| 25 Sep 2020 |
 kvark | If you have the budget left, will it be spent on those limits? And on which exactly? | 23:32:34 |
 kainino | if you have the budget left for all of them, it will be spent on all of them, and if not, then you're out of budget and device creation fails | 23:34:00 |
| kainino | or the UA can try to be smarter and figure out exactly how much budget you spent learning that device creation failed | 23:34:53 |
| kvark | That sounds tricky, non-obvious from the user perspective. | 23:39:41 |
| kainino | I think privacy budgeting is always tricky | 23:40:12 |
| kainino | the important thing is that it may be non-obvious from the user perspective, but from the application code's perspective nothing particularly weird has happened | 23:40:34 |
| kainino | it's just as if the adapter didn't support what you requested | 23:40:52 |
| kainino | so even if the dev doesn't think about it, the app will behave in a reasonable way | 23:41:16 |
 tojiro | I'm a little concerned about the behavior of frameworks with this pattern. For example: Three.js right now goes out and requests a bunch of extensions right away because they're used for various features in the library, even if you aren't using one of those features yet. (Because who knows! You might?) | 23:42:19 |
| tojiro | So they're going to be incentivised to spin through all possible extensions up front and enable as many of them as the library can possibly take advantage of. | 23:43:10 |
| kainino | in practice I think there are a lot of situations where you can enable quite a few extensions without spending much budget | 23:43:47 |
| tojiro | They could shift the responsibility to the devs, and say "You have to request what you want to use" but that's not always going to be an obvious GPU feature -> Library feature mapping | 23:43:52 |
| kainino | a lot of extensions will be available on almost every desktop device | 23:43:59 |
| tojiro | So the budget is a dynamic thing that depends on rarity? | 23:44:25 |
| kainino | in an ideal world, I assume so | 23:44:34 |
| tojiro | EG: The cost of checking for Anisotropic filtering (to pick a random GL extension) is ~0 because it's literally everywhere. | 23:45:07 |
| kainino | the budget is supposed to run out when k (as in k-anonymity) goes down to a certain value | 23:45:10 |
| tojiro | But the cost of checking for raytracing is presumably higher? | 23:45:21 |
| kainino | in practice I don't think any of these questions have been figured out yet | 23:45:35 |
| kainino | at least, conclusively | 23:45:53 |
| kvark | In reply to @kainino:matrix.org
so even if the dev doesn't think about it, the app will behave in a reasonable way No. Suppose the app uses 5 bind groups but doesn't check for it. Suddenly, it may work here but not there, depending on the privacy budget size, and what have been queried. | 23:53:48 |
| kainino | that's reasonable, because if they did that then it wouldn't work on system that can support 4 but it would on a system that can support 8 | 23:54:29 |
| kainino | they still have to requestDevice with 5 bind groups | 23:54:47 |
| kvark | True | 23:55:18 |
| kvark | Ok, I finally understand it:) | 23:56:07 |
| kainino | 😀 | 23:56:24 |
| 27 Sep 2020 |
|  michail | 17:48:56 |
| 28 Sep 2020 |
 cwfitzgerald | In reply to @kangz:matrix.org
That seems like a lot no? Creating descriptors should be cheap-ish. so I eliminated a bunch of issues and took a look in a proper sampling profiler, and now 256 textures shows 0.12ms, which is a lot better | 04:57:40 |
|  benkant joined the room. | 10:07:46 |
| 29 Sep 2020 |
|  kunigunde | 19:49:50 |
