Halo 4: Content Performance Tips & Techniques

hello everyone my name is Wes grammont I'm the lead technical art director at 343 Industries uh welcome to my talk Halo 4 content performance uh tips and techniques uh this afternoon I'm going to be talking about the work I did during the last year of production on Halo 4 uh to help get the game uh running at frame rate while uh maintaining the Integrity of the visuals that the uh the team worked so hard on so a little bit about me um double majored in computer graphics and illustration at Syracuse University and I've been in the uh the games industry for about 16 years uh the first three as a lead artist and uh the last 13 in various Tech art roles uh I spent nine years at Electronic Arts uh three years uh co-running a studio with uh three other guys and then the last four and a half years at Microsoft uh in the last two with uh 343 so Halo 4 was my 20th ship title

okay so this uh talk is really geared towards folks who have little to no experience with uh content performance optimization techniques um so I'm going to kind of make that assumption uh as I go through the talk um we're going to start with uh discussion around performance planning and the tools that make this kind of work uh go easier um then we're going to uh kind of get into the me of the talk which is common performance issues and ways to uh kind of address those issues and then we'll kind of wrap things up with um uh performance reports and performance scheduling so before we get too far in so what is performance um well in the context of this talk uh at at a very high level it's basically is our game running at frame rate and for Halo 4 uh our Target frame rate was 30 frames per second and so basically if we weren't running at 30 frames per second we had bad performance and we had to kind of then and at that point look a little bit

deeper so when I when I say deeper what I'm talking about is how how long is it taking the hardware to render a frame and we had 33 milliseconds on Halo 4 in which to accomplish rendering of a frame in order to maintain our frame rate and so there's various uh stages at which this can slow down and that we'll we'll talk about that in the next slide um the reason this is important which is probably obvious is that if your performance stinks then your interactivity goes down and your game's no fun um and the reason performance analysis is really important is because what we want to do is maintain that frame rate while at the same time making sure that we um kind of maintain the really awesome Rich environment that our uh art teams have spent so much time on so why Tech art um I think Tech art is uniquely positioned uh to kind of lead in this area uh we kind of have a background in both art and uh engineering so we have a what I think is a a little bit more of a holistic view on things where you know artists might be more focused on just the visuals and

the dev team might be more focused on just making sure the thing uh runs well so when I'm looking at performance analysis I find um having this kind of high level picture in mind is really helpful uh there's a lot more um to this but uh this is kind of the the key stuff so you've got two places where things can slow down you've got your CPU and your GPU and on Halo 4 uh we had two threads where you could potentially slow down on the CPU um or I would say the most common places we could slow down on the CPU um and they were our Sim thread and our render thread and so our Sim thread is basically where we do uh AI updating um Ray casts for Collision that sort of thing and our render thread is where we go through and sort of evaluate sort what's visible um kind of collect uh data up to uh bundle up and hand off to the GPU which we call feeding the GPU and the GPU would be responsible for unpackaging uh the the m and mesh data

Shader resources texture resources and and handle rendering the various elements that we're passing to it so at any point um in this in this pipeline things could slow down and your frame rate uh goes down as a result so before I get too far into uh kind of the nitty-gritty uh I thought I'd share a little story with you um that has some relevance towards this so I joined 343 in the fall of 2011 and up until that point uh the team had been in development for quite a while and they spent aot a lot of time um and energy making some really awesome visuals uh some really cool encounters uh working on the narrative uh but not a lot of time on uh performance so the the two projects that I'd worked on prior to coming to 343 I'd spent quite a bit of time on uh performance optimization so the uh engineering lead uh David berer uh came to me and he said hey Wes uh I'd love for you to kind of help lead the effort to get our game running um at 30 frames per second

and he said he was giving me the performance hammer and I said okay great um this kind of work I'd done in the past so I was like I definitely can help with that um and so uh a couple days later uh I showed up at work and there was this big brown box on my desk and I opened it up and there was literally this performance hammer in the box and so his his comment was basically and I like that this was a very squishy Hammer uh because I think it's very uh relevant um to kind of the performance uh balancing process but basically his comment to me was I need you to work with the art team to kind kind of get things uh running well and uh if you can't accomplish this with the performance Hammer um we'll come in with a sledgehammer so uh kind of with that information I went and approached the art leads let them know that uh we'd be partnering with them to try and maintain their awesome visuals and hit our uh frame rate goals and they were naturally

