0923_optimizing
user_3093867
c_cpp
18 days ago
14 kB
1
Indexable
Never
#include <iostream>
#include <cmath>
#include <algorithm>
#include <vector>
// #include <boost/sort/spreadsort/spreadsort.hpp>
#include <mpi.h>
#include <fstream>
#include <boost/sort/block_indirect_sort/block_indirect_sort.hpp>
#define ODD 1
#define ODD_CHECK 11
#define EVEN 2
#define EVEN_CHECK 22
int N; // element 個數
int parti; // 除 last process 以外，其餘 node 擁有的 element 個數
int read_index;
float *tmp;
bool strictLess(float a, float b) {
union { float f; uint32_t u; } ua = {a}, ub = {b};
// Handle sign bit
bool a_negative = ua.u >> 31;
bool b_negative = ub.u >> 31;
if (a_negative != b_negative) {
return a_negative;
}
// For negative numbers, invert the comparison
if (a_negative) {
return ua.u > ub.u;
} else {
return ua.u < ub.u;
}
}
void Merge(float *data, int data_Cnt, float *neighbor, int neighbor_Cnt, bool isLow) {
int i = isLow ? 0 : data_Cnt - 1;
int j = isLow ? 0 : neighbor_Cnt - 1;
int k = isLow ? 0 : data_Cnt - 1;
while (k >= 0 && k < data_Cnt) {
if (i >= 0 && i < data_Cnt && j >= 0 && j < neighbor_Cnt) {
if (isLow ? strictLess(data[i], neighbor[j]) : strictLess(neighbor[j], data[i])) {
tmp[k] = data[i];
i += isLow ? 1 : -1;
} else {
tmp[k] = neighbor[j];
j += isLow ? 1 : -1;
}
} else if (i >= 0 && i < data_Cnt) {
tmp[k] = data[i];
i += isLow ? 1 : -1;
} else {
tmp[k] = neighbor[j];
j += isLow ? 1 : -1;
}
k += isLow ? 1 : -1;
}
std::copy(tmp, tmp + data_Cnt, data);
}
// 取前半進行合併
// void Merge_low(float *data, int data_Cnt, float *neighbor, int neighbor_Cnt, bool *All_Sorted){
// if (strictLess(data[data_Cnt - 1], neighbor[0]) || !strictLess(neighbor[0], data[data_Cnt - 1]))
// return;
// int data_idx = 0, neighbor_idx = 0, tmp_index = 0;
// // 將比較後的 data 放入 tmp_buf
// while(tmp_index < data_Cnt){
// if(data_idx < data_Cnt && neighbor_idx < neighbor_Cnt){
// if(strictLess(data[data_idx], neighbor[neighbor_idx]) || !strictLess(neighbor[neighbor_idx], data[data_idx])){
// tmp[tmp_index++] = data[data_idx++];
// }
// else{ tmp[tmp_index++] = neighbor[neighbor_idx++];}
// }
// else{
// if(data_idx >= data_Cnt){tmp[tmp_index++] = neighbor[neighbor_idx++];}
// else{tmp[tmp_index++] = data[data_idx++];}
// }
// }
// // for(int i = 0; i < data_Cnt; i++){ data[i] = tmp[i]; } // 寫回 data
// // std::swap(data, tmp);
// std::copy(tmp, tmp+data_Cnt, data);
// return;
// }
// // 取後半進行合併
// void Merge_high(float *data, int data_Cnt, float *neighbor, int neighbor_Cnt, bool *All_Sorted){
// // 有可能要處理 last node 問題
// if (strictLess(neighbor[neighbor_Cnt - 1], data[0]) || !strictLess(data[0], neighbor[neighbor_Cnt - 1]))
// return;
// int data_idx = data_Cnt-1, neighbor_idx = neighbor_Cnt-1, tmp_index = data_Cnt-1;
// while(tmp_index >= 0){
// if(data_idx >= 0 && neighbor_idx >= 0){
// if(strictLess(neighbor[neighbor_idx], data[data_idx])){
// tmp[tmp_index--] = data[data_idx--];
// }
// else{ tmp[tmp_index--] = neighbor[neighbor_idx--];}
// }
// else{
// if(data_idx < 0){
// tmp[tmp_index--] = neighbor[neighbor_idx--];
// }
// else{tmp[tmp_index--] = data[data_idx--];}
// }
// }
// // for(int i = 0; i < data_Cnt; i++){ data[i] = tmp[i]; }
// // std::swap(data, tmp);
// std::copy(tmp, tmp+data_Cnt, data);
// return;
// }
int Element_Count(int rank){
if (parti*rank >= N){ return 0; } // wasted node
if (parti*(rank+1) >= N){ return N - rank*parti; } // true last node
else{ return parti; } // normal node
}
int main(int argc, char **argv){
// Declaration
double start, finish;
MPI_Init(&argc, &argv);
