[CPU] MVN nspc kernel optimization (#13268)

src/plugins/intel_cpu/src/nodes/mvn.h

@@ -14,8 +14,14 @@ namespace ov {
 namespace intel_cpu {
 namespace node {
 
+enum MVNLayoutType {
+    mvn_planar,
+    mvn_block,
+    mvn_by_channel
+};
+
 struct jit_mvn_config_params {
-    bool planar_layout;
+    MVNLayoutType layout;
     bool across_channels;
     bool normalize_variance;
     InferenceEngine::Precision src_prc;
@@ -69,6 +75,7 @@ struct jit_uni_mvn_kernel {
 
     jit_mvn_config_params jcp_;
     const dnnl_primitive_attr &attr_;
+    int optimized_scaleshift_num = 0;
 };
 
 class MVN : public Node {
@@ -102,11 +109,6 @@ public:
         OUTSIDE_SQRT
     };
 
-    enum MVNLayoutType {
-        planar,
-        block,
-        by_channel
-    };
     struct MVNAttrs {
         MVNLayoutType layout;
         std::tuple<size_t, size_t, size_t, size_t, size_t> shape5D;
@@ -152,6 +154,7 @@ private:
         private:
             void mvn_pln(const uint8_t *in_ptr_, uint8_t *out_ptr_, const void *post_ops_data_);
             void mvn_blk(const uint8_t *in_ptr_, uint8_t *out_ptr_, const void *post_ops_data_);
+            void mvn_nspc(const uint8_t *in_ptr_, uint8_t *out_ptr_, const void *post_ops_data_);
 
             std::shared_ptr<jit_uni_mvn_mean_variance_kernel> mvn_mean_kernel;
             std::shared_ptr<jit_uni_mvn_mean_variance_kernel> mvn_variance_kernel;

src/plugins/intel_cpu/tests/functional/single_layer_tests/mvn.cpp

@@ -433,13 +433,17 @@ const std::vector<ov::Shape> inputShapesStatic_3D = {
 const std::vector<ov::Shape> inputShapesStatic_4D = {
         {1, 7, 3, 5},
         {1, 15, 9, 5},
-        {4, 41, 6, 9}
+        {4, 41, 6, 9},
+        // cover channel case 4*16*2+16+3=147
+        {1, 147, 2, 2}
 };
 
 const std::vector<ov::Shape> inputShapesStatic_5D = {
         {1, 32, 8, 1, 6},
         {1, 9, 1, 15, 9},
-        {6, 64, 6, 1, 18}
+        {6, 64, 6, 1, 18},
+        // cover channel case 4*16*2+16+9=153
+        {6, 153, 2, 2, 2}
 };
 
 const auto Mvn2DStatic = ::testing::Combine(