skeptical um and I could definitely relate to that having been on other projects where um optimization often meant just kind of destroying the quality of stuff uh not because it was desired but because that was what ended up being needed to get to uh performance schools so you know I let them know hey my background is in art I really care about the Integrity of the work you guys have been doing and the reason I'm telling this story is I think it's really helpful to kind of build that trust uh going into the performance uh kind of optimization process I think having the art team know that you've got their uh interests at heart and that you're going to be working with them to try and maintain the all the um love and effort that they put into their work is a really important first step on this journey okay so when we started this uh process um one of the things that was really helpful was coming up with a perf frame budget and so as I mentioned a few slides back we had 33 milliseconds uh in

which to render a frame and right off the top we've got certain costs that we know we're always going to pay and I call those critical costs we didn't really call them that during the project but I needed to call them something for this slide so critical costs and those are things like your HUD if you happen to be uh in a situation where you have a HUD you're going to pay for those all the time um postprocess effects if you've got certain things that you just you know you're always going to do so we' kind of take that off the top we'd ALS also take uh give a kind of a a little overhead right off the bat like a couple milliseconds to make sure no matter what we always had a little bit of buffer in case things kind of went haywire and we just needed some extra wiggle room so our frame rate wouldn't drop and then what with what was left over um we sat down with all the art leads and we kind of mapped out and this is a spreadsheet from uh one of our campaign missions uh we basically mapped out for each area kind of what the classification was for that area so is it combat heavy uh is it a like a a beautiful Vista where we're not going to

be doing a lot so we can really put a lot more um of our frame budget into rendering uh the environment is in an area where the lighting team really wants to go bananas and um you know put a lot of like really interesting lighting there so we need more budget for doing that so we'd kind of go back and forth and negotiate these things and we come up with this uh per frame budget and so we had that for each each mission in each area and so I'm I'm going to mentioned this is more of like a best practice as I mentioned I came onto the project about a year out from ship so I didn't get a lot of opportunity um on Halo 4 to to do a lot of level review a lot of the levels were already kind of laid out but I think it's really important to mention this uh because if you are early on in a project uh getting involved getting the graphics team involved Tech art involved early on um can kind of kind of help head some things off that might uh be problematic or at least you'll know where your kind of problem spots are going to be so you can kind of come up with mitigation plans ahead of

time all right come on PowerPoint all right so this is the most obvious uh Performance Tool and hopefully everybody has access to this but your uh FPS counter um basically if you uh if you're running under your target frame rate this is your early warning sign that you've got a problem and uh just to kind of give you sort of a highle view of what this process L for us looked like for us and looks like for most people is um you find an area you're not running at 30 you go into that area in the game uh you do some timing captures which I'll talk about in a little bit and you kind of diagnose what's going on and you use very various debug modes in the game to kind of help uh sleuth out uh what what potentially could be could be the problem and ways to kind of address it so uh one of the things that was really helpful um I mentioned this Frame budget that we had uh we had a bunch of

in-game counters that would kind of give us the same sort of um information that we'd get from our timing captures and but they just weren't as accurate because the the uh the fact that you were actually measuring this stuff introduced a certain amount of overhead but it was good for ballp Parky things and it allowed our uh art leads to have a way to basically look at at their um timings for their area and see if they were aligning well with the uh per frame budgets um and again it was it was very ballp parkky but uh this was really helpful in kind of a triaging situation where we're like Hey we're not running at 30 um sometimes the the uh the artists could kind of self- diagnose the problem and correct it um but if they couldn't they'd escalated it to Tech art we'd do an analysis if we couldn't kind of come up with some uh ways that they could fix it in content and it was kind of deemed a engine side issue then we would escalate it to our engineering team uh uh I mentioned the debug views

