6 replies to this topic

[noizex]

Hello,
As I moved my rendering into centralized Renderer class, I can't use direct GL calls (glUniform*) to set uniforms where I know their value, but I have to pass them all as a part of draw command to renderer. Perfectly I'd have something like:

[source lang="cpp"]std::map< std::string, Uniform > uniforms;uniforms["mvp"] = Uniform(glm::mat4(0.0));uniforms["alpha"] = Uniform(0.56f);glm::vec4 sunpos = world->getsunpos();uniforms["sunpos"] = Uniform(sunpos);[/source]

For now I need support for: float, int, vec2, vec3, vec4, mat4, but later I may also need uniform structs. From what I know this is typical variant class that stores values of different types. I wanted to ask you for best solution, because the one I managed to hack so far is not perfect, is a bit bloated with nasty switches and requires me to initialize Uniform only by passing "new" into constructor (so this doesn't work: Uniform(mvp) or Uniform(glm::mat4(0.0)), only this: Uniform(new glm::mat4(mvp)).

I will show code for one sample value:

[source lang="cpp"]UniformValue::UniformValue(std::string name, glm::vec2* value){ m_Name = name; m_Data = static_cast< void* >(value); m_Type = UniformValue::Vec2;}UniformValue::~UniformValue(){ switch(m_Type) { case Vec2: delete static_cast< glm::vec2* >(m_Data); break; }}glm::vec2 UniformValue::ToVec2() const{ return *(static_cast< glm::vec2* >(m_Data));}void UniformValue::Apply(Shader* shader) const{ switch(m_Type) { case Vec2: shader->SetUniform(m_Name, ToVec2()); break; }}[/source]

I'm not entirely sure this is the best solution to my problem. Do you guys know any better? Especially if they allow me to pass local variables (Uniform(model) instead of Uniform(new glm::mat4(model))) and not have to use "new" in constructor.

I was thinking about using unions but I'm not sure they will work with glm::mat4 and other more complex objects (non-POD).

PS. Also solution should be pretty quick and not initialize unecessary things (like passing variable, copying it, then allocating some new one in the final step) - setting uniforms happens for all drawing commands every frame so it shouldn't be too unoptimized, because maybe its not much, but its done very often.

Edited by noizex, 24 September 2012 - 12:38 PM.

[Lightness1024]

my personal solution, but frankly I hate it, I think it is just a failed moment of having fun with C++, and that you can just go implement duplicated functions it is just a small part of the underlying layers of en engine you rarely need to put your nose into anyways. once its written its written.

anyways:

http://paste.blixt.org/8871931

PS note: the code=cpp tag in that forum is reaaaally buggy. impossible to post my snippet of code without losing entire lines !!

Edited by Lightness1024, 28 September 2012 - 03:15 PM.

[Lightness1024]

hm, I see maybe that is not totally what you wanted, maybe you wanted a 'deferred storage' to be able to get and mutate your uniforms multiple times in a frame and only once actually upload it to the driver, so you wanted a way of managing that store ? which my solution doesn't provide of course. but it can maybe give you some ideas for ways to handle multiple types.

Also about the "nasty switch", check your assembly code, but it gets more optimized than anything that can be written by hand. Most people wrongly assume switches are stupid comparators that tries every value until finding the good one, but actually it uses optimal hashing / research or binary selection during code generation since all constants are integers and known.

Edited by Lightness1024, 28 September 2012 - 03:25 PM.

[Katie]

The trick with stuff like this is to first off, write a bunch of overloaded version of glUniformXYZ() which takes C++ types and a slot and sets the uniform.

Then you can write templated setter classes which you can store in the map.

Then you can write a simple interface to the uniforms map which takes a string and an arbitrary type and puts a setter into the map.

Then you just iterate over the map calling the setter through an interface.

If you do this neatly, BTW, you can arrange to store either a value or a reference into the map. So you can express "Set this uniform to integer 7, but set that one to the value of this integer over here" in the same uniforms group.

If you make the setters a PIMPL interface with a templated inner class constructed inplace into a convenient array of spare memory and set up copy operators properly, then the setters are plain value objects and hence can be stored directly in the map. All this avoids nasty casts and will properly type-check things at compile time. (Because if you try and store an X and there ultimately isn't a suitable glUniformXYZ then the code won't build). And you can hotwire it to use your own vector/matrix classes and write them into the GLSL versions.

I've probably got some code somewhere if that doesn't make sense.

[Hodgman]

I've modelled my renderer around UBO's (cbuffers in D3D talk) instead of individual uniforms, which means you never set individual uniform values, but instead you set whole groups, based on where the data comes from -- e.g. you might have a camera UBO containing view/projection, a material UBO containing colours, an object UBO containing world matrices, etc... IMO, managing uniform buffers is extremely simple compared to managing individual uniforms.

Each UBO has a known byte size and layout so I just create a byte array of that size -- e.g. vector<byte> buffer(42) or new byte[42] or malloc(42) -- and can then memcpy uniforms around the place.

If a data source is C++ code, then I can write a C++ struct that exactly matches my GLSL UB layout, and then just cast my buffer to this struct and write the uniform values using regular C++ syntax.
If a data source is a user-generated file, with string variable names, then I've also got a reflection system so I can determine the byte-offset of a UB member from a name.
e.g. Mat4* proj = (Mat4*)&buffer[reflection["projection"].offset];

Then when I'm managing which UBO's to bind, there's no variant/etc's required because every slot holds the same thing, a UBO.

I was thinking about using unions but I'm not sure they will work with glm::mat4 and other more complex objects (non-POD).

You can only send POD data to GLSL, so it should be pretty safe to treat all uniforms as POD..

[noizex]

Wow, thanks! Thats a lot of very useful input, I was worried that this topic won't get any answer as it was few days and noone responded to it

@Lightness1024:
Yeah, <code> tag is broken unfortunately, it seems to strip template braces because they're similar to HTML tags - it should escape them but it doesnt. Putting space between braces and the template parameter helps, but it requires changing all code so probably better using some external pastebin for longer code. Thanks for code sample, I will try and comprehend it, although I see Boost there and I try to stay away from it as my code already has a lot of external libraries. And I completly agree about switch not being bad speed-wise, I meant it more from a programmer's perspective - adding another type requires you to fill few switch statements which is error-prone, rather than speed of actual code.

@Katie:
Ok, my shader already has overloaded functions that take glm::vecs and glm::mats, floats etc. But not sure what you mean about templated setter class and how to store different types in one map. Some code sample would be great, because it sounds like it could be a good and clean solution, but I'm not sure I understand 100% of it.

@Hodgman
Your solution is probably fastest because it works on simple structs and you can then copy it into UBO - only potential drawback I can see is that it requires you to have really well-thought-of concept of what uniforms you need, even on object/mesh-level. Right now I use UBO for "render pass" information like camera projection & view matrices. I also plan to have one lower level UBO (for things like material), then we go down to per-object state and here I'm not sure, because uniforms can really vary per object. I can be wrong of course, because for now I only have very similar objects rendered (they use the same uniforms and stuff like that) but I can't know what the future will bring. Thats the reason why I made the lowest level based on setting uniforms rather than a buffer, because then I don't have to rely on some static structure of UBO.

But now that I think about it more and more, I may just go for UBO till the moment where my uniforms can't fit one structure. Well, I can even have few structs for different objects - only thing it requires is making new struct definition, instead of freely setting any uniform I want through some SetUniform(name, val) call.

Edited by noizex, 29 September 2012 - 06:55 AM.

[Katie]

OK, so the usual trick about storing different types in a map is to go at it like this;

You create an envelope class which implements your interface and in turn delegates the work to an interface pointer. You create a suitable templated class into that interface pointer.

Like this;

class Inner
{
  virtual void doStuff(<params>)=0;
  virtual Inner *cloneAt(void *address)=0;
}


template<class DATA>
class InnerTempl
{
  DATA data;
  virtual void doStuff(<params>) { // in here you can use 'data' to work on it, and it has the right type. }

  InnerTempl(const DATA &d):data(d) {}
  virtual Inner *cloneAt(void *address)
  {
     return new(address) InnerTempl(data);
  }
}


class Outer
{
  Inner *pInner;
  void doStuff(<params>) { pInner->doStuff(<params>); }

  template <class X> assign(const X &x) { pInner=new InnerTempl<X>(x); }
}

So far, so good.

You can store Outers in a map. There's some issues about memory management and you'll need to flesh those out into real classes.

Avoiding the memory management can be done reasonably neatly if you're careful.
Start with this;

class Outer
{
  Inner *pInner;
  unsigned char space[32]; // you'll need to pick a size here that's big enough for your purposes.
  void doStuff(<params>) { pInner->doStuff(<params>); }
  template <class X> assign(const X &x) { pInner=new(space) InnerTempl<X>(x); } 
  template <class X> operator=(const X &x) { pInner=new(space) InnerTempl<X>(x); }
};

Now you don't need a separate inner object, so there's one less memory alloc and no loose pointers. It's a bit hacky and relies on picking a size that's 'big enough' but games development is ~1% genius and 99% hacking that genius to work practically.

STL mandates objects in maps must be copyable. And this then, is why we implemented cloneAt. Note that it needs the virtual dispatching to end up in a function which knows what type DATA really is.

Add to Outer an assignment operator.

Outer &operator=(const Outer &other) { pInner = other.pInner->cloneAt(space); }

Likewise, implement a copy constructor.

STL mandates objects in maps must have a null constructor. Make it have a null pInner to start with & add all the exception handling to cope with that and you're done.

Now you can store arbitrary types in the map by just assigning to them and iterate through them calling doStuff() on each one.

typedef map<string,Outer> MYMAP;
MYMAP mymap;
myMap["foo"]=1.5;
myMap["foo"]=vec3(0,0,7);

Job done.


Then in this case, doStuff() probably takes an integer slot and then calls an overloaded glUniform() function passing "data" in.

Mine looks like this;

void glUniform(int slot,int value)		  { glUniform1i (slot,value);}
void glUniform(int slot,unsigned int value) { glUniform1i (slot,value);}
void glUniform(int slot,float value)	    { glUniform1f (slot,value);}
void glUniform(int slot,const vec2 &value)  {glUniform2fv(slot,1,value.ptr());}
void glUniform(int slot,const vec3 &value)  { glUniform3fv(slot,1,value.ptr());}
void glUniform(int slot,const vec4 &value)  { glUniform4fv(slot,1,value.ptr());}
void glUniform(int slot,const ivec2 &value) { glUniform2iv(slot,1,value.ptr());}
void glUniform(int slot,const ivec3 &value) { glUniform3iv(slot,1,value.ptr());}
void glUniform(int slot,const ivec4 &value) { glUniform4iv(slot,1,value.ptr());}
void glUniform(int slot,const matrix &value){ glUniformMatrix4fv(slot,1,GL_FALSE,value.ptr());}

// I have my CPU vector classes named to match the GLSL ones because I got bored of accidentally using the wrong ones all the time.
// generally ptr() gets a base pointer to real data in the classes.