test: aarch64: Fix unimplemented error when --cpu-isa-hints=prefer_ymm #2208
Conversation
tests/benchdnn/dnnl_common.cpp
Outdated
| @@ -272,7 +272,7 @@ void init_isa_settings() { | |||
| if (hints.get() == isa_hints_t::no_hints) | |||
| DNN_SAFE_V(dnnl_set_cpu_isa_hints(dnnl_cpu_isa_no_hints)); | |||
| else if (hints.get() == isa_hints_t::prefer_ymm) | |||
| DNN_SAFE_V(dnnl_set_cpu_isa_hints(dnnl_cpu_isa_prefer_ymm)); | |||
| dnnl_set_cpu_isa_hints(dnnl_cpu_isa_prefer_ymm); | |||
There was a problem hiding this comment.
I feel that updating this function makes more sense. Aarch64 is fine to return success unconditionally.
There was a problem hiding this comment.
+1 on Dmitri's solution as it is the simplest solution.
Down the line, it might also make sense to guard the platform specific isa/isa_hint enum values in the headers. Currently, DNNL_X64 is not exposed there but we could add it to dnnl_config.h (similar to what we do for experimental features).
There was a problem hiding this comment.
Thank you for the review, I have update the code as suggested
|
Could you reformat the commit message to: The PR checks should have caught this but they are probably not working as intended. |
Ryo-not-rio
force-pushed
the
fix-ymm-unimplemented
branch
from
2fc6ff9
to
c6de06a
Compare
Ryo-not-rio
changed the title
…m on aarch64 When running with --cpu-isa-hints=prefer_ymm, tests fail on aarch64 due to an unimplmented error. This change will instead allow the code to run the reference implementations. Reproducible: ``` ./tests/benchdnn/benchdnn --conv --cpu-isa-hints=prefer_ymm --dt=bf16:bf16:bf16 --stag=axb --dtag=axb --attr-post-ops=sum+tanh:1:1 g1ic64ih1000oc64oh1000kh3ph128dh127 ``` Change-Id: Ibf075ad71da0b870002c8cbfc3fefcaa883ef9b6
Ryo-not-rio
force-pushed
the
fix-ymm-unimplemented
branch
from
c6de06a
to
b8d3cd8
Compare
When running with --cpu-isa-hints=prefer_ymm, tests fail on aarch64 due to an unimplmented error. This change will instead allow the code to run the reference implementations.
Reproducible: