*refactoring of SynetConvolution32f.

ermig1979 · Oct 1, 2024 · 12333d2 · 12333d2
1 parent ef1a1fe
commit 12333d2
Show file tree

Hide file tree

Showing 47 changed files with 5,326 additions and 4,776 deletions.
diff --git a/prj/vs2019/Avx2.vcxproj b/prj/vs2019/Avx2.vcxproj
@@ -91,7 +91,9 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2019/Avx2.vcxproj.filters b/prj/vs2019/Avx2.vcxproj.filters
@@ -238,9 +238,6 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNchw.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNhwc.cpp">
-      <Filter>Avx2</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdAvx2Transform.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
@@ -385,6 +382,15 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetDeconvolution16b.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fGemm.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Avx2">

diff --git a/prj/vs2019/Avx512bw.vcxproj b/prj/vs2019/Avx512bw.vcxproj
@@ -94,7 +94,9 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2019/Avx512bw.vcxproj.filters b/prj/vs2019/Avx512bw.vcxproj.filters
@@ -280,9 +280,6 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNchw.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNhwc.cpp">
-      <Filter>Avx512bw</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2f.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
@@ -397,6 +394,15 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fGemm.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Avx512bw">

diff --git a/prj/vs2019/Base.vcxproj b/prj/vs2019/Base.vcxproj
@@ -195,6 +195,11 @@
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution16bNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32f.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fDirectNchw.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDirect.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcGrouped.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution8i.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16b.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16bNhwcGemm.cpp" />

diff --git a/prj/vs2019/Base.vcxproj.filters b/prj/vs2019/Base.vcxproj.filters
@@ -358,6 +358,21 @@
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Base</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fDirectNchw.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fGemm.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcGrouped.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\..\src\Simd\SimdBase.h">

diff --git a/prj/vs2019/Neon.vcxproj b/prj/vs2019/Neon.vcxproj
@@ -96,7 +96,9 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConversion.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2019/Neon.vcxproj.filters b/prj/vs2019/Neon.vcxproj.filters
@@ -217,9 +217,6 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNchw.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNhwc.cpp">
-      <Filter>Neon</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdNeonGemm32fNN.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
@@ -301,6 +298,15 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonDescrIntCdu.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fGemm.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Neon">

diff --git a/prj/vs2019/Sse41.vcxproj b/prj/vs2019/Sse41.vcxproj
@@ -98,7 +98,9 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2019/Sse41.vcxproj.filters b/prj/vs2019/Sse41.vcxproj.filters
@@ -256,9 +256,6 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNchw.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNhwc.cpp">
-      <Filter>Sse41</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2f.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
@@ -409,6 +406,15 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fGemm.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Sse41">

diff --git a/prj/vs2022/Avx2.vcxproj b/prj/vs2022/Avx2.vcxproj
@@ -91,7 +91,9 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2022/Avx2.vcxproj.filters b/prj/vs2022/Avx2.vcxproj.filters
@@ -238,9 +238,6 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNchw.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fDirectNhwc.cpp">
-      <Filter>Avx2</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdAvx2Transform.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
@@ -385,6 +382,15 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx2SynetDeconvolution16b.cpp">
       <Filter>Avx2</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fGemm.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx2SynetConvolution32fNhwcDirect.cpp">
+      <Filter>Avx2</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Avx2">

diff --git a/prj/vs2022/Avx512bw.vcxproj b/prj/vs2022/Avx512bw.vcxproj
@@ -94,7 +94,9 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2022/Avx512bw.vcxproj.filters b/prj/vs2022/Avx512bw.vcxproj.filters
@@ -280,9 +280,6 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNchw.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fDirectNhwc.cpp">
-      <Filter>Avx512bw</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect2f.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
@@ -397,6 +394,15 @@
     <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Avx512bw</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fGemm.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdAvx512bwSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Avx512bw</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Avx512bw">

diff --git a/prj/vs2022/Base.vcxproj b/prj/vs2022/Base.vcxproj
@@ -195,6 +195,11 @@
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution16bNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32f.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fDirectNchw.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDirect.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcGrouped.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution8i.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16b.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16bNhwcGemm.cpp" />

diff --git a/prj/vs2022/Base.vcxproj.filters b/prj/vs2022/Base.vcxproj.filters
@@ -358,6 +358,21 @@
     <ClCompile Include="..\..\src\Simd\SimdBaseSynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Base</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fDirectNchw.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fGemm.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdBaseSynetConvolution32fNhwcGrouped.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\..\src\Simd\SimdBase.h">

diff --git a/prj/vs2022/Neon.vcxproj b/prj/vs2022/Neon.vcxproj
@@ -96,7 +96,9 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConversion.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2022/Neon.vcxproj.filters b/prj/vs2022/Neon.vcxproj.filters
@@ -217,9 +217,6 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNchw.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fDirectNhwc.cpp">
-      <Filter>Neon</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdNeonGemm32fNN.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
@@ -301,6 +298,15 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonDescrIntCdu.cpp">
       <Filter>Neon</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fGemm.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetConvolution32fNhwcDirect.cpp">
+      <Filter>Neon</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Neon">

diff --git a/prj/vs2022/Sse41.vcxproj b/prj/vs2022/Sse41.vcxproj
@@ -98,7 +98,9 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNchw.cpp" />
-    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNhwc.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2r.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect3r.cpp" />

diff --git a/prj/vs2022/Sse41.vcxproj.filters b/prj/vs2022/Sse41.vcxproj.filters
@@ -256,9 +256,6 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNchw.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fDirectNhwc.cpp">
-      <Filter>Sse41</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect2f.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
@@ -409,6 +406,15 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16bNhwcGemm.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fGemm.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDepthwise.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Sse41">