• cpu GEMM 之 微数据复用 寄存器优化

    一，三重for循环方式

    存储为 helloGemm_1x1.cpp

    编译：

    g++ hello_gemm_1x1.cpp -o hello_gemm_1x1.out -O3

    1. #include
    2. #include
    3.  
    4.  
    5. using namespace std;
    6.  
    7. void gemm_A_N_B_T_1x1(float* A, float* B, float* C, int M_, int N_, int K_){
    8.  
    9.         for(int i=0; i
    10.                 for(int j=0; j
    11.                         for(int k=0; k
    12.                                 C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    13.                         }
    14.                 }
    15.         }
    16.  
    17. }
    18.  
    19. void gemm_A_N_B_T_2x2(float* A, float* B, float* C, int M_, int N_, int K_){
    20.  
    21.         register float a1, a2, b1, b2;
    22.  
    23.         for(int i=0; i2; i++){
    24.                 for(int j=0; j2; j++){
    25.                         for(int k=0; k
    26.                                 a1 = A[2*i     + k*M_   ];//2*i   k
    27.                                 a2 = A[(2*i+1) + k*M_   ];//2*i+1 k
    28.                                 b1 = B[k*N_    + 2*j    ];//k     2*j
    29.                                 b2 = B[k*N_    + (2*j+1)];//k     2*j+1
    30.  
    31.                                 C[2*i     + 2*j*M_    ] += a1*b1;
    32.                                 C[2*i     + (2*j+1)*M_] += a1*b2;
    33.                                 C[(2*i+1) + 2*j*M_    ] += a2*b1;
    34.                                 C[2*i+1   + (2*j+1)*M_] += a2*b2;
    35. //                                C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    36.                         }
    37.                 }
    38.         }
    39.  
    40. }
    41.  
    42.  
    43.  
    44. void init_mat(float* A, int count, int mod){
    45.         for(int idx=0; idx
    46.                 A[idx] = idx%mod + 1;
    47.         }
    48. }
    49.  
    50.  
    51. void print_mat(float* A, int M, int N, bool colMajor){
    52.  
    53.         cout<
    54.  
    55.         for(int i=0; i
    56.                 for(int j=0; j
    57.                         cout<<"  "<<(colMajor? A[i + j*M]: A[i*N + j]);
    58.                 }
    59.                 cout<
    60.         }
    61. }
    62.  
    63.  
    64.  
    65. int main(){
    66.  
    67.  
    68.         int M = 1024;
    69.         int N = 1024;
    70.         int K = 1024;
    71.  
    72.         float* A = nullptr;
    73.         float* B = nullptr;
    74.         float* C_1x1 = nullptr;
    75.         float* C_2x2 = nullptr;
    76.  
    77.         A = (float*)malloc(M*K*sizeof(float));
    78.         B = (float*)malloc(K*N*sizeof(float));
    79.         C_1x1 = (float*)malloc(M*N*sizeof(float));
    80.         C_2x2 = (float*)malloc(M*N*sizeof(float));
    81.  
    82.         init_mat(A, M*K, 3);
    83.         init_mat(B, K*N, 4);
    84.  
    85. //      print_mat(A, M, K, true);
    86. //      print_mat(B, K, N, false);
    87.  
    88.         gemm_A_N_B_T_1x1(A, B, C_1x1, M, N, K);
    89. //      gemm_A_N_B_T_2x2(A, B, C_2x2, M, N, K);
    90.  
    91.         //print_mat(C_1x1, M, N, true);
    92.         //print_mat(C_2x2, M, N, true);
    93.  
    94.         cout<"C(M,N)"<< C_1x1[M*N-1]<
    95. //      cout<
    96.  
    97.  
    98.  
    99.  
    100.  
    101.         return 0;
    102. }

