*refactoring of SynetConvolution16bNhwcSpecV0 class.

Author: ermig1979

src/Simd/SimdAvx2SynetConvolution16b.cpp

@@ -33,8 +33,8 @@ namespace Simd
             ConvParam param(batch, conv, compatibility);
             if (!param.Valid(SimdTensorData32f, SimdTensorData16b))
                 return NULL;
-            if (SynetConvolution16bNhwcSpecV1::Preferable(param))
-                return new Avx2::SynetConvolution16bNhwcSpecV1(param);
+            //if (SynetConvolution16bNhwcSpecV1::Preferable(param))
+            //    return new Avx2::SynetConvolution16bNhwcSpecV1(param);
             if (SynetConvolution16bNhwcSpecV0::Preferable(param))
                 return new Avx2::SynetConvolution16bNhwcSpecV0(param);
             if (SynetConvolution16bNhwcGemm::Preferable(param))

src/Simd/SimdBaseSynetConvolution16bNhwcSpecV0.cpp

@@ -67,6 +67,7 @@ namespace Simd
             a.K = p.kernelX * p.kernelY;
 
             a.macroC = Simd::RestrictRange(AlignLo(L1 / a.microD / a.K / 2, a.microC), a.microC, a.srcC);
+            a.macroO = a.macroC * a.K / a.microC;
             a.batch = 1;
             size_t bufSize = a.srcC * a.srcH * a.srcW * 2;
             if (bufSize * 2 <= L2 && p.batch > 1)
@@ -84,6 +85,13 @@ namespace Simd
 
             _stepS = p.srcH * p.srcW * p.srcC * a.batch * _elemS;
             _stepD = p.dstH * p.dstW * p.dstC * a.batch * _elemD;
+
+            int dX = (int)a.microC, dY = (int)a.srcW * dX, dC = dY * int(a.srcH * a.batch);
+            _offset.Resize(DivHi(a.srcC, a.microC) * a.K);
+            for (size_t c = 0, offsS = 0, i = 0; c < a.srcC; c += dX, offsS += dC)
+                for (size_t y = 0, offsY = offsS; y < p.kernelY; y += 1, offsY += dY)
+                    for (size_t offsX = offsY, endX = offsY + p.kernelX * dX; offsX < endX; offsX += dX, i++)
+                        _offset[i] = (int)offsX;
         }
 
         size_t SynetConvolution16bNhwcSpecV0::ExternalBufferSize() const

src/Simd/SimdSynetConvolution16b.h

@@ -177,7 +177,7 @@ namespace Simd
             {
                 size_t batch, srcC, srcH, srcW, dstC, K;
                 size_t F, microD, microS, microC;
-                size_t macroD, macroH, macroC, numH;
+                size_t macroD, macroH, macroC, numH, macroO;
                 size_t bufS, bufD, elem;
                 Array32i offs;
             };
@@ -194,6 +194,7 @@ namespace Simd
             void Forward(const uint8_t* src, uint16_t* buf, float* sum, uint8_t* dst);
 
             AlgParam _alg;
+            Array32i _offset;
             PreprocessPtr _preprocess;
             ConvolutionPtr _convolution;
             PostprocessPtr _postprocess;

src/Test/TestSynetConvolution16b.cpp

@@ -363,7 +363,9 @@ namespace Test
 #endif
 #if 1
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 56, 48, 48, 56, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
+        result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 64, 48, 48, 56, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 112, 24, 24, 112, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
+        result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 128, 24, 24, 112, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 224, 12, 12, 224, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(1, 448, 6, 6, 448, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);
         result = result && SynetConvolution16bForwardAutoTest(eps, Param(10, 448, 6, 6, 448, _3, _1, _1, _1, _1, 1, aPr, tT, b16, b16), c, f1, f2);