Commit c733cd0
authored
[GPU] Optimize Dispatch Data Configuration for Mirror Padding Kernel (openvinotoolkit#32400)
## [GPU] Optimize Dispatch Data Configuration for Mirror Padding Kernel
### Description:
Added a new logic that optimizes the selection of GWS/LWS config for
Padding kernels by introducing predefined mapping for special cases.
This modification improves hardware resource utilization, resulting in a
significant increase in OpenCL kernel performance.
Detailed performance results: CVS-174882.
### Implementation-level details:
- Added an dispatch optimization for border kernels that improves kernel
performance.
- For selected GWS, the new logic applies a lookup table with predefined
optimal (GWS, LWS) pairs.
- The function introduces better local workgroup size (LWS) allocation.
- The optimization is stateless, encapsulated, and reusable across all
border kernel implementations.
- Additional predefined pairs can be added easily.
### Reproduction step and snapshot:
Description available at the foot of the description section of the JIRA
ticket : CVS-174882.
### Problematic graph:
n/a
### Checklist:
- [x] Is it a proper fix?
- [ ] Did you include test case for this fix, if necessary?
- [x] Did you review existing test that can be extended to cover this
scenario?
### Tickets:
- CVS-174882
- CVS-1738051 parent be8f36b commit c733cd0
File tree
1 file changed
+11
-0
lines changed- src/plugins/intel_gpu/src/kernel_selector/kernels/border
1 file changed
+11
-0
lines changedLines changed: 11 additions & 0 deletions
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
57 | 57 | | |
58 | 58 | | |
59 | 59 | | |
| 60 | + | |
| 61 | + | |
| 62 | + | |
| 63 | + | |
| 64 | + | |
| 65 | + | |
| 66 | + | |
| 67 | + | |
| 68 | + | |
| 69 | + | |
| 70 | + | |
60 | 71 | | |
61 | 72 | | |
62 | 73 | | |
| |||
0 commit comments