Removed unnecessary transposes

mryzhov · Mar 17, 2023 · f1b4990 · f1b4990
1 parent e57f26c
commit f1b4990
Show file tree

Hide file tree

Showing 4 changed files with 58 additions and 30 deletions.
diff --git a/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp b/src/plugins/intel_gna/src/transformations/gather_sinking_transpose_reshape.cpp
@@ -194,6 +194,13 @@ GatherSinkingTransposeReshapeForward::GatherSinkingTransposeReshapeForward() {
         auto transpose_const = as_type_ptr<Constant>(pattern_to_output.at(transpose_const_label).get_node_shared_ptr());
         auto reshape = pattern_to_output.at(reshape_label).get_node_shared_ptr();
 
+        const ov::Shape reshape_shape = pass::helper::SqueezeShape(reshape->get_shape());
+        const ov::Shape transpose_shape = pass::helper::SqueezeShape(transpose->get_shape());
+        if (reshape_shape == transpose_shape) {
+            pass::helper::RemoveSingleInputNodeFromFunction(transpose);
+            return true;
+        }
+
         const NodePair new_nodes = SinkForward(transpose, transpose_const, reshape);
 
         register_new_node(new_nodes.first);
@@ -220,6 +227,13 @@ GatherSinkingTransposeReshapeBackward::GatherSinkingTransposeReshapeBackward() {
         auto transpose_const = as_type_ptr<Constant>(pattern_to_output.at(transpose_const_label).get_node_shared_ptr());
         auto reshape = pattern_to_output.at(reshape_label).get_node_shared_ptr();
 
+        const ov::Shape reshape_shape = pass::helper::SqueezeShape(reshape->get_shape());
+        const ov::Shape transpose_shape = pass::helper::SqueezeShape(transpose->get_shape());
+        if (reshape_shape == transpose_shape) {
+            pass::helper::RemoveSingleInputNodeFromFunction(transpose);
+            return true;
+        }
+
         const NodePair new_nodes = SinkBackward(transpose, transpose_const, reshape);
         register_new_node(new_nodes.first);
         register_new_node(new_nodes.second);

diff --git a/src/plugins/intel_gna/src/transformations/remove_pre_post_processing.cpp b/src/plugins/intel_gna/src/transformations/remove_pre_post_processing.cpp
@@ -3,6 +3,7 @@
 //
 
 #include "transformations/remove_pre_post_processing.hpp"
+#include "transformations/utils/transformation_helper.hpp"
 
 #include <openvino/cc/ngraph/itt.hpp>
 #include <openvino/opsets/opset1.hpp>
@@ -19,14 +20,6 @@ using namespace ov::intel_gna::pass;
 
 namespace {
 
-ov::Shape SqueezeShape(const ov::Shape& shape) {
-    ov::Shape squeezed_shape;
-    std::copy_if(shape.begin(), shape.end(), std::back_inserter(squeezed_shape), [](size_t x) {
-        return x != 1;
-    });
-    return squeezed_shape;
-}
-
 bool IsPreprocessingLayerSuppported(std::shared_ptr<ngraph::Node>& layer) {
     // Gather layers are not supported by GNA and have to be executed on CPU
     if (std::dynamic_pointer_cast<ov::opset1::Gather>(layer) ||
@@ -37,7 +30,7 @@ bool IsPreprocessingLayerSuppported(std::shared_ptr<ngraph::Node>& layer) {
 
     // 2-d Transposes layers can be executed on GNA
     if (std::dynamic_pointer_cast<ov::opset1::Transpose>(layer)) {
-        const ov::Shape squeezed_shape = SqueezeShape(layer->get_shape());
+        const ov::Shape squeezed_shape = pass::helper::SqueezeShape(layer->get_shape());
         const size_t min_input_dim = std::min(squeezed_shape[0], squeezed_shape[1]);
         const size_t max_input_dim = std::max(squeezed_shape[0], squeezed_shape[1]);
 
