Tuesday, July 19, 2016

Is Anyone Generating PMREMs In-Game in Real-Time?

One of the standard solutions that has emerged for physically based rendering (PBR) is to use pre-filtered mipmapped radiance environment maps (PMREMs).

Put in very imprecise terms, a PMREM is a cube map that has a perfect reflection of the environment in the lowest (largest) mip and successively blurrier images in the lower mips. Typically the image is convolved with your material's distribution function at various roughnesses, and a cosine lobe is used at the lowest resolution.

From this, specular lighting is sampled from the LOD level that matches material roughness, and a low LOD is used for diffuse lighting.

What I haven't seen is any papers on games doing this entirely in-engine.

There's plenty written on baking these cubemaps ahead of time, and it appears that quite a few engines are augmenting PMREMs with screen space reflections (with various heuristics to cope with materials being rough and all of the weird things SSR can have).

But the only work I've seen on real-time PMREMs is the old GPU Gems 2 chapter that projects the original cube map into spherical harmonics (and then back into a cube map) as a way of getting a reasonable low frequency or diffuse cube map. But this was written back when a clean reflection and a diffuse map were good enough; it doesn't handle rough specular reflections at all.*

The problem that X-Plane faces in adopting modern game-engine graphics is that we can't bake. Our "level" is the entire planet, and it is built out of user-installed scenery packages that can be mixed and matched in real-time.  This includes adding a mix of surface objects onto a mesh from another pack. Baking is out the question because the final assembly of the level only exists when the user starts the app.

So I have been experimenting with both SH-based convolution of cube maps and simply importance sampling a distribution function on the GPU. It appears we're reaching the point where both of these can potentially function in real-time...at least, if you have a big GPU, a relatively low-res cube map (e.g. 256x256, not 1024x1024) and only one cube map.**

My question is: is anyone else already doing this? Is this a thing?

* You can add a lot more spherical harmonic coefficients, but it doesn't scale well in image quality; the amazing thing about SH are that the artifacts from having a very small number of bands are, perhaps by luck, very acceptable for low frequency lighting. The problem is that, as coefficients are added in, things get worse. The original image isn't reconstructed well (for the number of bands we can hope to use on a GPU) and the artifacts become significantly less desirable.

** To be clear: importance sampling is only going to work for a very, very small number of samples. I believe that for "tight" distributions it should be possible to find filter kernels that are equivalent to the importance-sampled result that can run in realtime.  For very wide distributions, this is out of the question, but in that case, SH convolution might provide a reasonable proxy.  What I don't know yet is what goes "in the middle".  My guess is: some kind of incorrect and hacky but tolerable blend of the two.

Friday, July 08, 2016

Worst. Lock. Ever

Here's some really good code:

    StLock take_lock(&global_asset_lock);
    delete this;

This was the code to release an art asset in X-Plane 10.45r2.

Can you see what's wrong with it? Don't over think it; it's not a relaxed vs sequential atomics issue or a bug in the RAII code. The logic is roughly this:

  • Decrement my reference count.
  • If I was the last one (and my count is now zero):
  • Lock the global table of art assets.
  • While the table is locked, clear out my entry.
  • Delete myself.
We should get "more optimal" and drop the lock before we delete ourselves, but that's not the issue either.

The issue is a data race!  The race exists between the completion of the atomic decrement and acquiring the table lock.  In this space another thread can come along and:

  1. Try to load this same art asset; it will acquire the table lock, look up our art asset, find it, increase the reference count back to 1.
  2. It will then drop the lock, leaving us exactly how we were except our reference count is now 0.
  3. When we proceed to nuke the art asset we will leave that other client with a bad pointer and crash.
I found this because I actually caught it in the debugger -- what are the odds?

Being too clever for my own good, I thought "let's be clever and just re-check the reference count after we take the global lock; in the rare load-during-delete we can then abort the load, and in most cases releasing our reference count stays fast.

That fixes the above race condition but doesn't fix this other race condition: in the space between the decrement and the lock:

  1. Another thread takes the global lock, finds our art asset, and increases its reference count, drops the table lock.
  2. That thread then releases its reference, which hits zero, takes the lock, deletes the art asset and the table entry and releases the lock.
Now we are the ones with a bad pointer, and we crash re-deleting the art asset.

