### Details:
- PRs addressing GFIs are getting difficult to track, this PR aims to
increase the community's awareness about this GitHub feature.
- The change to GFI template will also be introduced to GFI templates in
`openvino_contrib`, `openvino.genai` and `nncf`.
- An additional reminder has been added to the GFI board:
### Tickets:
- N/A
### Details:
- *Extended loader ADDV2 by propagating ComplexTypeMark from input to
output and to represent output complex type tensor as a floating-point
type tensor with an auxiliary dimension that concatenates real and
imaginary parts of complex tensor.*
- *Performed addition for complex numbers.*
- *Wrapped the complex result with ComplexTypeMark and returned the
result*
Fixes#22946
---------
Co-authored-by: Roman Kazantsev <roman.kazantsev@intel.com>
### Details:
- Added support for complex tensors and tests for it.
### Tickets:
- #23244
---------
Co-authored-by: Roman Kazantsev <roman.kazantsev@intel.com>
**Details:** Test RaggedTensorToTensor for rowids format and Equal for
string tensors. Needs to be merged after
https://github.com/openvinotoolkit/openvino_tokenizers/pull/70
**Tickets:** TBD
---------
Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>
### Details:
- It turned out that the `sccache` tool interfered with the Coverity
scans making it report 0 issues. This PR disables the caching
functionality in the Coverity workflow.
- It was tested using this run:
https://github.com/openvinotoolkit/openvino/actions/runs/8204259038 and
viewing on the Coverity platform.
### Tickets:
- *134999*
### Details:
- *Added a condition to defer reshape memory update*
- *If the reshape is static shaped and has null dep memory, its memory
update will be deferred*
### Tickets:
- *125236*
If TypeRelaxed's origin input type is undefined let's temporarily
override it with original input precision attribute value. During
ConstantFolding, some nodes can have temporarily mismatched input types
(e.g. Add(f16, f32)). If the node is TypeRelaxed - we're unable to clone
it since TypeRelaxed::clone_with_new_inputs creates a clone with 'fake'
inputs based on current inputs and that can trigger an exception for
certain nodes if the inputs have mismatched types.
Ticket: CVS-134604
- Updated paths in an entire documentation,
- updated scripts used in docs building process,
- updated docs cmake to handle new scripts,
- fixed links,
- fixed other errors found in docs.
---------
Co-authored-by: Maciej Smyk <maciejx.smyk@intel.com>
Co-authored-by: Sebastian Golebiewski <sebastianx.golebiewski@intel.com>
Co-authored-by: Karol Blaszczak <karol.blaszczak@intel.com>
Co-authored-by: Andrzej Kopytko <andrzejx.kopytko@intel.com>
Co-authored-by: Tatiana Savina <tatiana.savina@intel.com>
Co-authored-by: Vishniakov Nikolai <nikolai.vishniakov@intel.com>
### Details:
- Fix CreateStridedSliceOp() not to use fixed Rank 4 reshape for
new_shape_infer case
- Add a func test case
- Fix debug log in condition
### Tickets:
- 130775
By using a vector instead of a map.
To reduce an overhead of selecting an executor.
Since we know the exact mapping beforehand (by implementation Id): every
executor corresponds to
a particular implementation.
### Details:
- *Control flow pass `SoftmaxDecomposition` removed*
- *Introduced data flow `SoftmaxDecomposition` pass which decomposes
`Softmax` operation to a sequence of `Eltwise` ops + `Reduce` ops*
- *Introduced control flow `ReduceDecomposition` pass which decomposes
Reduce operation to a range of low-level operations on linear IR.
Limitation: only reduction by last dimension is supported*
### Tickets:
- *CVS-112847*
### Prerequisites:
- #21303
PR between branches for review (all the review discussions are there):
[Softmax decomposition moved to data flow
pipeline](https://github.com/v-Golubev/openvino/pull/24#top)
Closes#23042
### Details:
- Create function `validDeviceName` to
`src/inference/src/dev/core_impl.cpp` to check the validity of device
name.
### Tickets:
- CVS-124122
### Details:
- Previously, memory conflict check was done using std::string
(primitive_id), and it was time consuming
- Fixed to use unique_id as mem_dep, instead of std::string
### Tickets:
- 131916
### Details:
- Skip Depth To Space fusing when dynamic shape
- Skip broadcastable check in select typed_primitive_inst() when new
shape infer
### Tickets:
- 130775
### Details:
- Update logic of port name checking to check if queried port tensor
names is subset instead of op "==".
- *...*
### Tickets:
- CVS-122932
### Details:
- Throw exception when detecting wrong format like '-GPU' in BATCH.
- Throw exception with right information when getting multiple devices
for BATCH.
### Tickets:
- CVS-132276
### Related issues:
- Closes#23125
### Details:
constant folding been introduced to hetero, but this may bring other
issue in quantized models , for example, in gpu plugin, precision
conversion will be applied before constant folding, some ops can be
operated in fp16 if hardware support, while direct constant folding in
hetero doesn't take hardware capability into account and fold the
constants directly, this will result in memory concerns, for example,
OOM issue, or the performance issue because of different execution
precision
### ticket-id:
CVS-127934
### Details:
- Added internal `Convolution` op with bias and zero-points semantics
which covers both grouped and non-grouped case
- Added internal `Placeholder` op which indicates that port has no input
- Added `{v1::Convolution, v1::GroupConvolution} ->
internal::Convolution` conversion pass.
- Added fusion pass for convolution with asymmetric quantization
- Added pass for broadcasting and alignment of quantization parameters
- Removed `prepare_aymmetric_quantization` cldnn pass
### Details:
- OV Model API allows to create `ov::Model` with empty ResultVector. And
it may cause segmentation fault on `graph.dump()` e.g. when
`get_property("NETWORK_NAME")`
### Tickets:
- [CVS-134517](https://jira.devtools.intel.com/browse/CVS-134517)
[Transformations] Fix warning: implicit capture of ‘this’ via ‘[=]’
Details:
Implicit capture of 'this' for lambdas is deprecated since C++20. Fix it
by adding a special macro OV_CAPTURE_CPY_AND_THIS which expands into '='
or '=, this' depending on a C++ standard used.
Note:
Although C++20 identifier is 202002L, some compilers supporting C++20,
or its drafts like C++2a, producing the warning, are using 201402L
value. Also, MSVC requires a special check due to the __cplusplus value
compatibility issues.
Ticket: CVS-132688
Signed-off-by: Andrii Staikov andrii.staikov@intel.com
Signed-off-by: Andrii Staikov andrii.staikov@intel.com
### Details:
If ReadValue and Assign are not paired in model topology we have "Model
is incorrect. Assign and ReadValue operations must be in pairs on the
network."
Check for ReadValue-Assign pairs is useless and it is an obstacle for
model transformation debugging. Moreover, with more advanced techniques
of state modification there can be other ops (custom ones) that modify
state in place.
### Tickets:
- *CVS-133159*
Przemyslaw Wysocki
Rinne