+add SSE4.1 optimizations of class SynetDeconvolution16bNhwcGemm.

ermig1979 · Sep 10, 2024 · 645dfd7 · 645dfd7
1 parent efdefd3
commit 645dfd7
Show file tree

Hide file tree

Showing 8 changed files with 600 additions and 22 deletions.
diff --git a/docs/2024.html b/docs/2024.html
@@ -40,7 +40,7 @@ <h4>Algorithms</h4>
 <h5>New features</h5>
 <ul>
  <li>Base implementation of class SynetDeconvolution16bGemm.</li>
- <li>Base implementation of class SynetDeconvolution16bNhwcGemm.</li>
+ <li>Base implementation, SSE4.1 optimizations of class SynetDeconvolution16bNhwcGemm.</li>
 </ul>
 
 <h4>Test framework</h4>

diff --git a/prj/vs2022/Sse41.vcxproj b/prj/vs2022/Sse41.vcxproj
@@ -108,6 +108,8 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution8iNhwcDirect.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution8iNhwcDirect1x1.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution8iNhwcDirectAny.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16b.cpp" />
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16bNhwcGemm.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution32f.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetGridSample.cpp" />
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetGridSample2d32fBlZ.cpp" />

diff --git a/prj/vs2022/Sse41.vcxproj.filters b/prj/vs2022/Sse41.vcxproj.filters
@@ -418,6 +418,12 @@
     <ClCompile Include="..\..\src\Simd\SimdSse41SynetConvolution16bNchwGemm.cpp">
       <Filter>Sse41</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16b.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\Simd\SimdSse41SynetDeconvolution16bNhwcGemm.cpp">
+      <Filter>Sse41</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Sse41">

diff --git a/src/Simd/SimdBaseSynetDeconvolution16b.cpp b/src/Simd/SimdBaseSynetDeconvolution16b.cpp
@@ -210,14 +210,14 @@ namespace Simd
                     assert(p.group == 1);
                     GemmNN(_M, _N, _K, src16b, _ldS, wgt, _ldW, buf32f, _ldD);
                     if (!_is1x1)
-                        ImgToRow(buf32f, dst32f);
+                        RowToImg(buf32f, dst32f);
                 }
                 else
                 {
                     for (size_t g = 0; g < p.group; ++g)
                         GemmNN(_M, _N, _K, wgt + _grW * g, _ldW, src16b + _grS * g, _ldS, buf32f + _grD * g, _ldD);
                     if (!_is1x1)
-                        ImgToCol(buf32f, dst32f);
+                        ColToImg(buf32f, dst32f);
                 }
                 ConvolutionBiasAndActivation(_bias.data, p.dstC, p.dstH * p.dstW, p.activation, _params.data, p.trans, dst32f);
                 if (_dst16b)
@@ -227,7 +227,7 @@ namespace Simd
             }
         }
 
-        void SynetDeconvolution16bGemm::ImgToCol(const float* src, float* dst)
+        void SynetDeconvolution16bGemm::ColToImg(const float* src, float* dst)
         {
             const DeconvParam& p = _param;
             assert(!p.trans);
@@ -261,7 +261,7 @@ namespace Simd
             }
         }
 
-        void SynetDeconvolution16bGemm::ImgToRow(const float* src, float* dst)
+        void SynetDeconvolution16bGemm::RowToImg(const float* src, float* dst)
         {
             const DeconvParam& p = _param;
             assert(p.trans && p.group == 1);