So I've come to the conclusion that the only safe thing to do is to take the table lock first before doing any decrement. In the event that we hit zero reference count, since we did so while the table is locked, no one could have found our art asset by the table to increase the count (and clearly no one else already had it if we went from ref == 1 to ref == 0).  So now we know it's safe to delete.

This isn't great for performance; it means we take a global lock (per class of art asset) on any reference count decrement, but I think we can survive this; we have already moved not only most loading but most unloading to asynchronous worker threads, so we're eating the lock in a place where we can afford to take our time.

A More Asynchronous Design

I can imagine a more asynchronous design:
  1. The table of art assets itself holds one reference count.
  2. When we decrement the reference count to one, we queue up an asynchronous table "garbage collection" to run on a worker thread.
  3. The garbage collection takes the table lock once to find and clean out unused art assets, blocking loads for a small amount of time while the table is inspected.
Here the reference count decrement doesn't need to do any extra work unless it is "probably" finished (e.g. gets down to 1 reference count remaining) so transient reference count chagnes (e.g. we went from 3 to 4 back to 3) remain as fast as we can perform an atomic operation with some acq/rel semantics.

I have not had a chance to profile the simple solution since trying to make it correct; if it turns out to be a perf problem I'll post a follow-up. The only time we've seen a perf problem is when the table lock was required for all operations (this was a design that we fixed almost a decade ago -- a design that was okay when it was single threaded and thus lock free).

The rendering path in X-Plane requires no reference count changes at all, so it is unaffected by anything on this path.

Wednesday, July 06, 2016

This is Why ASAN Makes The Big Bucks

I've been meaning to post this: this is ASAN (Address Sanitizer) in X-Code 7 catching a memory scribble in an X-Plane beta.

I don't usually drink the Kool-Aid when it comes to Apple development tools. If you say "Swift Playground" my eyes roll all the way into the back of my head like in the exorcist.

My skepticism with Apple tools comes from X-Plane being a big heavy app that aims to use 100% of a gamer-class PC; when run with developer tools on the Mac, the results sometimes aren't pretty. Instruments took several versions to reach a point where we could trace X-Plane without the tool blowing up.

ASAN is something else. It's so fast that it can run X-Plane (fully unoptimized debug build) with real settings, like what users use, at 7-10 fps with full address checking. That's not even on the same planet as Valgrind. (When we tried Valgrind on Linux, we didn't have the patience to find the fps - it never finished auditing the load sequence.)

In this crash ASAN has not only shown me where I have utterly stomped on memory, but it has also provided a full backtrace to where I relinquished the memory that I am now stopping on and where I first allocated it. That's a huge amount of information to get in a real-time run.

In this case the underlying bug issue was: we have a geometry accumulator that takes a big pile of geometry and stuffs it in a VBO when done. (What makes the class interesting is that it takes input geometry from multiple sources and sequences them into one big VBO for efficiency.)

A participating client can't delete their chunk of the VBO until after accumulation is finished, but there was no assert protecting us from this programming mistake. When code does delete its geometry "too early", the removed reference isn't properly tracked and stale pointers are used during the VBO build-up process, trashing memory.

Suffice it to say, ASAN's view of this is about the best  you can hope for in this kind of scribble situation.

Saturday, May 07, 2016

Simultaneous Mipmap Level Generation

The novel trick deep in this post is the use of morton-number pixel addressing to generate mipmaps in parallel; I was surprised to not see anyone doing this so I figured I'd write it up. It is quite possible that no one does this because it's totally stupid, but the initial results looked promising under a narrow set of circumstances.

I was looking at CPU-based generation of mipmaps the other day. The traditional way to do this is with recursion: you calculate mip level 1 (quarter size of the original image) from the original image, then you calculate mip map level 2 from mip map level 1, again cutting down by a factor of four, and on and on.

This "recursive down-size" algorithm has a nice property: the total number of pixels read in the filtering operations is very low - about 33% more than the size of the original image. (We get this "efficiency" even though we are producing several new images because most of the mips are reading from much smaller images than the original.)

