Replies: 1 comment 1 reply
-
According to my understanding the runtime takes into account SLM size restrictions for work scheduling, but you can easily verify this with a simple app. Right? NOTE: by "runtime" here I mean low-level runtime as Level Zero, OpenCL, etc., so it's better to forward this question to https://github.com/intel/compute-runtime/ project. DPC++ runtime just propagates the information from user to the low-level runtime, but work-item/work-group mapping to HW is done by low-level runtime. |
Beta Was this translation helpful? Give feedback.
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
Take SIMD8 as an example, the max number of work groups on a DualSubSlice(DSS) is: (16EU * 8 Thread * SIMD8 / wg_size 512) = 2.
Say a DSS has 64KB SLM, and a workgroup needs 40KB SLM. Does the runtime know the max number of workgroups can be scheduled on an DSS is 1 (because 40 *2 > 64), or the developer must make sure not to use SLM size larger than 32KB for one workgroup on the application side?
Beta Was this translation helpful? Give feedback.
All reactions