Fix Vnode build on ICE and RPL

src/plugins/intel_cpu/CMakeLists.txt

@@ -47,6 +47,11 @@ ie_option(ENABLE_MLAS_FOR_CPU "Enable MLAS for OpenVINO CPU Plugin" ${ENABLE_MLA
 
 add_subdirectory(thirdparty)
 
+ie_dependent_option(ENABLE_CPU_PROFILER "enable CPU profiler" OFF "X86 OR X86_64" OFF)
+if (ENABLE_CPU_PROFILER)
+    add_definitions(-DOV_CPU_WITH_PROFILER)
+endif()
+
 if(WIN32)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNOMINMAX")
 endif()
@@ -136,7 +141,39 @@ cross_compiled_file(${TARGET_NAME}
         NAME        proposal_exec
         NAMESPACE   InferenceEngine::Extensions::Cpu::XARCH
 )
+cross_compiled_file(${TARGET_NAME}
+        ARCH AVX512F AVX2 ANY
+                    src/nodes/vnode_attn_softmax.cpp
+        API         src/nodes/vnode_attn_softmax.hpp
+        NAME        attn_softmax
+        NAMESPACE   InferenceEngine::Extensions::Cpu::XARCH
+)
+cross_compiled_file(${TARGET_NAME}
+        ARCH AVX2 ANY
+                    src/nodes/vnode_rms_norm.cpp
+        API         src/nodes/vnode_rms_norm.hpp
+        NAME        rms_norm
+        NAMESPACE   InferenceEngine::Extensions::Cpu::XARCH
+)
 
+# for avx2 compilation
+set_property(SOURCE src/nodes/vnode.cpp APPEND_STRING PROPERTY COMPILE_OPTIONS
+        "-mavx2")
+set_property(SOURCE src/nodes/vnode.cpp APPEND_STRING PROPERTY COMPILE_OPTIONS
+        "-mfma")
+
+check_cxx_source_compiles("#include <immintrin.h>\nint main() {auto v_scale = _mm512_set1_ps(0.0f);return 0;}" ENABLE_AVX512)
+if(ENABLE_AVX512)
+set_property(SOURCE src/nodes/vnode.cpp APPEND_STRING PROPERTY COMPILE_OPTIONS
+        "-mavx512f")
+set_property(SOURCE src/nodes/vnode.cpp APPEND_STRING PROPERTY COMPILE_OPTIONS
+        "-mfma")
+endif()
+check_cxx_source_compiles("#include <immintrin.h>\nint main() { __m256bh _s; __m512bh _a, _b; _s = _mm512_cvtneps_pbh(_mm512_dpbf16_ps(_mm512_cvtpbh_ps(_s), _a, _b)); return 0; }" ENABLE_AVX512_BF16)
+if(ENABLE_AVX512_BF16)
+set_property(SOURCE src/nodes/vnode.cpp APPEND_STRING PROPERTY COMPILE_OPTIONS
+        "-mavx512bf16")
+endif()
 # system dependencies must go last
 target_link_libraries(${TARGET_NAME} PRIVATE openvino::pugixml)
 set_ie_threading_interface_for(${TARGET_NAME})

src/plugins/intel_cpu/docs/chrome_tracing_profiler.md

@@ -0,0 +1,6 @@
+# Profiler based on Chrome Tracing
+Chrome tracing is a profiling tool which is easily accessible at URL `chrome://tracing/` in Chrome browser. CPU plugin compiled with `-DENABLE_CPU_PROFILER=ON` has the capability of generating such json-format tracing logs that can be loaded into & viewed with this powerful tool. 
+
+It cannot replace ITT & Vtune based profiling, but it provides a very convenient & customizable alternative.
+
+set `OV_CPU_PROFILE` environment variable to `1` will enable the profiling and tracing log generation.

src/plugins/intel_cpu/docs/debug_capabilities/README.md

@@ -5,14 +5,16 @@ They can be activated at runtime and might be used for analyzing issues, getting
 
 Use the following cmake option to enable debug capabilities:
 