um briefly but basically there was a couple of views that were really really helpful in sort of diagnosing problems uh during performance analysis uh probably the biggest one that I found helpful was a render freeze which basically you you uh freeze the camera it it kind of captures everything that is within the view Frost room that the GPU is currently dealing with and then it let you pop the camera out and you kind of look at it from another perspective so this is really helpful because um in this in this case here this is a shot from the first mission uh in uh Halo 4 where you're kind of uh going down the corridors of the Ford Unto Dawn after just coming out of uh kind of a deep freeze and those yellow lines are the view frustrum and so you can see that there was a lot of stuff being rendered here that wasn't actually visible from the point of view of Master Chief and so this was a case of where we we were able to you know quickly identify that we had some uh culling to do uh the the other kind of uh debug states were very helpful were we had

ways to um uh choose to draw or not draw certain types of objects so uh static meshes versus Dynamic meshes versus effects uh just changing to wireframe mode that sort of thing uh we also had this um Tool uh that would was more of a a kind of a option in when we ran our builds to log uh information to our database and so as we were progressing through the or actually as everybody was kind of progressing through the game uh we'd be periodically sampling uh FPS and then camera position orientation and logging this to a database and over time what this built out was this really nice uh heat map that we could then use a web we had a web tool called Cortana web um that could access this database and provide these heat Maps which provided much more targeted um uh kind of ways to identify where we need to go in and do real real deep performance work um without this we would have been wandering around levels and and really

kind of spending time in places we didn't need to be spending time so Halo 4 you know is a first-party uh exclusive um so we were developing on Xbox 360 and uh the xdk uh comes with this great tool called pix which is the performance investigator for Xbox 360 and this allows you to do CPU and GPU timing captures um and so uh I kind of mentioned our process is you know we find a place where we're running slow it was very quick to uh create timing captures because you could just kind of attach picks and it would just kind of run while the game was running and you could pause it it would pause the game you could just kind of select a region and generate one of these timing captures which basically would give you a uh breakdown um for the range of frames you've selected of kind of what the CPU was working on uh during that time period and the very top row um in this image is uh what the GP is doing

and so a real quick indicator that you've got a problem was uh in this case uh not a problem um is that if you have big white gaps and like if it does if that top GPU line didn't look like a solid black bar meant your GPU is probably waiting around for something to finish up on the Sim thread or render thread so it was a it was a really quick way to kind of say hey I need to Dig Down Deeper into the events that are happening on these threads and see where what the timings are um we'll talk a little bit more about that in a second but this was a first stop because you very quickly rule out that it was the CPU and then the GPU time and captures which talk about in the next slide take a little bit more time because we have to kind of gather up all the information that is used to render a single frame and uh it it it definitely uh it's it's more uh intensive for uh the dev kit so we'd kind of do this only once we' kind of identify that we have a a kind of a GPU related problem in an area and what what this does though is it it grabs everything kind of deconstructs

the frame into this nice uh spreadsheet where you can kind of Click through the spreadsheet you've got timing for each element that's being drawn and so it's very easy to kind of identify what's taking a long time and then you can either um there's some additional Tools in in the in the GPU timing uh where you can kind of analyze to see if you're kind of pixel Shader or vertex Shader bound or if you're you know installed on texture fetches that sort of thing and uh you can also then go into uh into the actual content then and sort of uh kind of analyze it as well to kind of see what what might be causing that to take a while okay so now we get to the meat and potatoes of the talk so these were some performance issues that came up over and over again uh during the uh development of Halo 4 and so I've kind of coated them all because they just they they kept happening over and over again I can only assume they will happen for other people as well um so large draw distances uh in Halo 4 we had some really big uh missions really expansive

spaces and we'd often have performance issues in these areas so generally this is due to uh one of a couple things poor loding uh related small triangles um and also too many draw calls so how do we deal with this stuff oh there's a kind of a bag of tricks some of them some of them are's that going all right there we go uh some of them are a little bit more straightforward than others so small triangles obvious to an LOD distance if we've got small triangles and we have lods we can swap to the next lowest LOD um in some cases we didn't have lods on some assets and that's what I'm showing here uh we had this this very complex model for the UNC Infiniti and it had been modeled kind of for cinematic sequences and we didn't actually know it was going to end up uh making an appearance um in the game pretty close to a combat area and so it was dropped in at the end of one of the missions and we ended up this was a case of small triangles where we just

