Rendering the World of Far Cry 4

[Music] so introduce myself I'm Steven McCauley I'm a technical leader Ubisoft Montreal and I'm sort of the lucky guy standing up here today with a speaker's badge that gets to talk to you about all the cool rendering stuff we did but really you know is that we have big teams Ubisoft Montreal and a game I'd far cry 4 is so big that it's a real team effort so I'm gonna try and make sure that I call out all the guys who worked on all the features I'm going to talk about today so am i join Ubisoft to work on a far cry 3 and this is a game that came out two in 2012 it was really great we actually found the loads people liked it ago 90% Metacritic score and you know when you're working on a game that's or the game design is really fun and you have a thing a really cool gameplay loop it's pretty awesome when you're gonna get to get the chance to make him kind of a sequel and do far cry 4 based on that so it's really cool as graphics team as well we're able to stick a lot of the same people so we knew with our

strengths our engine and our weaknesses and we knew what we need exactly what we want to do to improve the engine and it's obviously some of these improvements I'm going to go through today so far cry 4 if you don't know much about it it's an open world first-person shooter world about 10k by 10k it's pretty big we have a day/night cycle as you'd expect for an open-world game and with anything that's been a dynamic time of day as a graphics programmer you know there's kind of a really no hiding place and we're sending Kier out so it's actually kind of country based upon Nepal um we had cross platform and cross generation development so we shipped for five platforms a time and which is pretty tough I mean our mandate as a graphics team was to lead on kind of current gen on Xbox one and PlayStation 4 and try to keep the last gen version the same as what was shipped on Far Cry 3 to did a really good job it kind of gave us a lot of constraints because you had to keep the last gen version kind of running which have kind of affected a lot of our

decisions throughout the project and as the end of the project I was kind of inevitable like only had to go back and optimize a few lower bits about the ps3 you know the 360 in the end that kind of met we had a better product on the next III on all platforms the really given our mandate as a graphics team was developed their new features for the new consoles I'm really going to talk exclusively about that kind of work but I might threw a few tidbits in about the ps3 and a 360 along the way so I've gotta what areas did we improve well we got looked at five areas so materials lighting vegetation and anti-aliasing and a terrain so I'm going to cover the first four today I'm not gonna talk about terrain because hopefully you're all in a car chance excellent talk just beforehand in the opposite room and he talked about the virtual texturing work he developed and he didn't see it make sure you gonna check out the slides it was really awesome work um so let's get started with materials um so um I'll Far

Cry 3 I did a talk it's the graph 2012 talking about how we did some physically based shading stuff we could have had glossiness and reflectivity material parameters and we only had a monochrome specular and a bit of a side ranch icon notice nowadays there's quite definition of physically based keeps kind of changing like what I consider to be based back in 2012 is now the other old way of doing things and physically based now means you know LED material editors and other fancy things so maybe we should talk about more and less physically based and might be a bit more accurate terms um so the goal from Far Cry 3 was that we wanted to improve the appearance of our metals because we only had this monochrome specular reflectance that didn't look very good and we couldnt get color specular and if you look at something like a they're kind of concept art for a place like Nepal you'll see loads of gold and kind of bronze that we really wanted to bring out um and also in the first-person shooter you're finally a weapons cover quite a large portion of the screen so you can get your weapons looking really nice you know but at least part of the screen that's will look pretty awesome

so um look like everybody else we look to the Disney model nothing too original here which actually anisotropic ggx brdf and corresponding Smith's geometry term and just news to a standard and version diffuse we look at kind of other diffuse mode also to the Disney model but really I really started to see it's worth it cost to this moment in time so I'm looking we ended up with kind of four parameters in this zero to one range getting really important keep all your parameters in zero to one for kind of it artists and creativity I'm so bluffing this reflectance metallic and on isotropy so Disney call a reflectance specular and they called glossiness roughness or inverse bossiness roughness we stuck with glossiness for legacy reasons for keeping Far Cry 3 kind of naming the same I really won't go to go across is just yeah I want to that bit into how we did that in isotropy and how we did like an italic so like a metallic parameter I hopefully you all know about this by now so it's pretty simple so you have this

