Make some optimizations toggleable, so results can be compared easily

OpenMP/linear-algebra/kernels/atax/.bench.sh

@@ -1,14 +1,46 @@
 #!/bin/bash
 
-runs=9
+run_benchmarks() {
-totalt=0.0
+    runs=25
+    totalt=0.0
 
-for i in $(seq $runs)
+    for i in $(seq $runs)
+    do
+        exet=$(./atax_acc)
+        totalt=$(awk "BEGIN{print $totalt+$exet}")
+        # echo "Run #$i: " $(awk "BEGIN{printf(\"%.3g\", $exet)}") "seconds"
+    done
+
+    avgt=$(awk "BEGIN{print $totalt/$runs}")
+    echo "  Average of $runs runs: " $(awk "BEGIN{printf(\"%.3g\", $avgt)}") "seconds"
+}
+
+for c in $(seq 0 15)
 do
-    exet=$(./atax_acc)
+    cflags=""
-    totalt=$(awk "BEGIN{print $totalt+$exet}")
-    echo "Run #$i: " $(awk "BEGIN{printf(\"%.3g\", $exet)}") "seconds"
-done
 
-avgt=$(awk "BEGIN{print $totalt/$runs}")
+    if (( $c & 1 ))
-echo "Average: " $(awk "BEGIN{printf(\"%.3g\", $avgt)}") "seconds"
+    then
+        cflags="$cflags -DTOGGLE_INIT_ARRAY_1"
+    fi 
+
+    if (( $c & 2 ))
+    then
+        cflags="$cflags -DTOGGLE_INIT_ARRAY_2"
+    fi 
+
+    if (( $c & 4 ))
+    then
+        cflags="$cflags -DTOGGLE_KERNEL_ATAX_1"
+    fi 
+
+    if (( $c & 8 ))
+    then
+        cflags="$cflags -DTOGGLE_KERNEL_ATAX_2"
+    fi 
+
+    echo "Flags: $cflags"
+    make "EXTRA_CFLAGS=$cflags" clean all
+
+    run_benchmarks
+done

OpenMP/linear-algebra/kernels/atax/Makefile

@@ -15,12 +15,14 @@ CFLAGS+= -O3
 CFLAGS+= -g3
 # -DTHREAD_COUNT allows us to alter the number of threads used in the whole file
 CFLAGS+= -DTHREAD_COUNT=4
+# Extend CFLAGS with command line parameters
+CFLAGS+= ${EXTRA_CFLAGS}
 
 # Disable make output
 MAKEFLAGS+= --silent
 
+
 .PHONY: bench
 
 bench:
-	make clean all
 	./.bench.sh

OpenMP/linear-algebra/kernels/atax/atax.c

@@ -25,13 +25,17 @@ static void init_array(int nx, int ny,
   int i, j;
 
   /// Initialize the `x` array with PI and its multiples.
+  #ifdef TOGGLE_INIT_ARRAY_1
   #pragma omp parallel for num_threads(THREAD_COUNT) schedule(static)
+  #endif
   for (i = 0; i < ny; i++) {
     x[i] = i * M_PI;
   }
 
   /// Initialize the `A` matrix with [something?]
+  #ifdef TOGGLE_INIT_ARRAY_2
   #pragma omp parallel for num_threads(THREAD_COUNT) schedule(static)
+  #endif
   for (i = 0; i < nx; i++) {
     for (j = 0; j < ny; j++) {
       A[i][j] = ((DATA_TYPE)i * (j + 1)) / nx;
@@ -63,14 +67,18 @@ static void kernel_atax(int nx, int ny,
 {
   int i, j;
 
+  #ifdef TOGGLE_KERNEL_ATAX_1
   #pragma omp parallel for num_threads(THREAD_COUNT) schedule(static)
+  #endif
   for (i = 0; i < _PB_NY; i++)
     y[i] = 0;
   
   /// This computes... something? I guess whatever ATAX is?
   // Now this gives a nice speedup, especially with a lot more threads than the count!
   // THREAD_COUNT * 4 seems to be the best on my local computer. What's the best for the Jetson Nano?
+  #ifdef TOGGLE_KERNEL_ATAX_2
   #pragma omp parallel for num_threads(THREAD_COUNT) schedule(static)
+  #endif
   for (i = 0; i < _PB_NX; i++)
   {
     /// Every iteration has its own tmp variable