chundoong-lab-ta/SamsungDS22/submissions/HW4/ss1.eom/mat_mul.cpp

224 lines
6.4 KiB
C++
Raw Normal View History

2022-09-29 18:01:45 +09:00
#include "mat_mul.h"
#include <cstdio>
#include <cstdlib>
#include <mpi.h>
#include <omp.h>
#include "util.h"
#define FROM_MASTER 0
#define FROM_WORKER 1
#define MASTER 0
#define ITILESIZE (1024)
#define JTILESIZE (1024)
#define KTILESIZE (1024)
static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
/// for (int ii = is; ii < ie; ii += ITILESIZE) {
/// for (int jj = 0; jj < N; jj += JTILESIZE) {
/// for (int kk = 0; kk < K; kk += KTILESIZE) {
///
/// for (int k = kk; k < std::min(K, kk + KTILESIZE); k++) {
/// for (int i = ii; i < std::min(ie, ii + ITILESIZE); i++) {
/// float ar = A[i * K + k];
/// for (int j = jj; j < std::min(N, jj + JTILESIZE); j+=1) {
/// C[i * N + j] += ar * B[k * N + j];
static void mat_mul_omp(int num_row) {
int bs = 30;
int k_min;
int slice = num_row / num_threads;
omp_set_num_threads(num_threads);
#pragma omp parallel
{
int pid = omp_get_thread_num();
int start = pid * slice;
int end = pid == num_threads - 1 ? num_row : (pid + 1) * slice;
for (int ii = start; ii < end; ii+=ITILESIZE) {
for (int jj = 0; jj < N; jj += JTILESIZE) {
for (int kk = 0; kk < K; kk += bs) {
//for (int i = start; i < end; ++i) {
for (int i = ii; i < std::min( end, ii+ITILESIZE); ++i) {
for (int k = kk; k < std::min(kk+bs,K); ++k) {
//for (int j = 0; j < N; j=j+1) {
for (int j = jj; j< std::min(N, jj+JTILESIZE); j=j+1) {
C[i * N + j] += A[i*K +k] * B[k * N + j];
}
}
}
}
}
}
}
}
///static void mat_mul_omp(int num_row) {
///
/// int bs = 41;
/// int k_min;
/// int slice = num_row / num_threads;
/// omp_set_num_threads(num_threads);
///#pragma omp parallel
/// {
/// int pid = omp_get_thread_num();
/// //int start = pid * slice;
/// //int end = pid == num_threads - 1 ? num_row : (pid + 1) * slice;
///
/// int is = M / num_threads * pid + std::min(pid, M % num_threads);
/// int ie = M / num_threads * (pid + 1) + std::min(pid + 1, M % num_threads);
///
/// for (int ii = is; ii < ie; ii += ITILESIZE) {
/// for (int jj = 0; jj < N; jj += JTILESIZE) {
/// for (int kk = 0; kk < K; kk += KTILESIZE) {
///
/// for (int k = kk; k < std::min(K, kk + KTILESIZE); k++) {
/// for (int i = ii; i < std::min(ie, ii + ITILESIZE); i++) {
/// float ar = A[i * K + k];
/// for (int j = jj; j < std::min(N, jj + JTILESIZE); j+=1) {
/// C[i * N + j] += ar * B[k * N + j];
/// }
/// }
/// }
///
/// }
/// }
/// }
///
/// }
///}
///
/*
#define ITILESIZE (32)
#define JTILESIZE (1024)
#define KTILESIZE (1024)
static void* mat_mul_thread(void *data) {
int tid = (long)data;
int is = M / num_threads * tid + min(tid, M % num_threads);
int ie = M / num_threads * (tid + 1) + min(tid + 1, M % num_threads);
for (int ii = is; ii < ie; ii += ITILESIZE) {
for (int jj = 0; jj < N; jj += JTILESIZE) {
for (int kk = 0; kk < K; kk += KTILESIZE) {
for (int k = kk; k < min(K, kk + KTILESIZE); k++) {
for (int i = ii; i < min(ie, ii + ITILESIZE); i++) {
float ar = A[i * K + k];
for (int j = jj; j < min(N, jj + JTILESIZE); j+=1) {
C[i * N + j] += ar * B[k * N + j];
}
}
}
}
}
}
return NULL;
}
*/
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
// FIXME: for now, only root process runs the matrix multiplication.
//if (mpi_rank == 0)
// mat_mul_omp();
MPI_Status status;
MPI_Request request;
//int len_blk;
//int len_B, len_A, len_p_A, len_p_r_A;
//M_p = int(M/mpi_world_size);
//M_p_r = M_p + (M%mpi_world_size);
//len_A = M*K;
//len_B = K*N;
//len_C = M*N;
//len_p_A = M_p*K;
//len_p_C = M_p*N;
//len_p_r_A = M_p_r*K;
//len_p_r_C = M_p_r*N;
//
//printf("\n");
//printf("Seunsik\n");
//printf("Num node %d\n", mpi_world_size);
int offset;
int rows;
if (mpi_rank == 0)
{
int row_size = M/mpi_world_size;
int st, ed;
for (int i =1; i<mpi_world_size; i++)
{
st = offset = i * row_size;
ed = i == mpi_world_size -1 ? M : (i+1)*row_size;
rows = ed - st;
MPI_Isend(&offset, 1, MPI_INT, i, 0, MPI_COMM_WORLD, &request);
MPI_Isend(&rows, 1, MPI_INT, i, 0, MPI_COMM_WORLD, &request);
MPI_Isend(A+offset*K, rows*K, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
MPI_Isend(B, K*N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
//printf("rows: %d\n:", rows);
//printf("row_size: %d\n:", row_size);
//printf("offset: %d\n:", offset);
//printf("st: %d\n:", st);
//printf("ed: %d\n:", ed);
}
rows = row_size;
mat_mul_omp(row_size);
for (int i =1; i<mpi_world_size; i++)
{
MPI_Recv(&offset, 1, MPI_INT, i, FROM_WORKER, MPI_COMM_WORLD, &status);
MPI_Recv(&rows, 1, MPI_INT, i, FROM_WORKER, MPI_COMM_WORLD, &status);
MPI_Recv(C + offset*N, rows*N, MPI_FLOAT, i, FROM_WORKER, MPI_COMM_WORLD, &status);
}
//free(A);
//free(B);
//free(C);
}
else
{
//alloc_mat(&A, M_p, K);
//alloc_mat(&B, K, N);
//alloc_mat(&C, M_p, N);
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, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
MPI_Recv(&rows, 1, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
MPI_Recv(A, rows*K, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
MPI_Recv(B, K*N, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
mat_mul_omp(rows);
MPI_Isend(&offset, 1, MPI_INT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
MPI_Isend(&rows, 1, MPI_INT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
MPI_Isend(C, rows*N, MPI_FLOAT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
free(A);
free(B);
free(C);
}
}