src/dawn/updating_gpu_info.md - dawn.git - Git at Google

 # Updating gpu_info.json

 Dawn uses `gpu_info.json` to group GPUs by vendor and architecture when calling `adapter.info`. It uses
 vendor and masked PCI device IDs to identify the GPUs, and as such will need to be updated from time to time as new
 devices come on the market.

 ## Guidelines for adding a new device or set of devices

 There are no automated processes to add devices to `gpu_info.json`. Everything is inserted by hand. Fortunately the
 file is intended to be human readable, and should be relatively easy to update following the comments at the top of the
 file and referencing the existing entries.

 When adding new devices use of a `mask` is encouraged when applicable to capture more devices with less entries.
 However, it's important to ensure that the masked device IDs are not overly broad or capture devices which do not belong
 to a given architecture. Many GPU vendors do not use very consistent patterns in their IDs, or have strange exceptions
 to the patterns that do exist.

 When in doubt, prefer adding an explicit device ID without a `mask` over adding an masked ID that might accidentally
 capture more devices than intended.

 A GPUs `architecture` should be the vendor's terminology for a given family of devices when possible. An `architecture`
 group should not be detailed enough to identify specific devices, but should ideally group all devices made within a
 span of a couple of years or longer. It is OK to have significant performance disparities between devices in a single
 `architecture` group, as long as the capabilities and behavior of the devices are roughly similar.

 ## Rolling new devices

 Capturing new devices that need to be added to `gpu_info.json` right now is unfortunately a fairly awkward process. The
 method used so far has been to checkout https://github.com/kainino0x/gpuinfo-vulkan-query, update the data submodule
 with the latest records from https://gpuinfo.org using the script at https://github.com/kainino0x/gpuinfo-vulkan-slurp
 (Ask kainino@ or bajones@ for access), and then run `device_id.py` with the path to the `gpu_info.json` file you are
 updating. That script will compare all the records to the entries in `gpu_info.json` and print out how each of them are
 categorized followed by a list of all the uncategorized device IDs and descriptions sorted by vendor. You'll have to
 manually look at the device descriptions and research their architectural design in order to categorize them.

 Good luck!
	# Updating gpu_info.json

	Dawn uses `gpu_info.json` to group GPUs by vendor and architecture when calling `adapter.info`. It uses
	vendor and masked PCI device IDs to identify the GPUs, and as such will need to be updated from time to time as new
	devices come on the market.

	## Guidelines for adding a new device or set of devices

	There are no automated processes to add devices to `gpu_info.json`. Everything is inserted by hand. Fortunately the
	file is intended to be human readable, and should be relatively easy to update following the comments at the top of the
	file and referencing the existing entries.

	When adding new devices use of a `mask` is encouraged when applicable to capture more devices with less entries.
	However, it's important to ensure that the masked device IDs are not overly broad or capture devices which do not belong
	to a given architecture. Many GPU vendors do not use very consistent patterns in their IDs, or have strange exceptions
	to the patterns that do exist.

	When in doubt, prefer adding an explicit device ID without a `mask` over adding an masked ID that might accidentally
	capture more devices than intended.

	A GPUs `architecture` should be the vendor's terminology for a given family of devices when possible. An `architecture`
	group should not be detailed enough to identify specific devices, but should ideally group all devices made within a
	span of a couple of years or longer. It is OK to have significant performance disparities between devices in a single
	`architecture` group, as long as the capabilities and behavior of the devices are roughly similar.

	## Rolling new devices

	Capturing new devices that need to be added to `gpu_info.json` right now is unfortunately a fairly awkward process. The
	method used so far has been to checkout https://github.com/kainino0x/gpuinfo-vulkan-query, update the data submodule
	with the latest records from https://gpuinfo.org using the script at https://github.com/kainino0x/gpuinfo-vulkan-slurp
	(Ask kainino@ or bajones@ for access), and then run `device_id.py` with the path to the `gpu_info.json` file you are
	updating. That script will compare all the records to the entries in `gpu_info.json` and print out how each of them are
	categorized followed by a list of all the uncategorized device IDs and descriptions sorted by vendor. You'll have to
	manually look at the device descriptions and research their architectural design in order to categorize them.

	Good luck!