*improve Performance of SynetConvolution32f (NHWC, srcC=1, dstС=1).

ermig1979 · Dec 26, 2024 · 6f7b1e4 · 6f7b1e4
1 parent 97d2ed3
commit 6f7b1e4
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 2 deletions.
diff --git a/docs/2025.html b/docs/2025.html
@@ -55,6 +55,7 @@ <h5>Improving</h5>
  <li>Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k7p3d1s1w8.</li>
  <li>Extend using of AVX-512BW optimization of function Convolution32fNhwcDepthwise_k7p3d1s1w4.</li>
  <li>Extend using of AMX-BF16 optimization of function DepthwiseConvolution_k5p2d1s1w8.</li>
+ <li>Performance of SynetConvolution32f (NHWC, srcC=1, dstС=1).</li>
 </ul>
 <h5>Bug fixing</h5>
 <ul>

diff --git a/src/Simd/SimdAvx512bwSynetConvolution32fDirectNchw.cpp b/src/Simd/SimdAvx512bwSynetConvolution32fDirectNchw.cpp
@@ -426,6 +426,8 @@ namespace Simd
                 return false;
             if (!(p.IsStride(1) || p.IsStride(2) || p.IsStride(3)))
                 return false;
+            if (p.srcC == 1 && p.dstC == 1 && p.IsStride(1) && p.IsDilation(1))
+                return true;
             double k = double(p.srcC) / p.group * p.strideX * p.strideX * p.strideY / p.kernelX / p.kernelY;
             return k < 2.0 && ((p.IsStride(1) && p.IsKernel(1)) || p.IsKernel(2) || p.IsKernel(3)
 #if SIMD_ZMM_COUNT == 32 || 1

diff --git a/src/Simd/SimdSse41SynetConvolution32fDirectNchw.cpp b/src/Simd/SimdSse41SynetConvolution32fDirectNchw.cpp
@@ -332,6 +332,8 @@ namespace Simd
             if (!(p.IsStride(1) || p.IsStride(2) || p.IsStride(3)))
                 return false;
             double k = double(p.srcC) / p.group * p.strideX * p.strideX * p.strideY / p.kernelX / p.kernelY;
+            if (p.srcC == 1 && p.dstC == 1 && p.IsStride(1) && p.IsDilation(1))
+                return true;
             return k < 2.0 && ((p.IsStride(1) && p.IsKernel(1)) || p.IsKernel(2) || p.IsKernel(3)) && p.trans == 0;
         }
 

diff --git a/src/Test/TestSynetConvolution32f.cpp b/src/Test/TestSynetConvolution32f.cpp
@@ -244,7 +244,7 @@ namespace Test
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(12, 2000, 5, 5, 255, _1, _1, _1, _0, _0, 1, a, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(24, 2048, 6, 6, 255, _3, _1, _1, _0, _0, 1, a, t), f1, f2);
 #endif
-#if 1
+#if 0
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 816, 14, 14, 816, _5, _1, _1, _2, _2, 816, aRe, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 304, 16, 16, 304, _3, _1, _1, _1, _1, 304, aRe, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 304, 16, 16, 304, _7, _1, _1, _3, _3, 304, aPr, t), f1, f2);
@@ -253,6 +253,9 @@ namespace Test
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 152, 32, 32, 152, _7, _1, _1, _3, _3, 152, aRe, t), f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 76, 64, 64, 76, _7, _1, _1, _3, _3, 76, aRe, t), f1, f2);
 #endif
+#if 1
+        result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 1, 80, 240, 1, _3, _1, _1, _1, _1, 1, aRe, t), f1, f2);
+#endif
 #else
         result = result && SynetConvolution32fForwardAutoTest(eps, Param(1, 256, 44, 44, 256, _1, _1, _1, _0, _0, 1, a, t), f1, f2);
 #endif
@@ -270,7 +273,7 @@ namespace Test
         const SimdBool tF = SimdFalse, tT = SimdTrue;
 
 #ifdef NDEBUG
-        //result = result && SynetConvolution32fForwardAutoTest(eps, SimdConvolutionActivationGelu, tF, f1, f2);
+        result = result && SynetConvolution32fForwardAutoTest(eps, SimdConvolutionActivationGelu, tF, f1, f2);
         result = result && SynetConvolution32fForwardAutoTest(eps, SimdConvolutionActivationRelu, tT, f1, f2);
 #else
         //result = result && SynetConvolution32fForwardAutoTest(eps, SimdConvolutionActivationGelu, tF, f1, f2);