[GPU] Fix mvn unit test fails (openvinotoolkit#28372)

### Details:
- *When shape flattening is required, change blocked format to plain
format if reduction axis is not blocked axis.*

### Tickets:
 - *158306*

Author: jade-cho

src/plugins/intel_gpu/src/graph/graph_optimizer/add_required_reorders.cpp

@@ -246,18 +246,42 @@ void add_required_reorders::run(program& p) {
         }
 
         // Remove padded-inputs in spatial axes not to use ref kernel which causes huge perf drop
-        if (usr->is_type<mvn>() && usr->as<mvn>().input().is_padded_spatial()) {
-            auto out_layout = usr->get_output_layout();
-            // Check formats of implemented opt kernels without a spatial padding support
-            if (out_layout.format == format::b_fs_yx_fsv16 || out_layout.format == format::b_fs_zyx_fsv16 ||
-                out_layout.format == format::bs_fs_yx_bsv32_fsv16 || out_layout.format == format::bs_fs_yx_bsv32_fsv32) {
-                auto& dep = usr->as<mvn>().input();
-                cldnn::layout layout_wo_padding = dep.get_output_layout();
-                layout_wo_padding.data_padding = cldnn::padding{};
-                auto new_reorder = std::make_shared<reorder>(dep.id() + "_no_pad_reorder", dep.id(), layout_wo_padding);
-                auto& new_reorder_node = p.get_or_create(new_reorder);
-                p.add_intermediate(new_reorder_node, *usr, dep);
-                new_reorder_node.recalc_output_layout(false);
+        if (usr->is_type<mvn>()) {
+            if (usr->as<mvn>().input().is_padded_spatial()) {
+                auto out_layout = usr->get_output_layout();
+                // Check formats of implemented opt kernels without a spatial padding support
+                if (out_layout.format == format::b_fs_yx_fsv16 || out_layout.format == format::b_fs_zyx_fsv16 ||
+                    out_layout.format == format::bs_fs_yx_bsv32_fsv16 || out_layout.format == format::bs_fs_yx_bsv32_fsv32) {
+                    auto& dep = usr->as<mvn>().input();
+                    cldnn::layout layout_wo_padding = dep.get_output_layout();
+                    layout_wo_padding.data_padding = cldnn::padding{};
+                    auto new_reorder = std::make_shared<reorder>(dep.id() + "_no_pad_reorder", dep.id(), layout_wo_padding);
+                    auto& new_reorder_node = p.get_or_create(new_reorder);
+                    p.add_intermediate(new_reorder_node, *usr, dep);
+                    new_reorder_node.recalc_output_layout(false);
+                }
+            }
+
+            auto input_layout = usr->get_input_layout();
+            auto input_pshape = input_layout.get_partial_shape();
+            auto prim = usr->as<mvn>().get_primitive();
+
+            if (prim->requires_alignment(input_pshape)) {
+                auto block_sizes = format::block_sizes(input_layout.format);
+                auto axes = prim->reduction_axes;
+                if (input_layout.is_dynamic() || block_sizes.size() > 1
+                    || (block_sizes.size() == 1 &&
+                        input_pshape[block_sizes[0].first].get_length() % block_sizes[0].second != 0 &&
+                        std::count(axes.begin(), axes.end(), block_sizes[0].first) == 0)) {
+                    auto rank = input_pshape.size();
+                    input_layout.format = format::get_default_format(rank);
+                    auto& dep = usr->as<mvn>().input();
+                    auto new_reorder = std::make_shared<reorder>(dep.id() + "_to_plain", dep.id(), input_layout);
+                    auto& new_reorder_node = p.get_or_create(new_reorder);
+                    p.add_intermediate(new_reorder_node, *usr, dep);
+                    // Need to invalidate users because the output format of mvn follows input format.
+                    new_reorder_node.recalc_output_layout(true);
+                }
             }
         }
 

src/plugins/intel_gpu/src/graph/impls/ocl/mvn.cpp

@@ -63,7 +63,18 @@ struct mvn_impl : typed_primitive_impl_ocl<mvn> {
             ov::PartialShape shape = ov::PartialShape::dynamic(new_rank);
 
