C++性能优化系列——矩阵转置(七)Intrinsic 内存预取与OpenMP多线程并行化

冷不防 2022-11-28 15:18 61阅读 0赞

上一篇[C++性能优化系列——矩阵转置(六)Intrinsic转置实现与Core Bound优化][C_Intrinsic_Core Bound]中通过Intrinsic 实现了矩阵转置功能，同时针对寄存器使用导致的spilling问题，对每次处理的数据块尺寸进行调整。本篇将基于之前的优化版本，在内存访问方面与多线程并行化方面做进一步优化。

## 内存访问优化 ##

上一篇中主要针对寄存器使用问题优化代码，未针对内存访问问题 L3 Latency做处理。考虑Advance2版本中对内存访问的模式，写内存操作是连续执行的，读内存操作是交叉执行的。因此加载内存方式导致该问题的可能性最大。同时转置功能逻辑导致加载内存必须要交叉执行，因此要从别的方面考虑优化这个问题。  
SSE指令集中提供内存预取指令。考虑每次循环中加载下一次要用到的数据。  
代码实现：

void SimdTransposeBlock8RowPadPrefetch(unsigned char* pSource, unsigned char* pTarget)
    	{ 
    		clock_t begin = clock();
    		int iblock = 128;
    		int iwsize = NCOL / iblock;
    		int ihsize = NROW / iblock;
    		for (int i = 0; i < REPEAT; ++i)
    		{ 
    			for (int ibcol = 0; ibcol < iwsize; ++ibcol)
    			{ 
    				for (int ibrow = 0; ibrow < ihsize; ++ibrow)
    				{ 
    					for (int icol = 0; icol < iblock; icol += 8)
    					{ 
    						for (int irow = 0; irow < iblock; irow += 16)
    						{ 
    							__m128i A0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 0) * NREALCOL + ibcol * iblock + icol));
    							__m128i B0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 1) * NREALCOL + ibcol * iblock + icol));
    							__m128i C0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 2) * NREALCOL + ibcol * iblock + icol));
    							__m128i D0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 3) * NREALCOL + ibcol * iblock + icol));
    							__m128i E0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 4) * NREALCOL + ibcol * iblock + icol));
    							__m128i F0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 5) * NREALCOL + ibcol * iblock + icol));
    							__m128i G0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 6) * NREALCOL + ibcol * iblock + icol));
    							__m128i H0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 7) * NREALCOL + ibcol * iblock + icol));
    							__m128i I0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 8) * NREALCOL + ibcol * iblock + icol));
    							__m128i J0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 9) * NREALCOL + ibcol * iblock + icol));
    							__m128i K0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 10) * NREALCOL + ibcol * iblock + icol));
    							__m128i L0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 11) * NREALCOL + ibcol * iblock + icol));
    							__m128i M0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 12) * NREALCOL + ibcol * iblock + icol));
    							__m128i N0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 13) * NREALCOL + ibcol * iblock + icol));
    							__m128i O0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 14) * NREALCOL + ibcol * iblock + icol));
    							__m128i P0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 15) * NREALCOL + ibcol * iblock + icol));
    
    
    							//unpack2
    							__m128i AB0_7 = _mm_unpacklo_epi8(A0_7, B0_7);
    							__m128i CD0_7 = _mm_unpacklo_epi8(C0_7, D0_7);
    							__m128i EF0_7 = _mm_unpacklo_epi8(E0_7, F0_7);
    							__m128i GH0_7 = _mm_unpacklo_epi8(G0_7, H0_7);
    							__m128i IJ0_7 = _mm_unpacklo_epi8(I0_7, J0_7);
    							__m128i KL0_7 = _mm_unpacklo_epi8(K0_7, L0_7);
    							__m128i MN0_7 = _mm_unpacklo_epi8(M0_7, N0_7);
    							__m128i OP0_7 = _mm_unpacklo_epi8(O0_7, P0_7);
    
