104 lines
2.8 KiB
C++
104 lines
2.8 KiB
C++
|
#include "mat_mul.h"
|
||
|
|
#include "util.h"
|
||
|
|
|
||
|
|
#include <cstdio>
|
||
|
|
#include <cstdlib>
|
||
|
|
#include <mpi.h>
|
||
|
|
#include <omp.h>
|
||
|
|
|
||
|
|
static float *A, *B, *C;
|
||
|
|
static int M, N, K;
|
||
|
|
static int num_threads;
|
||
|
|
static int mpi_rank, mpi_world_size;
|
||
|
|
static int offset,rows;
|
||
|
|
|
||
|
|
int min(int a,int b)
|
||
|
|
{
|
||
|
|
return a<b?a:b;
|
||
|
|
}
|
||
|
|
|
||
|
|
static void mat_mul_omp() {
|
||
|
|
// TODO: parallelize & optimize matrix multiplication
|
||
|
|
// Use num_threads per node
|
||
|
|
int index, start, end;
|
||
|
|
|
||
|
|
int offset = M / (mpi_world_size) / num_threads;
|
||
|
|
int kk, i, k, j;
|
||
|
|
|
||
|
|
float A_calc;
|
||
|
|
int bs = 32;
|
||
|
|
|
||
|
|
#pragma omp parallel num_threads(num_threads) //private(index, offset, start, end)
|
||
|
|
{
|
||
|
|
index = omp_get_thread_num();
|
||
|
|
start = index * offset;
|
||
|
|
end = (index== num_threads -1) ? rows : (index+1) * offset;
|
||
|
|
for (kk = 0; kk < K; kk += bs) {
|
||
|
|
for (i = start; i < end; ++i) {
|
||
|
|
for (k = kk; k < min(kk + bs, K); k++) {
|
||
|
|
A_calc = A[i * K + k];
|
||
|
|
for(j=0;j<N;j++){
|
||
|
|
C[i * N + j] += A_calc * 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;
|
||
|
|
MPI_Status status;
|
||
|
|
MPI_Request request;
|
||
|
|
|
||
|
|
// TODO: parallelize & optimize matrix multiplication on multi-node
|
||
|
|
// You must allocate & initialize A, B, C for non-root processes
|
||
|
|
//
|
||
|
|
|
||
|
|
if(mpi_rank==0){
|
||
|
|
int row_size = M / mpi_world_size;
|
||
|
|
int st, ed;
|
||
|
|
|
||
|
|
for(int node =1;node<mpi_world_size;node++){
|
||
|
|
st = offset = node * row_size;
|
||
|
|
ed = node==mpi_world_size-1 ?M : (node+1)*row_size;
|
||
|
|
rows = ed-st;
|
||
|
|
|
||
|
|
MPI_Isend(&offset, 1, MPI_INT, node, 1, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(&rows, 1, MPI_INT, node, 1, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(&A[offset*K], rows*K, MPI_FLOAT, node, 1, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(B, K*N, MPI_FLOAT, node, 1, MPI_COMM_WORLD, &request);
|
||
|
|
}
|
||
|
|
rows = row_size;
|
||
|
|
mat_mul_omp();
|
||
|
|
|
||
|
|
for(int node =1;node<mpi_world_size;node++){
|
||
|
|
MPI_Recv(&offset, 1, MPI_INT, node, 2, MPI_COMM_WORLD, &status);
|
||
|
|
MPI_Recv(&rows, 1, MPI_INT, node, 2, MPI_COMM_WORLD, &status);
|
||
|
|
MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, node, 2, MPI_COMM_WORLD, &status);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else{
|
||
|
|
alloc_mat(&A, M, K);
|
||
|
|
alloc_mat(&B, K, N);
|
||
|
|
alloc_mat(&C, M, N);
|
||
|
|
zero_mat(C, M, N);
|
||
|
|
|
||
|
|
MPI_Recv(&offset, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
|
||
|
|
MPI_Recv(&rows, 1, MPI_INT, 0, 1, MPI_COMM_WORLD, &status);
|
||
|
|
MPI_Recv(A, rows*K, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
|
||
|
|
MPI_Recv(B, K*N, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
|
||
|
|
|
||
|
|
mat_mul_omp();
|
||
|
|
|
||
|
|
MPI_Isend(&offset, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(&rows, 1, MPI_INT, 0, 2, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(C, rows*N, MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &request);
|
||
|
|
}
|
||
|
|
|
||
|
|
}
|