Add NEON optimizations for SynetQuantizedMergedConvolution Cdc/Cd/Dc

docs/2026.html

@@ -63,6 +63,9 @@ <h5>New features</h5>
  <li>NEON, SVE2 optimizations of classes SynetMergedConvolution16bCdc.</li>
  <li>NEON, SVE2 optimizations of classes SynetMergedConvolution16bCd.</li>
  <li>NEON, SVE2 optimizations of classes SynetMergedConvolution16bDc.</li>
+ <li>NEON optimizations of classes SynetQuantizedMergedConvolutionCdc.</li>
+ <li>NEON optimizations of classes SynetQuantizedMergedConvolutionCd.</li>
+ <li>NEON optimizations of classes SynetQuantizedMergedConvolutionDc.</li>
 </ul>
 <h5>Improving</h5>
 <ul>

prj/vs2022/Neon.vcxproj

@@ -146,6 +146,10 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetMergedConvolution8iDepthwise.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetMergedConvolution8iInput.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetMergedConvolution8iOutput.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolution.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionDepthwise.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionInput.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionOutput.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetPermute.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetPooling.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetPoolingMax16b.cpp" />
@@ -250,6 +254,7 @@
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedActivation.h" />
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedAdd.h" />
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedAddCommon.h" />
+    <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedMergedConvolution.h" />
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizeLinear.h" />
     <ClInclude Include="..\..\src\Simd\SimdSynetScale8i.h" />
     <ClInclude Include="..\..\src\Simd\SimdTime.h" />

prj/vs2022/Neon.vcxproj.filters

@@ -364,6 +364,18 @@
     <ClCompile Include="..\..\src\Simd\SimdNeonSynetMergedConvolution8iOutput.cpp">
       <Filter>Neon\Synet\MergedConvolution</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolution.cpp">
+      <Filter>Neon\Synet\Quantized</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionDepthwise.cpp">
+      <Filter>Neon\Synet\Quantized</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionInput.cpp">
+      <Filter>Neon\Synet\Quantized</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdNeonSynetQuantizedMergedConvolutionOutput.cpp">
+      <Filter>Neon\Synet\Quantized</Filter>
+    </ClCompile>
     <ClCompile Include="..\..\src\Simd\SimdNeonSynet.cpp">
       <Filter>Neon\Synet\Other</Filter>
     </ClCompile>
@@ -779,6 +791,9 @@
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedAddCommon.h">
       <Filter>Inc</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\Simd\SimdSynetQuantizedMergedConvolution.h">
+      <Filter>Inc</Filter>
+    </ClInclude>
     <ClInclude Include="..\..\src\Simd\SimdSynetQuantizeLinear.h">
       <Filter>Inc</Filter>
     </ClInclude>

src/Simd/SimdLib.cpp

@@ -7480,7 +7480,7 @@ SIMD_API void* SimdSynetQuantizedMergedConvolutionInit(size_t batch, const SimdC
     SIMD_EMPTY();
 #if defined(SIMD_SYNET_ENABLE)
     typedef void* (*SimdSynetQuantizedMergedConvolutionInitPtr) (size_t batch, const SimdConvolutionParameters* convs, size_t count, int add);
-    const static SimdSynetQuantizedMergedConvolutionInitPtr simdSynetQuantizedMergedConvolutionInit = SIMD_FUNC5(SynetQuantizedMergedConvolutionInit, SIMD_AMXBF16_FUNC, SIMD_AVX512VNNI_FUNC, SIMD_AVX512BW_FUNC, SIMD_AVX2_FUNC, SIMD_SSE41_FUNC);
+    const static SimdSynetQuantizedMergedConvolutionInitPtr simdSynetQuantizedMergedConvolutionInit = SIMD_FUNC6(SynetQuantizedMergedConvolutionInit, SIMD_AMXBF16_FUNC, SIMD_AVX512VNNI_FUNC, SIMD_AVX512BW_FUNC, SIMD_AVX2_FUNC, SIMD_SSE41_FUNC, SIMD_NEON_FUNC);
 
     return simdSynetQuantizedMergedConvolutionInit(batch, convs, count, add);
 #else

src/Simd/SimdMath.h

@@ -1515,6 +1515,15 @@ namespace Simd
             return vcvtq_s32_f32(vaddq_f32(value, round));
         }
 