single or a float value that you can blend for your specular reflectance you just blend from a monochrome value to your diffuse albedo which really is any kind of a specular albedo at the same time you blend out your diffuse albedo to make sure you get no diffuse lighting and lots of cool colored specular lighting and simple and it works really really great because what it means you just need a single channel in your G buffer for colored specular rather having two keystore whole other kind of three color terms and actually this is great we could use this feature then for old gen consoles as well which got rid of the specular reflectance fixed it as 0.03 and replace the value in the G buffer with a metallic instead oh you'd do this approach so you do need to give your artist a little bit of training because behavioral your kind of diffuse textures which are no longer diffuse really a kind of changes with that use of metallic value and it can be a little bit confusing for people like you can't notice in your developing say a weapon and you've got like a metal Oh knife blade and you know I could say wooden handle we will find in the texture the

wood looks quite dark when the metals are all going to be really quite high in that range so your artist might kind of wanted to kind of adjust them and kind of make them closer together but we gave them some good color of Photoshop color swatches so to make sure they kind didn't do that and have some good guidelines and you know after a couple of weeks they really caught on pretty well and they did some awesome kind of work and we also introduced an austere choppy to get caller rush metals and also the certain types of cloth so I see how are the anis trouble you parameter increases there across those fears so let's look at the anisotropic ggx distribution formula which is kind of a long and confusing there's a bit scary because if you actually break this down and we look at what we have to store in our G buffer well we need to glossiness terms plus a normal tangent and by normal and that's when you're confusing them because we have to pack a full tangent space into the G buffer and yeah like that's gonna take a lot of values we think um actually we can actually

forward some ideas for animation programmers and they wrote some this auditioning papers from Crytek and also a Nicolas trickle on the bits wit blog about storing tangent space as a quaternion and they've kind of mastered the packing tunnel for this they want to get animation down nice and small so what we're going to store is just score and I quote Ernie in ten ten ten - we're gonna store the three smallest components and they're in the range minus 1 over root 2 to 1 over root 2 because if they're any bigger they would be the biggest component so we need to do that and that's great can you just pack those in to 0 to 1 range and we have get more precision and we also store naknak W component of the index of the component we've dropped the 0 1 2 or 3 you might be thinking well can't you just get away with it with the storing x y&z it'll say 11 11 10 getting a bit more precision out and we constructing the W well turns out it doesn't work you kind of lose a lot of precision so you really need to actually drop this biggest component and reconstruct that one and that says that's not too bad um

you kind of find the biggest component actually only 5 instructions if you use some kind GNCC embassy Victrix so come on in like a major axis problem so you have all you actually have the cube map face ID instruction where you can just use that to get the major axis out of the x y&z and a max 3 to quickly help you get to whether the W is any bigger and I'll put all the code for packing unpacking for turning into the G buffer in the appendix to this release line so you can go check all that out later when I restore slides next week um what's the quality like well I kind of worked out it's about quality of equipment to about a nice packing of normal than eight eight eight um you do get a bit of faceting on smooth surfaces and we're con la keying on Farquhar we don't really have that many smooth surfaces but if your game is kind of like that you might want to look again um I mean overall I mean the problems you're gonna get you gonna promise again what I said the corner to the normals they could be a bit higher I do think there's a bit more research to be done here whether we can use those ten bits a bit better to distribute the kind of values better to get precision where it's needed um you also can only store also orthonormal tangent space now

that's not so bad because actually we don't have alter normal kind of space in there on all maps in Sakai but what we construct for that sort of arabic shading is also normal you can kind of make those two different tangent spaces and it's actually works out okay um might be a problem if you want to do something on lean mapping which does call my initial plan that we weren't so that's okay obviously we'd like to make the speed of packing unpacking a little bit quicker but you know it never really came up a notch in our profiling more so you optimized it um the main problem really we had was how to blend decals because obviously now once you're doing some ragged quaternion you can't just do some simple alpha blending so we only supported one layer of decals and that was an art kind of restriction and we had to kind of read back the current like normal mode quaternion buffer and then reconstruct things from that was cool though is actually how that worked well is with the virtual textures because all the decals on the terrain was stored in a war virtual texturing so it wasn't actually a poem for us at all and that is only on kind of buildings there's kind of nice thing about two different

techniques working together and if our game was a little textured it kind of mend this problem went away for us um for reflections well it's a little bit trickier we'd love to do important something or something into a key map to get some awesome results but we can't so I'm kind of really grateful actually to him it was Matt Pitino who brought Don revie's article and good GPA from GPU Pro one or two to my attention about to doing so much dropping into fir shading they just did a simple trip to distort the reflection vector in a different direction actually no it's not perfect but it looks pretty good so no on the Left we've got Nana's toffee in the middle was put on what we had originally where we had this anisotropic highlight and a nice to my perfection which is a bit kind of broken and then on the right we have what happens when you actually there's no trick for the anisotropic reflection again it's kind of not perfect but you know it's kind of a hell a lot better than doing nothing at all um I really sure like showing pictures these knives I mean our weapons team did an awesome job and you know what they used a nice trophy everywhere they loved