had we were processing uh well not only just small triangles but too many triangles and so kind of removing a lot of the smaller details that you couldn't really appreciate at the distance it was at we basically generally you know created a one-off l for this case um similarly at the end of the second to last mission um there's this kind of Atrium area where UNC has kind of built this thing on an asteroid and it was modeled uh to kind of really look good when you're wandering around on foot but later on in the mission you kind of Traverse upward in an elevator and the art team had kind of just duplicated the mesh over to build this this other asset and hadn't really taken time to uh kind of evaluate the size of the triangles and so we had a lot of tiny little triangles which you can see I've circled all over the place saying hey you should probably delete these and they did and uh and our performance got back to where it needed to be in that area Okay uh another technique uh draw calls so so when you're dealing with big distances and you have a lot of draw

calls um this one may be again pretty obvious uh is you you want to call stuff and um but I was surprised at how many times we were drawing things that we didn't need to be drawing and nobody really noticed um so this was uh this was from the last mission you had we had this Trench Run um we were flying a broadsword uh down this trench and it was running really choppy and the team was very concerned about having to you know impact the visuals kind of get it into performance well it turned out we were just rendering um chunks of stuff just way way out into the distance and it was stuff that you couldn't even see um it stuff that wouldn't even render in the frame so this was a real simple case of just uh pulling in the uh the distance coling so we were dropping things out sooner to reduce draw calls um it didn't impact the visuals at all um and it got us back to where we needed to be for that area so sometimes when you've got these big expansive spaces though you don't you can't really call things you have to

draw them and that's where proxy Geo uh really kind of is a is a big help so in the first mission of uh Halo 4 you're again you're on this uh severed half of Ford Unto Dawn and uh probably about a third of the way into the mission you realize for the first time that there's these Covenant forces on the ship and you're looking out kind of over this Observatory decking onto this uh missile Bay kind of open area uh Beyond and it's actually going a combat space later on in the mission um but at the point where you you're viewing it we had some really bad performance issues because not only are you rendering this big space but we were doing it through a glass kind of surface and so any there's a lot of problems uh kind of compounding each other and so in this case we were able to uh get most of the way back into perf by um just creating a lower res uh version of this this uh big outer area and then taking an in-game uh screenshot of the highres geometry projection mapping it back onto

the lower res geometry and it looked the same um so we you know it didn't didn't drastically affect the visuals but it helped us get back into performance for that area Okay so sometimes you want to call things and you just they're right in front of your face and there's nothing you can do and we had uh but that's not true there is nothing you can do which is occlusion cards um so this is a technique I first uh started using on connect Adventures uh basically if you model the the heck out of an area and you've got you know a ton of Geo but you you kind of need to um you kind kind of need to get things back into scope um this is kind of a a way to approach that one of many but um so occlusion cards the basic idea is this hopefully this runs so you've got this Shader um that just Fades out as the camera approaches it so so when it's opaque you can call

everything behind it um and and then you just kind of start drawing it just before it starts to fade out um and I should also note that these don't have to be a plane it could be uh kind of a dome or it could be any kind of an arbitrary shape and it doesn't have to be kind of a a fog kind of plane here like this you could put textures on it you can put animated uh kind of shaders and I'm going to show you some examples of that so it doesn't it doesn't have to be a very obvious kind of just straight up fogging technique that that is um being used here so some flavors of this that were used on Halo 4 uh this is a bloom occlusion card um basically what we're doing here is we are uh using it to kind of fake a uh kind of a fake HDR exposure so we've turned up the color really bright and we let our post-processed Bloom kind of blow it out um but while we're coming up this tunnel we're basically rendering nothing behind that thing and as we come up uh just before it starts to fade is when we kind of bring in that huge complicated Vista