             auto& output_layout = updated_impl_params.output_layouts[0];
+
             if (input_pshape.is_static()) {
+                size_t flatten_axis = 0;
+                // Change flatten axis if the format is single fsv.
+                auto block_sizes = format::block_sizes(input_layout.format);
+                if (block_sizes.size() == 1
+                    && (input_pshape[block_sizes[0].first].get_length() % block_sizes[0].second == 0)
+                    && (std::count(axes.begin(), axes.end(), block_sizes[0].first) == 0)
+                    && block_sizes[0].first == 1) {
+                    flatten_axis = 1;
+                }
+
                 for (size_t i = 0; i < new_rank; i++) {
                     shape[i] = 1;
                 }
@@ -72,7 +83,7 @@ struct mvn_impl : typed_primitive_impl_ocl<mvn> {
                 // 1. normalized dimensions which are flattened and written to the last dim
                 // 2. not normalized dims which are flattened and written to the first dim
                 for (size_t i = 0; i < input_rank; i++) {
-                    shape[static_cast<int64_t>(i) < min ? 0 : (new_rank - 1)] *= input_pshape[i];
+                    shape[static_cast<int64_t>(i) < min ? flatten_axis : (new_rank - 1)] *= input_pshape[i];
                 }
             }
 

src/plugins/intel_gpu/tests/unit/fusions/mvn_fusion_test.cpp

@@ -80,6 +80,7 @@ class MVNFusingTest : public ::BaseFusingTest<mvn_test_params> {
 #define CASE_MVN_I8_6       { 2, 16, 8, 8 },    { 1, 1, 1, 1 },     data_types::i8, format::b_fs_yx_fsv16, {1, 2, 3}, true, data_types::f32, format::bfyx
 #define CASE_MVN_I8_7       { 2, 16, 1, 8 },    { 1, 1, 8, 1 },     data_types::i8, format::b_fs_yx_fsv16, {1, 2, 3}, true, data_types::f32, format::bfyx
 #define CASE_MVN_I8_8       { 2, 16, 3, 8 },    { 1, 1, 3, 8 },     data_types::i8, format::b_fs_yx_fsv16, {3}, true, data_types::f32, format::bfyx
+#define CASE_MVN_I8_8_NA    { 2, 15, 3, 8 },    { 1, 1, 3, 8 },     data_types::i8, format::b_fs_yx_fsv16, {3}, true, data_types::f32, format::bfyx
 #define CASE_MVN_3D_I8_1    { 1, 16, 8, 8, 8 }, { 1, 16, 8, 8, 8 }, data_types::i8, format::bfzyx, {2, 3, 4}, true, data_types::f32, format::bfzyx
 #define CASE_MVN_3D_I8_2    { 2, 16, 8, 8, 8 }, { 2, 16, 8, 8, 8 }, data_types::i8, format::bfzyx, {1, 2, 3, 4}, true, data_types::f32, format::bfzyx
 #define CASE_MVN_3D_I8_3    { 2, 16, 8, 8, 8 }, { 2, 1, 8, 8, 1 },  data_types::i8, format::bfzyx, {1, 2, 3, 4}, true, data_types::f32, format::bfzyx
@@ -170,7 +171,8 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, mvn_scale_quantize_i8, ::testing::ValuesIn
     mvn_test_params{ CASE_MVN_I8_2, 2, 2, 4 },
     mvn_test_params{ CASE_MVN_I8_3, 2, 2, 4 },
     mvn_test_params{ CASE_MVN_I8_4, 2, 2, 4 },
-    // mvn_test_params{ CASE_MVN_I8_8, 3, 3, 4 },   // TODO: It will be fix soon, test reference is wrong in new driver.
+    mvn_test_params{ CASE_MVN_I8_8, 3, 3, 4 },
+    mvn_test_params{ CASE_MVN_I8_8_NA, 3, 3, 4 },
     mvn_test_params{ CASE_MVN_3D_I8_1, 2, 2, 4 },
     mvn_test_params{ CASE_MVN_3D_I8_2, 2, 2, 4 },
     mvn_test_params{ CASE_MVN_U8_1, 2, 2, 4 },
@@ -221,7 +223,7 @@ INSTANTIATE_TEST_SUITE_P(fusings_gpu, mvn_scale_activation_eltwise_fp32_quantize
     mvn_test_params{ CASE_MVN_I8_5, 2, 4, 6 },
     mvn_test_params{ CASE_MVN_I8_6, 2, 4, 6 },
     mvn_test_params{ CASE_MVN_I8_7, 3, 4, 6 },
-    // mvn_test_params{ CASE_MVN_I8_8, 3, 5, 6 },   // TODO: It will be fix soon, test reference is wrong in new driver.
+    mvn_test_params{ CASE_MVN_I8_8, 3, 5, 6 },
     mvn_test_params{ CASE_MVN_3D_I8_1, 2, 4, 6 },
     mvn_test_params{ CASE_MVN_3D_I8_2, 2, 4, 6 },
     mvn_test_params{ CASE_MVN_3D_I8_3, 2, 4, 6 },