WIP: Add subgroups feature #4962
Conversation
* API
* new feature: subgroups
* new properties for adapter info:
* minSubgroupSize
* maxSubgroupSize
* WGSL
* new built-in values
* subgroup_invocation_id
* subgroup_size
* subgroup and quad built-in functions
* new uniformity diagnostic subgroup_uniformity
|
@kainino0x @toji I'd appreciate some early feedback on the API changes around properties. |
|
Closed in favour of #4963 |
| @@ -1929,6 +1931,29 @@ interface GPUAdapterInfo { | |||
| this value is not recommended. Applications which change their behavior based on the | |||
| {{GPUAdapterInfo}}, such as applying workarounds for known driver issues, should rely on the | |||
| other fields when possible. | |||
|
|
|||
| : <dfn>properties</dfn> | |||
There was a problem hiding this comment.
Is there a need/purpose for encapsulating these things in a new GPUAdapterProperties object?
Can the two new fields be included directly in GPUAdapterInfo?
Somewhat related in C-style APIs, I expect we would have a new chained struct, e.g. either GPUAdapterSubgroupInfo (or GPUAdapterPropertiesSubgroupInfo if we do introduce a new GPUAdapterProperties)
| while on other devices they may not. | ||
|
|
||
| Note: There is considerable non-portability among underlying devices when | ||
| operating in non-uniform control flow and device compilers often agressively |
| operating in non-uniform control flow and device compilers often agressively | ||
| optimize such code. | ||
| The result is that subgroups may contain different invocations than the shader author | ||
| reasons ought to be active in a particular operation. |
There was a problem hiding this comment.
The "reasons ought to" phrasing seems awkward.
Suggest replacing this sentence with something like one of these:
The result is that the set of active invocations executing a particular subgroup operation may not be what a shader author expects.
The result is that for a particular subgroup operation, the set of active invocations may not be what a shader author expects.
The result is that for a particular subgroup operation, the set of active invocations may differ from shader author expectations.
Missing subgroup_branching diagnostic and uniformity updates.