+add SSE4.1 optimizations of class SynetConvolution16bNhwcDepthwise.

ermig1979 · Oct 9, 2024 · 3a3a5ae · 3a3a5ae
1 parent b82f4f3
commit 3a3a5ae
Show file tree

Hide file tree

Showing 14 changed files with 322 additions and 19 deletions.
diff --git a/docs/2024.html b/docs/2024.html
@@ -39,7 +39,7 @@ <h3 id="R143">November X, 2024 (version X.X.143)</h3>
 <h4>Algorithms</h4>
 <h5>New features</h5>
 <ul>
- <li>Base implementation of class SynetConvolution16bNhwcDepthwise.</li>
+ <li>Base implementation, SSE4.1 optimizations of class SynetConvolution16bNhwcDepthwise.</li>
 </ul>
 <h5>Im
 <h5>Improving</h5>

diff --git a/prj/vs2019/Sse41.vcxproj b/prj/vs2019/Sse41.vcxproj
@@ -94,6 +94,7 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConversion.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16b.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNchwGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDeptwise.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32f.cpp" />

diff --git a/prj/vs2019/Sse41.vcxproj.filters b/prj/vs2019/Sse41.vcxproj.filters
@@ -415,6 +415,9 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDeptwise.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Sse41">

diff --git a/prj/vs2022/Sse41.vcxproj b/prj/vs2022/Sse41.vcxproj
@@ -94,6 +94,7 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConversion.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16b.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNchwGemm.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDeptwise.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32f.cpp" />

diff --git a/prj/vs2022/Sse41.vcxproj.filters b/prj/vs2022/Sse41.vcxproj.filters
@@ -415,6 +415,9 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution32fNhwcDirect.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNhwcDeptwise.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Sse41">

diff --git a/src/Simd/SimdAmxBf16SynetConvolution16b.cpp b/src/Simd/SimdAmxBf16SynetConvolution16b.cpp
@@ -40,7 +40,7 @@ namespace Simd
             if (SynetConvolution16bNchwGemm::Preferable(param))
                 return new AmxBf16::SynetConvolution16bNchwGemm(param);
             if (Base::SynetConvolution16bNhwcDepthwise::Preferable(param))
-                return new Base::SynetConvolution16bNhwcDepthwise(param);
+                return new Sse41::SynetConvolution16bNhwcDepthwise(param);
             return new Base::SynetConvolution16bGemm(param);
         }
     }

diff --git a/src/Simd/SimdAvx2SynetConvolution16b.cpp b/src/Simd/SimdAvx2SynetConvolution16b.cpp
@@ -40,7 +40,7 @@ namespace Simd
             if (Base::SynetConvolution16bNchwGemm::Preferable(param))
                 return new Avx2::SynetConvolution16bNchwGemm(param);
             if (Base::SynetConvolution16bNhwcDepthwise::Preferable(param))
-                return new Base::SynetConvolution16bNhwcDepthwise(param);
+                return new Sse41::SynetConvolution16bNhwcDepthwise(param);
             return new Base::SynetConvolution16bGemm(param);
         }
     }

diff --git a/src/Simd/SimdAvx512bwSynetConvolution16b.cpp b/src/Simd/SimdAvx512bwSynetConvolution16b.cpp
@@ -40,7 +40,7 @@ namespace Simd
             if (Base::SynetConvolution16bNchwGemm::Preferable(param))
                 return new Avx512bw::SynetConvolution16bNchwGemm(param);
             if (Base::SynetConvolution16bNhwcDepthwise::Preferable(param))
-                return new Base::SynetConvolution16bNhwcDepthwise(param);
+                return new Sse41::SynetConvolution16bNhwcDepthwise(param);
             return new Base::SynetConvolution16bGemm(param);
         }
     }

diff --git a/src/Simd/SimdSse41SynetConvolution16b.cpp b/src/Simd/SimdSse41SynetConvolution16b.cpp
@@ -39,8 +39,8 @@ namespace Simd
                 return new Sse41::SynetConvolution16bNhwcGemm(param);
             if (SynetConvolution16bNchwGemm::Preferable(param))
                 return new Sse41::SynetConvolution16bNchwGemm(param);
-            if (Base::SynetConvolution16bNhwcDepthwise::Preferable(param))
-                return new Base::SynetConvolution16bNhwcDepthwise(param);
+            if (SynetConvolution16bNhwcDepthwise::Preferable(param))
+                return new SynetConvolution16bNhwcDepthwise(param);
             return new Base::SynetConvolution16bGemm(param);
         }
     }

diff --git a/src/Simd/SimdSse41SynetConvolution16bNhwcDeptwise.cpp b/src/Simd/SimdSse41SynetConvolution16bNhwcDeptwise.cpp
diff --git a/src/Simd/SimdSse41SynetMergedConvolution16bDepthwise.cpp b/src/Simd/SimdSse41SynetMergedConvolution16bDepthwise.cpp
@@ -39,19 +39,7 @@ namespace Simd
         using AlgParam = Base::SynetMergedConvolution16b::AlgParam;
         using DepthwisePtr = Base::SynetMergedConvolution16b::DepthwiseConvolutionPtr;
 
-        template <class T> SIMD_INLINE __m128 LoadSrc(const T* src);
-
-        template <> SIMD_INLINE __m128 LoadSrc<float>(const float* src)
-        {
-            return _mm_loadu_ps(src);
-        }
-
-        template <> SIMD_INLINE __m128 LoadSrc<uint16_t>(const uint16_t* src)
-        {
-            return  _mm_castsi128_ps(UnpackU16<0>(K_ZERO, _mm_loadl_epi64((__m128i*)src)));
-        }
-
-        //---------------------------------------------------------------------
+        //-------------------------------------------------------------------------------------------------
 
         template<typename T, Term16bType term, SimdConvolutionActivationType type> void DepthwiseConvolution(const uint8_t* src8, const ConvParam& p, const AlgParam& a, 
             size_t maC, size_t yBeg, size_t yEnd, const float* weight, const float* bias, const float* params, uint8_t* dst)

diff --git a/src/Simd/SimdSynetConvolution16b.h b/src/Simd/SimdSynetConvolution16b.h
@@ -283,6 +283,14 @@ namespace Simd
             virtual String Ext() const { return "Sse41"; }
         };
 
+        class SynetConvolution16bNhwcDepthwise : public Base::SynetConvolution16bNhwcDepthwise
+        {
+        public:
+            SynetConvolution16bNhwcDepthwise(const ConvParam& p);
+
+            virtual String Ext() const { return "Sse41"; }
+        };
+
         class SynetConvolution16bNchwGemm : public Base::SynetConvolution16bNchwGemm
         {
         public:

diff --git a/src/Simd/SimdSynetConvolution16bCommon.h b/src/Simd/SimdSynetConvolution16bCommon.h
@@ -71,6 +71,20 @@ namespace Simd
 #ifdef SIMD_SSE41_ENABLE
     namespace Sse41
     {
+        template <class T> SIMD_INLINE __m128 LoadSrc(const T* src);
+
+        template <> SIMD_INLINE __m128 LoadSrc<float>(const float* src)
+        {
+            return _mm_loadu_ps(src);
+        }
+
+        template <> SIMD_INLINE __m128 LoadSrc<uint16_t>(const uint16_t* src)
+        {
+            return  _mm_castsi128_ps(UnpackU16<0>(K_ZERO, _mm_loadl_epi64((__m128i*)src)));
+        }
+
+        //-------------------------------------------------------------------------------------------------
+
         template <Term16bType term> struct Term16b
         {
             template<SimdConvolutionActivationType type, int index> static SIMD_INLINE void Save(uint16_t* ptr, __m128 value, const __m128* bias, const __m128* params);

diff --git a/src/Test/TestSynetConvolution16b.cpp b/src/Test/TestSynetConvolution16b.cpp
@@ -283,11 +283,13 @@ namespace Test
 #endif
 #if 1
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 76, 64, 64, 76, _7, _1, _1, _3, _3, 76, aId, tT, b16, f32), c, f1, f2);
+        result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 76, 64, 64, 76, _3, _1, _1, _1, _1, 76, aId, tT, f32, f32), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 152, 32, 32, 152, _7, _1, _1, _3, _3, 152, aPr, tT, b16, f32), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 304, 16, 16, 304, _7, _1, _1, _3, _3, 304, aMi, tT, f32, f32), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 304, 16, 16, 304, _3, _1, _1, _1, _1, 304, aRe, tT, f32, b16), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 304, 17, 15, 304, _7, _1, _1, _3, _3, 304, aGe, tT, b16, f32), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 608, 8, 8, 608, _7, _1, _1, _3, _3, 608, aSw, tT, b16, b16), c, f1, f2);
+        result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 152, 32, 32, 152, _5, _1, _2, _2, _2, 152, aPr, tT, b16, b16), c, f1, f2);
 #endif
 
 #else