    二，block2x2 register 优化

    存储为 helloGemm_2x2.cpp

    编译：

    g++ hello_gemm_2x2.cpp -o hello_gemm_2x2.out -O3

    1. #include
    2. #include
    3.  
    4.  
    5. using namespace std;
    6.  
    7. void gemm_A_N_B_T_1x1(float* A, float* B, float* C, int M_, int N_, int K_){
    8.  
    9.         for(int i=0; i
    10.                 for(int j=0; j
    11.                         for(int k=0; k
    12.                                 C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    13.                         }
    14.                 }
    15.         }
    16.  
    17. }
    18.  
    19. void gemm_A_N_B_T_2x2(float* A, float* B, float* C, int M_, int N_, int K_){
    20.  
    21.         register float a1, a2, b1, b2;
    22.  
    23.         for(int i=0; i2; i++){
    24.                 for(int j=0; j2; j++){
    25.                         for(int k=0; k
    26.                                 a1 = A[2*i     + k*M_   ];//2*i   k
    27.                                 a2 = A[(2*i+1) + k*M_   ];//2*i+1 k
    28.                                 b1 = B[k*N_    + 2*j    ];//k     2*j
    29.                                 b2 = B[k*N_    + (2*j+1)];//k     2*j+1
    30.  
    31.                                 C[2*i     + 2*j*M_    ] += a1*b1;
    32.                                 C[2*i     + (2*j+1)*M_] += a1*b2;
    33.                                 C[(2*i+1) + 2*j*M_    ] += a2*b1;
    34.                                 C[2*i+1   + (2*j+1)*M_] += a2*b2;
    35. //                                C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    36.                         }
    37.                 }
    38.         }
    39.  
    40. }
    41.  
    42.  
    43.  
    44. void init_mat(float* A, int count, int mod){
    45.         for(int idx=0; idx
    46.                 A[idx] = idx%mod + 1;
    47.         }
    48. }
    49.  
    50.  
    51. void print_mat(float* A, int M, int N, bool colMajor){
    52.  
    53.         cout<
    54.  
    55.         for(int i=0; i
    56.                 for(int j=0; j
    57.                         cout<<"  "<<(colMajor? A[i + j*M]: A[i*N + j]);
    58.                 }
    59.                 cout<
    60.         }
    61. }
    62.  
    63.  
    64.  
    65. int main(){
    66.  
    67.  
    68.         int M = 1024;
    69.         int N = 1024;
    70.         int K = 1024;
    71.  
    72.         float* A = nullptr;
    73.         float* B = nullptr;
    74.         float* C_1x1 = nullptr;
    75.         float* C_2x2 = nullptr;
    76.  
    77.         A = (float*)malloc(M*K*sizeof(float));
    78.         B = (float*)malloc(K*N*sizeof(float));
    79.         C_1x1 = (float*)malloc(M*N*sizeof(float));
    80.         C_2x2 = (float*)malloc(M*N*sizeof(float));
    81.  
    82.         init_mat(A, M*K, 3);
    83.         init_mat(B, K*N, 4);
    84.  
    85. //      print_mat(A, M, K, true);
    86. //      print_mat(B, K, N, false);
    87.  
    88. //      gemm_A_N_B_T_1x1(A, B, C_1x1, M, N, K);
    89.         gemm_A_N_B_T_2x2(A, B, C_2x2, M, N, K);
    90.  
    91.         //print_mat(C_1x1, M, N, true);
    92.         //print_mat(C_2x2, M, N, true);
    93.  
    94. //      cout<
    95.         cout<"C(M,N)"<< C_2x2[M*N-1]<
    96.  
    97.  
    98.  
    99.  
    100.  
    101.         return 0;
    102. }

    (m, n, k) = (1024, 1024, 1024)

    加速效果明显，  效果图： 2 second    VS   5 second

    加入 openmp后，并行效果更加明显：

    三，使用openmp对三重for循环优化

    将测试 size 全部改成：  ((m, n, k) = (1024, 1024, 4096))

