[GPU] Revert arg max min axis change and fix incorrect offset calculation#31956
Conversation
|
Able to fix the test failure by increase the size of the local memory when # of local threads increase. However, it is not reliable for dynamic support. During graph compilation, the dynamic shape is not known, choose between local/global memory to use for temp buffers are not reliable. There will be the case where if local memory has been choosen, but during runtime when dynamic shape already known, the size may exceed local memory limit. In this case, it is not able to revert back to use global memory. Will revert to use local memory, and fix the issue where data mismatch when using global memory temp buffer. |
In arg_max_min_axix, use sizeof to determine size of struct iav_type. Current method have issue to calculate for fp16 type as OpenCL compiler tend to add padding to improve memory access efficiency. Revert "[GPU] Use local memory for arg_max_min for dynamic shape (openvinotoolkit#31682)" This reverts commit 69a1121.
clee30
force-pushed
the
arg_max_min_increase_threads
branch
from
6697761 to
8d7be39
Compare
Add fix to original implementation using global memory to handle fp16 size correctly. |
clee30
changed the title
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
ddc697f
3 participants
In arg_max_min_axix, use sizeof to determine size of struct iav_type.
Current method have issue to calculate for fp16 type as OpenCL compiler tend to add padding to improve memory access efficiency.
Revert "[GPU] Use local memory for arg_max_min for dynamic shape (#31682)"
This reverts commit 69a1121.
CVS-172937, CVS-172939