But recursively down-sizing has one hidden problem: the smaller mips are not actually computed from the original image! If you have a 256x256 image, that last pixel is computed from data that has been re-sampled seven time! Your lower mips are on the wrong end of a game of telephone.

Does this matter? That all depends on what your filtering algorithm is. If you are doing a simple box filter average, then it almost certainly doesn't matter - the limits of that filter are a lot worse than the recursive error.

But there are algorithms that make approximate mipmaps. A common trick these days in physically based rendering engines is to increase the roughness of lower mips based on the divergence of normals in the higher mips. The idea is that at lower mips our normals (pointing in all different directions) should be scattering light, but they've averaged out. We increase roughness to say "some scattering happened under this pixel that you can't see any more."

The problem is that this roughness increase is very much an approximation; the last thing we want to do is approximate an approximation over and over by recursively down-sizing. Whatever the problems of the algorithm, let's limit ourselves to absorbing the error once by always working from the highest level mipmap.

Two Ways To Sample

A naive way to compute these mipmaps is to compute each mip level from the original image, with each higher number (and smaller) mip level sampling more source pixels for each destination pixel. This solves our quality problem, but it's slow - for a 256 x 256 image, we're going to do eight full passes over the source image - each time we make the source image twice as big, we eat another pass. This is way worse than 33% more samples that we had in the recursive algorithm.

(Quick side note: note that overflow must be re-examined. Recursive mipmaps sample at most four source pixels for each destination, so for example we could use a short to accumulate unsigned bytes. The "use the source" image might sample 1M pixels to build the last mip, overflowing accumulation data types.)

I then got an idea: what if we could build all of the mipmaps at once in a single iteration? The algorithm would go something like this:

  • For each mip level we will compute, start a "bucket" to accumulate pixels.
  • For each pixel in the source image, read that pixel and accumulate it into each bucket.
  • For each of those buckets, if it is full (that is, if it has accumulated enough samples to be done), compute the destination pixel and reset the bucket.
For the first mip, we are going to empty our bucket and move on every four samples; for the last level mip (1x1) the bucket empties only once at the very end.  Each bucket is an accumulation of the exact right number of source pixels, and each source pixel is read only once.

In order to make this work, we have to use a very particular iteration order: we have to cover all four pixels under the first pixel of the first mip before we can go on to the fifth pixel - this applies at all mip levels. It turns out that this is done by -interleaving- the digits of our pixel address in binary.

(If you are familiar with tiling and GPU hardware this is totally unsurprising in any way.)

In other words, if we are doing the 87th pixel of a 256 x 256 image, that pixel number is 01010111 in binary.  We split the digits to get an X address of 1111 and a Y address of 001, that is, this is the 15th X pixel and 1st Y pixel. The first 64 pixels are in the lower left 8x8 of the image - they must be because we have to read those first 64 pixels to compute the 3rd mip-level's first pixel.  We then move to the right by 8 pixels - the next 16 pixels are used for the next 4x4, so we move four more to the right.  The next four pixels are the next 2x2, leaving 3 pixels left at 14,0 - in that 2x2 we are in the upper right, giving us an address of 15,1.

(Note: if your image is not square, all bits to the left of the number of bits needed to compute an NxN square, where N is the shorter dimension, must be applied only to the longer dimension.)

There is one problem with this algorithm: it is really, really, really cache unfriendly. The reason it is so cache unfriendly is that our image is stored in linear order, not tiled order, so all of those Y hops are jumping way forward and backward and all over the place in memory.  Our images are too big to just sit in cache (e.g. 4 MB) but in our old algorithms we just read through them in perfect order. That meant that that for each cache line fetched, we used the whole thing, and more importantly, our access order was so bleeding obvious that the hardware prefetcher could know exactly what was going on and have the next bytes ready.

To performance test this, I put my down-sampling algorithms into traits classes and wrote templated down-samplers for recursive, linear, and parallel mipmap generation. To test, I used two down-sampling algorithms:

  • "Raw" down-sampling simply averaged pixel values, as integers.
  • "sRGB" down-sampling converted the pixels to floating-point linear space, averaged in linear space, and then converted to sRGB.  The sRGB conversion is a correct conversion (e.g. linear at the bottom, 2.4 power on the top) in floating point, rather than a table lookup or approximation.
Here's the numbers, in ms for a 2048 x 2048 RGBA image:
            Raw     sRGB
Recursive   18.4    374
Sequential  72.1    2732
Parallel    124.9   392

As you can see, recursive is always fastest, and parallel is worse than sequential for purely data-movement workflows like "raw".  But when we go to the sRGB work-flow, parallel is much faster than sequential, and almost as good as recursive.