diff --git a/src/Simd/SimdBaseSynetDeconvolution16bNhwcGemm.cpp b/src/Simd/SimdBaseSynetDeconvolution16bNhwcGemm.cpp
@@ -36,6 +36,9 @@ namespace Simd
         SynetDeconvolution16bNhwcGemm::SynetDeconvolution16bNhwcGemm(const DeconvParam& p)
             : SynetDeconvolution16b(p)
             , _convert(0)
+            , _gemm(0)
+            , _toImg(0)
+            , _biasAct(0)
         {
             assert(p.trans && p.group == 1);
         }
@@ -70,6 +73,7 @@ namespace Simd
             a.bufM = p.dstH * AlignHi(p.dstW, a.F);
             a.macroK = Simd::RestrictRange(AlignLo(L1 / a.microN / 2, a.microK), a.microK, a.bufK);
             a.macroH = Simd::RestrictRange(L2 / a.macroK / p.dstW / 2, size_t(1), p.dstH);
+            a.macroM = a.macroH * p.dstW;
             a.macroN = Simd::RestrictRange(AlignLoAny(L3 / a.macroK / 2, a.microN), a.microN, a.bufN);
             _stepS = p.srcH * p.srcW * p.srcC * _elemS;
             _stepD = p.dstH * p.dstW * p.dstC * _elemD;
@@ -112,24 +116,51 @@ namespace Simd
             uint16_t* bufS = _src16b && a.bufK == a.K ? NULL : Allocate<uint16_t>(buf, a.bufK * a.bufM); 
             float* bufB = _is1x1 ? NULL : Allocate<float>(buf, a.bufN * a.bufM);
             float* bufD = _dst16b ? Allocate<float>(buf, p.dstH * p.dstW * p.dstC) : NULL;
-            const uint16_t* wgt = _weight.data;
             for (size_t b = 0; b < p.batch; ++b)
             {
-                const uint16_t* src16b = _src16b ? (uint16_t*)src : bufS;
-                float* dst32f = _dst16b ? bufD : (float*)dst;
-                float* buf32f = _is1x1 ? dst32f : bufB;
-                if (!_src16b || a.bufK != a.K)
-                    _convert(src, p, a, 0, p.srcH, bufS);
-                //GemmNN(_M, _N, _K, src16b, _ldS, wgt, _ldW, buf32f, _ldD);
-                //if (!_is1x1)
-                //    ImgToRow(buf32f, dst32f);
-                _biasAct(dst32f, p, a, p.dstC, p.dstH, _bias.data, _params.data, dst);
+                ForwardCommon(src, bufS, bufB, bufD, dst);
                 src += _stepS;
                 dst += _stepD;
             }
         }
 
-        bool SynetDeconvolution16bNhwcGemm::Preferable(const ConvParam& p)
+        void SynetDeconvolution16bNhwcGemm::ForwardCommon(const uint8_t* src, uint16_t* bufS, float* bufB, float* bufD, uint8_t* dst)
+        {
+            const DeconvParam& p = _param;
+            const AlgParam& a = _alg;
+            const uint16_t* src16b = _src16b ? (uint16_t*)src : bufS;
+            float* dst32f = _dst16b ? bufD : (float*)dst;
+            float* buf32f = _is1x1 ? dst32f : bufB;
+            if (!_src16b || a.bufK != a.K)
+                _convert(src, p, a, 0, p.srcH, bufS);
+            GemmCommon(src16b, buf32f);
+            if (!_is1x1)
+                _toImg(buf32f, p, a, p.dstC, 0, p.dstH, dst32f);
+            _biasAct(dst32f, p, a, p.dstC, p.dstH, _bias.data, _params.data, dst);
+        }
+
+        void SynetDeconvolution16bNhwcGemm::GemmCommon(const uint16_t* src, float* dst)
+        {
+            const AlgParam& a = _alg;
+            for (size_t man = 0; man < a.N; man += a.macroN)
+            {
+                size_t macroN = Simd::Min(a.N, man + a.macroN) - man;
+                const uint16_t* wgt = _weight.data + man * a.bufK;
+                for (size_t mak = 0; mak < a.K; mak += a.macroK)
+                {
+                    size_t macroK = Simd::Min(a.bufK, mak + a.macroK) - mak;
+                    for (size_t mam = 0; mam < a.M; mam += a.macroM)
+                    {
+                        size_t macroM = Simd::Min(a.bufM, mam + a.macroM) - mam;
+                        _gemm(src + mam * a.bufK + mak, _param, a, macroM, macroN, macroK, mak == 0 ? 1 : 0, wgt, dst + macroM * a.bufN);
+                    }
+                    wgt += macroK * a.F;
+                }
+                dst += macroN;
+            }
+        }
+
+        bool SynetDeconvolution16bNhwcGemm::Preferable(const DeconvParam& p)
         {
             return false;
         }

diff --git a/src/Simd/SimdSse41SynetDeconvolution16b.cpp b/src/Simd/SimdSse41SynetDeconvolution16b.cpp
@@ -0,0 +1,42 @@
+/*
+* Simd Library (http://ermig1979.github.io/Simd).
+*
+* Copyright (c) 2011-2024 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/SimdSynetDeconvolution16b.h"
+
+namespace Simd
+{
+#if defined(SIMD_SSE41_ENABLE) && defined(SIMD_SYNET_ENABLE) 
+    namespace Sse41
+    {
+        void* SynetDeconvolution16bInit(size_t batch, const SimdConvolutionParameters* conv, SimdSynetCompatibilityType compatibility)
+        {
+            DeconvParam param(batch, conv, compatibility);
+            if (!param.Valid(SimdTensorData32f, SimdTensorData16b))
+                return NULL;
+            if (SynetDeconvolution16bNhwcGemm::Preferable(param))
+                return new Sse41::SynetDeconvolution16bNhwcGemm(param);
+            return new Base::SynetDeconvolution16bGemm(param);
+        }
+    }
+#endif
+}