Re: [5/8] WineD3D: Add a more formal framework for driver quirks
I accidentally deleted the mail in my inbox, so that reply is without a proper in-reply-to tag. Here's the archive for reference: http://www.winehq.org/pipermail/wine-devel/2009-April/074718.html
- I don't like how much information specific to constant fixups the table contains. There is a lot of intended redundancy here. The minimum we need is the flag to use the ARB or GLSL constant value. I want some additional safeguards to make the fixup noisy if one of the assumptions we make about the driver isn't true any more.
The patches here only handle constants, but I want to do the same for varyings. I don't like the idea of using a callback function to apply the fixups because it would repeat quite a bit of value checking and FIXME writing code for different cards, just with different constants in the code.
- The "apple" field is weird. If you're going to have a callback function you might as well match that information there, if appropriate. On the other hand, I don't think you really need the callback function, you only match on vendor and card anyway. All non-apple platforms have the same graphics driver, just the Apple drivers have their own weirdness that does not occur anywhere else. Its mostly many unsupported features that are advertised and software emulated which we have to filter out. All other platforms share the same driver anyway, even Windows to a certain extend.
I can replace this with a driver vendor enum or a driver vendor matching function. However, I am prepared to adjust the setup in the future if $NEWWEIRDRIVERVENDOR pops up(e.g. Gallium3D) that has different requirements. I want to keep the driver vendor match separated from the card match to allow defining fixups that apply to all cards run by a driver from a specific vendor. A later patch adds a fixup that uses the ARB constant limits on all cards if we're running on a apple driver. There are other ways to do this, like returning a matching score and using the best match, or merging two fixups into one. I think that my approach matches what we want to do best and is simpler than trying to merge fixup descs.
- What is the point of only allowing one fixup per card, and then adding QUIRK_NO_SANITY_CHECK to allow more fixups? You might as well just allow multiple fixups and apply them in the order they're defined. See above - generic fixups for a driver vendor that applies across cards. Obviously a Geforce 8 has different uniform or varying limits than a Intel GMA 950, but in both cases the GLSL uniform info is no good and we should use the ARB one instead.
- Why do you store the quirk table in gl_info? It's never used from there. (I realise the next couple of patches do use it from there). ... and just add a DWORD "quirk_flags" to gl_info. My patches add two fields that are accessed from other places in the code. There may or may not be more in the future. I prefer to have one pointer in the gl_info structure than multiple DWORDs containing quirk info, but I am flexible with this.
2009/4/12 Stefan Dösinger <stefan(a)codeweavers.com>:
I accidentally deleted the mail in my inbox, so that reply is without a proper in-reply-to tag. Here's the archive for reference: http://www.winehq.org/pipermail/wine-devel/2009-April/074718.html
- I don't like how much information specific to constant fixups the table contains. There is a lot of intended redundancy here. The minimum we need is the flag to use the ARB or GLSL constant value. I want some additional safeguards to make the fixup noisy if one of the assumptions we make about the driver isn't true any more.
The patches here only handle constants, but I want to do the same for varyings.
I don't like the idea of using a callback function to apply the fixups because it would repeat quite a bit of value checking and FIXME writing code for different cards, just with different constants in the code.
Well, the main fixup I see you doing is replacing the GLSL limit with the ARB one. "GLSL" and "HARDCODED" are never used, afaik. I don't see how you would need more than a single function for that. Even for "HARDCODED" you could just create a single function with a lookup table, although I don't see "HARDCODED" getting much use in the first place.
I want to keep the driver vendor match separated from the card match to allow defining fixups that apply to all cards run by a driver from a specific vendor. A later patch adds a fixup that uses the ARB constant limits on all cards if we're running on a apple driver. There are other ways to do this, like returning a matching score and using the best match, or merging two fixups into one. I think that my approach matches what we want to do best and is simpler than trying to merge fixup descs.
Well, it's not particularly hard to define a wildcard for the vendor and card fields to match eg. everything on a certain platform / driver vendor or everything from a certain hardware vendor.
- Why do you store the quirk table in gl_info? It's never used from there. (I realise the next couple of patches do use it from there). ... and just add a DWORD "quirk_flags" to gl_info. My patches add two fields that are accessed from other places in the code. There may or may not be more in the future. I prefer to have one pointer in the gl_info structure than multiple DWORDs containing quirk info, but I am flexible with this.
You can store the information in a structure if you like, but that's not the point. The point is that you don't actually use it in this patch, and that struct driver_quirk is the wrong structure to store. You want to know what fixups are active, not which entry in the quirks table caused them. Not storing the entry from the quirks table also avoids the problem you mention with "merging" fixups. You just apply the fixups in order, and add the appropriate flags.
participants (2)
-
Henri Verbeet -
Stefan Dösinger