for (Y = 0; Y < Height; Y++)
{for (X = 0; X < Width; X++){Index = .....;Sum = 0;for (XX = 0; XX < ConvW; XX++){for (YY = 0; YY < ConvH; YY++){Index1 = ..... ;Index2 = ..... ;Sum += Conv[Index1] * Pixel[Index2];}}Dest[Index] = Sum / Weight;}
}

float *Conv16 = (float *)_mm_malloc(PadConvLine * ConvH * sizeof(float), 16);                        //    保存16字节对齐的卷积矩阵，以方便使用SSEfor(Y = 0; Y < ConvH; Y++)
{memcpy (Conv16 + Y * PadConvLine, Conv->Data.F + Y * ConvW , ConvW * sizeof(float));            //    复制卷积矩阵的数据memset(Conv16 + Y * PadConvLine + ConvW, 0, (PadConvLine - ConvW) * sizeof(float));                //    把冗余部分的卷积数据设置为0
}

float MultiplySSE(float *Kernel, float *Conv, int Length)
{int Block;  const float *Data;                        // 将SSE变量上的多个数值合并时所用指针.float Sum = 0;if (Length > 16)                        //    可以进行四次SSE计算，测试表明，这个还是快些的    {const int BlockWidth = 4 * 4;        // 块宽. SSE寄存器能一次处理4个float，然后循环展开4次.Block = Length / BlockWidth;        // 块数.    float *KernelP = Kernel, *ConvP = Conv;                // SSE批量处理时所用的指针.__m128 Sum0 = _mm_setzero_ps();         // 求和变量。SSE赋初值0__m128 Sum1 = _mm_setzero_ps();__m128 Sum2 = _mm_setzero_ps();__m128 Sum3 = _mm_setzero_ps();for(int I = 0; I < Block; I++){Sum0 = _mm_add_ps(Sum0, _mm_mul_ps(_mm_load_ps(KernelP), _mm_load_ps(ConvP)));                    // SSE单精浮点紧缩加法Sum1 = _mm_add_ps(Sum1, _mm_mul_ps(_mm_load_ps(KernelP + 4), _mm_load_ps(ConvP + 4)));Sum2 = _mm_add_ps(Sum2, _mm_mul_ps(_mm_load_ps(KernelP + 8), _mm_load_ps(ConvP + 8)));Sum3 = _mm_add_ps(Sum3, _mm_mul_ps(_mm_load_ps(KernelP + 12), _mm_load_ps(ConvP + 12)));KernelP += BlockWidth;ConvP += BlockWidth;}Sum0 = _mm_add_ps(Sum0, Sum1);    // 两两合并(0~1).Sum2 = _mm_add_ps(Sum2, Sum3);    // 两两合并(2~3).Sum0 = _mm_add_ps(Sum0, Sum2);    // 两两合并(0~2).Data = (const float *)&Sum0;Sum = Data[0] + Data[1] + Data[2] + Data[3];Length = Length - Block * BlockWidth;            // 剩余数量.}if (Length != 0){const int BlockWidth = 4;                        //    程序已经保证了数量必然是4的倍数 Block = Length / BlockWidth;        float *KernelP = Kernel, *ConvP = Conv;                __m128 Sum0 = _mm_setzero_ps();        for(int I = 0; I < Block; I++){Sum0 = _mm_add_ps(Sum0, _mm_mul_ps(_mm_load_ps(KernelP), _mm_load_ps(ConvP)));        KernelP += BlockWidth;ConvP += BlockWidth;}Data = (const float *)&Sum0;Sum += Data[0] + Data[1] + Data[2] + Data[3];}return Sum;
}