-`-DENABLE_DEBUG_CAPS=ON`
+`-DENABLE_DEBUG_CAPS=ON -DENABLE_CPU_DEBUG_CAPS=ON`
 
 * [Verbose mode](verbose.md)
 * [Blob dumping](blob_dumping.md)
 * [Graph serialization](graph_serialization.md)
 * [Graph transformation disabling](feature_disabling.md#graph-transformations)
 * [Logging](logging.md)
 * [Inference Precision](infer_prc.md)
+* Dump runtime models
+    * set `OV_CPU_DUMP_MODELS` environment variable to `1` will enable runtime models to be dumpped as cpp-style text file and IR xml file.
 * Performance summary
     * set `OV_CPU_SUMMARY_PERF` environment variable to display performance summary at the time when model is being destructed.
     * Internal performance counter will be enabled automatically. 

src/plugins/intel_cpu/src/cpu_types.cpp

@@ -203,7 +203,9 @@ const InferenceEngine::details::caseless_unordered_map<std::string, Type> type_t
         { "Interaction", Type::Interaction},
         { "MHA", Type::MHA},
         { "Unique", Type::Unique},
-        { "Ngram", Type::Ngram}
+        { "Ngram", Type::Ngram},
+        { "VNode", Type::VNode},
+        { "RoPE", Type::RoPE}
 };
 
 Type TypeFromName(const std::string& type) {
@@ -313,6 +315,8 @@ std::string NameFromType(const Type type) {
         CASE(MHA);
         CASE(Unique);
         CASE(Ngram);
+        CASE(VNode);
+        CASE(RoPE);
         CASE(Unknown);
     }
 #undef CASE

src/plugins/intel_cpu/src/cpu_types.h

@@ -111,7 +111,9 @@ enum class Type {
     Interaction,
     MHA,
     Unique,
-    Ngram
+    Ngram,
+    VNode,
+    RoPE,
 };
 
 enum class Algorithm {

src/plugins/intel_cpu/src/exec_network.cpp

@@ -185,6 +185,7 @@ ExecNetwork::GraphGuard::Lock ExecNetwork::GetGraph() const {
                     ctx = std::make_shared<GraphContext>(_cfg, extensionManager, weightsCache, isQuantizedFlag);
                 }
                 graphLock._graph.CreateGraph(_network, ctx);
+                DEBUG_DUMP_GRAPH(graphLock._graph, "cpu_Graph");
             } catch (...) {
                 exception = std::current_exception();
             }

src/plugins/intel_cpu/src/extension.cpp

@@ -8,8 +8,10 @@
 #include "transformations/cpu_opset/common/op/power_static.hpp"
 #include "transformations/cpu_opset/common/op/swish_cpu.hpp"
 #include "transformations/cpu_opset/common/op/ngram.hpp"
+#include "transformations/cpu_opset/common/op/vnode.hpp"
 #include "transformations/cpu_opset/x64/op/mha.hpp"
 #include "transformations/cpu_opset/x64/op/interaction.hpp"
+#include "transformations/cpu_opset/x64/op/rope.hpp"
 #include "transformations/snippets/x64/op/load_convert.hpp"
 #include "transformations/snippets/x64/op/store_convert.hpp"
 #include "transformations/snippets/x64/op/brgemm_cpu.hpp"
@@ -60,6 +62,8 @@ std::map<std::string, ngraph::OpSet> Extension::getOpSets() {
         NGRAPH_OP(NgramNode, ov::intel_cpu)
         NGRAPH_OP_X64(MHANode, ov::intel_cpu)
         NGRAPH_OP_X64(InteractionNode, ov::intel_cpu)
+        NGRAPH_OP_X64(VNode, ov::intel_cpu)
+        NGRAPH_OP_X64(RoPENode, ov::intel_cpu)
 #undef NGRAPH_OP
 
         return opset;

src/plugins/intel_cpu/src/graph.cpp

@@ -43,6 +43,7 @@
 #include "utils/ngraph_utils.hpp"
 #include "utils/cpu_utils.hpp"
 #include "utils/verbose.h"
+#include "utils/profiler.hpp"
 #include "memory_desc/cpu_memory_desc_utils.h"
 
