Updated ConvertPrecision transformation to be executed for TI Body (#1673)

* Updated ConvertPrecision transformation to be executed for TI Body

* Added type fusion for GenericIE operation

* Added test for TensorIterator body precision conversion

diff --git a/inference-engine/src/mkldnn_plugin/mkldnn_memory.cpp b/inference-engine/src/mkldnn_plugin/mkldnn_memory.cpp
index e088146e8fd61ea616547724a5b5e9d5abfd1eb7..8d9985b615e5a11a0c6c184182cab3cbe4b46c29 100644 (file)
--- a/inference-engine/src/mkldnn_plugin/mkldnn_memory.cpp
+++ b/inference-engine/src/mkldnn_plugin/mkldnn_memory.cpp
@@ -1073,7 +1073,7 @@ MKLDNNMemoryDesc::MKLDNNMemoryDesc(const TensorDesc& tDesc):
             data_type = mkldnn::memory::data_type::bf16;
             break;
         default:
-            THROW_IE_EXCEPTION << "Cannot create MKLDNNMemoryDesc from TensorDesc. Unsupported precision!";
+            THROW_IE_EXCEPTION << "Cannot create MKLDNNMemoryDesc from TensorDesc. Unsupported precision: " << tDesc.getPrecision();
     }
 
     mkldnn::memory::format mkldnnFormat = memory::format::format_undef;

diff --git a/inference-engine/src/transformations/src/transformations/convert_precision.cpp b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
index 912bf4edb4d60e2658f9cf2ac8da061d715e5f96..ffd9956b10eee1deb3b50f56b5d2669727c18618 100644 (file)
--- a/inference-engine/src/transformations/src/transformations/convert_precision.cpp
+++ b/inference-engine/src/transformations/src/transformations/convert_precision.cpp
@@ -22,6 +22,7 @@ bool fuse_type_to_nms4(std::shared_ptr<ngraph::Node> & node, ngraph::element::Ty
 bool fuse_type_to_topk(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_nonzero(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 bool fuse_type_to_bucketize(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
+bool fuse_type_to_generic_ie(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx);
 
 static std::map<ngraph::NodeTypeInfo, std::function<bool(std::shared_ptr<Node>&, element::Type, size_t idx)>> type_to_fuse {
         {opset4::Parameter::type_info, fuse_type_to_parameter},
@@ -32,6 +33,7 @@ static std::map<ngraph::NodeTypeInfo, std::function<bool(std::shared_ptr<Node>&,
         {opset4::TopK::type_info, fuse_type_to_topk},
         {opset4::NonZero::type_info, fuse_type_to_nonzero},
         {opset4::Bucketize::type_info, fuse_type_to_bucketize},
+        {NodeTypeInfo("GenericIE", 1), fuse_type_to_generic_ie},
 };
 
 bool ngraph::pass::ConvertPrecision::run_on_function(std::shared_ptr<ngraph::Function> f) {
@@ -41,8 +43,12 @@ bool ngraph::pass::ConvertPrecision::run_on_function(std::shared_ptr<ngraph::Fun
     std::map<std::shared_ptr<Node>, std::vector<Input<Node>>> const_to_internal_output;
 
     std::function<void(const std::shared_ptr<Function> &)> register_constants =
-            [&const_to_internal_output](const std::shared_ptr<Function> & f) {
+            [&const_to_internal_output, &register_constants](const std::shared_ptr<Function> & f) {
         for (auto & node : f->get_ordered_ops()) {
+            // Recursively run for TensorIterator body function
+            if (auto ti = std::dynamic_pointer_cast<opset4::TensorIterator>(node)) {
+                register_constants(ti->get_body()->to_function());
+            }
             for (auto & input : node->inputs()) {
                 if (auto const_node = std::dynamic_pointer_cast<opset4::Constant>(input.get_source_output().get_node_shared_ptr())) {
                     const_to_internal_output[const_node].emplace_back(input);
@@ -83,11 +89,15 @@ bool ngraph::pass::ConvertPrecision::run_on_function(std::shared_ptr<ngraph::Fun
     };
 
     std::function<void(const std::shared_ptr<Function> &)> convert_function_precision =
-            [this, &const_to_internal_output, &convert_node_precision](const std::shared_ptr<Function> & f) {
+            [this, &const_to_internal_output, &convert_node_precision, &convert_function_precision](const std::shared_ptr<Function> & f) {
         // Iterate over all nodes in topological order and then iterate over node outputs.
         // If output type mismatch given type we try to fuse type into this operation
         // otherwise we insert Convert operation.
         for (auto &node : f->get_ordered_ops()) {
+            // Recursively run for TensorIterator body function
+            if (auto ti = std::dynamic_pointer_cast<opset4::TensorIterator>(node)) {
+                convert_function_precision(ti->get_body()->to_function());
+            }
             convert_node_precision(node);
         }
     };
@@ -177,6 +187,11 @@ bool fuse_type_to_bucketize(std::shared_ptr<ngraph::Node> & node, ngraph::elemen
     return false;
 }
 
