So I'm going through a ray tracing tutorial in an attempt to stretch my F# legs. Since the tutorial is written in C++, it a rather fun challenge to figure out how to apply the concepts in a functional manner. I'd like to write everything in as functional a manner as possible, because I intend to eventually run the ray tracer in parallel. But that got me to the following situation, the core of which I'm sure shows up in topics other than ray tracing.
We have an Engine object which (among other things) stores a Scene object (a collection of Primitive objects). Currently, the Scene and the Primitives within it are completely immutable.
To try improve the rendering time, we're implementing a regular three-dimensional grid. In order to traverse the grid in an effective manner, the tutorial references this paper for the both the traversal algorithm, and for reducing the number of intersection tests for primitives which lie across grid boundaries. Here's how the latter part works:
Each Ray is assigned a unique rayID
Each Primitive has a lastRayID field.
When a Ray checks for intersection with a Primitive,
If the rayID of the Ray equals the lastRayID of the Primitive, then the intersection test is skipped.
Otherwise, the test is performed, and the rayID of the Ray is stored in the lastRayID field of the Primitive.
This is a neat way of caching the intersection tests, but it's set up for a sequential ray tracer, and wouldn't work at all for even two concurrent rays. So although I could use mutable fields, and therefore be able to implement this algorithm, it wouldn't satify my end goals of an inherently parallelizable ray tracer.
I do have one idea. The problem with storing mutable state with each primitive is that, with respect to the rendering algorithm, the scene is a global entity - each ray could potentially strike any primitive. On the other hand, each ray is completely self contained. So in each ray, I figured I could build up a set of primitives which have already been tested against (this set would be the equivalent to the lastRayID field described above). The problem with this is that each ray has an extremely short life time, and there could potentially be many rays in existence at any given time, so setting up such a data structure in each ray could be costly ((de)allocation time and memory consumption costs could add up quickly)
Does anyone have advice on dealing with these kinds of situations? Even if it's a general idea about transforming shared mutable state into local mutable state or something like that, I'm sure that would help me out a lot. I'd be happy to clarify anything if necessary.
Related
There are a lot of questions online about allocating, copying, indexing, etc 2d and 3d arrays on CUDA. I'm getting a lot of conflicting answers so I'm attempting to compile past questions to see if I can ask the right ones.
First link: https://devtalk.nvidia.com/default/topic/392370/how-to-cudamalloc-two-dimensional-array-/
Problem: Allocating a 2d array of pointers
User solution: use mallocPitch
"Correct" inefficient solution: Use malloc and memcpy in a for loop for each row (Absurd overhead)
"More correct" solution: Squash it into a 1d array "professional opinion," one comment saying no one with an eye on performance uses 2d pointer structures on the gpu
Second link: https://devtalk.nvidia.com/default/topic/413905/passing-a-multidimensional-array-to-kernel-how-to-allocate-space-in-host-and-pass-to-device-/
Problem: Allocating space on host and passing it to device
Sub link: https://devtalk.nvidia.com/default/topic/398305/cuda-programming-and-performance/dynamically-allocate-array-of-structs/
Sub link solution: Coding pointer based structures on the GPU is a bad experience and highly inefficient, squash it into a 1d array.
Third link: Allocate 2D Array on Device Memory in CUDA
Problem: Allocating and transferring 2d arrays
User solution: use mallocPitch
Other solution: flatten it
Fourth link: How to use 2D Arrays in CUDA?
Problem: Allocate and traverse 2d arrays
Submitted solution: Does not show allocation
Other solution: squash it
There are a lot of other sources mostly saying the same thing but in multiple instances I see warnings about pointer structures on the GPU.
Many people claim the proper way to allocate an array of pointers is with a call to malloc and memcpy for each row yet the functions mallocPitch and memcpy2D exist. Are these functions somehow less efficient? Why wouldn't this be the default answer?
The other 'correct' answer for 2d arrays is to squash them into one array. Should I just get used to this as a fact of life? I'm very persnickety about my code and it feels inelegant to me.
Another solution I was considering was to max a matrix class that uses a 1d pointer array but I can't find a way to implement the double bracket operator.
