) call to invoke execFunction, such like the calling in void FrameGraph::execute(...):

std::invoke(*pass.m_exec, resources, context);

Is that means the std::invoke(*pass.m_exec, resources, context) will invoke the FrameGraphPass::operator()(...)?

[skaarj1989]

Once again it's type erasure, but this time for a function.

FrameGraph fg;

struct PassData {};
fg.addCallbackPass<PassData>("SimplePass",
  [&](auto &builder, PassData &data) {
    // setup callback, called immediatly via fg.addCallback
  },
  [=](const PassData &data, FrameGraphPassResources &resources, void *) {
    // this is an execution callback, it's execution is deferred.
    // (see the very last line of this post)
  }
);

Execution callback is wrapped:

FrameGraph/include/fg/FrameGraph.inl

Line 15 in 326e676

auto *pass = new FrameGraphPass<Data, Execute>(std::forward<Execute>(exec));

fg.execute()

FrameGraph/src/FrameGraph.cpp

Line 99 in 326e676

std::invoke(*pass.m_exec, resources, context);

Here, a pass Data is smuggled into the execution callback:

FrameGraph/include/fg/PassEntry.hpp

Line 23 in 326e676

execFunction(data, resources, context);

[FuXiii]

Is that means, When you call:

std::invoke(*pass.m_exec, resources, context); 

this std::invoke(...),it will call:

execFunction(data, resources, context); 

which is define of :

void operator()(FrameGraphPassResources &resources, void *context) override 

[skaarj1989]

Almost.

std::invoke(*pass.m_exec, resources, context); 
// -> FrameGraphPassConcept::operator() -> FrameGraphPass::operator() -> wrapped execution callback

[FuXiii]

Oh, Thank you. will your hint, I found this in cppreference:

Thank you very much 👍