it and it's kind of reenter esting because it's all the problem we had on Far Cry 3 where these guys his weapons team they keep coming up to us and say we don't want some fake specula on our weapons and weak I think no no no we want to be physically based no fake secular at all and then we gave them this ammo stroppy and that complaint just completely went away they always felt they saw some glint or some kind of like something interesting going on in the weapons so it's going to be perhaps this they expected they wanted after all was actually just brush metal that's actually what they were looking at and now we've kind of replicated the real world a little bit better we've supported a wider range of materials they can get what they want so next topic is there going to be our lighting so there's got a three improvements he wanted to make and the first was class sky occlusion system you want to kind of increase the resolution of that and the system mode Jaime wall worked on I also had an environment where we wanted to make sure that they could dealt better with time of day changes and there's something dead up by a Gabriella font and finally we looked at the indirect lighting we want to kind of extend our range beyond this call loading we have

so further out about far in the distance and want to fast update as well from Far Cry 3 to vent flickering errors so for Scott Lestat a sky lighting model well we used the Burnet on a sky model and the pretense our model this is something that she developed by assassin's creed unity that we managed to take and it'll work really well for us where you kind of really liked it and once you've got the sky on all we just need to generate the lighting inferred allosteric harmonics so that's pretty cool so one on your Far Cry 3 was our annex toward our sky lighting in light probes along with our indirect lighting from which their indirect lighting was kinda based on probes faced it 88 meters which is quite low resolution and we thought hey no gir kind of a difficult problem to solve so hey why don't we just increase the sky resolution or something we could definitely do see the Sun is separate out our kind of sky lighting direct sky lighting from me indirect lighting and we did something the simple trick think a lot of games have done the speed and the CryEngine has it in we have this

just top-down a sky occlusion map are generated from a height field over the kind of entire scene I'm actually actually generates a visibility and second orders forever monix from this height field so high field it's kind of rendered on-demand we do something with streaming so our world is split into 64 by 64 meter sectors and whenever a sector becomes visible we're just going to render the height field of that it's really really quick means we have no streaming cost and means we get quite high resolution as well of 25 centimeters protects all and we can't blur this then from before we generate the SH of visibility I mean the visibility is quite simple to do we has just taken s SEO kind of style approach we just need to sample our come a blurred height field whether I take to quest on disc I stall a proportion of the visible samples actually to some of the directions of the visible samples then to get the directionality out once we've got that we've got the direction and we've got the visibility we need to convert that into spherical harmonics so what I could do is just create a spherical cap based on looking on the up

to normal and let's go rotate into the direction that we need so how do we apply this well what we're gonna do in the deferred lighting we're going to sample this kind of SH sky losing texture do some tricks obviously the con 2d and we actually need to kind of modify things based upon our height above the terrain so remember crawl we have actually blurred height field sort of something at the terrain Heights we can have this full I use full SH occlusion and then actually blend up to something in the height of the blur type field well we're just going to kind of fade out to something fully visible gives us something looks kind of 3d then once we have the primary visibility direction from SH we can just instructor bent normal out of that how much neater then to sample our second order as visibility and the third order Shi lighting in this pen normal direction and just multiply them together and we didn't look at using Sh product here but we kind of found it was a bit low res so I think in the end

so just hear some code quick code to bending the normal so she need to get the primary Direction out scare that based upon the kind of the zeros band of the SH and then add that into our existing normal and we normalize so let's look at some pictures so this is kind of a scene without sky occlusion I mean even so I think it shows some nice stuff with having like kind of cool sky lighting right you've got look at the buildings you can see different kind of shades on different faces into Hammond's of you and the sky version in everything start looking a lot better it's still perfect I mean quite notice we actually have quite a harsh fall-off and that was actually art directed we had a few tweaks in there to get looking for our direction wanted and what's really cool is when we added the bent normal thin as well and what you might see if you look to see around the building just doing the building in the ground but you can see but very disparate lighting yeah the bent normal thing it really softens that out so something like that doesn't look quite as good on the big