    
    							//unpack4
    							__m128i ABCD0_3 = _mm_unpacklo_epi16(AB0_7, CD0_7);
    							__m128i EFGH0_3 = _mm_unpacklo_epi16(EF0_7, GH0_7);
    							__m128i IJKL0_3 = _mm_unpacklo_epi16(IJ0_7, KL0_7);
    							__m128i MNOP0_3 = _mm_unpacklo_epi16(MN0_7, OP0_7);
    							__m128i ABCD4_7 = _mm_unpackhi_epi16(AB0_7, CD0_7);
    							__m128i EFGH4_7 = _mm_unpackhi_epi16(EF0_7, GH0_7);
    							__m128i IJKL4_7 = _mm_unpackhi_epi16(IJ0_7, KL0_7);
    							__m128i MNOP4_7 = _mm_unpackhi_epi16(MN0_7, OP0_7);
    
    							//unpack8
    							__m128i A_H0_1 = _mm_unpacklo_epi32(ABCD0_3, EFGH0_3);
    							__m128i I_P0_1 = _mm_unpacklo_epi32(IJKL0_3, MNOP0_3);
    							__m128i A_H2_3 = _mm_unpackhi_epi32(ABCD0_3, EFGH0_3);
    							__m128i I_P2_3 = _mm_unpackhi_epi32(IJKL0_3, MNOP0_3);
    							__m128i A_H4_5 = _mm_unpacklo_epi32(ABCD4_7, EFGH4_7);
    							__m128i I_P4_5 = _mm_unpacklo_epi32(IJKL4_7, MNOP4_7);
    							__m128i A_H6_7 = _mm_unpackhi_epi32(ABCD4_7, EFGH4_7);
    							__m128i I_P6_7 = _mm_unpackhi_epi32(IJKL4_7, MNOP4_7);
    
    							//unpack16 
    							__m128i AP0 = _mm_unpacklo_epi64(A_H0_1, I_P0_1);
    							__m128i AP1 = _mm_unpackhi_epi64(A_H0_1, I_P0_1);
    							__m128i AP2 = _mm_unpacklo_epi64(A_H2_3, I_P2_3);
    							__m128i AP3 = _mm_unpackhi_epi64(A_H2_3, I_P2_3);
    							__m128i AP4 = _mm_unpacklo_epi64(A_H4_5, I_P4_5);
    							__m128i AP5 = _mm_unpackhi_epi64(A_H4_5, I_P4_5);
    							__m128i AP6 = _mm_unpacklo_epi64(A_H6_7, I_P6_7);
    							__m128i AP7 = _mm_unpackhi_epi64(A_H6_7, I_P6_7);
    
    							//store
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 0) * NROW + ibrow * iblock + irow), AP0);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 1) * NROW + ibrow * iblock + irow), AP1);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 2) * NROW + ibrow * iblock + irow), AP2);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 3) * NROW + ibrow * iblock + irow), AP3);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 4) * NROW + ibrow * iblock + irow), AP4);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 5) * NROW + ibrow * iblock + irow), AP5);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 6) * NROW + ibrow * iblock + irow), AP6);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 7) * NROW + ibrow * iblock + irow), AP7);
    
    							//prefetch
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 0) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 1) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 2) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 3) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 4) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 5) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 6) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 7) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 8) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 9) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 10) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 11) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 12) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 13) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 14) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 15) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    
    
    						}
    					}
    				}
    			}
    		}
    		clock_t end = clock();
    		std::cout << "SimdTransposeBlock8RowPadPrefetch 10240 Time " << (end - begin) << std::endl;
    		std::cout << "SimdTransposeBlock8RowPadPrefetch each Transpose (ms) " << ((float)(end - begin)) / (float)REPEAT << std::endl;
    	}

执行时间：

SimdTransposeBlock8RowPadPrefetch 10240 Time 686
    SimdTransposeBlock8RowPadPrefetch each Transpose (ms) 0.0669922

程序在执行时间上有了进一步提升。

## VTune分析 ##

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center]  
可以看到内存访问方面没有明显的性能问题。  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center 1]  
因为实现中添加了prefetch指令，因此总的执行指令数量增加了。但是CPI相对而言是下降的。  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center 2]  
之前的L3 Latency问题已经不是主要的性能热点。

