Merge branch 'master' into dependencies-update

ilya-lavrenov · Aug 12, 2024 · 3d1b5f3 · 3d1b5f3
2 parents 49521bc + e567c9e
commit 3d1b5f3
Show file tree

Hide file tree

Showing 29 changed files with 1,803 additions and 982 deletions.
diff --git a/src/plugins/intel_cpu/src/nodes/interpolate.cpp b/src/plugins/intel_cpu/src/nodes/interpolate.cpp
@@ -2021,7 +2021,13 @@ void Interpolate::initSupportedPrimitiveDescriptors() {
         return;
 
     ov::element::Type inputPrecision = getOriginalInputPrecisionAtPort(DATA_ID);
-    if ((inputPrecision != ov::element::i8) && (inputPrecision != ov::element::u8) && (inputPrecision != ov::element::bf16)) {
+
+#if defined(OV_CPU_WITH_ACL)
+    bool isInputPrecisionSupported = one_of(inputPrecision, ov::element::i8, ov::element::u8, ov::element::f16);
+#else
+    bool isInputPrecisionSupported = one_of(inputPrecision, ov::element::i8, ov::element::u8, ov::element::bf16);
+#endif
+    if (!isInputPrecisionSupported) {
         inputPrecision = ov::element::f32;
     }
 
@@ -2039,9 +2045,11 @@ void Interpolate::initSupportedPrimitiveDescriptors() {
         outputPrecision = fusedWith[fusedWith.size() - 1]->getOriginalOutputPrecisionAtPort(DATA_ID);
     }
 
+#if !defined(OV_CPU_WITH_ACL)
     if (!mayiuse(cpu::x64::sse41)) {
         inputPrecision = outputPrecision = ov::element::f32;
     }
+#endif
 
     auto targetShapeType = ov::element::i32;
     auto scalesType = ov::element::f32;
@@ -2115,6 +2123,8 @@ void Interpolate::initSupportedPrimitiveDescriptors() {
         canUseAclExecutor = !supportedPrimitiveDescriptors.empty();
         if (canUseAclExecutor)
             return;
+        //fallback to f32 if ref is used
+        inputPrecision = outputPrecision = ov::element::f32;
 #endif
 
         if (dataRank == 4) {
@@ -2147,6 +2157,8 @@ void Interpolate::initSupportedPrimitiveDescriptors() {
         canUseAclExecutor = !supportedPrimitiveDescriptors.empty();
         if (canUseAclExecutor)
             return;
+        //fallback to f32 if ref is used
+        inputPrecision = outputPrecision = ov::element::f32;
 #endif
 
         if (!mayiuse(cpu::x64::sse41) || interpAttrs.mode == InterpolateMode::linear) {

diff --git a/...unctional/shared_tests_instances/execution_graph_tests/duplicate_inputs_outputs_names.cpp b/...unctional/shared_tests_instances/execution_graph_tests/duplicate_inputs_outputs_names.cpp
@@ -0,0 +1,18 @@
+// Copyright (C) 2018-2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "execution_graph_tests/duplicate_inputs_outputs_names.hpp"
+
+#include "common_test_utils/test_constants.hpp"
+
+using namespace ExecutionGraphTests;
+
+namespace {
+
+INSTANTIATE_TEST_SUITE_P(smoke_duplicateInputsOutputsNames,
+                         ExecGraphDuplicateInputsOutputsNames,
+                         ::testing::Values(ov::test::utils::DEVICE_CPU),
+                         ExecGraphDuplicateInputsOutputsNames::getTestCaseName);
+
+}  // namespace
diff --git a/src/plugins/intel_gpu/src/graph/program_node.cpp b/src/plugins/intel_gpu/src/graph/program_node.cpp
@@ -988,6 +988,11 @@ void program_node::load(cldnn::BinaryInputBuffer& ib) {
         ib >> fused_prims_onednn[idx].dims;
         ib >> make_data(&fused_prims_onednn[idx].dt, sizeof(dnnl::memory::data_type));
     }
+
+    // added a dummpy onednn_attrs to prevent initializing it for non-onednn impls
+    if (impl_type != impl_types::onednn) {
+        onednn_attrs = std::make_shared<dnnl::primitive_attr>();
+    }
 #endif // ENABLE_ONEDNN_FOR_GPU
 }
 

diff --git a/src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp b/src/plugins/intel_gpu/src/plugin/transformations/convert_matmul_to_fc.cpp
@@ -41,10 +41,24 @@ ConvertMatMulToFullyConnected::ConvertMatMulToFullyConnected() {
             return false;
         }
 
-        // fc_input_a and fc_input_b - are the final inputs that will be set to FullyConnected.
-        // So in case of adding new operations that takes matmul inputs we need keep update fc_input_a and fc_input_b.
         auto fc_input_a = pattern_map.at(activations_m);
         auto fc_input_b = pattern_map.at(weights_m);
+
+        // Not to convert fc_input_b which should be fc_input_a of other sibling MatMul
+        auto input_b = fc_input_b.get_node_shared_ptr();
+        for (auto& user : input_b->get_users()) {
+            if (user != matmul && ov::is_type<ov::op::v0::MatMul>(user) && ov::is_type<ov::op::v0::Convert>(input_b)) {
+                auto other_matmul = std::dynamic_pointer_cast<ov::op::v0::MatMul>(user);
+                // Transpose for input_b generates invalid input for other sibling matmul
+                if (input_b == other_matmul->get_input_node_shared_ptr(0) || fc_input_b == fc_input_a ||
+                    (input_b == other_matmul->get_input_node_shared_ptr(1) && matmul->get_transpose_b() != other_matmul->get_transpose_b())) {
+                    return false;
+                }
+            }
+        }
+
+        // fc_input_a and fc_input_b - are the final inputs that will be set to FullyConnected.
+        // So in case of adding new operations that takes matmul inputs we need keep update fc_input_a and fc_input_b.
         bool is_convert = false;
         if (auto convert_node = std::dynamic_pointer_cast<ov::op::v0::Convert>(fc_input_b.get_node_shared_ptr())) {
             is_convert = true;