IS_RET __stdcall Conv2(TImage *Src, TMatrix *Conv, TImage *Dest, EdgeMode Edge)
{if (Src == NULL || Dest == NULL || Conv == NULL) return IS_RET_ERR_PARA;if (Src->Width != Dest->Width || Src->Height != Dest->Height || Src->BitCount != Dest->BitCount || Src->Stride != Dest->Stride) return IS_RET_ERR_PARA;if (Src->Scan0 == NULL || Dest->Scan0 == NULL || Conv->Data.F == NULL) return IS_RET_ERR_MEM;if (Conv->Width < 1 || Conv->Height < 1) return IS_RET_ERR_PARA;int Width = Src->Width, Height = Src->Height, Stride = Src->Stride;int ConvW = Conv->Width, ConvH = Conv->Height;unsigned char *PtSrc = Src->Scan0, *PtDest = Dest->Scan0;if (Src->BitCount == 24){}else{int Left = ConvW / 2, Top = ConvH / 2, Right = ConvW - Left - 1, Bottom = ConvH - Top - 1, ExpHeight = Height + ConvH - 1;        //    注意核中心那个元素不用扩展，比如核的宽度为3，则只要左右各扩展一个像素就可以了int PadConvLine = Pad4(ConvW), Length = PadConvLine * ConvH;int X, Y, IndexD, IndexE, IndexK, ExpStride;float *CurKer, Inv, Sum = 0;unsigned char *PtExp, *PtDest;TImage *Expand;IS_RET Ret = GetPadImage(Src, &Expand, Left, Right, Top, Bottom, Edge);                                //    得到扩展后的数据，可以提速和方便编程，但是多占用一份内存if (Ret != IS_RET_OK) return Ret;PtExp = Expand->Scan0; PtDest = Dest->Scan0; ExpStride = Expand->Stride;for (X = 0; X < ConvH * ConvW; X ++) Sum += Conv->Data.F[X];Inv = (Sum == 0 ? 1: 1 / Sum);                                                                        //    如果卷积举证的和为0，则设置为1float *Conv16 = (float *)_mm_malloc(PadConvLine * ConvH * sizeof(float), 16);                        //    保存16字节对齐的卷积矩阵，以方便使用SSEfloat *Kernel = (float *)_mm_malloc(PadConvLine * ExpHeight * sizeof(float), 16);                    //    保存16字节对齐的卷积核矩阵，以方便使用SSEfor(Y = 0; Y < ConvH; Y++) {memcpy (Conv16 + Y * PadConvLine, Conv->Data.F + Y * ConvW , ConvW * sizeof(float));            //    复制卷积矩阵的数据memset(Conv16 + Y * PadConvLine + ConvW, 0, (PadConvLine - ConvW) * sizeof(float));                //    把冗余部分的卷积数据设置为0}for (Y = 0; Y < ExpHeight; Y++){IndexE = Y * ExpStride;CurKer = Kernel + Y * PadConvLine;                        //    计算第一列所有像素将要取样的卷积核数据for (X = 0; X < ConvW; X++){CurKer[X] = PtExp[IndexE++];}}for (X = 0 ; X < Width ; X ++){if (X != 0)                                                //    如果不是第一列，需要更新卷积核的数据{memcpy(Kernel, Kernel + 1, (PadConvLine * ExpHeight - 1) * sizeof(float));    //    往前移动一个数据IndexK = ConvW - 1 ;IndexE = IndexK + X;for (Y = 0; Y < ExpHeight; Y++){Kernel[IndexK] = PtExp[IndexE];        //    只要刷新下一个元素IndexK += PadConvLine;IndexE += ExpStride;}}CurKer = Kernel;    IndexD = X;for (Y = 0; Y < Height; Y ++)                            //    沿列的方向进行更新{PtDest[IndexD] = Clamp((int)( MultiplySSE(Conv16, CurKer, Length) * Inv + 0.5));        //    直接把函数放在这里也没有啥提速的，注意改函数不会被内联的CurKer += PadConvLine;IndexD += Stride;}}_mm_free(Conv16);_mm_free(Kernel);FreeImage(Expand);return IS_RET_OK;}
}