[SYCL][Graph] Query for backend command-buffer support

[EwanC]

Rather than the user deciding whether to use the emulation mode or backend command-buffer mode, it is less error-prone to do it programmatically as a user can't select an unsupported config.

This also fixes #90 where currently the SYCL_EXT_ONEAPI_GRAPH macro isn't defined when in emulation mode.

Done by using a PI device info query in finalization. This also corresponds to a new UR device info query until the extension mechanism is decided oneapi-src/unified-runtime#458 but will be superceeded by whatever extension reporting mechanism is decided on.

The only PI backend which reports support for command-buffer implementation is Level Zero, the other backends/adapters report no support.

[julianmi]

Looks good to me in general. One drawback I see is that this would prevent the explicit use of the emulation mode on L0, right? This might be useful for testing or comparing both approaches.

[Bensuo]

LGTM

[EwanC]

Looks good to me in general. One drawback I see is that this would prevent the explicit use of the emulation mode on L0, right? This might be useful for testing or comparing both approaches.

Yup, I think an L0 emulation mode still does have some value. I guess we'd want to agree on whether the L0 emulation mode should be exposed to users, or it's just for us developing the code. If emulation mode on L0 is just for our own comparison & testing, then I can add an in source #ifdef to force emulation mode that can be changed, and then we recompile. Alternatively, if for an end user forcing of emulation is useful, then I can repurpose the existing CMake/configure option from "enable graphs" to "emulate graph", or something that will force the mode to be emulated.

[reble]

LGTM

[EwanC]

Looks good to me in general. One drawback I see is that this would prevent the explicit use of the emulation mode on L0, right? This might be useful for testing or comparing both approaches.

Yup, I think an L0 emulation mode still does have some value. I guess we'd want to agree on whether the L0 emulation mode should be exposed to users, or it's just for us developing the code. If emulation mode on L0 is just for our own comparison & testing, then I can add an in source #ifdef to force emulation mode that can be changed, and then we recompile. Alternatively, if for an end user forcing of emulation is useful, then I can repurpose the existing CMake/configure option from "enable graphs" to "emulate graph", or something that will force the mode to be emulated.

I went with the second approach and added a FORCE_EMULATION_MODE define to the top of graph_impl.cpp. If a developer flips this from 0 to 1 and recompiles, it will force emulation node for targets with native support, i.e L0 right now.