@@ -56,24 +49,6 @@ bool IsPreprocessingLayerSuppported(std::shared_ptr<ngraph::Node>& layer) {
 
     return false;
 }
-
-/*
-  works only if we have one date input and one output
- */
-void RemoveSingleInputNodeFromFunction(std::shared_ptr<ov::Node> node) {
-    const ov::Shape input_node_shape = node->get_input_shape(0);
-    const ov::Shape output_node_shape = node->get_output_shape(0);
-
-    std::shared_ptr<ov::Node> node_parent = node->get_input_node_shared_ptr(0);
-    if (!std::equal(input_node_shape.begin(), input_node_shape.end(), output_node_shape.begin())) {
-        auto reshape_const_node =
-            std::make_shared<Constant>(ov::element::i64, ov::Shape{output_node_shape.size()}, output_node_shape);
-        node_parent = std::make_shared<Reshape>(node_parent, reshape_const_node, false);
-    }
-
-    ov::replace_output_update_name(node->output(0), node_parent->output(0));
-}
-
 /*
   Support only one data node as 0 input
  */
@@ -110,7 +85,7 @@ bool RemoveInputsProcessing::run_on_model(const std::shared_ptr<ov::Model>& mode
                     m_subgraph_cpu_map->emplace(param_node.get_node_shared_ptr()->get_friendly_name(),
                                                 CopySingleInputNodeFromFunction(target_node));
                 }
-                RemoveSingleInputNodeFromFunction(target_node);
+                pass::helper::RemoveSingleInputNodeFromFunction(target_node);
                 result = true;
             }
         }
@@ -130,7 +105,7 @@ bool RemoveOutputsProcessing::run_on_model(const std::shared_ptr<ov::Model>& mod
                     m_subgraph_cpu_map->emplace(r_input_node->get_friendly_name(),
                                                 CopySingleInputNodeFromFunction(r_input_node));
                 }
-                RemoveSingleInputNodeFromFunction(r_input_node);
+                pass::helper::RemoveSingleInputNodeFromFunction(r_input_node);
                 result = true;
             }
         }

diff --git a/src/plugins/intel_gna/src/transformations/utils/transformation_helper.cpp b/src/plugins/intel_gna/src/transformations/utils/transformation_helper.cpp
@@ -4,12 +4,14 @@
 
 #include "transformation_helper.hpp"
 
-#include <ngraph/opsets/opset7.hpp>
+#include "openvino/opsets/opset7.hpp"
 #include <ngraph/pattern/op/wrap_type.hpp>
 #include <ngraph/rt_info.hpp>
 #include "ops/gna_convolution.hpp"
 #include "ops/gna_max_pool.hpp"
 
+using namespace ov::opset7;
+
 namespace ov {
 namespace intel_gna {
 namespace pass {
@@ -105,6 +107,29 @@ std::shared_ptr<ngraph::Node> InsertFQLayer(const std::shared_ptr<ngraph::opset7
     }
     return last_node;
 }
+
+void RemoveSingleInputNodeFromFunction(std::shared_ptr<ov::Node> node) {
+    const ov::Shape input_node_shape = node->get_input_shape(0);
+    const ov::Shape output_node_shape = node->get_output_shape(0);
+
+    std::shared_ptr<ov::Node> node_parent = node->get_input_node_shared_ptr(0);
+    if (!std::equal(input_node_shape.begin(), input_node_shape.end(), output_node_shape.begin())) {
+        auto reshape_const_node =
+            std::make_shared<Constant>(ov::element::i64, ov::Shape{output_node_shape.size()}, output_node_shape);
+        node_parent = std::make_shared<Reshape>(node_parent, reshape_const_node, false);
+    }
+
+    ov::replace_output_update_name(node->output(0), node_parent->output(0));
+}
+
+ov::Shape SqueezeShape(const ov::Shape& shape) {
+    ov::Shape squeezed_shape;
+    std::copy_if(shape.begin(), shape.end(), std::back_inserter(squeezed_shape), [](size_t x) {
+        return x != 1;
+    });
+    return squeezed_shape;
+}
+
 }  // namespace helper
 }  // namespace pass
 }  // namespace intel_gna

diff --git a/src/plugins/intel_gna/src/transformations/utils/transformation_helper.hpp b/src/plugins/intel_gna/src/transformations/utils/transformation_helper.hpp
@@ -110,6 +110,20 @@ std::shared_ptr<ngraph::Node> VerifyBiasGetConst(std::shared_ptr<ngraph::Node> c
 std::shared_ptr<ngraph::Node> InsertFQLayer(const std::shared_ptr<ngraph::opset7::FakeQuantize> fq_layer,
                                             std::shared_ptr<ngraph::Node> last_node);
 
+/**
+ * @brief removes single node from the function and insert Reshape if input and outpur shapes are different
+ * @param node the node to be deleted
+ * @return void
+ */
+void RemoveSingleInputNodeFromFunction(std::shared_ptr<ov::Node> node);
+
+/**
+ * @brief remove all 1 dimentions from the shape vector
+ * @param shape original tensor shape
+ * @return shape without 1 dimentions
+ */
+ov::Shape SqueezeShape(const ov::Shape& shape);
+
 }  // namespace helper
 }  // namespace pass
 }  // namespace intel_gna