src: cpu: aarch64: Enable matmul and convolution static quantisation through ACL. #2198
Conversation
renato-arantes
changed the title
github-actions
bot
added
the
platform:cpu-aarch64
renato-arantes
force-pushed
the
master
branch
2 times, most recently
from
efad4f6 to
d728790
Compare
| } | ||
|
|
||
| status_t acl_lowp_matmul_sq_t::execute(const exec_ctx_t &ctx) const { | ||
| std::lock_guard<std::mutex> _lock {this->mtx}; |
There was a problem hiding this comment.
Maybe add a comment on which part below is not thread safe. That would help remove the lock later in time.
There was a problem hiding this comment.
For some reason, I do not see any new comments. Did you forget to push your changes?
There was a problem hiding this comment.
I removed the code from src/common/convolution_pd.hpp as you requested. Now the bias is set to f32 from inside src/cpu/aarch64/acl_gemm_convolution.cpp, and consequently, the types match and so the #ifdef inside convolution_pd.hpp is not needed anymore. The change is there, but it is mostly about removing the previous one!
renato-arantes
force-pushed
the
master
branch
2 times, most recently
from
25271ff to
91524da
Compare
|
@mgouicem Can we please close this? |
|
@renato-arantes you have put in a merge commit. You need to squash and rebase. |
renato-arantes
force-pushed
the
master
branch
from
9c53fdc to
360a144
Compare
|
@mgouicem, I have addressed all the comments. |
|
@theComputeKid, @renato-arantes, Looks like this PR passed benchdnn Full log is here. |
|
As the result I had to disable all required checks for AArch64 to avoid blocking promotion of PRs for other platforms. |
|
Thought I was doing a good deed by expanding CI but alas no good deed goes unpunished. @Sqvid do you know why we didn't catch this internally? We need to fix this ASAP so we can get our public CI working again. |
Change-Id: Ifaecea0c53d270e7dd3e4c38b4261456e40c7418 Signed-off-by: Renato Arantes <[email protected]> Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/c/oncpuml/oneDNN/+/706947 Tested-by: svc_mongoosetron <[email protected]> Reviewed-by: Radu Salavat <[email protected]> IP-Review: Hamza Butt <[email protected]>
5 participants
Description
This is to enable
matmulandconvstatic quantization operations through ACL. Currently, the supported data type combinations ares8:s8:s8andu8:s8:u8.Checklist
General
make testandmake test_benchdnn_*) pass locally for each commit?