diff --git a/filelist.json b/filelist.json
index 37c0f1d..d6b6f47 100644
--- a/filelist.json
+++ b/filelist.json
@@ -873,22 +873,24 @@
           "common": [
             "src/cpu/operators/CpuAdd.cpp",
             "src/cpu/kernels/CpuAddKernel.cpp",
-            "src/runtime/NEON/functions/NEArithmeticAddition.cpp",
-            "src/cpu/kernels/add/generic/neon/qasymm8.cpp",
-            "src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp",
-            "src/cpu/kernels/add/generic/neon/qsymm16.cpp"
+            "src/runtime/NEON/functions/NEArithmeticAddition.cpp"
           ],
           "neon": {
-            "common": [ "src/cpu/kernels/add/generic/neon/impl.cpp" ],
+            "common": ["src/cpu/kernels/add/generic/neon/impl.cpp"],
             "fp32":["src/cpu/kernels/add/generic/neon/fp32.cpp"],
             "fp16":["src/cpu/kernels/add/generic/neon/fp16.cpp"],
-            "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"]
+            "integer":["src/cpu/kernels/add/generic/neon/integer.cpp"],
+            "qasymm8": ["src/cpu/kernels/add/generic/neon/qasymm8.cpp"],
+            "qasymm8_signed": ["src/cpu/kernels/add/generic/neon/qasymm8_signed.cpp"],
+            "qsymm16": ["src/cpu/kernels/add/generic/neon/qsymm16.cpp"]
           },
           "sve": {
             "common": [ "src/cpu/kernels/add/generic/sve/impl.cpp" ],
-            "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"],
-            "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"],
             "integer":["src/cpu/kernels/add/generic/sve/integer.cpp"],
+            "fp32":["src/cpu/kernels/add/generic/sve/fp32.cpp"],
+            "fp16":["src/cpu/kernels/add/generic/sve/fp16.cpp"]
+          },
+          "sve2": {
             "qasymm8": [ "src/cpu/kernels/add/generic/sve2/qasymm8.cpp" ],
             "qasymm8_signed": [ "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp" ],
             "qsymm16": [ "src/cpu/kernels/add/generic/sve2/qsymm16.cpp" ]
@@ -1024,38 +1026,42 @@
             "src/cpu/kernels/CpuCol2ImKernel.cpp",
             "src/cpu/kernels/CpuIm2ColKernel.cpp",
             "src/cpu/kernels/CpuWeightsReshapeKernel.cpp",
-            "src/core/NEON/kernels/convolution/common/padding.cpp",
-            "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
-            "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
-            "src/core/NEON/kernels/convolution/common/utils.cpp",
-            "src/core/NEON/kernels/convolution/winograd/padding.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_1x8_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_4x4_fp16_fp16_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_4x4_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_6x6_fp16_fp16_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_6x6_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2_7_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_5x5_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4_5_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp16_fp16_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_6_3_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2_7_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2x2_3x3_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2x2_5x5_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4_5_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp32_fp32_integers.cpp",
-            "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_6_3_fp32_fp32_integers.cpp",
             "src/runtime/NEON/functions/NEConvolutionLayer.cpp",
             "src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp",
             "src/runtime/NEON/functions/NEFFTConvolutionLayer.cpp",
             "src/runtime/NEON/functions/NEGEMMConv2d.cpp",
             "src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp",
             "src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp"
-          ]
+          ],
+          "neon": {
+            "common": [
+              "src/core/NEON/kernels/convolution/common/padding.cpp",
+              "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
+              "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
+              "src/core/NEON/kernels/convolution/common/utils.cpp",
+              "src/core/NEON/kernels/convolution/winograd/padding.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_1x8_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_4x4_fp16_fp16_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_4x4_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_6x6_fp16_fp16_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/input_6x6_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2_7_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_3x3_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_2x2_5x5_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4_5_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp16_fp16_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_4x4_3x3_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/output_6_3_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2_7_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2x2_3x3_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_2x2_5x5_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4_5_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp16_fp16_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_4x4_3x3_fp32_fp32_integers.cpp",
+              "src/core/NEON/kernels/convolution/winograd/winograd_transforms/weights_6_3_fp32_fp32_integers.cpp" 
+            ]
+          }
         }
       },
       "Copy": {
@@ -1079,7 +1085,7 @@
             "fp32":    [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ],
             "fp16":    [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ],
             "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp"   ]
-          }  
+          }
         }
       },
       "Deconv2d": {
@@ -1114,68 +1120,72 @@
             "src/cpu/operators/CpuDepthwiseConv2dAssemblyDispatch.cpp",
             "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.cpp",
             "src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp",
-            "src/core/NEON/kernels/convolution/common/padding.cpp",
-            "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
-            "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
-            "src/core/NEON/kernels/convolution/common/utils.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp",
-            "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
-            "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp"
+            "src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp"
           ],
+          "neon": {
+            "common": [
+              "src/core/NEON/kernels/convolution/common/padding.cpp",
+              "src/core/NEON/kernels/convolution/common/qasymm8.cpp",
+              "src/core/NEON/kernels/convolution/common/qsymm8.cpp",
+              "src/core/NEON/kernels/convolution/common/utils.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp16.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8q.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_u8s8u8q.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/8b_mla.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_s8q.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/depthwise_fp32.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_3x3_s2_with_multiplier_output3x3_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_5x5_s1_with_multiplier_output2x4_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_direct.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8qs_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_5x5_s1_output2x2_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_generic_output9_mla_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp",
+              "src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp"
+              ]
+          },
           "sve": {
             "common": [
               "src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_8b_mla.cpp",
@@ -1388,92 +1398,94 @@
             "src/cpu/operators/CpuGemm.cpp",
             "src/cpu/operators/CpuGemmLowpOutputStage.cpp",
             "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp",
-            "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp",
-            "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp",
-            "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp",
-            "src/core/NEON/kernels/arm_gemm/mergeresults.cpp",
-            "src/core/NEON/kernels/arm_gemm/misc.cpp",
-            "src/core/NEON/kernels/arm_gemm/quantized.cpp",
-            "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp",
-            "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp",
-            "src/core/NEON/kernels/arm_gemm/transform.cpp",
             "src/runtime/NEON/functions/NEGEMM.cpp",
             "src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp",
-            "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
-            "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp"
+            "src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp"
           ],
           "neon": {
+            "common": [
+              "src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_int16.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_int8.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_uint16.cpp",
+              "src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp",
+              "src/core/NEON/kernels/arm_gemm/interleave_indirect.cpp",
+              "src/core/NEON/kernels/arm_gemm/mergeresults-fp16.cpp",
+              "src/core/NEON/kernels/arm_gemm/mergeresults.cpp",
+              "src/core/NEON/kernels/arm_gemm/misc.cpp",
+              "src/core/NEON/kernels/arm_gemm/quantized.cpp",
+              "src/core/NEON/kernels/arm_gemm/rowsum_indirect_s8.cpp",
+              "src/core/NEON/kernels/arm_gemm/rowsum_indirect_u8.cpp",
+              "src/core/NEON/kernels/arm_gemm/transform.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/a55r1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_8x12/x1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/a55r1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12/x1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/a55r1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24/x1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_4x24/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_mmla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_dot_4x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8qa_mmla_4x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_mmla_6x16/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/x1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/a510.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/a510.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a53.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/a55r1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12/x1.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x6/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_6x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_fp32_mla_8x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_6x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_s8s32_dot_8x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_6x4/generic.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/a55.cpp",
+              "src/core/NEON/kernels/arm_gemm/kernels/a64_smallK_hybrid_u8u32_dot_8x4/generic.cpp"    
+            ],
             "estate32": [
               "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a53.cpp",
               "src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6/a55r1.cpp",
@@ -1677,36 +1689,38 @@
             "src/cpu/operators/CpuPool2d.cpp",
             "src/cpu/kernels/CpuPool2dKernel.cpp",
             "src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
-            "src/runtime/NEON/functions/NEPoolingLayer.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
-            "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp",
             "src/cpu/kernels/pool2d/neon/qasymm8.cpp",
-            "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp"
+            "src/cpu/kernels/pool2d/neon/qasymm8_signed.cpp",
+            "src/runtime/NEON/functions/NEPoolingLayer.cpp"
           ],
           "neon": {
+            "common": [
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/cpp_nhwc_1x1_stride_any_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_fp16.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_fp32.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_s8.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_s8q.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_u8.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/pooling_u8q.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp16_nhwc_max_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_3x3_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_fp32_nhwc_max_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8_nhwc_max_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_s8q_nhwc_max_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_2x2_s1_output2x2_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8_nhwc_max_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_avg_generic_depthfirst/generic.cpp",
+              "src/core/NEON/kernels/arm_conv/pooling/kernels/a64_u8q_nhwc_max_generic_depthfirst/generic.cpp"  
+            ],
             "nchw": [ "src/cpu/kernels/pool2d/neon/nchw/all.cpp" ],
             "fp16": [ "src/cpu/kernels/pool2d/neon/fp16.cpp" ],
             "fp32": [ "src/cpu/kernels/pool2d/neon/fp32.cpp" ]