-
Notifications
You must be signed in to change notification settings - Fork 0
/
LogisticRegressionMPISSE.c
218 lines (191 loc) · 7.32 KB
/
LogisticRegressionMPISSE.c
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
/*
* All changes to code are copyright, 2017, Zhu Li, zhuli@unm.edu
*/
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <time.h>
#include <x86intrin.h>
#include <mpi/mpi.h>
#include <sys/time.h>
#define SAMPLE_NUMBER (1024 * sample_number)
#define SAMPLE_ATTRIBUTE_NUMBER 32
#define INITIAL_WEIGHTS_RANGE 0.01
#define SAMPLE_VALUE_RANGE 50
#define CONVERGE_RATE 0.0001
#define ITERATION_NUMBER 6000 * 2
#define DATA_NUMBER 4
#define BYTE_NUMBER 16
#define MICROSEC_IN_SEC 1000000
//#define DEBUG
int comm_sz; //Number of Processes
int my_rank; //My process rank
int sample_number = 0;
/**
*
* @param n Length of the array.
* @param range Range of the numbers in the array is [0, range].
* @return An array filled with random numbers.
*/
float* generateRandomVectorFloat(int n, float range) {
float* ptr = (float*)aligned_alloc(BYTE_NUMBER, sizeof(float) * n);
if (ptr != NULL) {
for (int i = 0; i < n; i++) {
ptr[i] = (range * rand() / RAND_MAX) - range / 2;
}
}
return ptr;
}
/**
* return dot product of vector x and w.
*/
float dotProduct(float* x, float* w, int n) {
__m128 product = _mm_set1_ps(0);
__m128* xSSE = (__m128*)x;
__m128* wSSE = (__m128*)w;
for (int i = 0; i < n / DATA_NUMBER; i++) {
product = _mm_add_ps(product, _mm_mul_ps(xSSE[i], wSSE[i]));
//product = _mm256_fmadd_ps(xSSE[i], wSSE[i], product);
}
float* dp = (float*)&product;
float sum = 0;
for (int i = 0; i < DATA_NUMBER; i++) {
sum += *dp++;
}
return sum;
}
float logisticFunction(float* x, float* w, int n, float w0) {
return 1 / (1 + exp(w0 + dotProduct(x, w, n)));
}
float* temp;
float* tempSum;
void updateDelta(float **x, float *difference, __m128* weightsSSE) {
__m128* tempSSE = (__m128*)temp;
__m128* tempSumSSE = (__m128*)tempSum;
float converge_rate = CONVERGE_RATE;
// int splitSize = SAMPLE_ATTRIBUTE_NUMBER / DATA_NUMBER / comm_sz;
// int start = splitSize * my_rank;
// int end = start + splitSize;
int splitSize = SAMPLE_NUMBER / comm_sz;
int start = my_rank * splitSize;
int end = start + splitSize;
for (int j = 0; j < SAMPLE_ATTRIBUTE_NUMBER / DATA_NUMBER; j++) {
__m128 *xiSSE = (__m128*)x[start];
const __m128 multiplier = _mm_set1_ps(difference[start] * converge_rate);
tempSSE[j] = _mm_mul_ps(xiSSE[j], multiplier);
for (int i = start + 1; i < end; i++) {
__m128 *xiSSE = (__m128*)x[i];
const __m128 multiplier = _mm_set1_ps(difference[i] * converge_rate);
tempSSE[j] = _mm_add_ps(tempSSE[j], _mm_mul_ps(xiSSE[j], multiplier));
//weightsSSE[j] = _mm256_fmadd_ps(xiSSE[j], multiplier, weightsSSE[j]);
}
}
MPI_Allreduce(temp, tempSum, SAMPLE_ATTRIBUTE_NUMBER, MPI_FLOAT, MPI_SUM, MPI_COMM_WORLD);
for (int i = 0; i < SAMPLE_ATTRIBUTE_NUMBER / DATA_NUMBER; i++) {
weightsSSE[i] = _mm_add_ps(weightsSSE[i], tempSumSSE[i]);
}
//MPI_Allgather(weights + start * DATA_NUMBER, splitSize * DATA_NUMBER, MPI_FLOAT, weights, splitSize * DATA_NUMBER, MPI_FLOAT, MPI_COMM_WORLD);
}
void updateWeights(float *difference, float* weights, float** x, float* y, float w0) {
const __m128 minusOne = _mm_set1_ps(-1);
__m128 *diffSSE = (__m128 *) difference;
__m128 *ySSE = (__m128 *) y;
__m128* weightsSSE = (__m128*)weights;
// Divide workload among different processes
int splitSize = SAMPLE_NUMBER / comm_sz;
int start = my_rank * splitSize;
int end = start + splitSize;
// Calculate the difference according to logistic regression update formula
for (int i = start; i < end; i++) {
difference[i] = logisticFunction(x[i], weights, SAMPLE_ATTRIBUTE_NUMBER, w0);