uh that we reveal as you kind of Crest over this rise another example here this is at the end of the second mission in Halo 4 um we had this very complicated Forerunner um building and a kind of a pretty heavy encounter going on in this area so when you first enter this space we're having performance problems CU not only were we rendering this this really complicated exterior um there was a hallway at the base of this structure in uh let me see if I can get this to work here right in this area and so what we did was we created what I call an imposter occlusion card in this space where uh we took a screenshot um similar to what we did with the proxy Geo of the interior of that space before we called it and then we just used that as a texture map on one of these occlusion cards and we do a kind of a cross Spade um and in the context of the gameplay you're kind of in this lower level so you don't really

tend to notice this I'm where I'm standing in the in the in the level is probably the where it is most prominent you really need to be kind of keeping your eye on it to really notice it but uh it got us back into perf for this area and then another kind of variation on this is a fog occlusion card and uh this is actually it's not just a fog occlusion card it's it's kind of like an animated fog kind of spilling over effect so it's just a card like we I've been saying where it kind of Fades out but what we did here was this big space off to the left which is has a ton of geometry in it uh we were having some performance issues when we came down the hall here so we just C all that and put this kind of fog kind of animated spilling over and it kind of masks the fact that we basically uh if we took that fog away we're just looking outside the world um there's nothing rendering there so it it kind of just masks the culling happening in that area

okay let's go power point okay there we go um so the next performance issue uh kind of related to the draw call stuff is uh poor culling so this is um can sometimes be stuff that's actually in view in the camera view frustrum but a lot of times poor cing can be do to things that are not even in the camera view frustrum or or seem like they're not in the view frustrum but we're still paying for them um sometimes well usually this is due to really big meshes um it can also be due to uh if we aren't um rendering the right stuff into the Z prepass so that we can kind of early out um both of these can be contributing factors uh this next video is going to kind of show you the problem with big meshes so if you got a really complicated mesh tends to be like a Terrain mesh or like maybe like a big building or something um even when a little polygon and view we have to kind of uh have the whole thing uh

being processed where if we chop it up we can immediately start culling things out as they move out of the camera frust them so it's this doesn't really affect visuals at all it's really more of a Content fix again that's what the like the focus of this talk is basically how can we modify content to kind of get it back to where it needs to be and this is really a kind of a lwh hanging fruit uh an example of this uh is there was a uh Mission uh on Halo 4 where we're kind of in this desert environment we're riding on this really big vehicle called the mammoth and we come to the end of it and um at the end of the mammoth ride and we jump off and we're in kind of this this Grotto that we called sniper alley which you can see here is this little area right here and we were having some performance issues and it turned out it was exactly this big mesh problem so when we went into the render freeze you could see that we had this entire giant terrain mesh this is a actually a a problem at a couple different spots on this Mission and um like I've just said

chopping it up uh solved the problem uh also had the added byproduct of giving us a little bit better management over our light Maps um another thing that we could do in these areas too that helped with uh with culling is um we use a uh a portaling system to uh do a calling on the CPU and so tightening up those portals and making sure that they were really crafted as well as they could be uh helped a lot too cuz when we first build these spaces out we keep this stuff kind of loose cuz the designers would still be adding and removing things but as we got closer to ship getting those things really aligned with the the content was important um so another big area performance issue uh expensive shaders and I'm probably all of you have experienced this um so lots of Samplers lots of math instructions all contribute to shaders being expensive so the solution is Shader optimization um this image is a picture of some Forerunner metal um we had a fairly complicated metal

Shader that we used for Forerunner surfaces and the last mission of uh Halo 4 had basically metal everywhere and so we had to create a more optimized version for that mission and this one was made about 20% cheaper just by switching out the specular model um which reduced the math instruction significantly um kind of going along the Shader optimization theme we had this idea of production shaders versus ship shaders so production Shader is kind of like a generic Shader uh that the artists can kind of experiment with try different things out but when it came time to ship and we knew a Shader how a Shader was going to be used in a specific context uh we would dup that Shader off and just hack the hell out of it to get it like just mashed down and so but maintain the look as much as possible so uh we would do that we would kind of pack uh channels differently uh we'd reduce or uh removed shading terms

