Search this blog
04 October, 2011
03 October, 2011
Don't lie.
A task is not done when its code gets checked-in. In production, "done" is:
- Checked-In
- Stable (passes automated tests)
- In-budget (memory and performance)
- Usable (tools, parameters)
- Verified (art-directior or designer or lead programmer)
If you ignore the last four, you're living in a dream. A dream in which you will ship the game. Instead, you'll die under the pressure of crunch time and ship a flawed product that does not match the quality standards it should.
That also means that in production we start with a stable, in-budget product. And that we do have means of verifying that this is true for the entire production (tests).
Yes, it will take twice as much time to finish a task. Yes, it will mean that some tasks can't be declared done until you make more space somewhere else, to fit them in.
Of course you can avoid all this. All it takes, is to lie.
We are a creative industry. We have to deal with change, we don't plan things up ahead and then waterfall until they are all done (not _most_ of the times at least, there are situations in which that applies).
We don't craft a product by following a plan, we make drawings and sketches (prototypes) and then take a canvas and start painting.
You don't have a person detailing a finger, another one refining an eye, another working on the nose and then hope that everything will stick together just right. Or hope that you will have all the parts done by a given date! And what if it then misses one of the ears? What do you do, take ten artists working on that ear till midnight every day near the deadline?
Only an idiot would do that, and still many game companies work like that, they don't consider that everything has to fit just right, and that change is not local, every change has to fit with the entire context, every brush stroke has to make a sense in the entire painting.
You start with a painting, a rough one, and then refine it, and at all time the painting is a painting, it's not a collection of unrelated pieces. You can stop at any moment and it will be still a painting, maybe not that refined, maybe not as intricate and detailed as you wanted, but it's a painting.
A game is such only if it runs. If it crashes or goes out of memory on the target platform, it's not a game, it's some binary that crashes. If it does not hold its performance, it's not a game, we can't burn it to a disc and call it a game, it won't be shipped, it won't pass certification. Iteration should not break this quality, it should go from a "shippable" game to a "shippable" game. Especially during production.
We are a creative industry. We have to deal with change, we don't plan things up ahead and then waterfall until they are all done (not _most_ of the times at least, there are situations in which that applies).
We don't craft a product by following a plan, we make drawings and sketches (prototypes) and then take a canvas and start painting.
You don't have a person detailing a finger, another one refining an eye, another working on the nose and then hope that everything will stick together just right. Or hope that you will have all the parts done by a given date! And what if it then misses one of the ears? What do you do, take ten artists working on that ear till midnight every day near the deadline?
Only an idiot would do that, and still many game companies work like that, they don't consider that everything has to fit just right, and that change is not local, every change has to fit with the entire context, every brush stroke has to make a sense in the entire painting.
You start with a painting, a rough one, and then refine it, and at all time the painting is a painting, it's not a collection of unrelated pieces. You can stop at any moment and it will be still a painting, maybe not that refined, maybe not as intricate and detailed as you wanted, but it's a painting.
A game is such only if it runs. If it crashes or goes out of memory on the target platform, it's not a game, it's some binary that crashes. If it does not hold its performance, it's not a game, we can't burn it to a disc and call it a game, it won't be shipped, it won't pass certification. Iteration should not break this quality, it should go from a "shippable" game to a "shippable" game. Especially during production.
30 September, 2011
Rendering 101
This is how I explain (videogame) rendering to non-rendering (engineers... even if hopefully it's still understandable for some non-programmers as well).
If you have suggestions on how to make it more clear, post a comment!
28 September, 2011
Fight Night Champion Rendering @ GDC
I wrote so much for Fight Night Champion over the years, but not much can be found on the internet. I crafted at least five presentations for it, three mostly internal for the team on Fight Night 4 analysis and pre-production, one on the various prototypes and ideas done to exit pre-production, to share ideas with the other EA teams, and one for GDC after the game was done, which was refined and presented by my (awesome!) colleague Vicky Ferguson when I left EA (you should see the recording of her presentation, and she also presented for FN4 at the previous GDC which was very cool too).
Unfortunately the slides of this last one were not published (a couple could be found here, but they are pretty useless)... until now.
At the beginning I thought about writing more detalied blog articles about many of the experiments that were done, but many of them do not really need that much space, some others were rediscovered and published by others (i.e. pre-blurred SSS, bent normals SSAO), and I think FNC was more about the process than the final tech (in fact, the lengthy presentation never goes into the detail of what at the end were the final FNC shaders!), so...
This is a "director's cut" of the GDC slides, it's way more verbose (not meant for presentation) and with much more material.
Enjoy! [update: moved out of Scribd]
[update] Bonus links:
27 September, 2011
Bad ideas don't require much explanation: Caching Stable Cascaded Shadowmaps
Note:
Ok, some explanation (as I was asked recently about this). The idea is that for static CSM and static objects, you just shift every frame your fixed "window" into an "infinite" plane. So why don't we just do that in the actual CSM? We can shift the old data (or just implicitly shift by wrapping around and keeping an origin marker in the shadowmap coordinates) and just render the tiny strip that corresponds to the part of the window that is novel (corresponding to the intersection between the old clip volume and the new one).
Culling is easy too, it's just a boolean test between the two volumes. The only real issue is that the near-far planes change every frame in a stable CSM, thus the old shadow zbuffer data would be in a different coordinate space than the new one (not only a shift but also a change in the projection matrix).
This can be solved in a number of ways, the worse is to "reinterpret" the old values in the new projection as that loses precision. The best would probably be to write an index in the stencil part and use that index to address an array keeping the last 256 near-far values for the previous frames. 256 values will go pretty fast, so you might need to re-render the entire shadowmap when you're out of them... But really I think there could be stupid tricks to avoid that, like adding some "borders" to the CSM and shifting it every N-texels instead of every single one. Or computing the near-far for volumes that are snapped at shadowmap-sized positions, so every actual shadowmap rendered would at worse interesect four of them, thus requiring a maximum of four stencil indices...
Let me say that again. Our stable CSM shadow is just a fixed-size, shifting window inside a infinite projection plane aligned with our light direction, with a near-far clip that corresponds to the min-max interesection between the projection window and the entire scene. Now we can imagine that we precompute the near-far values (we won't but, let's imagine) in a grid sized as the shadow window. Then, no matter how we shift the window in this infinite plane, it will intersect only four (maxium) near-far precomputed grid cells... Thus, four values in the stencil are all that we need to index cached z-values.
Subscribe to:
Posts (Atom)