* Fix in the transformation PreserveRuntimeInfo: now Transpose is inserted before input port 0 of Result only, not after data node of layer before Result layer.
* Deleted commented code.
* Added more tests for the MO transformation PreserveRuntimeInfo.
* Use fp16-int8 mixed precision, instead of fp32-int8 mixed precision for onednn
* Allow quantization fusion into bsv32_fsv16 conv
* For conv, do not select bsv16_fsv16. Select bsv32_fsv16 for mixed-layout
* depthwise conv is supported even though it is not fp16
* Allow resample kernel to work as cross-layout
* test case for cross-layout of resample_opt kernel
* Select onednn-friendly format from cldnn conv
* Optimization for fp16 mixed precision
* Choose mixed layout in case of mixed precision from reorder_inputs
* Support for mixed precision from depth_to_space
* Do not convert first conv format
* Use onednn for FC output of fp16
* Choose bsv8_fsv4 from quantization even when conv kernel size is not 7
* Select cldnn for first conv when input feature depth is 1
* For first conv, use onednn only when kernel size is 7x7
* Use short variable name and added is_i8_u8 helper function
Co-authored-by: Kim,SungEun <sungeun.kim@intel.com>
* [LPT] Documentation
* 1) ToC was removed 2) SVG => PNG temporary conversion
* [LPT] Refactoring + developer guide
* [LPT] attribute doxygen documentation was added
* [LPT] Developer Guide to Reference API links were added
* [LPT] comments fixes
* [LPT] Reference API to Developer Guide links were added
* [LPT] titles were changed
* [LPT] comments fixes#2
* [LPT] root document was moved to Plugin DG
* [LPT] Documentation: image link quick fix
* [LPT] Docummentation: PrecisionsAttribute description quick fix
* fix comments from Karol
* fixes
* movement
* directive was added
* movement #2
* LPT reference in Executable Network rollback
* snippets were updated ini accordance with new API
* Handle names collisions for old IR with new API
* Fixed load model
* Try to fix tests
* Try to fix tests
* Try to fix build
* Try to fix tests
* Fixed tests
* Revert "Fixed tests"
This reverts commit 35da307210.
* Refactoring
* Fixed functional test
* Try to fix CPU tests
Co-authored-by: Ilya Lavrenov <ilya.lavrenov@intel.com>
* [Python API] Remove offline transformations from old python api
* try to fix import error
* try to fix pylint
* try to fix pylint2
* Use new api in graph_utils
* Fix pylint
* Try to fix pylint
* Use serialize from pass manager
* try to skip tests
* try to use new ir
Co-authored-by: AlexeyLebedev1 <alexey.lebedev@intel.com>
* [GPU] Enable unet2d enable on DG2
Add to support is_os_yx_isa2_osa8_isv8_osv2 format, which is used in
weight reorder.
Signed-off-by: hyunback <hyunback.kim@intel.com>
* [GPU] Enable implicit concat batch1 in oneDNN.
* Use gpu_usm memory offset, enable implicit concat batch1 in oneDNN.
And optimized_out node doesn't always have to be mutable input,
so add to check whether mutable input is existed in optimized node.
* Update to check use_usm condition in implicit concat.
* Add the condition for implicit concat.
* implicit concat's dependency should not be fused_op with eltwise.
* Buffer reuse is required for onednn sum post operation, output padding
did the buffer reuse failure.
Signed-off-by: hyunback <hyunback.kim@intel.com>
+ cldnn supports hard sigmoid activation function but onednn doesn't.
+ split it into eltwise linear and eltwise clip in
add_onednn_optimization_attributes pass.
* Add MatMulMultiplyFusion
MatMulMultiplyFusion replaces following subgraph:
MatMul->Multiply (with const)
to following:
Multiply->MatMul
where Multiply is applied to MatMul's second input.
