From: Yu-Chin <guestwalk@gmail.com>
Date: Fri, 25 Oct 2013 15:22:32 +0000 (+0800)
Subject: .
X-Git-Tag: v194~6
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=4ba18c24e9dc9f526bd4b363aa0cc632e18a2b0b;p=liblinear

.
---

diff --git a/Makefile b/Makefile
index 7a74e26..166308b 100644
--- a/Makefile
+++ b/Makefile
@@ -1,6 +1,6 @@
 CXX ?= g++
 CC ?= gcc
-CFLAGS = -Wall -Wconversion -O3 -fPIC
+CFLAGS = -Wall -Wconversion -O3 -fPIC -fopenmp
 LIBS = blas/blas.a
 SHVER = 1
 OS = $(shell uname)
diff --git a/linear.cpp b/linear.cpp
index ca835ff..b7e1c88 100644
--- a/linear.cpp
+++ b/linear.cpp
@@ -4,6 +4,8 @@
 #include <string.h>
 #include <stdarg.h>
 #include <locale.h>
+#include <unistd.h>
+#include <time.h>
 #include "linear.h"
 #include "tron.h"
 typedef signed char schar;
@@ -44,6 +46,8 @@ static void info(const char *fmt,...)
 static void info(const char *fmt,...) {}
 #endif
 
+extern int nr_threads;
+
 class l2r_lr_fun: public function
 {
 public:
@@ -64,6 +68,8 @@ private:
 	double *z;
 	double *D;
 	const problem *prob;
+    double **XTv_sub;
+	double *wa;
 };
 
 l2r_lr_fun::l2r_lr_fun(const problem *prob, double *C)
@@ -75,12 +81,19 @@ l2r_lr_fun::l2r_lr_fun(const problem *prob, double *C)
 	z = new double[l];
 	D = new double[l];
 	this->C = C;
+
+    XTv_sub = new double*[nr_threads];
+    for(int ti=0;ti<nr_threads;ti++)
+        XTv_sub[ti] = new double[prob->n];
+
+    wa = new double[l];
 }
 
 l2r_lr_fun::~l2r_lr_fun()
 {
 	delete[] z;
 	delete[] D;
+    delete[] wa;
 }
 
 
@@ -97,6 +110,7 @@ double l2r_lr_fun::fun(double *w)
 	for(i=0;i<w_size;i++)
 		f += w[i]*w[i];
 	f /= 2.0;
+//#pragma omp parallel for  private(i) reduction(+:f)
 	for(i=0;i<l;i++)
 	{
 		double yz = y[i]*z[i];
@@ -116,6 +130,7 @@ void l2r_lr_fun::grad(double *w, double *g)
 	int l=prob->l;
 	int w_size=get_nr_variable();
 
+//#pragma omp parallel for private(i)
 	for(i=0;i<l;i++)
 	{
 		z[i] = 1/(1 + exp(-y[i]*z[i]));
@@ -138,24 +153,31 @@ void l2r_lr_fun::Hv(double *s, double *Hs)
 	int i;
 	int l=prob->l;
 	int w_size=get_nr_variable();
-	double *wa = new double[l];
+
+    for(i=0; i<l; ++i)
+        wa[i] = 0;
 
 	Xv(s, wa);
+//#pragma omp parallel for private(i)
 	for(i=0;i<l;i++)
 		wa[i] = C[i]*D[i]*wa[i];
 
 	XTv(wa, Hs);
+//#pragma omp parallel for private(i)
 	for(i=0;i<w_size;i++)
 		Hs[i] = s[i] + Hs[i];
-	delete[] wa;
 }
 
 void l2r_lr_fun::Xv(double *v, double *Xv)
 {
+    time_t start,end;
+    time(&start);
+
 	int i;
 	int l=prob->l;
 	feature_node **x=prob->x;
 
+#pragma omp parallel for private(i)
 	for(i=0;i<l;i++)
 	{
 		feature_node *s=x[i];
@@ -166,26 +188,45 @@ void l2r_lr_fun::Xv(double *v, double *Xv)
 			s++;
 		}
 	}
+
+    time(&end);
+    double diff = difftime(end,start);
+    //printf("Xv time: %.2lf\n",diff);
 }
 
 void l2r_lr_fun::XTv(double *v, double *XTv)
 {
+    time_t start,end;
+    time(&start);
+
 	int i;
 	int l=prob->l;
 	int w_size=get_nr_variable();
 	feature_node **x=prob->x;
 
+    for(int ti=0;ti<nr_threads;ti++)
+        for(i=0;i<w_size;i++)
+            XTv_sub[ti][i]=0;
 	for(i=0;i<w_size;i++)
 		XTv[i]=0;
-	for(i=0;i<l;i++)
-	{
-		feature_node *s=x[i];
-		while(s->index!=-1)
-		{
-			XTv[s->index-1]+=v[i]*s->value;
-			s++;
-		}
-	}
+#pragma omp parallel for private(i)
+    for(int ti=0; ti<nr_threads; ti++) 
+        for(i=ti*(l/nr_threads);i<(ti+1)*(l/nr_threads);i++)
+        {
+            feature_node *s=x[i];
+            while(s->index!=-1)
+            {
+                XTv_sub[ti][s->index-1]+=v[i]*s->value;
+                s++;
+            }
+        }
+    for(int ti=0; ti<nr_threads; ti++) 
+        for(i=0;i<w_size;i++)
+            XTv[i] += XTv_sub[ti][i];
+
+    time(&end);
+    double diff = difftime(end,start);
+    //printf("XTv time: %.2lf\n",diff);
 }
 
 class l2r_l2_svc_fun: public function
diff --git a/predict.c b/predict.c
index c5b3f1d..c110dbc 100644
--- a/predict.c
+++ b/predict.c
@@ -11,6 +11,7 @@ static int (*info)(const char *fmt,...) = &printf;
 
 struct feature_node *x;
 int max_nr_attr = 64;
+int nr_threads;
 
 struct model* model_;
 int flag_predict_probability=0;
diff --git a/train.c b/train.c
index d388175..aa493b5 100644
--- a/train.c
+++ b/train.c
@@ -4,6 +4,9 @@
 #include <string.h>
 #include <ctype.h>
 #include <errno.h>
+#include <unistd.h>
+#include <time.h>
+#include <omp.h>
 #include "linear.h"
 #define Malloc(type,n) (type *)malloc((n)*sizeof(type))
 #define INF HUGE_VAL
@@ -92,6 +95,7 @@ struct model* model_;
 int flag_cross_validation;
 int nr_fold;
 double bias;
+int nr_threads;
 
 int main(int argc, char **argv)
 {
@@ -115,7 +119,12 @@ int main(int argc, char **argv)
 	}
 	else
 	{
+        time_t start,end;
+        time(&start);
 		model_=train(&prob, &param);
+        time(&end);
+        double diff = difftime(end,start);
+        printf("training time: %.2lf\n",diff);
 		if(save_model(model_file_name, model_))
 		{
 			fprintf(stderr,"can't save model to file %s\n",model_file_name);
@@ -188,6 +197,7 @@ void parse_command_line(int argc, char **argv, char *input_file_name, char *mode
 	param.weight = NULL;
 	flag_cross_validation = 0;
 	bias = -1;
+    nr_threads = 2;
 
 	// parse options
 	for(i=1;i<argc;i++)
@@ -238,6 +248,11 @@ void parse_command_line(int argc, char **argv, char *input_file_name, char *mode
 			case 'q':
 				print_func = &print_null;
 				i--;
+
+			case 't':
+				nr_threads = atoi(argv[i]);
+				break;
+
 				break;
 
 			default: