118 lines
3.0 KiB
C++
118 lines
3.0 KiB
C++
|
#include "mat_mul.h"
|
||
|
|
|
||
|
|
#include <cstdio>
|
||
|
|
#include <cstdlib>
|
||
|
|
#include <mpi.h>
|
||
|
|
|
||
|
|
#include <omp.h>
|
||
|
|
#include <algorithm>
|
||
|
|
#include "util.h"
|
||
|
|
|
||
|
|
#define NODE0 0
|
||
|
|
|
||
|
|
using namespace std;
|
||
|
|
|
||
|
|
static float *A, *B, *C;
|
||
|
|
static int M, N, K;
|
||
|
|
static int num_threads;
|
||
|
|
static int mpi_rank, mpi_world_size;
|
||
|
|
|
||
|
|
#define ITILE 32
|
||
|
|
#define JTILE 1024
|
||
|
|
#define KTILE 1024
|
||
|
|
|
||
|
|
static void mat_mul_omp() {
|
||
|
|
// TODO: parallelize & optimize matrix multiplication
|
||
|
|
// Use num_threads per node
|
||
|
|
// A[M*K] * B[K*N] = C[M*N];
|
||
|
|
|
||
|
|
int node_rows = M / mpi_world_size;
|
||
|
|
|
||
|
|
int start, end, rows;
|
||
|
|
start = mpi_rank * node_rows;
|
||
|
|
end = (mpi_rank == mpi_world_size - 1) ? M : (mpi_rank+1)*node_rows;
|
||
|
|
rows = end - start;
|
||
|
|
|
||
|
|
int tid = omp_get_thread_num();
|
||
|
|
|
||
|
|
int slice = rows / num_threads;
|
||
|
|
start = tid * slice + min(tid, rows % num_threads);
|
||
|
|
end = (tid + 1) * slice + min(tid + 1, rows % num_threads);
|
||
|
|
|
||
|
|
#pragma omp parallel for
|
||
|
|
for (int ii = start; ii < end; ii += ITILE) {
|
||
|
|
for( int jj = 0; jj < N; jj += JTILE) {
|
||
|
|
for (int kk = 0; kk < K; kk += KTILE) {
|
||
|
|
|
||
|
|
for (int k = kk; k < min(K, kk + KTILE); ++k) {
|
||
|
|
for (int i = ii; i < min(end, ii + ITILE); ++i) {
|
||
|
|
float Aik = A[i * K + k];
|
||
|
|
for (int j = jj; j < min(N, jj + JTILE); ++j) {
|
||
|
|
C[i * N + j] += Aik * 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
|
||
|
|
MPI_Request request;
|
||
|
|
MPI_Status status;
|
||
|
|
|
||
|
|
// FIXME: for now, only root process runs the matrix multiplication.
|
||
|
|
if (mpi_rank == 0){
|
||
|
|
int node_rows = M / mpi_world_size;
|
||
|
|
int start, end, rows;
|
||
|
|
|
||
|
|
for(int node = 1; node < mpi_world_size; node++) {
|
||
|
|
start = node * node_rows;
|
||
|
|
end = (node == mpi_world_size - 1) ? M : (node + 1) * node_rows;
|
||
|
|
rows = end - start;
|
||
|
|
|
||
|
|
MPI_Isend(&A[start*K], rows*K, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &request);
|
||
|
|
MPI_Isend(B, K*N, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &request);
|
||
|
|
}
|
||
|
|
|
||
|
|
#pragma omp parallel num_threads(num_threads)
|
||
|
|
mat_mul_omp();
|
||
|
|
|
||
|
|
for(int node = 1; node < mpi_world_size; node++) {
|
||
|
|
start = node * node_rows;
|
||
|
|
end = (node == mpi_world_size - 1) ? M : (node + 1) * node_rows;
|
||
|
|
rows = end - start;
|
||
|
|
|
||
|
|
MPI_Recv(&C[start*N], rows*N, MPI_FLOAT, node, 0, MPI_COMM_WORLD, &status);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
int node_rows = M / mpi_world_size;
|
||
|
|
int start, end, rows;
|
||
|
|
start = mpi_rank * node_rows;
|
||
|
|
end = (mpi_rank == mpi_world_size - 1) ? M : (mpi_rank + 1) * node_rows;
|
||
|
|
rows = end - start;
|
||
|
|
|
||
|
|
alloc_mat(&A, M, K);
|
||
|
|
alloc_mat(&B, K, N);
|
||
|
|
alloc_mat(&C, M, N);
|
||
|
|
zero_mat(C, M, N);
|
||
|
|
|
||
|
|
MPI_Recv(A, rows*K, MPI_FLOAT, NODE0, 0, MPI_COMM_WORLD,&status);
|
||
|
|
MPI_Recv(B, K*N, MPI_FLOAT, NODE0, 0, MPI_COMM_WORLD,&status);
|
||
|
|
|
||
|
|
#pragma omp parallel num_threads(num_threads)
|
||
|
|
mat_mul_omp();
|
||
|
|
|
||
|
|
MPI_Isend(C, rows*N, MPI_FLOAT, NODE0, 0, MPI_COMM_WORLD,&request);
|
||
|
|
}
|
||
|
|
}
|