Also according to this link: Copy an object to device?
and the sub link answer: cudaMemcpy segmentation fault
This gets a little iffy.
The classes I want to use CUDA with all have 2/3d arrays and wouldn't there be a lot of overhead in converting those to 1d arrays for CUDA?
I know I've asked a lot but in summary should I get used to squashed arrays as a fact of life or can I use the 2d allocate and copy functions without getting bad overhead like in the solution where alloc and cpy are called in a for loop?
Since your question compiles a list of other questions, I'll answer by compiling a list of other answers.
cudaMallocPitch/cudaMemcpy2D:
First, the cuda runtime API functions like cudaMallocPitch and cudaMemcpy2D do not actually involve either double-pointer allocations or 2D (doubly-subscripted) arrays. This is easy to confirm simply by looking at the documentation, and noting the types of parameters in the function prototypes. The src and dst parameters are single-pointer parameters. They could not be doubly-subscripted, or doubly dereferenced. For additional example usage, here is one of many questions on this. here is a fully worked example usage. Another example covering various concepts associated with cudaMallocPitch/cudaMemcpy2d usage is here. Instead the correct way to think about these is that they work with pitched allocations. Also, you cannot use cudaMemcpy2D to transfer data when the underlying allocation has been created using a set of malloc (or new, or similar) operations in a loop. That sort of host data allocation construction is particularly ill-suited to working with the data on the device.
general, dynamically allocated 2D case:
If you wish to learn how to use a dynamically allocated 2D array in a CUDA kernel (meaning you can use doubly-subscripted access, e.g. data[x][y]), then the cuda tag info page contains the "canonical" question for this, it is here. The answer given by talonmies there includes the proper mechanics, as well as appropriate caveats:
there is additional, non-trivial complexity
the access will generally be less efficient than 1D access, because data access requires dereferencing 2 pointers, instead of 1.
(note that allocating an array of objects, where the object(s) has an embedded pointer to a dynamic allocation, is essentially the same as the 2D array concept, and the example you linked in your question is a reasonable demonstration for that)
Also, here is a thrust method for building a general dynamically allocated 2D array.
flattening:
If you think you must use the general 2D method, then go ahead, it's not impossible (although sometimes people struggle with the process!) However, due to the added complexity and reduced efficiency, the canonical "advice" here is to "flatten" your storage method, and use "simulated" 2D access. Here is one of many examples of questions/answers discussing "flattening".
general, dynamically allocated 3D case:
As we extend this to 3 (or higher!) dimensions, the general case becomes overly complex to handle, IMO. The additional complexity should strongly motivate us to seek alternatives. The triply-subscripted general case involves 3 pointer accesses before the data is actually retrieved, so even less efficient. Here is a fully worked example (2nd code example).
special case: array width known at compile time:
Note that it should be considered a special case when the array dimension(s) (the width, in the case of a 2D array, or 2 of the 3 dimensions for a 3D array) is known at compile-time. In this case, with an appropriate auxiliary type definition, we can "instruct" the compiler how the indexing should be computed, and in this case we can use doubly-subscripted access with considerably less complexity than the general case, and there is no loss of efficiency due to pointer-chasing. Only one pointer need be dereferenced to retrieve the data (regardless of array dimensionality, if n-1 dimensions are known at compile time for a n-dimensional array). The first code example in the already-mentioned answer here (first code example) gives a fully worked example of that in the 3D case, and the answer here gives a 2D example of this special case.
doubly-subscripted host code, singly-subscripted device code:
Finally another methodology option allows us to easily mix 2D (doubly-subscripted) access in host code while using only 1D (singly-subscripted, perhaps with "simulated 2D" access) in device code. A worked example of that is here. By organizing the underlying allocation as a contiguous allocation, then building the pointer "tree", we can enable doubly-subscripted access on the host, and still easily pass the flat allocation to the device. Although the example does not show it, it would be possible to extend this method to create a doubly-subscripted access system on the device based off a flat allocation and a manually-created pointer "tree", however this would have approximately the same issues as the 2D general dynamically allocated method given above: it would involve double-pointer (double-dereference) access, so less efficient, and there is some complexity associated with building the pointer "tree", for use in device code (e.g. it would necessitate an additional cudaMemcpy operation, probably).