## OpenMP并行化 ##

基于本文中之前实现版本，对程序做多线程并行化优化。因为每次循环中处理的数据块没有交集，因此对for循环做并行化可以保证结果的正确性。  
代码实现：

void SimdTransposeBlock8RowPPP(unsigned char* pSource, unsigned char* pTarget)
    	{ 
    		//PPP -> Pad Prefetch Parallel
    		clock_t begin = clock();
    		int iblock = 128;
    		int iwsize = NCOL / iblock;
    		int ihsize = NROW / iblock;
    		for (int i = 0; i < REPEAT; ++i)
    		{ 
    #pragma omp parallel for num_threads(8) schedule(dynamic)
    			for (int ibcol = 0; ibcol < iwsize; ++ibcol)
    			{ 
    				for (int ibrow = 0; ibrow < ihsize; ++ibrow)
    				{ 
    					for (int icol = 0; icol < iblock; icol += 8)
    					{ 
    						for (int irow = 0; irow < iblock; irow += 16)
    						{ 
    							__m128i A0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 0) * NREALCOL + ibcol * iblock + icol));
    							__m128i B0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 1) * NREALCOL + ibcol * iblock + icol));
    							__m128i C0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 2) * NREALCOL + ibcol * iblock + icol));
    							__m128i D0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 3) * NREALCOL + ibcol * iblock + icol));
    							__m128i E0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 4) * NREALCOL + ibcol * iblock + icol));
    							__m128i F0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 5) * NREALCOL + ibcol * iblock + icol));
    							__m128i G0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 6) * NREALCOL + ibcol * iblock + icol));
    							__m128i H0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 7) * NREALCOL + ibcol * iblock + icol));
    							__m128i I0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 8) * NREALCOL + ibcol * iblock + icol));
    							__m128i J0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 9) * NREALCOL + ibcol * iblock + icol));
    							__m128i K0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 10) * NREALCOL + ibcol * iblock + icol));
    							__m128i L0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 11) * NREALCOL + ibcol * iblock + icol));
    							__m128i M0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 12) * NREALCOL + ibcol * iblock + icol));
    							__m128i N0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 13) * NREALCOL + ibcol * iblock + icol));
    							__m128i O0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 14) * NREALCOL + ibcol * iblock + icol));
    							__m128i P0_7 = _mm_loadl_epi64((__m128i*)(pSource + (ibrow * iblock + irow + 15) * NREALCOL + ibcol * iblock + icol));
    
    
    							//unpack2
    							__m128i AB0_7 = _mm_unpacklo_epi8(A0_7, B0_7);
    							__m128i CD0_7 = _mm_unpacklo_epi8(C0_7, D0_7);
    							__m128i EF0_7 = _mm_unpacklo_epi8(E0_7, F0_7);
    							__m128i GH0_7 = _mm_unpacklo_epi8(G0_7, H0_7);
    							__m128i IJ0_7 = _mm_unpacklo_epi8(I0_7, J0_7);
    							__m128i KL0_7 = _mm_unpacklo_epi8(K0_7, L0_7);
    							__m128i MN0_7 = _mm_unpacklo_epi8(M0_7, N0_7);
    							__m128i OP0_7 = _mm_unpacklo_epi8(O0_7, P0_7);
    
    
    							//unpack4
    							__m128i ABCD0_3 = _mm_unpacklo_epi16(AB0_7, CD0_7);
    							__m128i EFGH0_3 = _mm_unpacklo_epi16(EF0_7, GH0_7);
    							__m128i IJKL0_3 = _mm_unpacklo_epi16(IJ0_7, KL0_7);
    							__m128i MNOP0_3 = _mm_unpacklo_epi16(MN0_7, OP0_7);
    							__m128i ABCD4_7 = _mm_unpackhi_epi16(AB0_7, CD0_7);
    							__m128i EFGH4_7 = _mm_unpackhi_epi16(EF0_7, GH0_7);
    							__m128i IJKL4_7 = _mm_unpackhi_epi16(IJ0_7, KL0_7);
    							__m128i MNOP4_7 = _mm_unpackhi_epi16(MN0_7, OP0_7);
    
