* Use Gna2DeviceCreateForExport when GNA_EXEC_TARGET is != detected
* Update detected GNA device version field in GNA Device helper
* Use EXEC instead of COMPILE TARGET to append
CNN Legacy enforcement (GNA1)
* Apply review
Quote: The Skylake microarchitecture implements a different state
machine than prior generations to manage the YMM state transition
associated with mixing SSE and AVX instructions.
It no longer saves the entire upper YMM state when executing
an SSE instruction when in “Modified and Unsaved” state,
but saves the upper bits of individual register.
As a result, mixing SSE and AVX instructions will experience
a penalty associated with partial register dependency of
the destination registers being used and additional blend
operation on the upper bits of the destination registers.
Such type of penalties have a huge impact on openvino's and oneDNN's kernels.
Basically the mixing of VEX and non-VEX instructions should be avoided.
* [GNA] Port padding and 2d convolution support from master
* [GNA] Add separate fixes required after porting
* [GNA] Add support for activation without bias in padding and 2d convolution decomposition
* [GNA] Fix scale factors propogation for Eltwise with very different inputs ranges
* [GNA] Added test
* [GNA] Added exception for scale factor <= 0
* [GNA] Disable tests with integer weights
* [GNA] Added assert for CNNLayer in getScaleFactor()
* [GNA] Added check if scale factor is inf
* [GNA] Fixed legacy tests
* install_NEO_OCL_driver: Added detecting current driver via intel-opencl-icd package in case of newest one on ubuntu20. Added removing intel-opencl-icd package for ubuntu.
* install_NEO_OCL_driver: Fixed pattern for parsing driver version for newest drivers
* [GNA] Fixed search of the next layer for FQ
* [GNA] Fixed calculation of input scale factor for POT-quantized model in the case if the first layer after input is activation
* [GNA] Use stride instead of window for pooling (#5946)
* Use pool stride instead of window size where applicable
* Add test for pooling stride not equal to wnd
* Add more tests and cleanup
* Fix SW_FP32 legacy cnn
* [WIP] Refactor CNN1D
* Remove unused (commented out) code
* Add tests
* Gna split align convert to conv filter (#6347)
* Make unaligned split based on Conv instead of Affine
* Dump Gna2Tensor.Data pointer for debugging
* Apply suggestions from code review
* Reuse conv helpers
* Cleanup CNN fields
* Disable weights reducer on ConvolutionFilter
# Conflicts:
# inference-engine/src/gna_plugin/backend/am_intel_dnn.cpp
# inference-engine/src/gna_plugin/optimizer/gna_pass_manager.cpp
Co-authored-by: Krzysztof Bruniecki <krzysztof.bruniecki@intel.com>
* [GNA] Remove transposes around MatMul
* Added tests for transformation HandleTransposesAroundMatMul
* Move IsTransposeSupported function to GNA limitations file
* added TransposeAfterMatmul tests and moved InsertTransposeBeforeMatmul tests to handle_transposes_around_matmul.cpp
* added inifitiry loop checker and memory concat test
* fixed build errors
* changed the conditions for selecting an input of Concat for ScaleFactor calculation when entering an infinite loop
* fixed after review
* s/INSTANTIATE_TEST_SUITE_P/INSTANTIATE_TEST_CASE_P
* .ignore
Co-authored-by: Elizaveta Lobanova <elizaveta.lobanova@intel.com>
