Ferdig med 12

lab_2/12/Makefile

@@ -16,3 +16,6 @@ clear: clean
 
 zip: clean
 	zip $(folder) Makefile *.cpp *.h
+
+profile:
+	g++ -pg *.cpp -o $@ $^

lab_2/12/chain-random-walk.cpp

@@ -19,7 +19,9 @@ namespace {
    {
        double sum = 0.0;
        for(long i = 0; i < size; i++) sum += config[i];
-       for(long i = 0; i < size; i++) config[i] -= sum/size;
+
+       const float ratio = sum/size;
+       for(long i = 0; i < size; i++) config[i] -= ratio;
    }
    
    const int el_out_max = 9;  // show at most the first nine elements
@@ -37,6 +39,7 @@ float crw::elongation(long size, float config[])
    return max - min;
 }
 
+
 float* crw::step(long size, float previous[])
 {
    // allocate the next configuration
@@ -44,8 +47,7 @@ float* crw::step(long size, float previous[])
     
    // first, let the chain contract: each element is attracted by its neighbours
    for(long i = 0; i < size; i++)
-      config[i] = 0.5*previous[i] + 0.25*previous[(i-1) % size]
+      config[i] = 0.5*previous[i] + 0.25*previous[(i-1) % size]*2;
-                                  + 0.25*previous[(i+1) % size];
 
    stochastic_unit_step(size, config);  // actual random walk step
    shift_centre_to_origin(size, config);  // shift such that the average remains zero

lab_2/12/const_Nm_timings.csv

@@ -0,0 +1,10 @@
+1024, 36.21
+512, 36.88
+256, 43.77
+128, 40.27
+64, 40.47
+32, 43.17
+16, 47.71
+8, 50.55
+4, 49.56
+2, 54.86

lab_2/12/const_m_timings.csv

@@ -0,0 +1,10 @@
+1024, 36.24
+512, 18.53
+256, 9.72
+128, 4.94
+64, 2.56
+32, 1.36
+16, 0.75
+8, 0.40
+4, 0.19
+2, 0.11

lab_2/12/memleak.cpp

@@ -49,7 +49,7 @@ int main(int argc, char** argv)
       // Simple fix to memleak:
       float* old_configuration = present_configuration;
       present_configuration = crw::step(size, present_configuration);
-      delete old_configuration;
+      delete[] old_configuration;
       float present_elongation = crw::elongation(size, present_configuration);
 
       if(present_elongation > extreme_elongation)

lab_2/12/plot_speedup.py

@@ -0,0 +1,38 @@
+import matplotlib.pyplot as plt
+import matplotlib.cbook as cbook
+
+import numpy as np
+import pandas as pd
+
+f1 = open('const_m_timings.csv', 'r')
+f2 = open('../11/const_m_timings.csv', 'r')
+x1 = []
+y1 = []
+for l,p in zip(f1.readlines(), f2.readlines()):
+    l = l.split(',')
+    p = p.split(',')
+    x1.append(p[0])
+    y1.append(float(p[1])/float(l[1]))
+
+f1 = open('const_Nm_timings.csv', 'r')
+f2 = open('../11/const_Nm_timings.csv', 'r')
+x2 = []
+y2 = []
+for l,p in zip(f1.readlines(), f2.readlines()):
+    l = l.split(',')
+    p = p.split(',')
+    x2.append(p[0])
+    y2.append(float(p[1])/float(l[1]))
+
+plt.plot(x1, y1, label = 'constant m')
+plt.plot(x2, y2, label = 'constant Nm')
+for i in range(len(x1)):
+        plt.text(i,y1[i],f'{y1[i]:.2f}', color="blue")
+for i in range(len(x2)):
+        plt.text(i,y2[i],f'{y2[i]:.2f}', color="orange")
+plt.title('Speedup after optimization')
+plt.xlabel('N')
+plt.ylabel('xspeedup')
+plt.grid()
+plt.legend()
+plt.show()

lab_2/12_answer.md

@@ -0,0 +1,24 @@
+# What did work:
+## Removing a double calculation when step
+I changed `config[i] = 0.5*previous[i] + 0.25*previous[(i-1) % size]+ 0.25*previous[(i+1) % size];` to `config[i] = 0.5*previous[i] + 0.25*previous[(i-1) % size]*2;` in step
+This made all the improvement as far as i can tell
+
+
+## In shift_centre_to_origin compute the ratio before the for loop
+```
+const float ratio = sum/size;
+for(long i = 0; i < size; i++) config[i] -= ratio;
+```
+This did not make a noticable difference
+
+# What did not work and actually made it slower:
+- Trying to find a fancy way to get a random bool in stochastic_unit_step
+- Using the std::minmax_element function in elongation
+- Changing the step function to be a void that changed the array directly made it about 8 secounds slower
+
+# Compiler optimizations
+I treid these flags: -Ofast -march=native -flto -fno-signed-zeros -fno-trapping-math
+The improvements were in the margin of error
+
+# Speedup
+I got at best a 1.1 times speedup. See 12_speedup.png