* Fix fusing Multiply node with Convolution in case group != 1
* Add transformation test
* Do not fuse if not possible to reshape const
* Update fuse_linear_ops.py
* Make pybind more verbose in debug on windows
* Remove the NDEBUG flag everywhere
* Code complexity reduction...
* Missing colon
* And now the missing empty line...
* Reusable functions
* Now the mood of the sentence was wrong...
* Free functions instead of methods
* Add host_tesnor_2_vector() implementation and unit tests. One reference OP refactored to use it.
* Ngraph assertion message refactored.
Co-authored-by: Tomasz Dołbniak <tomasz.dolbniak@intel.com>
* Fix style.
Co-authored-by: Tomasz Dołbniak <tomasz.dolbniak@intel.com>
- U16toF32 conversion kernel converted to more generic ConvDepth one
- U16 <-> F32 conversion only are supported for now
- kernel is not used in the preprocessing graph yet
- tests are extended
* Moves splitLargeKernelConv tests to unit tests
Originally, file with tests has been placed in a wrong place
so it was not integrated into any testing application.
Now it is a part of unit tests on VPU.
Test itself has been disabled due to issue with NCE unit usage
described in #-33366
* Introduces pass I/O memory types annotation of stages
It is useful to see where inputs and outputs are located in
performance report for analysing possible issues.
* Introduces endsWith and tuple2Vector utilities
endsWith checks if source has suffix equals to second
argument. tuple2Vector converts tuple of arbitrary size
containing the same type to vector. It could be useful
working with gtest parameter generators that have
std::tuple as return type.
* Introduces unit tests on annotating stages memory types
* Introduces missing format placeholders
* Makes memory types annotation optional
Enables private option "enableMemoryTypesAnnotation" which
disabled by default. Disabling annotation by default allows
avoid issues with tests which rely on stages names.
Signed-off-by: Gladilov, Gleb <gleb.gladilov@intel.com>
* Mish activation calculation costs more time than memory copy, so
allocate more shaves mish activation.
Co-authored-by: Jiang, Renzhi <renzhi.jiang@intel.com>
* Fixed visitor for Interpolate-1 and Interpolate-4
* Code style fix
* Remove unnecessary changes
* Fixed compilation on Linux for Atttribute visitor of vector<size_t>
* Added unit test for IE IR Reader for Interpolate-4
* Updated unit test for IR Reader for Interpolate-4
* Updated unit test
[GNA] Fixed case of unconnected output of split layer
[GNA] Fixed case of unconnected output of split layer
test
[GNA] Fixed case of unconnected output of split layer
fixed
* Updated ConcatOptimization transformation to work when one dimension of input to Concat is 0D
* Fixed ConcatOptimization transformation to reconnect input edges to Concat
* Completely re-written ConcatOptimization
* Updated Concat0D optimization transformation
* Fixed order of traversing Concat input ports
* Refactored ConcatOptimization transformation to use `delete_input_port` function
* Detele trailing unconnected ports in the ConcatOptimization.py
* Cleaner implementation of ConcatOptimization + unit test