Why? I think the answer is that the conversion from 8-bit sRGB to floating point linear is expensive - there's branching, power functions and data unpacking in that conversion.  And the parallel algorithm absolutely minimizes this work - every source pixel is decoded exactly once - that's even 33% less than the recursive algorithm.  (Every algorithm has to write the same number of pixels in the end, so there's no change in absolute work then.)

My gut feeling is that this "win" isn't as good as it looks - the more data limited and the less CPU limited our algorithm, the more like 'raw' and the less like 'sRGB' we are.  In other words, there's no real replacement for recursion if you can afford it, and they wouldn't look close if I spent some time tuning my sRGB conversion.* And if there's any long term trend, it's code becoming more input data limited and less ALU bound.

You Know We Have GPUs, Right??

At this point you should have only one question: Ben, why in the hell are you not using the GPU to compute your mipmaps???

There are a few conditions particular to the X-Plane engine that make on-CPU mipmapping interesting in X-Plane.
  • Most of our mipmaps are precomputed in DDS files. Therefore most of the time, our interaction with the driver is to feed the entire mip pyramid to the driver from the CPU side. By computing mipmaps for PNG files on the CPU, we ensure that every one of our textures goes into the driver in the same way. 
  • Since we are OpenGL based, the driver needs to manage residency of textures; if the driver decides to prioritize PNG files because they are more expensive to swap off the GPU because the mipmaps aren't already resident in system memory too, that's not good for us.
  • Some day when we use a lower level API, we will need to manage residency, and it will be easier to have only code flow (system-pool for all resources, page them to the GPU when they are in the working set) for all textures.
  • Disk-texture upload is done entirely in background threads - the sim never waits, and pre-loading is N-way threaded during load, so the cost of CPU time is actually pretty cheap in a multicore world.
Anyway, if it turns out that parallel mip generation is "a thing" and my Google Fu is just weak, I'd be curious to hear about it.

* I did not try a table-based sRGB sRGB decode, which would be an interesting comparison - I'll have to try that on another day.  I did try RYG's SSE + table-based re-encode, and it does speed things up compared to a non-SSE implementation. I was able to modify it for alpha support but haven't had time to write a matched decoder.  (I didn't see any back story to the post, so I don't know if decode should always just be table-based.)

Friday, January 08, 2016

Work Stealing and Lock Free Chaos

Pablo Halpern explains work stealing at cppcon 2015:

A good introduction to more advanced parallel task scheduling strategies.

X-Plane's task scheduler for multi-core work is a central work-pool, similar to libdispatch and pretty much any basic worker-pool design that isn't trying to get too clever with continuations.  Our assumption is that the amount of background work is (hopefully) relatively large; typically the work is "cold" when it's scheduled and is going to a core other than the one running the main rendering lop.

And Fedor Pikus on lock-free programming.  There's some mind-bending stuff in here. For example, is your concurrent FIFO really a FIFO when used with multiple threads?  No one knows, because the only way to test it (from a correct race free program) disallows testing concurrent ordering.

(I view that as a win - one more thing I don't have to code...)

One good point from the talk:
Design only what you need, avoid generic designs.
Avoid writing generic lock-free code.
In other words, making a fully general lock free thingie with no fine print and no strings attached is really, really hard. After you spend more time engineering your super-data-structure than you should have, you will end up with something that has subtle concurrency bugs.

Instead, look for chances to simplify the problem in the application space.

An example from X-Plane: the APIs around our art assets are not lock-free. But the API to use them is lock-free, and that's actually all we care about.  Loading/unloading happens in the background on worker threads and is always asynchronous.  It has to be, because the actual work of loading/unloading the art assets is often non-trivial.

But once you have a reference count and pointer to an art asset, you can use it with no further checks; the reference count guarantees it isn't going anywhere. This means the rendering loop can proceed and never gets locked out by loading threads, which is what we care about.

This isn't a "fair" design, and it's not lock-free or wait-free, but it's unfair in exactly the way we need it to be and it's faster on the path that actually matters.

(It also has a subtle race condition due to an implementation error...but that's for another post.)

Saturday, December 19, 2015

The Dangers of Super Smart Compilers

For the first time today, I ran an optimized DSF render using RenderFarm (the internal tool we use to make the global scenery) compiled by Clang.
The result was a segfault, which was a little bit surprising (and very disheartening) because the non-optimized debug build worked perfectly, and the optimized build works perfectly when compiled by GCC. When -O0 revealed no bug (meaning the bug wasn’t some #if DEV code) it was time for a “what did the optimizer do this time session.”
After a lot of printf and trial and error, it became clear that the optimizer had simply skipped an entire block of code that went roughly like this:
for(vector<mesh_mash_vertex_t>::iterator pts = 
   ioBorder.vertices.begin(); pts != 
   ioBorder.vertices.end(); ++pts)
if(pts->buddy == NULL)
   /* do really important stuff */
The really important stuff was being skipped, and as it turns out, it was really important.
So…WTF? Well, buddy isn’t a pointer - it’s a smart handle, so operator== isn’t a pointer compare it’s code. We can go look at that code, let’s see what’s in it.
The handle turns out to just be a wrapper around a pointer - it’s operator* returns *m_ptr. Operator== is defined out of line and has a case specifically designed to make comparison-with-null work.
  template < class DSC, bool Const >
  bool operator==(const CC_iterator<DSC, Const> &rhs,
                  Nullptr_t CGAL_assertion_code(n))
    CGAL_assertion( n == NULL);
    return &*rhs == NULL;
Of course, Clang is way smarter than I am, and it actually has commentary about this very line of code!
Reference cannot be bound to dereferenced null pointer in well-defined C++ code; comparison may be assumed to always evaluate to false.
Oh @#. Well, there’s our problem. This operator==, like plenty of other semi-legit code, is “unpacking” the handle wrapper by using &* to get a bare pointer to the thing being wrapped. In practice, the & and * cancel each other out and you get the bare pointer that is secretly inside whatever you’re working with.
Except that Clang is sooooo clever. It goes “hrm - if &*rhs == NULL then what was *rhs? It’s a NULL reference (because rhs is NULL and we dereferenced it). And since NULL objects by reference are illegal, this must never have happened - our code is in undefined behavior land as soon as *rhs runs.
Since our code is in undefined behavior land (if and only if *rhs is a “null object” if such a thing exists, which it doesn’t) then the compiler can do whatever it wants!
If *rhs is not a NULL object, &*rhs won’t ever equal NULL, and the result is false. So if one side of the case returns false and the other side is undefined, we can just rewrite the whole function.
  template < class DSC, bool Const >
  bool operator==(const CC_iterator<DSC, Const> &rhs,
                  Nullptr_t CGAL_assertion_code(n))
    return false; /* there I fixed it! */
and that is exactly what Clang does. Thus if(pts->buddy == NULL) turns into if(false) and my important stuff never runs.
The short term “fix” (and I use the term loosely) is to do this:
for(vector<mesh_mash_vertex_t>::iterator pts = 
   ioBorder.vertices.begin(); pts != 
   ioBorder.vertices.end(); ++pts)
if(pts->buddy == CDT::Vertex_handle())
   /* do really important stuff */
Now we have operator== between two handles:
  template < class DSC, bool Const1, bool Const2 >
  bool operator!=(const CC_iterator<DSC, Const1> &rhs,
                  const CC_iterator<DSC, Const2> &lhs)
    return &*rhs != &*lhs;
This one is also doing illegal undefined stuff (&* on a null ptr = bad) but Clang can’t tell in advance that this is bad, so the optimizer doesn’t hammer our code. Instead it shortens this to a pointer compare and we win.
Newer versions of CGAL* have fixed this by taking advantage of the fact that a custom operator->() returns the bare pointer underneath the iterator, avoiding the illegal null reference case. (This technique doesn’t work in the general case, but the CGAL template is specialized for a particular iterator.)
In Clang’s defense, the execution time of the program was faster until it segfaulted!
  • You can make fun of me for not updating to the latest version of every library every time it comes out, but given the time it takes to update libraries on 3 or 4 compilers/build systems and then deal with the chain of dependencies if they don’t all work together, you’ll have to forgive me for choosing to get real work done instead.