if they were subtle or not used at all um this image here is uh an example of on um was the third mission of Halo that the art team had spent quite a bit of time kind of crafting the look of these sort of volcanic rocks on that mission and they were very concerned about maintaining the look um but we'd reached a point then the kind of the optimization process where these rocks were definitely they were occupying a lot of screen space and they were quite expensive to render um but we were just a little bit over so I you know I asked the art lead hey would you mind uh if I took a crack at optimizing um this rock Shader and his response was sure but it needs to look exactly the same and so I said okay fair enough um I was able to get it about 15% faster and what I did was rather than do like an AB comparison when I showed them the results um I was inspired by Sesame Street you guys ever heard like one of these things is not like the other yeah okay so I built a scene with about

five of these rocks and only one of them was the optimized one and I sat them down in front of it and I said if you can tell me which one is the optimized uh one I'll go back to the drawing board and he sat there for about a minute and then pointed out the wrong Rock so the optimization stuck and the reason I tell that story is because I think uh it's definitely a technique I'll use in the future um I think when you've got uh some maybe some built-in bias um I think making things a little more obis gated when it comes to kind of these sort of subjective judgment calls uh it makes sure that people are really being honest about what's important visually um another thing I want to mention is our art team did a lot of Kit bashing to kind of build assets so a lot of times we'd end up with just gobs of materials on what were otherwise very simple assets and those materials would often be sampling uh a ton of textur so um doing some material optimization was also uh an important part of this Shader optimization

process so overdraw um rfx team was actually very sensitive to this uh this problem this is where we're drawing just too many transparent things on top of each other uh and it starts to bog things down so despite uh some of our our efforts in this area you still have cases where a particular effect might end up being placed in a certain context and it might be just the straw that breaks the camel's back so um we would solution in these areas would be to either kind of tune the emission rates to minimize the uh the amount of overdraw that was occurring um sometimes we'd go back ask the effects team to go back to um sort of do some additional work on the materials that are applied to these Sprites to add complexity in those materials rather than achieve that complexity by overlapping a lot of Sprites um another area kind of related uh we had a lot of lens flares in Halo 4 um a lot and and so we had we had one uh

area in one of our missions where it was just really running terrible and it turned out there was just all these overlapping quads from all these lens FL lens flirt elements and so the solution here was creating a more optimized lens flare um for that specific use case and that's really a common theme in a lot of this uh kind of performance work is when you're trying to kind of find this sweet spot with the visuals is rather than trying to create a solution that is um going to fix it for your entire uh game is just really just tune things in the places where they need to be tuned so that way your visuals don't get affected in areas where there is no problem um the other thing we did uh our Graphics team would occasionally render to a kind of a quarter reses buffer uh so that wouldn't take as long to render our transparence that helped as well so skyboxes were another area where we had lots of transparent pixels um when we first started doing the skyboxes were gorgeous um when they first started doing them

though they were just kind of putting them on these big concentric quad strips and so we were we were paying you know they were just textures with Alpha in the sky so all those pixels in the sky we were paying uh overdraw cost for nothing visually so the solution there was to go in and kind of cut the geometry so that we aren't paying for basically fully transparent pixels and then taking it a step further to go in and actually build a quad strip along the um the right along the edge of the kind of the mountain ranges and stuff so we only pay transparent cost right along that edge line and then put an opaque Shader on the on the lower portions where we don't need the transparency um this uh same technique can be extended to water um if you if you're doing kind of depth blending along a shoreline and you're kind of fading out to kind of deeper water um as soon as you get into that deeper opaque water you can just cut the geometry there and put a fully opaque Shader in that area so you're basically you've got opaque pixels in the center of of your your water masses and you're not you're only paying for transparency along the

shorelines we use that on connected Ventures not on Halo okay another performance issue lighting costs um so Hala 4 uh used a deferred renderer and so we had uh two types of Lights we had screen space lights and we had Dynamic lights and there was good uh use cases uh for using either one but screen space lights were definitely uh the more efficient of the two and so um but we also got immediate feedback with the dynamic lights so the lighters would often work with fully Dynamic lights and only switch things to screen space when they were ready to uh to uh kind of call it final um so occasionally we'd have lights that just slipped through the Gap that where they they forgot to flip that switch and turn them into screen space that was a common thing we just have to check for uh some other cases though even with screen space we could have some uh performance hits and this is a this image here is a good example of that we had these little uh Covenant light sticks that had these