    在上面个的两个版本基础上，加入openmp加速指令后的代码为：

    存储为 helloGemm_1x1_omp.cpp

    编译：

     g++ hello_gemm_1x1_omp.cpp -o hello_gemm_1x1_omp.out -O3 -fopenmp

    1. #include
    2. #include
    3. #include 
    4.  
    5. using namespace std;
    6.  
    7. void gemm_A_N_B_T_1x1(float* A, float* B, float* C, int M_, int N_, int K_){
    8.  
    9. #pragma omp parallel for num_threads(omp_get_num_procs())
    10.         for(int i=0; i
    11.                 for(int j=0; j
    12.                         for(int k=0; k
    13.                                 C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    14.                         }
    15.                 }
    16.         }
    17.  
    18. }
    19.  
    20. void gemm_A_N_B_T_2x2(float* A, float* B, float* C, int M_, int N_, int K_){
    21.  
    22. //      register float a1, a2, b1, b2;
    23. #pragma omp parallel for num_threads(omp_get_num_procs())
    24.         for(int i=0; i2; i++){
    25.                 for(int j=0; j2; j++){
    26.                         for(int k=0; k
    27.                                 register float a1, a2, b1, b2;
    28.  
    29.                                 a1 = A[2*i     + k*M_   ];//2*i   k
    30.                                 a2 = A[(2*i+1) + k*M_   ];//2*i+1 k
    31.                                 b1 = B[k*N_    + 2*j    ];//k     2*j
    32.                                 b2 = B[k*N_    + (2*j+1)];//k     2*j+1
    33.  
    34.                                 C[2*i     + 2*j*M_    ] += a1*b1;
    35.                                 C[2*i     + (2*j+1)*M_] += a1*b2;
    36.                                 C[(2*i+1) + 2*j*M_    ] += a2*b1;
    37.                                 C[2*i+1   + (2*j+1)*M_] += a2*b2;
    38. //                                C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    39.                         }
    40.                 }
    41.         }
    42.  
    43. }
    44.  
    45.  
    46.  
    47. void init_mat(float* A, int count, int mod){
    48.         for(int idx=0; idx
    49.                 A[idx] = idx%mod + 1;
    50.         }
    51. }
    52.  
    53.  
    54. void print_mat(float* A, int M, int N, bool colMajor){
    55.  
    56.         cout<
    57.  
    58.         for(int i=0; i
    59.                 for(int j=0; j
    60.                         cout<<"  "<<(colMajor? A[i + j*M]: A[i*N + j]);
    61.                 }
    62.                 cout<
    63.         }
    64. }
    65.  
    66.  
    67.  
    68. int main(){
    69.  
    70.  
    71.         int M = 1024;
    72.         int N = 1024;
    73.         int K = 4096;
    74.  
    75.         float* A = nullptr;
    76.         float* B = nullptr;
    77.         float* C_1x1 = nullptr;
    78.         float* C_2x2 = nullptr;
    79.  
    80.         A = (float*)malloc(M*K*sizeof(float));
    81.         B = (float*)malloc(K*N*sizeof(float));
    82.         C_1x1 = (float*)malloc(M*N*sizeof(float));
    83.         C_2x2 = (float*)malloc(M*N*sizeof(float));
    84.  
    85.         init_mat(A, M*K, 3);
    86.         init_mat(B, K*N, 4);
    87.  
    88. //      print_mat(A, M, K, true);
    89. //      print_mat(B, K, N, false);
    90.  
    91.         gemm_A_N_B_T_1x1(A, B, C_1x1, M, N, K);
    92. //      gemm_A_N_B_T_2x2(A, B, C_2x2, M, N, K);
    93.  
    94.         //print_mat(C_1x1, M, N, true);
    95.         //print_mat(C_2x2, M, N, true);
    96.  
    97.         cout<"C(M,N)"<< C_1x1[M*N-1]<
    98. //      cout<
    99.  
    100.  
    101.  
    102.  
    103.  
    104.         return 0;
    105. }