+bool fuse_type_to_generic_ie(std::shared_ptr<ngraph::Node> & node, ngraph::element::Type to, size_t idx) {
+    node->set_output_type(idx, to, node->output(idx).get_partial_shape());
+    return true;
+}
+
 template <element::Type_t PREC_FROM, element::Type_t PREC_TO>
 std::shared_ptr<Node> change_constant_precision(std::shared_ptr<opset4::Constant> & constant) {
     using src_type = typename element_type_traits<PREC_FROM>::value_type;

diff --git a/inference-engine/tests/functional/inference_engine/transformations/convert_precision.cpp b/inference-engine/tests/functional/inference_engine/transformations/convert_precision.cpp
index 95b0d6ba0ee9d8eaac2a9bfe8681216c2ff06a2d..70b8a5a684a92ef6cd72ba3cffcc0e64764cd649 100644 (file)
--- a/inference-engine/tests/functional/inference_engine/transformations/convert_precision.cpp
+++ b/inference-engine/tests/functional/inference_engine/transformations/convert_precision.cpp
@@ -234,4 +234,57 @@ TEST(TransformationTests, ConvertPrecision_Roundings) {
 
     ASSERT_FALSE(has_type<ngraph::element::Type_t::f16>(f));
     ASSERT_FALSE(has_type<ngraph::element::Type_t::i64>(f));
+}
+
+TEST(TransformationTests, ConvertPrecision_TIBody) {
+    std::shared_ptr<ngraph::Function> f(nullptr), f_ref(nullptr);
+    {
+        auto X = std::make_shared<opset4::Parameter>(element::f16, Shape{2, 1, 16});
+        auto Y = std::make_shared<opset4::Parameter>(element::f16, Shape{1, 128});
+
+        auto Xi = std::make_shared<opset4::Parameter>(element::f16, Shape{1, 1, 16});
+        auto Yi = std::make_shared<opset4::Parameter>(element::f16, Shape{1, 128});
+
+        // Body
+        auto axis = ngraph::opset4::Constant::create(ngraph::element::i64, ngraph::Shape{}, {0});
+        auto squeeze = std::make_shared<opset4::Squeeze>(Xi, axis);
+
+        auto w_val = std::vector<float>(384 * 16, 0);
+        auto r_val = std::vector<float>(384 * 128, 0);
+        auto b_val = std::vector<float>(384, 0);
+        auto W = ngraph::opset4::Constant::create(ngraph::element::f16, ngraph::Shape{384, 16}, w_val);
+        auto R = ngraph::opset4::Constant::create(ngraph::element::f16, ngraph::Shape{384, 128}, r_val);
+        auto B = ngraph::opset4::Constant::create(ngraph::element::f16, ngraph::Shape{384}, b_val);
+
+        auto gru_cell = std::make_shared<opset4::GRUCell>(squeeze, Yi, W, R, B, 128);
+        auto res_1 = std::make_shared<opset4::Result>(gru_cell);
+        auto unsqueeze = std::make_shared<opset4::Unsqueeze>(gru_cell, axis);
+        auto res_2 = std::make_shared<opset4::Result>(unsqueeze);
+        auto body = std::make_shared<opset4::TensorIterator::BodyLambda>(OutputVector{res_1, res_2},
+                                                                         ParameterVector{Xi, Yi});
+
+        auto tensor_iterator = std::make_shared<opset4::TensorIterator>();
+        tensor_iterator->set_body(body);
+
+        tensor_iterator->set_sliced_input(Xi, X, 0, 1, 1, -1, 0);
+        tensor_iterator->set_merged_input(Yi, Y, res_1);
+
+        auto out0 = tensor_iterator->get_iter_value(res_1, -1);
+        auto out1 = tensor_iterator->get_concatenated_slices(res_2, 0, 1, 1, -1, 0);
+
+        auto res_ti_1 = std::make_shared<opset4::Result>(tensor_iterator->output(1));
+        //auto res_ti_2 = std::make_shared<opset4::Result>(tensor_iterator->output(0));
+        f = std::make_shared<ngraph::Function>(ngraph::NodeVector{res_ti_1},
+                                               ngraph::ParameterVector{X, Y});
+
+        ngraph::pass::Manager manager;
+        manager.register_pass<ngraph::pass::ConvertPrecision>(ngraph::element::i64, ngraph::element::i32);
+        manager.register_pass<ngraph::pass::ConvertPrecision>(ngraph::element::f16, ngraph::element::f32);
+        manager.run_passes(f);
+
+        ASSERT_FALSE(has_type<ngraph::element::Type_t::f16>(f));
+        ASSERT_FALSE(has_type<ngraph::element::Type_t::i64>(f));
+        ASSERT_FALSE(has_type<ngraph::element::Type_t::f16>(tensor_iterator->get_body()->to_function()));
+        ASSERT_FALSE(has_type<ngraph::element::Type_t::i64>(tensor_iterator->get_body()->to_function()));
+    }
 }
\ No newline at end of file