From the above methods, you'll need to choose one that fits your appetite and needs. There is not one single recommendation that fits every possible case.
I am in the process of coding a level design tool in Qt with OpenGL (for a relevant example see Valve's Hammer, as Source games are what I'm primarily designing this for) and have currently written a few classes to represent 3D objects (vertices, edges, faces). I plan to implement an "object" class which ties the three together, keeps track of its own vertices, etc.
After having read up on rendering polygons on http://open.gl, I have a couple of questions regarding the most efficient way to render the content. Bear in mind that this is a level editor, so I am anticipating needing to render a large number of objects with arbitrary shapes and numbers of vertices/faces.
Edit: Updated to be less broad.
At what point would be the best point to create the VBO? The Qt OpenGL example creates a VBO when a viewport is initialized, but I'd expect it to be inefficient to create a close for each viewport.
Regarding the submitted answer, would it be a sensible idea to create one VBO for geometry, another for mesh models, etc? What happens if/when a VBO overflows?
VBOs should be re-/initialized whenever there's a need for it. Consider VBOs as memory pools. So you'd not allocate one VBO per object, but group similar objects into a single VBO. When you run out of space in one VBO you allocate a another one.
Today's GPUs are optimized for rendering indexed triangles. So GL_TRIANGLES will suffice in 90% of all cases.
Frankly modern OpenGL implementations completely ignore the buffer object access mode. So many programs did made ill use of that parameter, that it became more efficient to profile the usage pattern and adjust driver behavior toward that. However it's still a good idea to use the right mode. And in your case it's GL_STATIC_DRAW.
Does anyone have any good implementation strategies or resources for putting together a b-rep modeling system?
OpenCascade is an apparently good library for b-rep modeling (used by FreeCad and PythonOCC are both very cool) but the library is huge, complicated and may not be a good starting point to learn about b-rep modeling 'engines'.
I've done quite a bit of research paper reading, and while the fundamental math is useful for understanding why everything works, its left me with some implementation questions.
The halfedge data-structure seems to be the preferred way to store information about a body in b-rep implementations.
So a handful of questions in no particular order:
Using the halfedge data-structure how is rendering typically implemented? Triangulation based on the solid's boundaries?
How are circular faces/curved surfaces typically implemented? For instance a cylinder in one basic introduction to b-rep's I read, was internally stored as a prism. IE an extruded triangle and meta-data was stored about the cap faces denoting that they were indeed circular.
How are boolean operations typically implemented? I've read about generating BSP-Tree's along the intersection curves then combining those tree's to generate the new geometry. Are there other ways to implement boolean operations and what sort of pro's/con's do they have?
Thanks!
If you'd like to provide a code example don't worry about the language -- the questions are more about algorithmic/data-structure implementation details
I'm working on a B-Rep modeler in C# (I'm in a very early stage: it's an huge project) so I ask myself the same questions as you. Here is my answers:
Triangulation: I've not done this step, but the strategy I'm thinking about is as follow: project the face boundaries in parameter space to obtain 2D polygons (with holes), triangulate that with the ear clipping algorithm and then reproject triangle vertices in 3D space. For curved surfaces, I need to split the polygons with a grid in order to follow the surface;
For a cylinder, there is 3 edges : two circulars and one line segment. I have classes for each type of curves (Segment3d, Circle3d...) and each half-edge hold an instance of one of theses classes. Each face hold an instance of a surface object (plane, cylinder, sphere...);
There is an interesting project here based on BSP-Tree, but it uses CSG method, not B-rep. I'm still researching how to do this, but I don't think I will need a BSP tree. The difficulty is in computing intersections and topology.
The best books I've found on this subject:
3D CAD - Principles and Applications (old but still relevant)
Geometric Modeling: The mathematics of shapes (more recent than the previous one, but less clear)
What should be the measures that should be used to identify that code is over abstracted and very hard to understand and what should be done to reduce over abstraction?
