Allow BGFX manual selection of GPU to use

[angelosa]

BGFX backend currently autoselects the first GPU it finds by using BGFX_PCI_ID_NONE in vendorId.
User should really be able to select whatever adapter it wants instead, apparently the best way is to set up deviceId if provided, fallback to current auto behaviour otherwise.

BGFX documentation about vendorId/deviceId
MAME code where BGFX init happens

[MooglyGuy]

I've taken a look at the BGFX API documentation, and I can't seem to find a call that will produce a list of available devices.

For Windows at least, the user can run dxdiag, save out DxDiag.txt, and then look up the appropriate vendorId and deviceId from it, but that's a pretty awful user experience.

This seems like the sort of thing that should be solved upstream by having some sort of API endpoint that can provide a list of the relevant devices with which BGFX can be brought up. If there already is such an API endpoint, I'm interested in knowing what it is.

Overall, I'd like to hear @bkaradzic's opinion on this. It's a pretty solid suggestion and one I'd be happy to add to MAME, but not if it requires the user to have to go screwing around with external tools. From an ideal UX standpoint, the user should be able to ask MAME for a list of viable video output devices, and subsequently specify "device 0" or "device 1" to MAME, letting MAME handle arbitrating with BGFX to decide which device is which.

[bkaradzic]

@MooglyGuy After initialization inside Caps you'll have numGPUs and gpu[4]: https://bkaradzic.github.io/bgfx/bgfx.html#_CPPv4N4bgfx4Caps7numGPUsE

If you want to convert GPU vendorId/deviceId into user readable text you should use this: https://pci-ids.ucw.cz/ on your end.

[MooglyGuy]

@bkaradzic Sorry to harass you with out-of-band questions, but do I understand the process correctly?

• Call bgfx::getCaps to be returned a const Caps* struct
• Look at the Caps::gpu array to determine which GPUs were enumerated
• Ignore the specific vendorId and deviceId returned within the Caps struct itself
• Continue bringing up BGFX via bgfx::init with the appropriate members of the Init struct set to the vendorId and deviceId corresponding to the user-selected GPU index (or, presumably, bail with an error if the user has selected a GPU index which does not exist)

Is that about right?

[bkaradzic]

You could do something like:

if (bgfx::init())
{
    const bgfx::Caps* caps = bgfx::getCaps();
    // ... copy info somewhere else or make decissions based on caps ...
    bgfx::shutdown();
}
else
{
    // ... this code path could be used to inform user that they don't have minimum required GPU to run bgfx.
}

Or you can make that part with UI, like settings selector where GPU doesn't matter, and then when user clicks OK, you would shutdown bgfx, and initialize with selected GPU/API combo.

[MooglyGuy]

Alright, but is it possible to call getCaps without having first called Init?

At the moment, the user specified whatever backend he or she wants to run with from the command-line, and we pass that in when calling bgfx::Init.

If it's not possible to call getCaps without having called Init first, then would it be reasonable to call Init, then call getCaps on success, then shut BGFX back down and call Init again with the user's desired deviceId and vendorId if it isn't what BGFX has already selected?

MAME's options are supplied from the command line (which is a garbage UX that I intend to eventually solve), but in the meantime, that's what we've got to work with.

[bkaradzic]

It's not possible to obtain caps without calling bgfx::init.

You could also call bgfx::init with renderer and device user specified since that will be backend that will enumerate GPUs. Device will be selected only if it exists. If both matched exactly to what user specified, you could continue without shutdown.