Merge branch 'master' of github.com:pc2/StencilStream

examples/conway/build_nalla.sh

@@ -6,4 +6,4 @@
 #SBATCH --mem=90GB
 #SBATCH --time=3-00:00:00
 
-../build.sh nalla_pcie 10 conway
+../build.sh nalla_pcie/19.4.0_max 10 conway

examples/fdtd/build_nalla.sh

@@ -6,4 +6,4 @@
 #SBATCH --exclusive
 #SBATCH --time=3-00:00:00
 
-../build.sh nalla_pcie 35 fdtd
+../build.sh nalla_pcie/19.4.0_max 35 fdtd

examples/fdtd/src/fdtd.cpp

@@ -95,20 +95,18 @@ int main(int argc, char **argv)
             std::cout << "Run " << i << " with " << (i + 1) * parameters.n_sample_steps << " passes took " << runtimes.back() << " seconds" << std::endl;
         }
 
-        double delta_passes = parameters.n_sample_steps / pipeline_length;
-
         // Actually the mean delta seconds per pass.
         double delta_seconds_per_pass = 0.0;
         for (uindex_t i = 0; i < runtimes.size() - 1; i++)
         {
-            delta_seconds_per_pass += abs(double(runtimes[i + 1]) - double(runtimes[i])) / delta_passes;
+            delta_seconds_per_pass += abs(double(runtimes[i + 1]) - double(runtimes[i]));
         }
-        delta_seconds_per_pass /= runtimes.size() - 1;
+        delta_seconds_per_pass = (delta_seconds_per_pass / (runtimes.size() - 1)) * (pipeline_length / paraemeters.n_sample_steps);
 
         std::cout << "Time per buffer pass: " << delta_seconds_per_pass << "s" << std::endl;
 
         double loops_per_pass = n_buffer_rows * n_buffer_columns;
-        double seconds_per_loop = delta_seconds_per_pass * loops_per_pass;
+        double seconds_per_loop = delta_seconds_per_pass / loops_per_pass;
         double cycles_per_loop = seconds_per_loop * clock_frequency;
 
         std::cout << "Cycles per Loop, aka II.: " << cycles_per_loop << std::endl;

examples/hotspot/analyze_benchmark.py

@@ -7,14 +7,12 @@
 automatically. The output is formatted like a Python dict, but you need to remove error and output
 messages. Lastly, you need to copy the cycle frequency and the pipeline length from the reports into this script. 
 """
-from matplotlib import pyplot
-import numpy as np
 
 from benchmark import runtime
-from verify_benchmark import value_derivation
 
 cycle_frequency = 79.63 * 10**6
 pipeline_length = 225
+step_size = 900
 
 height = width = 1024
 radius = 1
@@ -29,7 +27,7 @@
     offset_runtime = [original_runtime[0]] + original_runtime
 
     delta = [current_sample - next_sample for (current_sample, next_sample) in zip(original_runtime, offset_runtime)]
-    mean_delta = sum(delta) / len(delta)
+    mean_delta = (sum(delta) / len(delta)) * (pipeline_length / step_size)
 
     print("\tDelta Time per Grid Pass: {:.2f} milliseconds".format(mean_delta * 10**3))
 
@@ -47,19 +45,4 @@
 
     throughput = (2 * cell_size * width * height) / mean_delta * 10**-9
 
-    print("\tGlobal Memory Throughput: {:.2f} GB/s".format(throughput))
-
-pyplot.title("Ratio of correct values")
-for length in value_derivation:
-    pyplot.plot([deriv[0] for deriv in value_derivation[length]])
-pyplot.show()
-
-pyplot.title("Mean result derivation")
-for length in value_derivation:
-    pyplot.plot([deriv[1] for deriv in value_derivation[length]])
-pyplot.show()
-
-pyplot.title("Maximal absolute result derivation")
-for length in value_derivation:
-    pyplot.plot([deriv[2] for deriv in value_derivation[length]])
-pyplot.show()
+    print("\tGlobal Memory Throughput: {:.2f} GB/s".format(throughput))

examples/hotspot/build_nalla.sh

@@ -6,4 +6,4 @@
 #SBATCH --exclusive
 #SBATCH --time=3-00:00:00
 
-../build.sh nalla_pcie 225 hotspot
+../build.sh nalla_pcie/19.4.0_max 200 hotspot

examples/hotspot/hotspot.cpp

@@ -276,20 +276,18 @@ int main(int argc, char **argv)
             std::cout << "Run " << i << " with " << (i + 1) * sim_time << " passes took " << runtimes[i] << " seconds" << std::endl;
         }
 
-        double delta_passes = sim_time / pipeline_length;
-
         // Actually the mean delta seconds per pass.
         double delta_seconds_per_pass = 0.0;
         for (uindex_t i = 0; i < n_simulations - 1; i++)
         {
-            delta_seconds_per_pass += abs(double(runtimes[i + 1]) - double(runtimes[i])) / delta_passes;
+            delta_seconds_per_pass += abs(double(runtimes[i + 1]) - double(runtimes[i]));
         }
-        delta_seconds_per_pass /= n_simulations - 1;
+        delta_seconds_per_pass = (delta_seconds_per_pass / (n_simulations - 1)) * (pipeline_length / sim_time);
 
         std::cout << "Time per buffer pass: " << delta_seconds_per_pass << "s" << std::endl;
 
         double loops_per_pass = max_width * max_height;
-        double seconds_per_loop = delta_seconds_per_pass * loops_per_pass;
+        double seconds_per_loop = delta_seconds_per_pass / loops_per_pass;
         double cycles_per_loop = seconds_per_loop * clock_frequency;
 
         std::cout << "Cycles per Loop, aka II.: " << cycles_per_loop << std::endl;

examples/main/build_nalla.sh

@@ -6,4 +6,4 @@
 #SBATCH --mem=60GB
 #SBATCH --time=3-00:00:00
 
-../build.sh nalla_pcie 10 main
+../build.sh nalla_pcie/19.4.0_max 10 main