"Simplicity over complexity, complexity over complicatedness"
So - there's a benefit to abstract something only if You are "de-leveling" complicatedness to complexity. Reasons to do that can vary: better modularity, better encapsulation etc.
Identifying over abstraction is a chicken and egg problem. In order to reduce over abstraction You need to understand actual reason behind code lines. That includes understanding idea of particular abstraction itself (in contrast to calling it over abstracted cause of lack of understanding). And that's not enough - You need to know a better, simpler solution to prove that it's over abstracted.
If You are looking for tool that could do it in Your place - look no more, only mind can reliably judge that.
I will give an answer that will get a LOT of down votes!
If the code is written in an OO language .. it is necessarily heavily over-abstracted. The purer the language the worse the problem.
Abstraction should be used with great caution. If in doubt always use concrete data structures. (You can always abstract later, this is easier than de-abstraction :)
You must be very certain you have the right abstraction in your current context, and you must be very sure that concept will stand the test of change. Abstraction has a high price in performance of both the code and the coder.
Some weak tests for over-abstraction: if the data structure is a product type (struct in C) and the programmer has written get and set method for each field, they have utterly failed to provide any real abstraction, disabled operators like C increment, for no purpose, and simply not understood that the struct field names are already the abstract representation of a product. Duplicating and laming up the interface is not a good idea.
A good test for the product case is whether there exist any data invariants to maintain. For example a pair of integers representing a rational number is almost sufficient, there's little need for any abstraction because all pairs are valid except when the denominator is zero. However for performance reasons one may choose to maintain an invariant, typically the denominator is required to be greater than zero, and the numerator and denominator are relatively prime. To ensure the invariant, the product representation is encapsulated: the initial value protected by a constructor and methods constrained to maintain the invariant.
To fix code I recommend these steps:
Document the representation invariants the abstraction is maintaining
Remove the abstraction (methods) if you can't find strong invariants
Rewrite code using the method to access the data directly.
This procedure only works for low level abstraction, i.e. abstraction of small values by classes.
Over abstraction at a higher level is much harder to deal with. Ideally you'd refactor the code repeatedly, checking to see after each step it continues to work. However this will be hard, and sometimes a major rewrite is required, rather than a refinement. It's probably not worth it unless the abstraction is so far off base it is not tenable to continue to maintain it.
Download Magento and have a look at the code, read some documents on it and have a look at their ERD: http://www.magentocommerce.com/wiki/_media/doc/magento---sample_database_diagram.png?cache=cache
I'm not joking, this is over-abstraction.. trying to please everyone and cover every base is a terrible idea and makes life extremely difficult for everyone.
Personally I would say that "What is the ideal level of abstraction?" is a subjective question.
I don't like code that uses a new line for every atomic operation, but I also don't like 10 nested operations within one line.
I like the use of recursive functions, but I don't appreciate recursion for the sole sake of recursion.
I like generics, but I don't like (nested) generic functions that e.g. use different code for each specific type that's expected...
It is a matter of personal opinion as well as common sense. Does this answer your question?
I completely agree with what #ArnisLapsa wrote:
"Simplicity over complexity, complexity over complicatedness"
And that
an abstraction is used to "de-level" those, from complicated to complex
(and from complex to simpler)
Also, as stated by #MartinHemmings a good abstraction is quite subjective because we don't all think the same way. And actually our way of thinking change with time. So Something that someone find simple might looks complex to others, and even become simpler with more experiences. Eg. A monadic operation is something trivial for functional programmer, but can be seriously confusing for others. Similarly, a design with mutable object communicating with each other can be natural for some and feel un-trackable for others.
That being said, I would like to add a couple of indicators. Note that this applies to abstractions used in code-base, not "paradigm abstraction" such as everything-is-a-function, or everything-is-designed-as-objects. So:
To the people it concerns, the abstraction should be conceptually simpler than other alternatives, without looking at the implementation. If you find that thinking of all possible cases is simpler that reasoning using the abstraction, then this abstraction is not suitable (for you)
Its implementation should reason only about the abstraction, not the specific cases that it will be used for. As soon as the abstraction implementation has parts made for specific cases, it indicates an "unfit" abstraction. And increasing generalization to cope with each new case, is going the wrong way (and tends to fall to the next issue).