screen as doesn't on the monitor so look at the slides later and you'll see but I'll do that again so just kind of gives a little bit of softness and little bit unifies everything little bit more in the scene I'm not a direction really light back is they felt for things sometimes where the buildings doing the ground just they felt this things didn't look quite right so environment maps so let's ask ourselves questions so how can we get a single key map to have the right intensity at every single time of day and well what idea of course is only just rely to our key map every frame so what we need to do to do that well relighting is some writing some we're kind of good at doing we can install G buffers in I deferred lighting so what do we do something like that for our cue maps where what we'll do Rob we're going to bake off line key maps install and serve a fully lit kind of thing we're gonna store a G buffer we go store our B dough and we restore normal then what we need to do in our kind of frame we're gonna insert those along with their kind of a real-time generated sky texture to fill in those black out be no pictures

pixels and we observe that into a lighting pass light based upon our Sun on our sky lighting then we need to take that and do our pre-filtering so we can have the brdf to your GX POF they'll built in there and then we're getting our final a really icky map as a result so how do you the lighting well we do our Sun and sky lighting without any shadowing or occlusion because this sort of required jet buffer we decided to kind of save the memory I think in the future we probably ought to add that Dettol back into I mean get wall position out and I think would help definitely help improve things um we also get the indirect lighting from the nearest probe and we normalize everything again with illuminance for the ambient term and a far cry 3 we were normalizing on key maps by the kind of the color of the ambient it turned out actually that introduced loads are kind of banding and color distortion and just real normalizing on the luminance works a lot better so here's some results what's really cool you know seated we

have a real-time sky there's no clouds in there but you pick up some fancy clouds moving you can see the Sun moving across the time moving across at this sky you can see that kind of in the mountains there in the distance you can see how things the lighting shifting on those and of course obviously you can see how the ground changes lighting at different times um also you know the Philippines talk about the filtering because it's really important we do this so what we need to do for successive MIT levels you to sample our GD X distribution and do some important sampling just on its own it's a little bit kind of slow it's really really key to use a filter in a porter sampling that was done in GPU gems too I turned out that I think several Gardiner shells DeRussy area dice had the same idea and they kind of talked about this before me at SIGGRAPH last year we came up with the same things the right thing to do if n was really important to match your key map faces together in your computer better occupancy obviously they're doing a four by four came out facing you're sending that individually to the GPU you can have a lot of wasted performance um and also if your hardware supports it is really calling backup levels together as

well and you're going to really fill the GPU up with work so now that all this all was actually pretty quick they're filtering results there you go it works so yeah really really quick so for we had want to write by one to eight key maps with HDR key maps lighting again really quick generation quick filtering little bit slower but still really really reasonable we're bandwidth bound here daily from Microsoft it's kind of suggested that we look at using kind of our eleven DLF and B ten formats it might help speed things up that's true everything deal with that for the precision issues when I hope these times we go down a lot further as well so indirect lighting our third kind of lighting improvement we wanted to make 4 farcry 4 um so I start off with a really cool system if here at GDC 3 years ago you might have seen nikolai stefan often McCullar McCullar sheila bair talked about deferred radius transfer volumes

and they develop for Far Cry 3 um so these will like probes placed at various points in the scene that stalled radiance transfer information we came to looking at what wanted to do for this system we're the number of goals um gave me this limitation of being this kind of cross generational game we decided right we won't actually want to stick with the same light probe set for kind of last gen and current gen we really want to extend the range of the indirect lighting outs in Far Cry 3 was kind of stuck it around on our 64 meters around the camera and we really wanted to go further you can really see that kind of destroying I really want to do you have faster updates as well I'm moving all the CPU work on to the GPU so oh here's an overview of what we do so offline we're gonna bake some probes we're gonna bake this they got stall radius transfer information the second order Sh so that's going to be 3 4 by 4 matrices on the CPU we're going to stream this probate in we're going to upload it to the GPU and we'll update our kind of page table on the GPU we seem to calculate all readiness transfer and I'm

gonna inject our probes into the clip map and something the deferred lighting so I'm gonna step through some of these in a little more detail so our world as I said before splitting 64 by 64 meter sectors and probes are stored in cells and in each cell we have no creative a PI probes so we're going to bake a highest level data so just the first top level the probes and where we have warm sector per cell and then we're going to join it's from Mac so you have a MIT data five levels emit data so what's in a Cell well it's a radius transfer of propolis so at 64 probes again this 8 by 8 regular grid we also need to store our sh probe list as well so this is going to be the final probes often wondered well through the radians transfer they've come for this week on a real lighting process we just saw that as well a stone was per frame it also need to stall the dimensions of the cell so what maple oval are we you know what size are dips is it where are you in the world um obviously once we've got all these cells