    							//unpack8
    							__m128i A_H0_1 = _mm_unpacklo_epi32(ABCD0_3, EFGH0_3);
    							__m128i I_P0_1 = _mm_unpacklo_epi32(IJKL0_3, MNOP0_3);
    							__m128i A_H2_3 = _mm_unpackhi_epi32(ABCD0_3, EFGH0_3);
    							__m128i I_P2_3 = _mm_unpackhi_epi32(IJKL0_3, MNOP0_3);
    							__m128i A_H4_5 = _mm_unpacklo_epi32(ABCD4_7, EFGH4_7);
    							__m128i I_P4_5 = _mm_unpacklo_epi32(IJKL4_7, MNOP4_7);
    							__m128i A_H6_7 = _mm_unpackhi_epi32(ABCD4_7, EFGH4_7);
    							__m128i I_P6_7 = _mm_unpackhi_epi32(IJKL4_7, MNOP4_7);
    
    							//unpack16 
    							__m128i AP0 = _mm_unpacklo_epi64(A_H0_1, I_P0_1);
    							__m128i AP1 = _mm_unpackhi_epi64(A_H0_1, I_P0_1);
    							__m128i AP2 = _mm_unpacklo_epi64(A_H2_3, I_P2_3);
    							__m128i AP3 = _mm_unpackhi_epi64(A_H2_3, I_P2_3);
    							__m128i AP4 = _mm_unpacklo_epi64(A_H4_5, I_P4_5);
    							__m128i AP5 = _mm_unpackhi_epi64(A_H4_5, I_P4_5);
    							__m128i AP6 = _mm_unpacklo_epi64(A_H6_7, I_P6_7);
    							__m128i AP7 = _mm_unpackhi_epi64(A_H6_7, I_P6_7);
    
    							//store
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 0) * NROW + ibrow * iblock + irow), AP0);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 1) * NROW + ibrow * iblock + irow), AP1);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 2) * NROW + ibrow * iblock + irow), AP2);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 3) * NROW + ibrow * iblock + irow), AP3);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 4) * NROW + ibrow * iblock + irow), AP4);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 5) * NROW + ibrow * iblock + irow), AP5);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 6) * NROW + ibrow * iblock + irow), AP6);
    							_mm_storeu_si128((__m128i*)(pTarget + (ibcol * iblock + icol + 7) * NROW + ibrow * iblock + irow), AP7);
    
    							//prefetch
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 0) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 1) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 2) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 3) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 4) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 5) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 6) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 7) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 8) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 9) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 10) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 11) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 12) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 13) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 14) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    							_mm_prefetch((const char*)(pSource + (ibrow * iblock + irow + 15) * NREALCOL + ibcol * iblock + icol + 8), _MM_HINT_T0);
    
    
    						}
    					}
    				}
    			}
    		}
    		clock_t end = clock();
    		std::cout << "SimdTransposeBlock8RowPPP 1024 Time " << (end - begin) << std::endl;
    		std::cout << "SimdTransposeBlock8RowPPP each Transpose (ms) " << ((float)(end - begin)) / (float)REPEAT << std::endl;
    	}

执行时间

SimdTransposeBlock8RowPPP 1024 Time 205
    SimdTransposeBlock8RowPPP each Transpose (ms) 0.0200195

通过多线程加速，获得的加速比为3.43.

## 总结 ##

本文通过内存预取技术，对矩阵转置功能中内存加载的L3 Cache问题进行优化。同时，通过OPenMP对程序做多线程并行化优化，获得3.43倍的加速比。

[C_Intrinsic_Core Bound]: https://blog.csdn.net/yan31415/article/details/107464744
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center]: /images/20221124/19e2aec201f5418c86d4a800e85cb4af.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center 1]: /images/20221124/cac0a1672046447ea59e3401be8d06af.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3lhbjMxNDE1_size_16_color_FFFFFF_t_70_pic_center 2]: /images/20221124/a3aaf9055ffa41518e4c7c8f24716a10.png