Thursday, December 10, 2015

Source Control for Art Assets - This Must Exist

I've been thinking a lot lately about revision control for art assets. As X-Plane has grown, our art team has grown, and as the art team has grown, our strategy for dealing with art assets is coming under strain.

Currently we use GIT for source code and SVN for art assets in a single shared repo. No one likes SVN - it was selected as the least bad alternative:

  • Since it's centralized, it's much more in line with what artists expect for revision control - no explaining distributed source control to non-programmers.
  • It doesn't replicate the entire history of an art asset, which is too much data.
  • Parts of a tree can be checked out without paying for the entire tree.
  • There are decent GUIs for every platform.
  • It's scriptable for integration flexibility.
SVN still has some real problems:

  • It is just so slow. You can look at your wire speed and SVN's speed and you're just not getting a fast transfer.Update: this finding is wrong! SVN's speed at transferring binary files is about the same as your wire speed to the server. I'll write up a separate post on speed tests. Many of us are using GUI clients and it is possible that some of them are adding a tax, but the command line SVN client is similar in up/down transfer speed to GIT and rsync for basic data transfer.
  • SVN can't do an incremental update without a working repo, which means having a .svn directory even for the art assets you're not working on. That means at least 2x the disk space on the entire art asset pile, just to be able to get latest.

GIT's Not It

Since I am a programmer, my first thought was: well, clearly GIT can be made to do this, because GIT is the answer to all problems involving files. I spent some time trying to figure out how to shoe-horn GIT into this roll and have concluded that it's not a good idea. GIT simply makes too many fundamental assumptions that are right for source trees and wrong for art asset piles. We'd be fighting GIT's behavior all of the time.

We Kind of Want Rsync

There are two parts of art asset version control: letting the guys who are doing the work make revisions, and letting the people not doing the work get those revisions. It's easy to overlook that second task, but for any given person working on X-Plane, that artist is not working on most of the airplanes, scenery packs, etc.  And the programming team is working on none of them.

For the task of getting art without revision control, rsync would be just great.

  • It can work incrementally.
  • It only gets what you need.
  • It's reasonably fast.
  • It doesn't waste any disk space.
One of the main problems with SVN is performance - if I have to change a branch, having SVN take half an hour to get the new art asset pack I need is pretty painful. So it's at least interesting to look at the architecture rsync implies:

  • Files live on the server.
  • We fetch only the files we want.
  • We basically do a straight network transfer and we don't try anything to clever.
Hrm....I know another program like that.

We Kind of Want The X-Plane Installer/Updater

We solved the problem of getting the latest art assets for all of our users - it's called the X-Plane updater. In case you haven't spent your copious free time wire-sharking our updater, it's really, really simple:

  • All files live on an HTTP server, pre-compressed.
  • A manifest lives on the HTTP server.
  • The client downloads the manifests, compares what it has to what's on the server, then fetches the missing or newer files and decompresses them.
Our installer is (sadly) not content-addressed (meaning a file's name is what is inside it, which naturally removes dupes). If I could redesign it now it would be, but in my defense, GIT wasn't a round when we did the original design. (As a side note, it's way easier to debug server side problems when you are not content addressed. :-)

But we can imagine if it was. If it was, we wouldn't keep a fresh mirror of every version of X-Plane on the server - we'd just have a big pool of content-addressed files (a la GIT) and fetch the subset we need.

Let's Version Control the Manifest

So naively my thinking is that all we need to do is version control our file manifest and we have our art asset management solution.
  • Each atomic revision of a version-controlled art asset pack (at whatever granularity that is) creates a new manifest describing exactly what art assets we have.
  • Art assets are transferred from a loose file dump by syncing the manifest with the local machine.
Here's what is interesting to me: we could use pretty much any source control system and get away with it, because the manifest files are going to be relatively small.

Does This Really Not Exist

I feel like I must be missing something...does a tool like this not already exist?  Please point me in the right direction and call me an idiot in the comments section if someone has already done this!