the thing is we can't have these kind of buffers stored individually in GPU memory we can't have I could buffer put just one dr. buffer for radius travel approach for cell I mean it's not only really slow to be working on small sets of data all you do is come from virtualization to have one big list so as we're going to do we're going to allocate this big fit size list of cells and we're gonna break that into a fixed number of cells allocated for each MIT so in our case we have 80 cells in total five mid-levels and a 16 cells permit as we're saying a lot using the virtual buffer stop here we're not using any PRT hardware so don't worry about that it's all done in software so as a seller streamed in on the CPU well no ever free sellers available for that may play ball on the GPU well we're just gonna allocate well if no sellers available then we just need to discard a more distant cell then as I said have streamed out on the CPU while wishing to mark the sellers are free on the GPU it was pretty simple and what's called

clear had this big long bit and virtual buffer list we now come a form I'm like the radius transfer all at once really fill up the GPU it's really really quick so just insert our sunlight from our skylight take a list and output some SH probes at the end and these are the final probes going to use for lighting so let's talk a bit about our pace table so each MIT level actually has a page table and these are kind of spatially indexed in a 2d grid on the XY plane and this kind of page table covers the entire world what we're gonna do on the CPU we have all these page tables myth and permit and we're going to combine them into one kind of filling in all the gaps so in the end we have this kind of page table which contains all the data data for every cell in the world and so the final page table is 160 by 160 sectors at the highest MIT problem resolution and what are you to do if you want to sample a life rave well wait do we see the sample into the

page table to find our index into the cell list so what cell we currently need as a lightning data currently store that we don't care what mid blob of it is from the cell also Novosel rin while we can calculate from our wall position what probe index in that cell we need to look at now we just need to fetch the probe now that's great in all but obviously that's not gonna have any filtering taking not between mid levels and not watching individual probes so it's not too great and actually quite slow to do that lookup for every pixel on the screen so we don't wanna do that it's all we're going to do instead we're gonna inject our probes into a 32 by 32 by a 5 map so and that's like a 32 by 32 texture array 5 array levels and obviously the same size because they each the 32 by 32 is covering a larger area in the world these are centered about the camera and then the three of these textures having installed us 12 Sh coefficients we need

so to sample in the end of the day we need to find what cascade our pixel is in you should calculate the UVs for that pixel for the Cascades the story cascade and also the next level cascade down and we can sample both of them and just blend together based upon the distance so let's look at kind of what performance we kind of have so memory is pretty low obviously there's a page table is 25 kilobytes total cell memory is about 640 K and you probably thinking well you're doing an exchange aim we've got all these massive load gigs of data to use why is your memory so low well again as I said we're showing data sets between kind of current gen and like last gen so actually what would happen here we actually managed to optimize our last gen memory a lot actually on lighting system and we saved two megabytes or so in total across the last gen consoles so obviously people were very very happy about that performance again radius transfers were really really quick and so is our injection into the clip map and the full-screen pass here the full-screen past is actually what does our sky lighting so

all the sky illusion stuff I showed you earlier flying the third-order attic sky lighting and something to click map in getting the indirect lighting out we store that in a separate buffer now it's ready in the final the third lighting pass so what was great about this we solved all our problems in Far Cry 3 we have fast updates because now we are no kind of CPU GPU synchronization it was all going to be done on the GPU all the radius transfer rather than before we had a much large indirect lighting range because now he had all these nipped levels of the lighting data where you'd go all the way into the distance and we had a lot lower memory requirements which as I just said was awesome for our last gen consoles um obviously it has property quality issues so we have lower frequency lighting my eight-pack meets every 8 by 8 meters a kind of at best and also temporally as well because kind of second order and SH spherical harmonics and rate so radius transfer it doesn't really pick up enough data and also finally we don't have our local lights kind of taken into account so

these are things obviously we want to look at and kind of improve for the future so vegetation now legislation will actually one area that we decided to completely redo for next-gen consoles there's a really important part of our a far cry game so well give credit to Phillip Daniel and is also asking gay from working on this system and again it was a completely different it's data set for iconic current gen on last gen build so really we did two complete different sets of trees our goals are thrown up in this new system were at well we want to improve our little fidelity up close some we get really close to trees I'm wanted to get improved loading and imposters as those kind of trees then go further away I'm also get some cool simulation a physics simulation done so then while talking about where all the vegetation can mean quite a few things or trees bushes grass really our new system only touched Connie's trees and