A very common indicator of over-abstraction I have found (and actually fell for) are abstractions that represent more than what is needed, now. As much as possible, they should allow to do exactly what is required, but nothing more. For example, say you're thinking of, or already have, a "2d point" abstraction for which you can define many operators you need. Then you have another need that could really be a "4d point" similar to the 2d. Don't start to use a "Ndimensionnal point" abstraction, especially thinking that you might later need it. Maybe you'll never have anything else than 2 and 4d (because it stays as "a good idea" in the backlog forever) but instead some requirements pops to convert 4d points into pairs of 2d points. That's going to be hard to generalize to n-dimensions. So, each abstraction can be checked to cover and only cover the actual needs. In my point example, the complexity "n-dimensional" is actually only used to cope with the 2 and 4d cases (and the 4d might not even be used that much).
Finally, in a more global point of view, a code-base that has many not related abstractions, is an indicator that the dev team tends to abstract every little issues. So probably many of them are or became over-abstracted.
I'm in the process of trying to 'learn more of' and 'learn lessons from' functional programming and the idea of immutability being good for concurrency, etc.
As a thought exercise I imagined a simple game where Mario-esq type character can run and jump around with enemies that shoot at him...
Then I tried to imagine this being written functionally using immutable objects.
This raised some questions that puzzled me (being an Imperative OO programmer).
1) If my little guy at position x10,y100 moves right 1 unit do I just re-instantiate him using his old values with a +1 to his x position (e.g x11,y100)?
2) (If my first assumption is correct)
If my input thread moves little guy right 1 unit and my enemy AI thread shoots little guy and enemy-ai-thread resolves before input-thread then my guy will loose health, then upon input thread resolving, gain it back and move right ...
Does this mean I can't fire-&-forget my threads even with immutability?
Do I need to send my threads off to do their thing then new()up little guy synchronously when I have the results of both threaded operations? or is there a simple 'functional' solution?
This is a slightly different threading problem than I face on a day to day basis.
Usually I have to decide if I care about what order threads resolve in or not. Where as in the above case I technically don't care if he takes damage or moves first. But I do care if race conditions during instantiation cause one threads data to be totally lost.
3) (Again if my first assumption is correct) Does constantly instantiating new instances of an object (e.g Mario guy) have a horrible overhead that makes it a very serious/important design decision ?
EDIT
Sorry for this additional edit, I wasn't what good practice is on here about follow up questions...
4) If immutability is something I should strive for and even jump though hoops of instantiating new versions of objects that have changed...And If I instantiate my guy every time he moves (only with a different position) don't I have exactly the same problems as I would if he was mutable? in as much that something that referenced him at one point in time is actually looking at old values?.. The more I dig into this the more my head's spinning as generating new versions of the same thing with differing values just seems like mutability, via hack. :¬?
I guess my question is: How should this work? and how is it beneficial over just mutating his position?
for(ever)//simplified game-loop update or "tick" method
{
if(Keyboard.IsDown(Key.Right)
guy = new Guy(guy){location = new Point(guy.Location.x +1, guy.Location.y)};
}
Also confusing is: The above code means that guy is mutable!(even if his properties are not)
4.5) Is that at all possible with a totally immutable guy?
Thanks,
J.
A couple comments on your points:
1) Yes, maybe. To reduce overhead, a practical design will probably end up sharing a lot of state between these instances. For example, perhaps your little guy has an "Equipment" structure which is also immutable. The new copy and the old copy can reference the same "equipment" structure safely, since it's immutable; so you only have to copy a reference, not the whole thing. This is an common advantage you only get thanks to immutability -- if "equipment" was mutable, you couldn't share the reference, since if it changed, your "old" version would change too.
2) In a game, the most practical solution to this issue would probably be to have a global "clock" and have this sort of processing happen once, at a clock tick. Note that your exact scenario would still be a problem if you didn't write it in a functional style: Suppose H0 is the health at time T. If you passed H0 to a function which made a decision about health at time T, you took damage at time T+1, and then the function returned at time T+5, it might have made the wrong decision based on your current health.
