[apps, scripts] Adjust kernel performance

apps/fft/main.c

@@ -192,7 +192,7 @@ int main() {
   stop_timer();
   runtime = get_timer();
 
-  float perf = (float)10.0 * NFFT * (31 - __builtin_clz(NFFT)) / runtime;
+  float perf = (float)5.0 * NFFT * (31 - __builtin_clz(NFFT)) / runtime;
   float max_perf = 6.0 / 5.0 * NR_LANES * 8.0 / sizeof(float);
 
   printf("Performance: %f. Max perf: %f. Actual performance is %f%% of max.\n",

apps/jacobi2d/kernel/jacobi2d.c

@@ -74,6 +74,7 @@ WITH ACCESS OR USE OF THE SOFTWARE.
 // Author: Matteo Perotti, ETH Zurich, <[email protected]>
 
 #include "jacobi2d.h"
+#define DOUBLE_BUFFERING
 
 void j2d_s(uint64_t r, uint64_t c, DATA_TYPE *A, DATA_TYPE *B,
            uint64_t tsteps) {

apps/roi_align/main.c

@@ -22,7 +22,7 @@
 
 // Execute only the central kernel with fake data
 // Execute only on the channels
-//#define V_KERNEL_ONLY 1
+#define V_KERNEL_ONLY 1
 
 #define EXTRAPOLATION_VALUE 0
 

scripts/performance.py

@@ -62,7 +62,7 @@ def fconv3d(args, cycles):
 def jacobi2d(args, cycles):
   size        = int(args[0])
   trash_0     = args[1]
-  performance = 5 * (size-1) * (size-1) / cycles
+  performance = 2 * 5 * (size-1) * (size-1) / cycles
   return [size, performance]
 def dropout(args, cycles):
   size        = int(args[0])
@@ -71,7 +71,7 @@ def dropout(args, cycles):
 def fft(args, cycles):
   size        = int(args[0])
   dtype       = args[1]
-  performance = 10 * size * np.log2(size) / cycles
+  performance = 5 * size * np.log2(size) / cycles
   return [size, performance]
 def dwt(args, cycles):
   size        = int(args[0])