these bushes we didn't really look at the grass at all so what we're going to do we're going to create our trees using speed tree and we're going to put them into the engine we're not going to use anything at the speed rerun time we're going to take the mesh data generated and pull that in a format that we like um I will actually how he built a tree because of this because this is a tree system a trunk and these leaf clusters and also have a lot of leaf clusters and put them on a trunk we got a tree um and these are all modern individually and we have three logs for the trunks and three logs of these clusters before a tree becomes an impostor what's really cool is how we do the leaf cluster loading Cory acting aduba generate there when we want to switch loads off line and I can calculate the percentage of each leaf cluster visible from 12 distances and nine viewpoints around the camera around the tree so let's take a look at this so let's look at in the left hand image this is what how many's looking at the tree the red is going to be your kind of lot zero clusters the green is going to be your

Lord one and your blue is can be a lot too and then if you kind of move the camera around to fix the kind of cunning camera and you're going to look at sneak a look around the side on the right-hand image what you're able to see is that obviously the cost is closest to the camera or the highest Lord and I say get further away more more occluded by the clusters in fronts they're gonna be using a lower logs and obviously of course as the tree moves into the distance everything's going to turn blue here and you're gonna have the lowest laden in total before everything goes to an impostor a simulation so we have all our physics and movement to the tree simulated on GPU and what we do there we have instance bones transformed into a wall space bones so what's going to happen is we have a model bone list of all the models in the scene and obviously a list of all our instances three instances in the scene and they're going to go into a computer simulation and we going to spit out a much longer list then of all the wall space bones every tree which goes into our vertex shader and this is great because it allows some really cool physics effects so this is the first point where I try

to play a video I can see where you're flying a buzzer and it comes down I can see all the trees moving individually these bushes interacting to the buzzer and even see all the grass waving as well for the grass your plashy applies to achieve water like a simulation to get that going as well and this is really cool not many games something carefully it's kind of kind of effects so that's great for close to the camera what happened to me it far away um well what could some cool things generating have good imposters we're actually gonna do we're gonna take a screenshot some nine view points around the camera eight perpendicular to the tree and one top down so I'm having a posture we can easily blend depending a lot of you angle what the tree should look like but we're gonna go further than that there's actually what we're going to do good deal which could take G buffer kind of screenshots we want to capture the albedo the normals and the material

properties and that gives this a lot of Texas for a tree and again we're not going to stop there and even more we're going to add this kind of depth billboard effect so we have one Capitol low whereas depth better as well during on to Bothell paths and when we went our billboards are actually tessellated into a sixteen by sixteen grid and we're gonna displace our vertices depend according to the original tree depth what are the screenshots also great but this is a front view on the left and the side view is this of the same tree as you see all the depth applied I've been told this actually really helps a lot when a tree is lit from the side and also in two tree Impostors intersect so like it's much better blending between them so it's worth doing tech good listening of this technique and I was given a few benefits later as well as I'll explain I mean occlusion standard thing everyone did a OH volumes you will seen this before but it only has a benefit to our tree lighting and let's

give some scary looking for the text numbers so for our Rosewood tree on the right and one hundred seventy thousand vertices for all of the leaves it's worth noticing here because we have our loading system we're never going to render that many it's only going to be a tower eighty thousand max for that the causal loading that we do which allows was per half is going to high detail of course it's also probably fair to say that I'm not gonna let artists make a tree line but again I'm performance we're all mostly pretty good as well our tree simulation again on compute is like really really quick and given for this scene there's a lot of trees on the scene now shadows about millisecond G buffer about five and a half rendering of all this kind of alpha test geometry it's pretty good so obviously we like to optimize because trees are always expensive for everybody everywhere um visually however I think

meditation is a really tricky thing to get right because you know there's loads of things we're not simulating and you know light bounces between blazars grass or the light scattering between leaves what's really great is when you have some really cool technical artists you can kind of come in at the end of the project and just kind of do a few little cool tweaks to get things looking cool I'm gonna share a few of those with you suddenly wanted is a kind of a per cluster for a leaf cluster random color tint so this example the first colors green the second colors are red and we just generate tints varying between those two randomly really simple really effective and this is a bit of extreme example but you know it's not two different two trees you might see in Vermont in the autumn um loitering in to give them postures to like bring horses to life because you handle this cool simulation upfront but then we didn't imposters they looked a bit flat and boring so what we're going to do well we just seemed like some branch motion first by doing some per vertex sine waves this is break because we have those tessellated billboards but actually the vertices are going to move when the tree is going to shift a bit and we just see