+        SIMD_INLINE int32x4_t NearbyInt(float32x4_t value)
+        {
+#if defined(__aarch64__)
+            return vcvtnq_s32_f32(value);
+#else
+            return Round(value);
+#endif
+        }
+
         SIMD_INLINE uint32x4_t RoundPositive(float32x4_t value)
         {
             return vcvtq_u32_f32(vaddq_f32(value, vdupq_n_f32(0.5f)));

src/Simd/SimdNeonSynetQuantizedMergedConvolution.cpp

@@ -0,0 +1,93 @@
+/*
+* Simd Library (http://ermig1979.github.io/Simd).
+*
+* Copyright (c) 2011-2026 Yermalayeu Ihar.
+*
+* Permission is hereby granted, free of charge, to any person obtaining a copy
+* of this software and associated documentation files (the "Software"), to deal
+* in the Software without restriction, including without limitation the rights
+* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+* copies of the Software, and to permit persons to whom the Software is
+* furnished to do so, subject to the following conditions:
+*
+* The above copyright notice and this permission notice shall be included in
+* all copies or substantial portions of the Software.
+*
+* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+* SOFTWARE.
+*/
+#include "Simd/SimdSynetQuantizedMergedConvolution.h"
+#include "Simd/SimdSynetQuantizeLinear.h"
+#include "Simd/SimdSynetConvolution8iCommon.h"
+#include "Simd/SimdSynet.h"
+#include "Simd/SimdMath.h"
+#include "Simd/SimdBase.h"
+#include "Simd/SimdCpu.h"
+#include "Simd/SimdLog.h"
+
+namespace Simd
+{
+#if defined(SIMD_NEON_ENABLE) && defined(SIMD_SYNET_ENABLE)   
+    namespace Neon
+    {
+        typedef Base::SynetQuantizedMergedConvolution::AlgParam AlgParam;
+
+        //------------------------------------------------------------------------------------------------
+
+        SynetQuantizedMergedConvolutionCdc::SynetQuantizedMergedConvolutionCdc(const MergConvParam& p)
+            : Base::SynetQuantizedMergedConvolutionCdc(p)
+        {
+            SetSize(F, 4, 2);
+            SetInputConvolution(p.conv[0], _alg, _inputConvolution);
+            SetDepthwisePreprocess(p.conv[1], _alg, _depthwisePreprocess);
+            SetDepthwiseConvolution(p.conv[1], _alg, _depthwiseConvolution);
+            SetOutputConvolution(p.conv[2], _alg, _outputConvolution);
+            SetAddInputToOutput(p.conv[2], _alg, _addInputToOutput);
+        }
+
+        //------------------------------------------------------------------------------------------------
+
+        SynetQuantizedMergedConvolutionCd::SynetQuantizedMergedConvolutionCd(const MergConvParam& p)
+            : Base::SynetQuantizedMergedConvolutionCd(p)
+        {
+            SetSize(F, 4, 2);
+            SetInputConvolution(p.conv[0], _alg, _inputConvolution);
+            SetDepthwisePreprocess(p.conv[1], _alg, _depthwisePreprocess);
+            SetDepthwiseConvolution(p.conv[1], _alg, _depthwiseConvolution);
+        }
+
+        //------------------------------------------------------------------------------------------------
+
+        SynetQuantizedMergedConvolutionDc::SynetQuantizedMergedConvolutionDc(const MergConvParam& p)
+            : Base::SynetQuantizedMergedConvolutionDc(p)
+        {
+            SetSize(F, 4, 2);
+            SetDepthwisePreprocess(p.conv[0], _alg, _depthwisePreprocess);
+            SetDepthwiseConvolution(p.conv[0], _alg, _depthwiseConvolution);
+            SetOutputConvolution(p.conv[1], _alg, _outputConvolution);
+            SetAddInputToOutput(p.conv[1], _alg, _addInputToOutput);
+        }
+
+        //------------------------------------------------------------------------------------------------
+
+        void* SynetQuantizedMergedConvolutionInit(size_t batch, const SimdConvolutionParameters* convs, size_t count, int add)
+        {
+            MergConvParam param(batch, convs, count, add);
+            if (!param.Valid(SimdTensorData8u, SimdTensorData8u))
+                return NULL;
+            else if (SynetQuantizedMergedConvolutionCdc::Preferable(param))
+                return new SynetQuantizedMergedConvolutionCdc(param);
+            else if (SynetQuantizedMergedConvolutionCd::Preferable(param))
+                return new SynetQuantizedMergedConvolutionCd(param);
+            else if (SynetQuantizedMergedConvolutionDc::Preferable(param))
+                return new SynetQuantizedMergedConvolutionDc(param);
+            return new Base::SynetQuantizedMergedConvolutionRef(param);
+        }
+    }
+#endif
+}