From 8546af9eb5ebfda40316be13727750f0a1c61b57 Mon Sep 17 00:00:00 2001
From: Peter Schaefer <peter.schaefer@tuwien.ac.at>
Date: Fri, 3 Aug 2012 12:48:35 +0200
Subject: [PATCH] [src] doppelte geschwindigkeit

---
 src/mex_build_V.cpp | 29 +++++++++++++++++++----------
 src/test_sol.m      |  2 +-
 2 files changed, 20 insertions(+), 11 deletions(-)

diff --git a/src/mex_build_V.cpp b/src/mex_build_V.cpp
index d02a66b..00e0d56 100644
--- a/src/mex_build_V.cpp
+++ b/src/mex_build_V.cpp
@@ -211,19 +211,24 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
 	double d[3] = { 0, 0, 0 };
 	double dt[3] = { 0, 0, 0 };
 
-#define V_MINROW_PER_WORKER 3
+#define MINSIZE_PER_WORKER 3
+#define MAX_WORKER 10
 
 	int actualNumberOfThreads = omp_get_max_threads();
+
+	if(MAX_WORKER<actualNumberOfThreads)
+		actualNumberOfThreads = MAX_WORKER;
+
 	int firstRow = 0, lastRow = -1;
 	int targetSize = em * (em + 1) / (2 * actualNumberOfThreads);
 
-	if (targetSize < V_MINROW_PER_WORKER) {
-		actualNumberOfThreads = (int) em * (em + 1) / (2 * V_MINROW_PER_WORKER);
+	if (targetSize < MINSIZE_PER_WORKER) {
+		actualNumberOfThreads = (int) em * (em + 1) / (2 * MINSIZE_PER_WORKER);
 		if (actualNumberOfThreads < 1) {
 			actualNumberOfThreads = 1;
 			targetSize = em;
 		} else {
-			targetSize = V_MINROW_PER_WORKER;
+			targetSize = MINSIZE_PER_WORKER;
 		}
 	}
 #ifdef DEBUG
@@ -234,13 +239,12 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
 #endif
 
 //Ausrechnen
-#ifndef DEBUG
 #ifdef PARALLEL
 #pragma omp parallel private(i,j,k,tmp,x,y,xa,xb,ya,yb,d,dt,rx, rxa, rxb, ry, rya, ryb,lastRow,firstRow) shared(C,E,ctypeO,ctypeP,em,targetSize,actualNumberOfThreads)
 	{
-#pragma omp parallel for
-#endif
+#pragma omp for
 #endif
+
 	for (i = 0; i < actualNumberOfThreads; ++i) {
 		firstRow = (int) sqrt(2*i*targetSize);
 		if (i == actualNumberOfThreads - 1) {
@@ -352,17 +356,22 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
 					}
 
 				}
-
+#ifdef PARALLEL
+#pragma omp critical
+#endif
 				A[(k * em) + j] = 1. / (4 * M_PI) * tmp;
 				if (k != j)
 					A[(j * em) + k] = 1. / (4 * M_PI) * tmp;
+#ifdef PARALLEL
+#pragma omp end critical
+#endif
 			}
 		}
 	}
-#ifndef DEBUG
+
 #ifdef PARALLEL
 }
 #endif
-#endif
+
 	return;
 }
diff --git a/src/test_sol.m b/src/test_sol.m
index 16bf4e4..2a8e973 100644
--- a/src/test_sol.m
+++ b/src/test_sol.m
@@ -9,7 +9,7 @@ mex mex_build_V.cpp slpRectangle.cpp CFLAGS="\$CFLAGS -fopenmp" CXXFLAGS="\$CXXF
 % Test ausführen
 
 %Anzahl der Schritte
-steps = 15;
+steps = 16;
 
 %Art der Berechnungen
 type = [1];
-- 
2.47.3