[CPU][ARM] Enable subgraph tests #25399
Conversation
alvoron
force-pushed
the
alvoron_subgraph_tests2
branch
from
e9c1c60 to
d727e6a
Compare
EgorDuplensky
left a comment
EgorDuplensky
left a comment
There was a problem hiding this comment.
Can we try to enable causal mask preprocess fusion for ARM as well?
@usstq What do you think?
| }; | ||
|
|
||
| #if defined(OPENVINO_ARCH_ARM) | ||
| const ExpectedResult successfull_fuse_result{1, 1, 3}; |
There was a problem hiding this comment.
It seems the reason is an additional reorder appears in the graph because the Eltwise (bias?) is not fused into groupconvolution node
There was a problem hiding this comment.
Yes, EltwiseAdd is not in the list of supported postops: https://github.com/alvoron/openvino/blob/37ab2c6abe2519e1ae2494e163798dfce7d8fce4/src/plugins/intel_cpu/src/dnnl_extension_utils.cpp#L227-L234
EgorDuplensky
left a comment
EgorDuplensky
left a comment
There was a problem hiding this comment.
@maxnick Can we use a planar layout for ReLuConcatConvSumInPlaceTest subgraph test? Since currently the test forces blocking layout it is impossible to run it on ARM
EgorDuplensky
left a comment
EgorDuplensky
left a comment
There was a problem hiding this comment.
let's remove layout specialization from interaction subgraph test and enable it for ARM as well.
| @@ -0,0 +1,90 @@ | |||
| // Copyright (C) 2018-2024 Intel Corporation | |||
There was a problem hiding this comment.
Why we have to make those instances x86 specific?
There was a problem hiding this comment.
Reference implementation (ref_any) does not support nchw/ndchw layouts on ARM. I've got the error No supported primitive descriptors matched the provided input / output memory format filters while pushing these layouts into the test.
nspc works well for reference on both ARM and x64. Can we replace nchw/ndchw with nhwc/ndhwc and keep conv tests common?
There was a problem hiding this comment.
Discussed with Dmitry, there is no need to enable ncsp descriptor for ARM in Convolution node. We will keep ncsp tests as x64-specific.
|
|
||
| TEST_P(FuseTransposeAndReorderTest3, CompareWithRefs) { | ||
| run(); | ||
| check_transpose_count(1); |
There was a problem hiding this comment.
what result do we get for ARM?
There was a problem hiding this comment.
There is no Transpose on ARM.
Yes, sure, but please use nspc layout instead as the main one for optimized convolution implementations. |
As discussed offline, Interaction node and interaction transformations are enabled on x64 only, so Interaction tests are x64 specific as well. |
alvoron
force-pushed
the
alvoron_subgraph_tests2
branch
from
5a9b362 to
37ab2c6
Compare
| selectedType += execType.get_type_name(); | ||
| } | ||
|
|
||
| std::vector<CPUSpecificParams> filterCPUSpecificParams(const std::vector<CPUSpecificParams>& paramsVector) { |
There was a problem hiding this comment.
Let's revert this change and fix the particular test cases which cause the issue
There was a problem hiding this comment.
This change is reverted (this method has been moved to utils/x64/filter_cpu_info.cpp
2 methods - getCPUParamsBlocked4D and getCPUParamsBlocked5D were implemented in single_layer_tests/slice.cpp to do x64-specific fintering
…openvino into alvoron_subgraph_tests2
### Details: - Subgraph tests are enabled on ARM platforms ### Tickets: - CVS-112267
4 participants
Details:
Tickets: