From 90588c3474fd87afbf3e7984d6503be2a06f3aa8 Mon Sep 17 00:00:00 2001
From: Matteo Perotti <mperotti@iis.ee.ethz.ch>
Date: Thu, 7 Sep 2023 17:21:28 +0200
Subject: [PATCH] [apps, scripts] Adjust kernel performance

---
 apps/fft/main.c                 | 2 +-
 apps/jacobi2d/kernel/jacobi2d.c | 1 +
 apps/roi_align/main.c           | 2 +-
 scripts/performance.py          | 4 ++--
 4 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/apps/fft/main.c b/apps/fft/main.c
index 6283287ed..b74037f73 100644
--- a/apps/fft/main.c
+++ b/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",
diff --git a/apps/jacobi2d/kernel/jacobi2d.c b/apps/jacobi2d/kernel/jacobi2d.c
index b72ff7759..a815645a6 100644
--- a/apps/jacobi2d/kernel/jacobi2d.c
+++ b/apps/jacobi2d/kernel/jacobi2d.c
@@ -74,6 +74,7 @@ WITH ACCESS OR USE OF THE SOFTWARE.
 // Author: Matteo Perotti, ETH Zurich, <mperotti@iis.ee.ethz.ch>
 
 #include "jacobi2d.h"
+#define DOUBLE_BUFFERING
 
 void j2d_s(uint64_t r, uint64_t c, DATA_TYPE *A, DATA_TYPE *B,
            uint64_t tsteps) {
diff --git a/apps/roi_align/main.c b/apps/roi_align/main.c
index 2e97d04f6..709d6fec4 100644
--- a/apps/roi_align/main.c
+++ b/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
 
diff --git a/scripts/performance.py b/scripts/performance.py
index ae9bbbece..c25b380eb 100755
--- a/scripts/performance.py
+++ b/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])