}
for (int i = start / DATA_NUMBER; i < end / DATA_NUMBER; i++) {
diffSSE[i] = _mm_add_ps(diffSSE[i], _mm_add_ps(ySSE[i], minusOne));
}
//MPI_Allgather(difference + start, splitSize, MPI_FLOAT, difference, splitSize, MPI_FLOAT, MPI_COMM_WORLD);
// update and add delta to the original weights
// function name is not changed for consistency
updateDelta(x, difference, weightsSSE);
}
int main(int argc, char **argv) {
char* sample_number_string = argv[1];
sample_number += sample_number_string[0] - '0';
if (sample_number_string[1]) {
sample_number = 10 * sample_number + sample_number_string[1] - '0';
}
srand(time(NULL));
// initialize the weights randomly
float w0 = (INITIAL_WEIGHTS_RANGE * rand() / RAND_MAX) - INITIAL_WEIGHTS_RANGE / 2;
float* weights = generateRandomVectorFloat(SAMPLE_ATTRIBUTE_NUMBER, INITIAL_WEIGHTS_RANGE);
// TODO: load real data into x and y;
// Generate random data for x
float** x = (float**)malloc(SAMPLE_NUMBER * sizeof(float*));
for (int i = 0; i < SAMPLE_NUMBER; i++) {
x[i] = generateRandomVectorFloat(SAMPLE_ATTRIBUTE_NUMBER, SAMPLE_VALUE_RANGE);
}
// Set all benchmark weights as 0.5 or -0.5 randomly and generate the corresponding labels.
// So we could test the effectiveness of the program according to whether
// the program could predict the labels generated with benchmark weights
float* y = (float*)aligned_alloc(BYTE_NUMBER, SAMPLE_NUMBER * sizeof(float));
float* benchMarkWeights = (float*)aligned_alloc(BYTE_NUMBER, SAMPLE_ATTRIBUTE_NUMBER * sizeof(float));
float benchMarkWeight0 = rand() % 2 - 0.5;
for (int i = 0; i < SAMPLE_ATTRIBUTE_NUMBER; i++) {
benchMarkWeights[i] = rand() % 2 - 0.5;
}
for (int i = 0; i < SAMPLE_NUMBER; i++) {
y[i] = logisticFunction(x[i], benchMarkWeights, SAMPLE_ATTRIBUTE_NUMBER, benchMarkWeight0) > 0.5 ? 0 : 1;
}
struct timeval tv;
gettimeofday(&tv, NULL);
long start = tv.tv_usec + tv.tv_sec * MICROSEC_IN_SEC;
//clock_t start = clock(), diff;
MPI_Init(NULL, NULL);
MPI_Comm_size(MPI_COMM_WORLD, &comm_sz);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
float *difference = (float *) aligned_alloc(BYTE_NUMBER, sizeof(float) * SAMPLE_NUMBER);
temp = (float*)aligned_alloc(BYTE_NUMBER, sizeof(float) * SAMPLE_ATTRIBUTE_NUMBER);
tempSum = (float*)aligned_alloc(BYTE_NUMBER, sizeof(float) * SAMPLE_ATTRIBUTE_NUMBER);
for (int i = 0; i < ITERATION_NUMBER; i++) {
updateWeights(difference, weights, x, y, w0);
}
free(difference);
free(temp);
free(tempSum);
#ifdef DEBUG
for (int i = 0; i < SAMPLE_ATTRIBUTE_NUMBER; i++) {
printf("Benchmark weight: %lf Estimated weight:%lf\n", benchMarkWeights[i], weights[i]);
}
#endif
// Predict the labels with weights estimated with logistic regression.
float error = 0;
// Split workload among several processes
int splitSize = SAMPLE_NUMBER / comm_sz;
int splitStart = splitSize * my_rank;
int splitEnd = splitStart + splitSize;
for (int i = splitStart; i < splitEnd; i++) {
float predict = logisticFunction(x[i], weights, SAMPLE_ATTRIBUTE_NUMBER, w0) > 0.5 ? 0 : 1;
#ifdef DEBUG
printf("y[%d]: %lf Predicted: %lf\n", i, y[i], predict);
#endif
error += fabs(predict - y[i]);
}
float totalError = 0;
MPI_Reduce(&error, &totalError, 1, MPI_FLOAT, MPI_SUM, 0, MPI_COMM_WORLD);
if (my_rank == 0) {
printf("Average error:%lf\n", totalError / SAMPLE_NUMBER);
//int diff = gettimeofday() - start;
//diff = clock() - start;
gettimeofday(&tv, NULL);
long diff = (tv.tv_sec * MICROSEC_IN_SEC + tv.tv_usec - start) / 1000;
printf("Time taken: %ld seconds %ld milliseconds\n", diff / 1000, diff % 1000);
}
MPI_Finalize();
for (int i = 0; i < SAMPLE_NUMBER; i++) {
free(x[i]);
}
free(x);
free(y);
free(weights);
free(benchMarkWeights);
return 0;
}