 #include <openvino/core/model.hpp>
@@ -53,6 +54,7 @@
 #include "memory_desc/dnnl_blocked_memory_desc.h"
 #include <common/primitive_desc.hpp>
 #include <common/primitive_desc_iface.hpp>
+#include "transformations/cpu_opset/common/op/vnode.hpp"
 #if (OV_THREAD == OV_THREAD_TBB || OV_THREAD == OV_THREAD_TBB_AUTO)
 #   include <tbb/task.h>
 #endif
@@ -247,8 +249,13 @@ void Graph::Replicate(const CNNNetwork &network) {
 
         if (op->get_type_info() == op::v0::Result::get_type_info_static()) {
             const auto &input = op->input_value(0);
-            const auto name = op::util::get_ie_output_name(input);
-
+            std::string name;
+            // VNode may produces many results
+            if (auto vnode = ov::as_type_ptr<ov::intel_cpu::VNode>(input.get_node_shared_ptr())) {
+                name = vnode->get_output_name(input.get_index());
+            } else {
+                name = op::util::get_ie_output_name(input);
+            }
             if (outputsInfo.count(name) != 0) {
                 outputNodesMap[name] = node;
             }
@@ -647,6 +654,15 @@ void Graph::InitEdges() {
 
                     for (auto node : vecConsumers) {
                         if (node->getExecIndex() >= execIndex) {
+                            DEBUG_LOG(edge->name(),
+                                      " needReorder due to execIndex: ",
+                                      node->getName(),
+                                      "  ",
+                                      node->getExecIndex(),
+                                      ">",
+                                      modifyingNode->getName(),
+                                      " ",
+                                      execIndex);
                             return true;
                         }
                     }
@@ -1000,7 +1016,7 @@ bool Graph::ProcessDynNodes() {
 
 void Graph::PushInputData(const std::string& name, const InferenceEngine::Blob::Ptr &in) {
     if (!IsReady()) IE_THROW()<< "Wrong state. Topology not ready.";
-
+    PROFILE(_prof, "Graph::PushInputData");
     auto input = inputNodesMap.find(name);
     if (input != inputNodesMap.end()) {
         auto& inTensorDesc = in->getTensorDesc();
@@ -1035,6 +1051,7 @@ void Graph::PushInputData(const std::string& name, const InferenceEngine::Blob::
 
 // suppose always being shared infer_request intel_cpu::Tensor to Graph if isDynamic.
 void Graph::PullOutputData(BlobMap &out) {
+    PROFILE(_prof, "Graph::PullOutputData");
     if (!IsReady())
         IE_THROW() << "Wrong state. Topology not ready.";
 
@@ -1121,11 +1138,11 @@ void Graph::PullOutputData(BlobMap &out) {
 
 void Graph::InferStatic(InferRequestBase* request) {
     dnnl::stream stream(getEngine());
-
+    PROFILE(_prof0, std::string("Graph::InferStatic_#") + std::to_string(infer_count), {});
     for (const auto& node : executableGraphNodes) {
         VERBOSE(node, getConfig().debugCaps.verbose);
         PERF(node, getConfig().collectPerfCounters);
-
+        PROFILE(_prof1, node->getTypeStr(), node->getName());
         if (request)
             request->ThrowIfCanceled();
         ExecuteNode(node, stream);
@@ -1330,7 +1347,7 @@ class UpdateNodes : public UpdateNodesBase {
 
 void Graph::InferDynamic(InferRequestBase* request) {
     dnnl::stream stream(getEngine());
-
+    PROFILE(_prof0, std::string("Graph::InferDynamic_#") + std::to_string(infer_count));
     std::set<size_t> syncIndsWorkSet;
     for (const auto& nodeIndx : syncNodesInds) {
         syncIndsWorkSet.insert(nodeIndx.second);
@@ -1349,12 +1366,15 @@ void Graph::InferDynamic(InferRequestBase* request) {
     size_t inferCounter = 0;
 
     for (auto stopIndx : syncIndsWorkSet) {
-        updateNodes->run(stopIndx);
+        {
+            PROFILE(_prof, "updateNodes");
+            updateNodes->run(stopIndx);
+        }
         for (; inferCounter < stopIndx; ++inferCounter) {
             auto& node = executableGraphNodes[inferCounter];
             VERBOSE(node, getConfig().debugCaps.verbose);
             PERF(node, getConfig().collectPerfCounters);
-
+            PROFILE(_prof, node->getTypeStr(), node->getName());
             if (request)
                 request->ThrowIfCanceled();
             ExecuteNode(node, stream);
@@ -1387,7 +1407,7 @@ void Graph::Infer(InferRequestBase* request) {
         IE_THROW() << "Unknown ov::intel_cpu::Graph state: " << static_cast<size_t>(status);
     }
 
-    if (infer_count != -1) infer_count++;
+    infer_count++;
 }
 
 void Graph::VisitNode(NodePtr node, std::vector<NodePtr>& sortedNodes) {