just huge falloff radiuses to the lights and this is actually a frame from pix and again pix is really helpful it's got some um some cool uh visualization modes and this one allows you to see the wireframe for the fall off of the light and so when when we saw that these lights were expensive and usually screen space lights were quite cheap um it was a good uh way to kind of say hey we've there's you know we've got to um got to reduce the radius on these guys and that was a real quick fix um kind of speaking to the kind of Lights being left behind we had we had like we were over in this area and we nobody could figure it out turned out it was there was these tiny little spot Dynamic spotlights in the middle of nowhere that were costing 3 milliseconds which is a huge part of your 30 m 33 millisecond budget and just uh deleting these solve the problem so um just kind

of being aware of what your content's doing is a a huge part of this so cinematics um we did a lot of work on cinematics to get them into performance uh uh we had some very complicated characters uh Cortana was by far one of our most expensive things to render uh in the game uh she clocked in around uh 5 to six milliseconds due to the the various render targets we had to uh great uh to achieve that effect um so we had this sequence at the end of uh the game where Cortana um let's see I'll play it while I talk uh where Cortana is kind of she comes up and we've got this pretty complicated effect behind her and uh she's doing some stuff and we were very concerned about this uh particular shot and so because we were she's expensive uh she was draw drawing over other transparent things in a really huge environment um

kind of all the problems I've just uh kind of called out all in one shot and so uh some of the ways we dealt with this uh one we kind of pulled the plin back a little bit so that we could and and tilted the camera down so that we could get more opaque pixels covering the screen so you'll notice this this uh kind of platform you're standing on occupies about half the frame that helped improve uh the render time you also notice we put this um barricade behind her in between her and that uh that kind of orange orb effect and what that allowed us to do is again this is to address overdraw issues we basically put some opaque pixels back there and so where she's rendering we're not paying all that overdraw cost in addition of the cost of just rendering her so um those are some techniques that helped uh ensure that we uh we could get that shot achieve that shot for narrative um finally I just want to mention um I kind of touched on this at

the beginning of the talk uh you definitely want to make sure you optimize your post effects uh on Halo 4 we had two types of fog we had a a patchy fog and a screen space fog and a lot of times we had both of those turned on at the same time we didn't really need it um so kind of being aware of what you're what you're actually using and whether or not it's really contributing uh significantly to the final look of the of the frame and then um having your making sure your uh all the things you're paying critical costs for uh things that we rendering all the time like the HUD are as efficient as they can be and uh early on in this process our HUD actually wasn't very efficient and our UI team team did a great job kind of getting that down as far as they could uh so that it wasn't eating too much of our budget up so I mentioned at the beginning of the talk GPU starvation um there are a lot of cases where your GPU is not bottleneck it's your CPU and so this can

be due to a variety of issues uh slow Collision um poor culling uh just too many AI simulating all kind of contribute to GPU starvation um so ways to fix that we T we've already talked about some of the culling techniques um going in and optimizing collusion meshes uh helps uh get your your ray cast costs down and we did quite a bit of that on Halo 4 um we had two types of collision we had a play collision and a bullet collision and we would go in and optimize those as necessary to uh to get those raycast times down we also um had times like the one shown here uh where we were spawning in way too many Dynamic objects and or AI um and we would the solution here was to go talk with the design team and they would have to uh kind of stagger the spawning of objects um which usually wasn't a problem because it was usually stuff that wasn't really in view anyway it was just and

and they hadn't intentionally meant it to spawn in as early as it did it was just kind of something that happened and so uh at the end of this whole process uh you know the the end goal here is we want to create a bunch of actionable items for the team to go uh address so creating these performance reports at the end are uh probably the most important part of the entire process so so we would create uh kind of call outs per team um environment lighting effects all the ones listed here and we would give them multiple suggestions uh for ways to fix uh each of their areas where they're over budget again these are all referenced back to that per frame budget I talked about at the beginning of the whole uh talk so they're they would get kind of a call out for each area like hey you know you're maybe a millisecond over here's some ways you could uh kind of get that millisecond back and we'd give them you know maybe half a dozen different uh options and to me that was very important because it let them own the