missed them by smooth triangle waves as they did of a Crytek many years ago and we're also going to add some per pixel noise as well from a texture to simulate leaf motion so you have the branch motion from these big kind of distortions and then little per pixel noise to fake this leaf motion worked really well again I'm gonna play this video I'm not hopefully about to see the difference between the noise on the noise off if not again take a look offline I was working pretty well then you can see that things look a lot a lot more dynamic when the noise is on so I got a final topic improve today ISM attic aliasing now I hope you all saw the hold robots talk at SIGGRAPH last year where he went into a big overview of all the anti-aliasing techniques we tried so really all the credit to this work a

toast to my house and also all the difficult questions at the end please yeah so um one matching a present here is a stripped down of that talk and I'm going to talk about the things what we actually finally shipped with so really this highways reconstruction anti-aliasing is looking at about kind of three different techniques they were going to apply together to get the real anti-aliasing result so the first is edge anti-aliasing so this is pretty self-explanatory your order well this is what MSA is tries to deal with when you have these edges in your image you know see to look at temporal anti-aliasing as well because images move in time so you've got this kind of third dimension we just think about a finally we have temporal or super sampling so obviously super sampling is when we take a big image and we go down sample it to something smaller and we wanted lighten that temporarily where we take a different sample of alternative frames and you have to combine this bigger image that would down sampling so the payment overview of kind of how things going to work on don't go through this in a bit more detail so you start with a

current frame it's going to go into our educator sink that feeds in along with our previous frame into the temporal super sampling a temporal aunty editing shader which one outputs our final frame and that then goes back fed in again or frame n minus one unlike human history buffer it's fed back in again do all that final temporal super sampling and temporal anti-aliasing for the next frame so edge idea this thing well we tried a lot of techniques so we looked at analytic edge anti-aliasing for alpha tested geometry we also really wanted what we really wanted to do with the coverage reconstruction anti-aliasing that Howell presented a cig graph and this gave us by far the best performance but we had a lot of content issues the small triangle also much we couldn't do this but I'd arise checking it out for your game so we actually ended up with with SMA this is just temporary stabilized and we had normal depth and luma predicated thresholding so for the temple stupid sampling so this is based on work done for a Killzone Shadowfall so we're going to do

this is two types of sampling so I'm going to use the current frame of the previous frame for our two samples the previous frame sample is valid only valid if the motion flow between frames and minus one is coherent and also the coral color flow between frames and n minus 2 is Garant so those we need to look at those two frames so they have the same sub pixel jitter that's going to prevent flickering except that we didn't actually do the frame n minus 2 stuff for far cry 4 and you can actually do can see some flickering present in the game and thanks for those who are already pointed out to me this GDC so again what we're going to do so I just go over these metrics again so if we're on this geometric metric which is looking at a motion buffer to make sure that we kind of stick with the same geometry with chocolate between the same geometric objects so two plating for our n minus 1 to our n sample if the motion ends up being incoherent and we're still gonna do some kind of color metric on art and honor the N minus 1 sample so we're gonna look at the color bounding box around three my 3 by 3 neighborhood around our sample if it's close to the mean we take no new information it was close

to the min and Max we think we have new information we want to accumulate that in its outside this window and we think it's probably a fluctuation we should get rid of it and just to visualize that this gives a good graph of how we're actually doing things hopefully that kind of makes a bit more sense in the visual so we accept things around the min and the max of these Peaks actually worth looking at we didn't actually use the mean in the end we actually use the center of the pixels so also it's worth checking out Brian kerasys talk at SIGGRAPH last year as well - and when loads of things on this acceptance metric so I definitely want to worth checking that out so we're gonna need to look at what samples we're gonna use well sample pattern we're going to use to maximize our two samples to get the best image out so there's various sample patterns we can try let's take a look at what they are well obviously just for a one-time centroid well we're gonna get some early 80s seeing any secret of the edge there i

standed warms we kind of two times MSA there's two times rotated grid well you know just be better you have to actually have some form anti-aliasing there it's not great we could do two times quincunx now Quinn doctors interesting because it optimizes patterns by kind of sharing from a corner samples between various pixels see I've got three unique rows three unique columns but of course this doesn't reduce a 0.5 a radius blur all we can look at doing something like 4 times rotated grid and again that'll be four times samples there we can't really afford that but it does get a better looking result but there is a better pattern and that's actually this thing called a fit quad the kind of combines these ideas from the four times native grid and also you're kind of quincunx pattern so what happens you have four unique grows your four unique columns and it does also like quincunx have this kind of 0.5 pixel blur but you can't recover backed in a post processing with unsharp mask so it's actually gonna be okay actually you just use comparison

