[LPT] FQ decomposition in common pass manager - exporation #5

eshoguli · Jun 24, 2021 · c73b161 · c73b161
1 parent 6bb43a7
commit c73b161
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 4 deletions.
diff --git a/inference-engine/src/low_precision_transformations/include/low_precision/fake_quantize.hpp b/inference-engine/src/low_precision_transformations/include/low_precision/fake_quantize.hpp
@@ -25,7 +25,7 @@ class LP_TRANSFORMATIONS_API FakeQuantizeTransformation : public LayerTransforma
 private:
     std::shared_ptr<opset1::FakeQuantize> fuseElementwise(
             TransformationContext& context,
-            MatcherPass& matcherPass,
+            MatcherPass* matcherPass,
             const std::shared_ptr<opset1::FakeQuantize>& fakeQuantize) const;
 };
 

diff --git a/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/fake_quantize.cpp
@@ -42,7 +42,7 @@ bool FakeQuantizeTransformation::transform(TransformationContext& context, ngrap
     bool wasHandled = false;
     std::shared_ptr<opset1::FakeQuantize> fakeQuantize = layer;
     do {
-        fakeQuantize = fuseElementwise(context, *this, fakeQuantize);
+        fakeQuantize = fuseElementwise(context, this, fakeQuantize);
         wasHandled = wasHandled || (fakeQuantize != nullptr);
     } while (fakeQuantize != nullptr);
 
@@ -127,7 +127,7 @@ bool FakeQuantizeTransformation::checkElementwise(const std::shared_ptr<Node>& e
 
 std::shared_ptr<opset1::FakeQuantize> FakeQuantizeTransformation::fuseElementwise(
     TransformationContext& context,
-    MatcherPass& matcherPass,
+    MatcherPass* matcherPass,
     const std::shared_ptr<opset1::FakeQuantize>& fakeQuantize) const {
     const std::shared_ptr<Node> eltwise = fakeQuantize->get_input_node_shared_ptr(0);
 
@@ -198,7 +198,7 @@ std::shared_ptr<opset1::FakeQuantize> FakeQuantizeTransformation::fuseElementwis
         foldConvert(fakeQuantize->input_value(3), deqPrecision),
         foldConvert(fakeQuantize->input_value(4), deqPrecision) }));
 
-    matcherPass.register_new_node(newFakeQuantize);
+    matcherPass->register_new_node(newFakeQuantize);
 
     replace_node(fakeQuantize, newFakeQuantize);
     ngraph::copy_runtime_info({ fakeQuantize, eltwise }, newFakeQuantize);

diff --git a/inference-engine/src/low_precision_transformations/src/fake_quantize_decomposition.cpp b/inference-engine/src/low_precision_transformations/src/fake_quantize_decomposition.cpp
@@ -157,6 +157,7 @@ DataPrecision getDataPrecisionByOutputPort(std::shared_ptr<opset1::FakeQuantize>
 
 // TODO: LPT: refactor: use one way to decompose FakeQuantize
 std::shared_ptr<ngraph::Node> decomposeFakeQuantize(
+    MatcherPass* matcherPass,
     std::shared_ptr<opset1::FakeQuantize>& layer,
     const std::shared_ptr<IntervalsAlignmentAttribute>& intervalsAlignment,
     const DataPrecision& dataPrecision,
@@ -196,6 +197,7 @@ std::shared_ptr<ngraph::Node> decomposeFakeQuantize(
             roundf(updatedOutputLowValue),
             roundf(updatedOutputHighValue),
             false);
+        matcherPass->register_new_node(newFakeQuantizeLayer);
         newFakeQuantizeLayer->set_levels(levels);
 
         auto dequantization = ngraph::pass::low_precision::NetworkHelper::makeDequantization(
@@ -232,6 +234,7 @@ std::shared_ptr<ngraph::Node> decomposeFakeQuantize(
             dataPrecision.hasZeroPoint,
             updatePrecisions);
 
+        matcherPass->register_new_node(std::get<0>(QDQ));
         dequantize = std::get<1>(QDQ);
     }
 
@@ -375,6 +378,7 @@ bool FakeQuantizeDecompositionTransformation::transform(TransformationContext& c
     }
 
     std::shared_ptr<ngraph::Node> dequantize = fq_decomposition::decomposeFakeQuantize(
+        this,
         layer,
         intervalsAlignment,
         dataPrecision,