3) In a language that encourages functional programming, object instantiation is often made as cheap as possible. I know that on the JVM, creating small objects on the heap is so fast that it's rarely a performance consideration in any practical situation at all, and in C# I've never encountered a situation where it was a concern either.
If my little guy at position
x10,y100 moves right 1 unit do I just
re-instantiate him using his old
values with a +1 to his x position
(e.g x11,y100)?
Well, not necessarily. You could instantiate the guy once, and change its position during play. You may model this with agents. The guy is an agent, so is the AI, so is the render thread, so is the user.
When the AI shoots the guy, it sends it a message, when the user presses an arrow key that sends another message and so on.
let guyAgent (guy, position, health) =
let messages = receiveMessages()
let (newPosition, newHealth) = process(messages)
sendMessage(renderer, (guy, newPosition, newHealth))
guyAgent (guy, newPosition, newHealth)
"Everything" is immutable now (actually, under the hood the agent's dipatch queue does have some mutable state probably).
If immutability is something I
should strive for and even jump though
hoops of instantiating new versions of
objects that have changed...And If I
instantiate my guy every time he moves
(only with a different position) don't
I have exactly the same problems as I
would if he was mutable?
Well, yes. Looping with mutable values and recurring with immutable ones is equivalent.
Edit:
For agents, the wiki is always helpful.
Luca Bolognese has an F# implementation of agents.
This book (called by some The Intelligent Agent Book), though targeting the AI applications (instead of having a SW engineering point of view) is excellent.
If everything in the global system state, outside the current stack frame, is immutable, unless gives another thread a reference to something on the stack (VERY DANGEROUS) there won't be any way for a threads to do anything to affect each other. You could fire and forget, or simply not bother firing in the first place, and the effect would be the same.
Assuming there are some parts of the global state that are mutable, one useful pattern is:
Do
Latch a mutable reference to an immutable object
Generate a new object based upon the latched reference
Loop While CompareExchange fails.
The compare exchange should update the mutable reference to the new one if it still points to old one. This avoids the overhead of locking if there is no concurrent access, but may perform worse than locking if many threads are try to update the same object and generating a new instance from the latched one is slow. One advantage of this approach is that there is no danger of deadlock, though in some situations liveLock could occur.
Another functional approach to this sort of problem is to take a step back and separate out the idea of state from the idea of your little guy.
Your state will include your little guy's position, as well as the position of your baddy and it's shot, and then you have some functions that take some or all of the state and do things like generating the next state and drawing the screen.
The timing issues you're talking about when things you want to parallelize depend on each other are real problems that won't magically go away, although the solutions may be more or less convenient in different languages.
Several suggestions have already been made, and there are a variety of concurrency solutions. The central clock and agents would work, as would Software Transactional Memory, Mutexes or CSP (go style channels), and probably others. The best approach is going to depend on the specifics of the problem, and to a certain extent on personal taste.
As for the head-spinning, try not to get too caught up in whether a thing is changing or not. The point of immutability is not that things don't change, it's that you can create pure functions so that your program is easier to reason about.
For example, an OO program might have a drawing function that iterates over all the objects in a scene, and asks them all to draw themselves, where a functional program might have a function that takes a state and draws a frame.
The end result would be the same scene, but the way the logic and the state is organised is very different.
I, for one, find that it's much easier to work on when you have all the data over here, in one big input lump, and all the drawing logic there, encapsulated in some functions. There are some pretty clear architectural wins too - serialization, testing, and swapping out front ends gets a lot easier with this sort of structure.
Not everything in your program should be immutable. A player's position is something you would expect to be mutable. His name, maybe not.
Immutability is good, but you should perhaps rethink your approach to use more concurrent solutions than simple "immutable"ize everything. Consider this
Thread AI gets copy of your position
You move three units to the left.
AI shoots you based on your old position, and hits... shouldn't happen!
Also, most gaming is done in "game ticks" - there's not much multithreading going on!