Updated the changelog; refactored the database-get-bests code a bit

3 files changed, 28 insertions, 20 deletions

diff --git a/CHANGELOG b/CHANGELOG
index 8fce1969..852b734f 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -8,6 +8,7 @@ Development version (next release)
 - Fixed a performance issue (caused by fp16 support) by optimizing alpha/beta parameter passing to kernels
 - Added an option (-warm_up) to do a warm-up run before timing in the performance clients
 - Improved performance significantly of rotated GEMV computations
+- Improved performance of unseen/un-tuned devices by a better default tuning parameter selection
 - Various minor fixes and enhancements
 - Added tuned parameters for various devices (see README)
 
diff --git a/scripts/database/database/bests.py b/scripts/database/database/bests.py
index edb81733..e6239258 100644
--- a/scripts/database/database/bests.py
+++ b/scripts/database/database/bests.py
@@ -18,3 +18,26 @@ def get_best_results(df):
         best_parameters = database_entry[database_entry["time"] == best_time].iloc[0]
         database_bests = database_bests.append(best_parameters, ignore_index=True)
     return database_bests
+
+
+def get_relative_bests(df, parameter_column_names, name, verbose=False):
+    """Retrieves the relative best execution time over different devices"""
+
+    # Computes the sum of the execution times over the different devices
+    def sum_performance(x):
+        x["group_performance"] = x["relative_performance"].sum()
+        return x
+    df = df.groupby(parameter_column_names).apply(sum_performance)
+
+    # Retrieves the entries with the highest performance
+    best_performance = df["group_performance"].max()
+    df_bests = df[df["group_performance"] == best_performance]
+
+    # Retrieves one example only (the parameters are the same anyway)
+    df_bests = df_bests.drop_duplicates(["group_performance"])
+
+    # Completed, report and return the results
+    if verbose:
+        print("[database] " + str(name) + " with performance " + str(best_performance) + " " + str(df_bests.shape))
+    assert len(df_bests) == 1
+    return df_bests
diff --git a/scripts/database/database/defaults.py b/scripts/database/database/defaults.py
index 985f24bd..3428d9a9 100644
--- a/scripts/database/database/defaults.py
+++ b/scripts/database/database/defaults.py
@@ -97,25 +97,9 @@ def get_common_best(database, group_name, verbose):
 
     # Fall back to another method in case there are no shared entries at all across devices
     if len(database_common) == 0:
-        # print("[database] Skipping: " + str(group_name) + " with devices: %d %d " % (num_devices, len(database)))
+        if verbose:
+            print("[database] No common kernels for: " + str(group_name) + " with devices: %d " % num_devices)
         return get_smallest_best(database)
 
-    # Computes the sum of the execution times over the different devices
-    def sum_performance(x):
-        x["group_performance"] = x["relative_performance"].sum()
-        return x
-    database_common = database_common.groupby(parameter_column_names).apply(sum_performance)
-
-    # Retrieves the entries with the highest performance
-    best_performance = database_common["group_performance"].max()
-    database_bests = database_common[database_common["group_performance"] == best_performance]
-
-    # Retrieves one example only (the parameters are the same anyway)
-    database_bests = database_bests.drop_duplicates(["group_performance"])
-
-    # Completed, report and return the results
-    if verbose:
-        print("[database] " + str(group_name) + " with performance " + str(best_performance) + " with devices: " +
-        str(num_devices) + " " + str(database_bests.shape))
-    assert len(database_bests) == 1
-    return database_bests
+    # Retrieves the entries with the highest relative performance
+    return bests.get_relative_bests(database_common, parameter_column_names, group_name, verbose)