final decision on what got optimized and we we definitely could have gone in and just said hey this is what we're doing live with it um but you get a lot more buyin from the team by involving them uh in the process and and allowing them to help be the decision makers for how their their content gets optimized so uh performance analysis takes quite a bit of time and I often felt basically like Master Chief here in this image where we had our ship date kind of coming in and there was lots of performance to do and there was not a lot of time um it usually took about a day or more uh per pass on a on a mission to do a run through and that was with our automated uh kind of uh heat map creation process so we could do kind of more targeted analysis um so definitely uh for us it was important to make sure that we had time scheduled not only for the initial analysis but then uh those those uh reports that I just mentioned would get turned into um bugs in our bug tracking software and then uh

when they were resolved it would be on us to go back and follow up with folks and make sure uh or follow up in those areas that we' called out and see if they'd actually uh addressed the issue and then if not we had to you know generate a follow-up report and more bugs um the other thing we did uh towards the end uh was making sure we had scheduled playthroughs um on release kind of builds um when we're when we're kind of in the midst of production uh we run our our builds that we were running with had a lot of extra counters in them to uh give us more um information on timing um but those uh similar to the case where you've got the in-game uh debug menu that kind of introduces some overhead even with the uh the timing captures just having those counters in there also um introduces a little bit of overhead so really we got our true numbers when we actually did our full release build and then we would have very minimal um kind of instrument mentation in there to kind of get a sense of where frame rate was uh and we

would make sure we were doing playthroughs in both uh full screen and split screen because that was one of our modes and we wanted to make sure that um that was a good experience on both those uh those modes so before I wrap up here uh I want to call out this is a talk that's happening tomorrow if you uh have a lot of interest in this uh performance stuff uh there's going to be a talk focused specifically on Xbox One um on the performance tools and uh definitely suggest checking that out and I just want to say thank you to the entire 343 team to Bonnie Ross Josh Holmes and Neil Harrison for allowing me to share these stories with you and for to David burer for entrusting me with the perf hammer and that's it thanks [Applause] [Music] everybody so I can take question do we have time for a question yeah so if anybody's got

questions hey uh so with the occlusion cards did you just let it rendered to your duck buffer when it was fully op or did you do some other way to call the aore behind it no so the uh so yeah actually I'm glad you mentioned that um so there's there's a couple ways you can do this so the depth the the the depth uh occlusion cards uh are part of the transparent path they don't render the to the depth buffer as you as you pointed out but one thing you can do is if you want to you can put like a really big uh you could dupe the the the uh clusion geometry and maybe move it back a little bit and put an opaque Shader on there and that can write to your like your Z pre pass and that can be like another way to uh do your calling um and you just turn that off before you start fading so would that would that do Dynamic austion queries then for the actual objects though to stop the draw calls or did you like actually like have somebody hook up like hey these objects will turn off yeah no so so in our engine we had a basically a flag you could set that just says past this distance just stop evaluating this

object and so we just set that and we had some tools uh that would just very easily uh let us set that on a bunch of of them all at once yeah all so in this you were talking all about your career missions so I was wondering how much if any of this applied to the multiplayer maps that's a great question uh it very much applies to it um we couldn't use as many of the tricks because multiplayer is kind of like you've got uh just everybody everywhere all the time um so we we were a lot more strict about our uh kind of requirements for those maps about um kind of how complicated they could be CU um there was just too many situations where you just you never knew if there was going to be a bazillion grenades going off or what the you know kind of the the situation was so uh we were we were we could be a little bit more crafted in the campaign and multiplayer we' had to be a lot more strict so especially with the Forge Maps where you know you have a huge amount of content that

players can put into the maps was there any kind of optimization that you could do for those uh not a lot um other than put caps on the number of uh things you can place in the in the in the Forge environment that's really uh where the where the optimization kind of is is just kind of creating a a built-in kind of budget for for players to kind of be able to adhere to so we don't don't create a scene that will be unplayable okay yeah thanks sure all right thanks [Applause] [Music] [Applause] everybody