[GPU] updates to build some selected kernels in separate batches #24499
Conversation
| return unique_kernel_name.substr(0, pos); | ||
| }; | ||
|
|
||
| auto get_target_batch = [&]() -> batch_program& { |
There was a problem hiding this comment.
Does the issue really happen due to multiple instances of same kernel in the batch or it's just related to batch size? As I remember, if program source is too large, then igc may produce worse binary
There was a problem hiding this comment.
It's not clear what the root cause is as of now. BTW, it looks like that it is not due to program source size. In my test, the issue was gone when I commented out just one line.
e-ddykim
force-pushed
the
gpu-wa-kernels_cache
branch
from
1215cb9
to
8e6b519
Compare
e-ddykim
changed the title
e-ddykim
force-pushed
the
gpu-wa-kernels_cache
branch
from
8e6b519
to
e55b60a
Compare
e-ddykim
force-pushed
the
gpu-wa-kernels_cache
branch
from
e55b60a
to
9b45712
Compare
| // check if the current kernel name is in special_kernels | ||
| auto target_base_kernel_name = get_base_kernel_name(entry_point); | ||
| if (std::count(special_kernels.begin(), special_kernels.end(), target_base_kernel_name) > 0) | ||
| return true; |
There was a problem hiding this comment.
If currerent entryu has gemm_tiled_opt => it will need_seperate_batch : Is this the intention?
(Current behavior seems so)
If it is so, why not just simply check :
if (entry_point.find("gemm_tiled_opt") != string::npos)
=> need_separate_batch?
There was a problem hiding this comment.
Oh! I updated it as you reviewed. Thank you!
e-ddykim
force-pushed
the
gpu-wa-kernels_cache
branch
from
9b45712
to
7cbd233
Compare
|
I believe this will be reverted once the driver issue is resolved. Could you please add the ticket numbers to the PR? |
I added it. Thank you. |
github-merge-queue
bot
removed this pull request from the merge queue due to failed status checks
Details:
kernels_cacheto build the selected kernels in separate batches.gemm_tiled_opt.Tickets: