Add SVE2 optimizations for SynetConvolution32fDirectNchw (#810)

* Add SVE2 optimizations for SynetConvolution32fDirectNchw

Port DirectNchw convolution kernels to SVE2 with predicated loads/stores
and gather-based strided access, wire Sve2 dispatch, extend AutoTest
coverage for NCHW DirectNchw shapes, and document the change in release
7.2.165.

Co-authored-by: Ihar Yermalayeu <ermig1979@gmail.com>

* Fix SVE2 DirectNchw for sizeless vector types

Avoid arrays and pointer arithmetic on svfloat32_t; broadcast kernel
weights from scalar float pointers and pass activation params by value.

Co-authored-by: Ihar Yermalayeu <ermig1979@gmail.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>

Author: ermig1979

docs/2026.html

@@ -46,6 +46,7 @@ <h5>New features</h5>
  <li>SVE2 optimizations of class SynetConvolution32fWinograd.</li>
  <li>NEON optimizations of class SynetConvolution32fNhwcGroupedBlock1x2.</li>
  <li>SVE2 optimizations of class SynetConvolution32fNhwcGroupedBlock1x2.</li>
+ <li>SVE2 optimizations of class SynetConvolution32fDirectNchw.</li>
 </ul>
 
 <a href="#HOME">Home</a>

prj/vs2022/Sve2.vcxproj

@@ -106,6 +106,7 @@
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetActivation.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution8i.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32f.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32fDirectNchw.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32fGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32fNhwcGrouped.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConversion.cpp" />

prj/vs2022/Sve2.vcxproj.filters

@@ -557,6 +557,9 @@
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32f.cpp">
       <Filter>Sve2\Synet\Convolution</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32fDirectNchw.cpp">
+      <Filter>Sve2\Synet\Convolution</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdSve2SynetConvolution32fGemm.cpp">
       <Filter>Sve2\Synet\Convolution</Filter>
     </ClCompile>

src/Simd/SimdSve2SynetConvolution32f.cpp

@@ -40,8 +40,8 @@ namespace Simd
                 return new Neon::SynetConvolution32fDepthwiseDotProduct(param);
             else if (SynetConvolution32fWinograd::Preferable(param))
                 return new SynetConvolution32fWinograd(param);
-            else if (Neon::SynetConvolution32fDirectNchw::Preferable(param))
-                return new Neon::SynetConvolution32fDirectNchw(param);
+            else if (SynetConvolution32fDirectNchw::Preferable(param))
+                return new SynetConvolution32fDirectNchw(param);
             else if (SynetConvolution32fGemmNT::Preferable(param))
                 return new SynetConvolution32fGemmNT(param);
             else if (Neon::SynetConvolution32fNhwcDirect::Preferable(param))

src/Simd/SimdSve2SynetConvolution32fDirectNchw.cpp

@@ -760,6 +760,18 @@ namespace Simd
             virtual String Ext() const { return "Sve2"; }
         };
 
+        class SynetConvolution32fDirectNchw : public Neon::SynetConvolution32fDirectNchw
+        {
+        public:
+            SynetConvolution32fDirectNchw(const ConvParam & p);
+            virtual String Ext() const { return "Sve2"; }
+
+            static bool Preferable(const ConvParam & p);
+
+        protected:
+            virtual ConvolutionBiasActivationPtr SetConvolutionBiasActivation();
+        };
+
         void * SynetConvolution32fInit(size_t batch, const SimdConvolutionParameters * conv);
     }
 #endif//SIMD_SVE2_ENABLE

src/Simd/SimdSynetConvolution32f.h

@@ -267,6 +267,10 @@ namespace Test
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 128, 16, 16, 1, _1, _1, _1, _0, _0, 1, a, tT), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 3, 64, 64, 16, _3, _1, _1, _1, _1, 1, aPr, tF), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 3, 64, 64, 16, _3, _1, _1, _1, _1, 1, aPr, t), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 1, 48, 48, 8, _3, _1, _1, _1, _1, 1, aRe, tF), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 3, 40, 40, 16, _2, _1, _1, _0, _0, 1, aId, tF), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 1, 33, 33, 4, _3, _1, _2, _1, _1, 1, aRe, tF), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 1, 64, 64, 8, _1, _1, _1, _0, _0, 1, aRe, tF), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 64, 128, 128, 256, _3, _1, _1, _1, _1, 1, a, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 32, 16, 16, 32, _3, _1, _1, _1, _1, 1, aId, tT), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 32, 6, 15, 32, _3, _1, _1, _1, _1, 1, aId, tT), f1, f2);
@@ -296,6 +300,8 @@ namespace Test
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 48, 4, 16, 48, Size(1, 3), _1, _1, Size(0, 1), Size(0, 1), 1, aId, tT), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 16, 9, 9, 16, _3, Size(2, 2), _1, _2, _2, 1, aRe, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 16, 9, 9, 16, _3, Size(2, 2), _2, _2, _2, 1, aRe, tF), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 1, 48, 48, 8, _3, _1, _1, _1, _1, 1, aRe, tF), f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 3, 40, 40, 16, _2, _1, _1, _0, _0, 1, aId, tF), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 96, 14, 14, 192, _3, _1, _1, _1, _1, 96, a, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 49, 29, 29, 98, _7, _1, _2, _3, _3, 49, a, t), f1, f2);
 #endif