portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-14 01:35:36 +00:00
Code Releases Activity

Synchronise decouple from the launch

Browse Source 
Speeds up multileg stencils
This commit is contained in:
Peter Boyle 2023-12-22 18:06:13 -05:00
parent 78ab955fec
commit 3d517bbd2a
1 changed files with 14 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
Grid/algorithms/multigrid/BatchedBlas.h
View File

@ -204,13 +204,12 @@ public:
} }
} }
#endif #endif
synchronise();
RealD t1=usecond(); RealD t1=usecond();
RealD flops = 8.0*m*n*k*batchCount; RealD flops = 8.0*m*n*k*batchCount;
RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount; RealD bytes = 1.0*sizeof(ComplexD)*(m*k+k*n+m*n)*batchCount;
std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
} }
void gemmBatched(int m,int n, int k, void gemmBatched(int m,int n, int k,
@ -279,13 +278,12 @@ public:
} }
} }
#endif #endif
synchronise();
RealD t1=usecond(); RealD t1=usecond();
RealD flops = 8.0*m*n*k*batchCount; RealD flops = 8.0*m*n*k*batchCount;
RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount; RealD bytes = 1.0*sizeof(ComplexF)*(m*k+k*n+m*n)*batchCount;
std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
} }
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////
@ -357,13 +355,12 @@ public:
} }
} }
#endif #endif
synchronise();
RealD t1=usecond(); RealD t1=usecond();
RealD flops = 8.0*m*n*k*batchCount; RealD flops = 2.0*m*n*k*batchCount;
RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount; RealD bytes = 1.0*sizeof(RealF)*(m*k+k*n+m*n)*batchCount;
std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
} }
@ -452,13 +449,12 @@ public:
} }
} }
#endif #endif
synchronise();
RealD t1=usecond(); RealD t1=usecond();
RealD flops = 8.0*m*n*k*batchCount; RealD flops = 2.0*m*n*k*batchCount;
RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount; RealD bytes = 1.0*sizeof(RealD)*(m*k+k*n+m*n)*batchCount;
std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas copy "<<(t0-t2)/1.e3 <<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< flops/(t1-t0)/1.e3 <<" GF/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl; // std::cout <<GridLogPerformance<< " batched Blas call "<<m<<","<<n<<","<<k<<" "<< bytes/(t1-t0)/1.e3 <<" GB/s "<<(t1-t0)/1.e3<<" ms "<<std::endl;
} }