// double io_elapsed = 0, comm_elapsed = 0, computation_elapsed = 0;
// double elapsed; // total time
// double io_start, io_finish;
// double comm_start, comm_finish;
// double computation_start, computation_finish;
start = MPI_Wtime();
int rank, size;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
N = atoi(argv[1]); // total elementsd in the array: N
// parti = std::ceil(N/(double)size);
parti = N/size + 1;
read_index = parti * rank; // 要讀寫哪一個位置
// double *io_buf = (double*) malloc(size*sizeof(double));
// double *com_buf = (double*) malloc(size*sizeof(double));
// double *comput_buf = (double*) malloc(size*sizeof(double));
MPI_File input_file, output_file;
char *input_filename = argv[2]; // input file
char *output_filename = argv[3]; // output file
// Create local data structure & buffer for process array by given rank
int Cnt = Element_Count(rank);
float *data = (float*) malloc(Cnt*sizeof(float));
float *buf = (float*) malloc(parti*sizeof(float));
bool isSorted = true;
bool bufSorted = true;
// 依照指定位置讀取檔案（if last node, 則僅能讀取 count 個 element）
MPI_File_open(MPI_COMM_WORLD, input_filename, MPI_MODE_RDONLY, MPI_INFO_NULL, &input_file);
if (Cnt != 0){
MPI_File_read_at(input_file, sizeof(float) * read_index, data, Cnt, MPI_FLOAT, MPI_STATUS_IGNORE);}
MPI_File_close(&input_file);
if (Cnt != 0){
//* try different sort algorithm, pdqsort works best
// first local sort（使用 spreadsort 加速：參考 github）
// boost::sort::spreadsort::float_sort(data, data + Cnt, strictLess);
// std::sort(data, data + Cnt, strictLess);
boost::sort::pdqsort(data, data + Cnt, strictLess);
// boost::sort::spinsort(data, data + Cnt, strictLess);
// boost::sort::block_indirect_sort(data, data + Cnt, strictLess);
tmp = (float*) malloc(Cnt*sizeof(float));
}
// computation_finish = MPI_Wtime();
// computation_elapsed += computation_finish - computation_start;
// get neighbor count
int leftCnt, rightCnt;
if (rank-1 >= 0){leftCnt = Element_Count(rank-1);}
else{ leftCnt = 0;}
if (rank+1 > size){rightCnt = 0;}
else { rightCnt = Element_Count(rank+1);}
int iter = 0; // iteration 次數，一但超過 process 數量表示 sort 完成
// bool lastProc = read_index N && rank*(parti+1) >= N;
bool lastProc = read_index <= N && read_index + parti >= N;
while(true){
if(Cnt != 0){
// Even Phase - main node (even: 0, 2, 4, 6, 8)
if(rank%2 == 0){
if(!lastProc){
// 確認是否兩邊需要交換（把 main node 最大的交換過去）
MPI_Sendrecv(&data[Cnt-1], 1, MPI_FLOAT, rank+1, EVEN, buf, 1, MPI_FLOAT, rank+1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// buf[0] 是 neighbor 最小的值
// data[Cnt-1] 是 main node 最大的值
if(strictLess(buf[0], data[Cnt-1])){
int send_start = std::lower_bound(data, data+Cnt , buf[0], strictLess) - data;
float send_count = Cnt - send_start;
//* send send_count to neighbor, and get the send_count from neighbor
MPI_Sendrecv(&send_count, 1, MPI_FLOAT, rank+1, EVEN_CHECK, buf,1, MPI_FLOAT, rank+1, EVEN_CHECK, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
float recv_count = buf[0];
MPI_Sendrecv( data+send_start, static_cast<int>(send_count), MPI_FLOAT, rank+1, EVEN, buf, static_cast<int>(recv_count), MPI_FLOAT, rank+1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// MPI_Sendrecv(data, Cnt, MPI_FLOAT, rank+1, EVEN, buf, rightCnt, MPI_FLOAT, rank+1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
Merge(data, Cnt, buf, static_cast<int>(recv_count), true);
// Merge_low(data, Cnt, buf, rightCnt, &isSorted);
}
}
}
// Even Phase - supplement node (odd: 1, 3, 5, 7, 9)
else{