and actually or some images well it looks kind of is if like the fit quad is looking pretty as good as the four times to take your grid and if we look at some kind of studies it turns out actually fit quad is probably the best of them all I don't know all the things you looked at so let's go with that so how is that going to work all we need to do when you split the pattern in half so on frame a look a look at sample say 0 & 1 2 & 3 & on frame B you look at the red 0 1 2 & 3 & 4 say 4 pixels 0 then all we need to do is average of 0 1 and blue and your 0 2 in red I asked you to give this our kind of super sample image um there are obviously some tricks with this it doesn't work out quite as easy as you hope because we need to do actually interpret your UV the sample positions you just use them some hlsl interpolation modifiers of course then you have the second problem where you

have incorrect derivative calculations because your spatial differences between the distances between the rasterization samples differ and that causes some quite bad problems because let's look at term say frame a well this is render with the correct MIT and then what happens when we render with frame B the D DX D dy changes and we get this over sharpened myth and obviously you're going to blend between those two well we can have a few issues so what do we do well assets need to we could do if you do a few things we could do submit flop bias between frames seems a bit tricky we could obviously trying to do things manually ourselves but that's gonna be expensive in a bit of a pain actually the thing we can just do is I just reorder our samples to make sure that these derivatives match and it turns out it's pretty straightforward to do and it works pretty well so this is same age before frame a still looks great don't change anything there with a frame B now everything looks pretty much the same as a frame a and that's great um now so the temporal anti-aliasing so

rashly temporarily stabilized like a load of our buffers takes up loads of memory that's cool we've got all the next-gen right but actually really gives a really good image quality so we look at the final frame buffer the SMA there are some you say spent cones motion vectors bloom they're all temporarily stabilized so we have this history exponential buffer and always do is here's our multi sudden visual changes so make sure we only accumulate new important data into it we actually teach the same acceptance metric so I just talked about for the temporal super sampling so the geometric metrics to compare whether the motion vectors are the same I call a metric ting pair of the colors were in the same bounding box so now we're doing this not only for a 10-1 frame but we're good separately to our full history buffer um you might need to eat those values separately though like is some good different results um however after all this we're gonna use a blue lose a bit of detail with some wider and kind of more complicated kind of down sampling kernel

and also we have this kind of par pixel blur from flipped quad so this is pretty pretty okay actually we need to approximate a four taps in kernel by doing a half pixel blur hey we've already done that with a foot board and we asked you to do a half pixel radius a shot masking as well and then to match our super sample resolution because now our resolution is a lot higher than what we initially was we just need to emit bias and all textures and overall we have then some really high quality anti-aliasing and actually the performance is really really good SMA is a millisecond and look at the kind of temporal flip part plus the temporal ITA this thing takes about 0.65 I combine that with a shot off mask as well we've got two milliseconds then pretty much higher and for all anti-aliasing we did the manage to do the coverage reconstruction anti-aliasing we probably get a little bit back off that again something look out for the future so what's the

conclusions gonna come to you well first of all I'll show you a bit of a typical form of a typical Far Cry scene it's awesome doing the cross chain game sometimes because this is when in frame thank you to all the artists who make good data so as I said they actually they sing lead to milliseconds about three milliseconds to post-processing G buffer is obviously most expensive path followed up by our lighting and there's a few miscellaneous things in there as well versus our tree simulation and a various second buffer copying so the far cry 4 we match to make some significant rendering improvements to our engine while whilst maintaining the last gen build we enjoyed some and I stropping metallic materials might have some higher fidelity sky occlusion is an indirect lighting we had a whole new and vegetation system and finally we built a new MTA as aliasing algorithm that give us some high quality image quality at the end really thanks to really everyone the far cry 4 team especially kind of these guys who worked on things for the

Ike mention in the presentation especially thanks Stephen Hill for all the conversations we've had about to ACU and far cry 4 and Julian Mercer on as well for being the GDC advisor I have a lot of references please check them out they're all actually pretty good they're worth your time and I'm just anyone have any questions [Applause] no one okay don't talk to me afterwards if you need thanks again for coming everyone [Applause] you