chundoong-lab-ta/SamsungDS22/submissions/HW4/jongho62.kim/mat_mul.cpp

119 lines
3.1 KiB
C++

#include "mat_mul.h"
#include <cstdio>
#include <cstdlib>
#include <mpi.h>
#include <omp.h>
#include "util.h"
//#define min(a, b) (((a) < (b)) ? (a) : (b))
#define MASTER 0
#define FORM_MASTER 1
#define FORM_WORKER 2
static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int rows[4]={0,};
static int offset[4]={0,};
#define SLICEM 32
#define SLICEK 32
#define SLICEN 2048
static void mat_mul_omp() {
// TODO: parallelize & optimize matrix multiplication
// #pragma omp parallel num_threads(num_threads)
// #pragma omp parallel for
// for (int i = 0; i < rows; ++i) {
// for (int k = 0; k < K; ++k) {
// float arr = A[i*K+k];
// for (int j = 0; j < N; j+=1) {
// C[i * N + j] += arr * B[k * N + j];
// }
// }
// }
int start=0;
int ed=rows[mpi_rank];
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
for(int i2=start;i2<ed;i2+=SLICEM) {
for(int k2=0;k2<K;k2+=SLICEK) {
for(int j2=0;j2<N;j2+=SLICEN) {
int edk=k2+SLICEK<K?(k2+SLICEK):K;
int edm=i2+SLICEM<M?(i2+SLICEM):M;
int edn=j2+SLICEN<N?(j2+SLICEN):N;
for(int i=i2;i<edm;++i) {
for(int k=k2;k<edk;++k) {
for(int j=j2;j<edn;++j) {
C[i*N+j]+=A[i*K+k]*B[k*N+j];
}
}
}
}
}
}
}
void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K,
int _num_threads, int _mpi_rank, int _mpi_world_size) {
A = _A, B = _B, C = _C;
M = _M, N = _N, K = _K;
num_threads = _num_threads, mpi_rank = _mpi_rank,
mpi_world_size = _mpi_world_size;
// TODO: parallelize & optimize matrix multiplication on multi-node
// You must allocate & initialize A, B, C for non-root processes
//int numworkers = mpi_world_size -1;
//int averow, extra;
MPI_Status status;
MPI_Request request;
// FIXME: for now, only root process runs the matrix multiplication.
// if (mpi_rank == 0)
// mat_mul_omp();
int nrows = M / mpi_world_size;
//averow = M/numworkers;
//extra = M%numworkers;
//offset = 0;
//calcurate rows numbers
for(int i=0;i<mpi_world_size;i++) {
rows[i]=(i==mpi_world_size-1)?(M-(nrows *(mpi_world_size-1))):nrows;
}
//calculate offsets
for(int i=0;i<mpi_world_size-1;i++) {
offset[i+1]=offset[i]+rows[i];
}
if(mpi_rank != 0) {
M=rows[mpi_rank];
alloc_mat(&A, rows[mpi_rank], K);
alloc_mat(&B, K, N);
alloc_mat(&C, rows[mpi_rank], N);
}
MPI_Bcast(B,K*N,MPI_FLOAT,0,MPI_COMM_WORLD);
//isend / recv
if(mpi_rank == 0) {
for(int i=1;i<mpi_world_size;i++)
MPI_Isend(&A[offset[i]*K],rows[i]*K,MPI_FLOAT,i,0,MPI_COMM_WORLD,&request);
} else {
MPI_Recv(A,rows[mpi_rank]*K,MPI_FLOAT,0,0,MPI_COMM_WORLD,&status);
}
mat_mul_omp();
if(mpi_rank != 0) {
MPI_Isend(C,rows[mpi_rank]*N,MPI_FLOAT,0,0,MPI_COMM_WORLD,&request);
} else {
for(int i=1;i<mpi_world_size;i++)
MPI_Recv(&C[offset[i]*N],rows[i]*N,MPI_FLOAT,i,0,MPI_COMM_WORLD,&status);
}
}