    四，对block2x2 register 优化加入 openmp加速

    存储为 helloGemm_2x2_omp.cpp

    g++ hello_gemm_2x2_omp.cpp -o hello_gemm_2x2_omp.out -O3 -fopenmp

    1. #include
    2. #include
    3. #include 
    4.  
    5. using namespace std;
    6.  
    7. void gemm_A_N_B_T_1x1(float* A, float* B, float* C, int M_, int N_, int K_){
    8.  
    9. #pragma omp parallel for num_threads(omp_get_num_procs())
    10.         for(int i=0; i
    11.                 for(int j=0; j
    12.                         for(int k=0; k
    13.                                 C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    14.                         }
    15.                 }
    16.         }
    17.  
    18. }
    19.  
    20. void gemm_A_N_B_T_2x2(float* A, float* B, float* C, int M_, int N_, int K_){
    21.  
    22. //      register float a1, a2, b1, b2;
    23. #pragma omp parallel for num_threads(omp_get_num_procs())
    24.         for(int i=0; i2; i++){
    25.                 for(int j=0; j2; j++){
    26.                         for(int k=0; k
    27.                                 register float a1, a2, b1, b2;
    28.  
    29.                                 a1 = A[2*i     + k*M_   ];//2*i   k
    30.                                 a2 = A[(2*i+1) + k*M_   ];//2*i+1 k
    31.                                 b1 = B[k*N_    + 2*j    ];//k     2*j
    32.                                 b2 = B[k*N_    + (2*j+1)];//k     2*j+1
    33.  
    34.                                 C[2*i     + 2*j*M_    ] += a1*b1;
    35.                                 C[2*i     + (2*j+1)*M_] += a1*b2;
    36.                                 C[(2*i+1) + 2*j*M_    ] += a2*b1;
    37.                                 C[2*i+1   + (2*j+1)*M_] += a2*b2;
    38. //                                C[i + j*M_] += A[i + k*M_]*B[k*N_ + j];
    39.                         }
    40.                 }
    41.         }
    42.  
    43. }
    44.  
    45.  
    46.  
    47. void init_mat(float* A, int count, int mod){
    48.         for(int idx=0; idx
    49.                 A[idx] = idx%mod + 1;
    50.         }
    51. }
    52.  
    53.  
    54. void print_mat(float* A, int M, int N, bool colMajor){
    55.  
    56.         cout<
    57.  
    58.         for(int i=0; i
    59.                 for(int j=0; j
    60.                         cout<<"  "<<(colMajor? A[i + j*M]: A[i*N + j]);
    61.                 }
    62.                 cout<
    63.         }
    64. }
    65.  
    66.  
    67.  
    68. int main(){
    69.  
    70.  
    71.         int M = 1024;//2048;
    72.         int N = 1024;//2048;
    73.         int K = 4096;
    74.  
    75.         float* A = nullptr;
    76.         float* B = nullptr;
    77.         float* C_1x1 = nullptr;
    78.         float* C_2x2 = nullptr;
    79.  
    80.         A = (float*)malloc(M*K*sizeof(float));
    81.         B = (float*)malloc(K*N*sizeof(float));
    82.         C_1x1 = (float*)malloc(M*N*sizeof(float));
    83.         C_2x2 = (float*)malloc(M*N*sizeof(float));
    84.  
    85.         init_mat(A, M*K, 3);
    86.         init_mat(B, K*N, 4);
    87.  
    88. //      print_mat(A, M, K, true);
    89. //      print_mat(B, K, N, false);
    90.  
    91. //      gemm_A_N_B_T_1x1(A, B, C_1x1, M, N, K);
    92.         gemm_A_N_B_T_2x2(A, B, C_2x2, M, N, K);
    93.  
    94.         //print_mat(C_1x1, M, N, true);
    95.         //print_mat(C_2x2, M, N, true);
    96.  
    97. //      cout<
    98.         cout<"C(M,N)"<< C_2x2[M*N-1]<
    99.  
    100.  
    101.  
    102.  
    103.  
    104.         return 0;
    105. }

    编译图：

    效果图：

    优化效果统计：

    三重for循环：                       43 s

    block2x2 register：              11 s

    openmp 三重for循环：          2 s

    openmp block2x2 register:    1s

  • 相关阅读:
    智汀家庭云究竟有什么魔力？实现跨品牌联动
    [阶段4 企业开发进阶] 3. 消息队列--RabbitMQ
    yolov5 tensorrt 精度对齐总结
    Transformer为什么如此有效 | 通用建模能力，并行
    chmod的用法，及几个权限的宏
    网络安全漏洞分析之远程代码执行
    23种设计模式之单例模式
    CDC实时数据同步
    将文字转为UTF-8编码(即HTML可识别的)
    基于MetaTown构建数字资产平台
  • 原文地址：https://blog.csdn.net/eloudy/article/details/126698509