MPI_Sendrecv(data, 1, MPI_FLOAT, rank-1, EVEN, buf, 1, MPI_FLOAT, rank-1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// buf[0] 是 neighbor 最大的值
// data[0] 是 main node 最小的值
if(strictLess(data[0], buf[0])){
int send_start = std::lower_bound(data, data+Cnt, buf[0], strictLess) - data;
float send_count = send_start;
//* send send_count to neighbor, and get the send_count from neighbor
MPI_Sendrecv(&send_count, 1, MPI_FLOAT, rank-1, EVEN_CHECK, buf, 1, MPI_FLOAT, rank-1, EVEN_CHECK, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
float recv_count = buf[0];
MPI_Sendrecv(data, static_cast<int>(send_count), MPI_FLOAT, rank-1, EVEN, buf, static_cast<int>(recv_count), MPI_FLOAT, rank-1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// MPI_Sendrecv(data, Cnt, MPI_FLOAT, rank-1, EVEN, buf, leftCnt, MPI_FLOAT, rank-1, EVEN, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// Merge_high(data, Cnt, buf, leftCnt, &isSorted);
Merge(data, Cnt, buf, static_cast<int>(recv_count), false);
}
}
// Odd Phase - main node (odd: 1, 3, 5, 7, 9) - 和右邊的比較
if(rank%2 == 1){
if(!lastProc){
MPI_Sendrecv(&data[Cnt-1], 1, MPI_FLOAT, rank+1, ODD, buf, 1, MPI_FLOAT, rank+1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// buf[0] 是 neighbor 最小的值
// data[Cnt-1] 是 main node 最大的值
if(strictLess(buf[0], data[Cnt-1])){
int send_start = std::lower_bound(data, data+Cnt, buf[0], strictLess) - data;
float send_count = Cnt - send_start;
//* send send_count to neighbor, and get the send_count from neighbor
MPI_Sendrecv(&send_count, 1, MPI_FLOAT, rank+1, ODD_CHECK, buf, 1, MPI_FLOAT, rank+1, ODD_CHECK, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
float recv_count = buf[0];
MPI_Sendrecv(data+send_start, static_cast<int>(send_count), MPI_FLOAT, rank+1, ODD, buf, static_cast<int>(recv_count), MPI_FLOAT, rank+1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// MPI_Sendrecv(data, Cnt, MPI_FLOAT, rank+1, ODD, buf, rightCnt, MPI_FLOAT, rank+1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// Merge_low(data, Cnt, buf, rightCnt, &isSorted);
Merge(data, Cnt, buf, static_cast<int>(recv_count), true);
}
}
}
// Odd Phase - supplement node (odd: 0, 2, 4, 6, 8)
else{
if(rank != 0){ // rank 0 沒有左邊 process 可以進行交換
MPI_Sendrecv(data, 1, MPI_FLOAT, rank-1, ODD, buf, 1, MPI_FLOAT, rank-1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
if(strictLess(data[0], buf[0])){
int send_start = std::lower_bound(data, data+Cnt, buf[0], strictLess) - data;
float send_count = send_start;
//* send send_count to neighbor, and get the send_count from neighbor
MPI_Sendrecv(&send_count, 1, MPI_FLOAT, rank-1, ODD_CHECK, buf, 1, MPI_FLOAT, rank-1, ODD_CHECK, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
float recv_count = buf[0];
MPI_Sendrecv(data, static_cast<int>(send_count), MPI_FLOAT, rank-1, ODD, buf, static_cast<int>(recv_count), MPI_FLOAT, rank-1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// MPI_Sendrecv(data, Cnt, MPI_FLOAT, rank-1, ODD, buf, leftCnt, MPI_FLOAT, rank-1, ODD, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
// Merge_high(data, Cnt, buf, leftCnt, &isSorted);
Merge(data, Cnt, buf, static_cast<int>(recv_count), false);
}
}
}
}
iter += 2;
if(iter > size){break;}
}
MPI_File_open(MPI_COMM_WORLD, output_filename, MPI_MODE_CREATE|MPI_MODE_WRONLY, MPI_INFO_NULL, &output_file);
MPI_File_write_at(output_file, sizeof(float) * read_index, data, Cnt, MPI_FLOAT, MPI_STATUS_IGNORE);
MPI_File_close(&output_file);
MPI_Finalize();
// free(data);
// free(buf);
// if (Cnt != 0)(free(tmp));
return 0;
}
Leave a Comment