[cpu] Remove custom shape inference factories #27924
Conversation
Signed-off-by: Raasz, Pawel <[email protected]>
Signed-off-by: Raasz, Pawel <[email protected]>
to inputs data as constant if available Signed-off-by: Raasz, Pawel <[email protected]>
Signed-off-by: Raasz, Pawel <[email protected]>
Signed-off-by: Raasz, Pawel <[email protected]>
Signed-off-by: Raasz, Pawel <[email protected]>
praasz
force-pushed
the
shape-infer/remove-cpu-custom-shape-infer-factories
branch
from
14ee211 to
75dfef5
Compare
| /** | ||
| * @brief Shape inference for v14 MaxPool. | ||
| * | ||
| * It requires dedicated port mask for CPU output shape data dependency but is not used by inference function. | ||
| * Review shape_infer function to include this dependency. | ||
| */ | ||
| template <> | ||
| class ShapeInferPaddingTA<ov::op::v14::MaxPool, EMPTY_PORT_MASK> : public ShapeInferPaddingBase { | ||
| public: | ||
| using ShapeInferPaddingBase::ShapeInferPaddingBase; | ||
|
|
||
| ov::optional<std::vector<StaticShape>> infer(const std::vector<StaticShapeRef>& input_shapes, | ||
| const ov::ITensorAccessor&) override { | ||
| return {shape_infer(static_cast<ov::op::v14::MaxPool*>(m_node.get()), input_shapes, m_pads_begin, m_pads_end)}; | ||
| } | ||
|
|
||
| port_mask_t get_port_mask() const override { | ||
| return util::bit::mask(0); | ||
| } | ||
| }; | ||
|
|
||
| /** | ||
| * @brief Shape inference for v8 MaxPool. | ||
| * | ||
| * It requires dedicated port mask for CPU output shape data dependency but is not used by inference function. | ||
| * Review shape_infer function to include this dependency. | ||
| */ | ||
| template <> | ||
| class ShapeInferPaddingTA<ov::op::v8::MaxPool, EMPTY_PORT_MASK> : public ShapeInferPaddingBase { | ||
| public: | ||
| using ShapeInferPaddingBase::ShapeInferPaddingBase; | ||
|
|
||
| ov::optional<std::vector<StaticShape>> infer(const std::vector<StaticShapeRef>& input_shapes, | ||
| const ov::ITensorAccessor&) override { | ||
| return {shape_infer(static_cast<ov::op::v8::MaxPool*>(m_node.get()), input_shapes, m_pads_begin, m_pads_end)}; | ||
| } | ||
|
|
||
| port_mask_t get_port_mask() const override { | ||
| return util::bit::mask(0); | ||
| } | ||
| }; | ||
|
|
There was a problem hiding this comment.
Pooling::Pooling CPU node has been using NgraphShapeInferFactory without any additional mask modifications. And it seems that the Pooling node doesn't require input data dependency for shape inference. Such change may negatively impact performance, as the artificial data dependency introduce additional synchronizations to inference.
There was a problem hiding this comment.
There is no issue when pooling is created as Pooling node.
The issue is visible if is not possible to make as Pooling and the Reference node, its case for some dynamic cases
in tests/layer_tests/pytorch_tests/test_pooling.py::TestPooling tests.
The previous Reference node always used FULL_MASK regardless if shape inference is defined or not (fallback path). I think this approach has masked issues. Now reference use mask defined in shape inference (if defined) or FULL_MASK for fallback path.
Then issue appears in when the input(0) has dynamic shape but data at port(0) has static shape (exception thrown when get static shape). When dada dependency is set then static shape from data is taken and pooling output shape can be calculated.
These specialization (can be removed if problem solved) shows that issue could be fixed in:
- shape inference add additional data dependency (it looks like is not required now)
- for dynamic cases the data dependency should be always set for port 0 to get input static shape.
For FakeQuantize is similar issue and there are test failing if data dependency not set to port 0
…cpu-custom-shape-infer-factories
Details:
Related PR
Tickets: