Added a first version of the database's common-best default calculation

2 files changed, 64 insertions, 9 deletions

diff --git a/scripts/database/database.py b/scripts/database/database.py
index e115d68c..0ce89d03 100755
--- a/scripts/database/database.py
+++ b/scripts/database/database.py
@@ -98,7 +98,8 @@ def main(argv):
     database_best_results = bests.get_best_results(database)
 
     # Determines the defaults for other vendors and per vendor
-    database_defaults = defaults.calculate_defaults(database_best_results)
+    print("[database] Calculating the default values...")
+    database_defaults = defaults.calculate_defaults(database)
     database_best_results = db.concatenate_database(database_best_results, database_defaults)
 
     # Outputs the database as a C++ database
diff --git a/scripts/database/database/defaults.py b/scripts/database/database/defaults.py
index 357c3a3a..fca793ea 100644
--- a/scripts/database/database/defaults.py
+++ b/scripts/database/database/defaults.py
@@ -6,7 +6,9 @@
 #   Cedric Nugteren <www.cedricnugteren.nl>
 
 import pandas as pd
+
 import clblast
+import bests
 
 
 def set_default_device(database_entry):
@@ -23,16 +25,18 @@ def set_default_time(database_entry):
     return database_entry
 
 
-def calculate_defaults(df):
-    """# Sets defaults for devices of the same type/vendor based on the smallest values of all known entries. The average
-    might be better for performance but some parameters might not be supported on other devices."""
+def calculate_defaults(database, calculate_common_best=True):
+    """Sets defaults for devices of the same type/vendor. An option determines how to compute the defaults."""
     database_defaults = pd.DataFrame()
 
     # Defaults per combination of device vendors and device types (e.g. AMD GPU)
-    database_type_vendor = df.groupby(clblast.DEVICE_TYPE_ATTRIBUTES + clblast.KERNEL_ATTRIBUTES + ["kernel"] +
-                                      clblast.ARGUMENT_ATTRIBUTES)
+    database_type_vendor = database.groupby(clblast.DEVICE_TYPE_ATTRIBUTES + clblast.KERNEL_ATTRIBUTES + ["kernel"] +
+                                            clblast.ARGUMENT_ATTRIBUTES)
     for group_name, database_group in database_type_vendor:
-        default_values = database_group.min(axis=0)
+        if calculate_common_best:
+            default_values = get_common_best(database_group, group_name)
+        else:
+            default_values = get_smallest_best(database_group)
         default_values = set_default_device(default_values)
         default_values = set_default_time(default_values)
         database_defaults = database_defaults.append(default_values, ignore_index=True)
@@ -45,9 +49,9 @@ def calculate_defaults(df):
             print("[WARNING] Entries for a single kernel with multiple argument values: " + description)
 
     # Defaults over all device types and vendors
-    groups = df.groupby(clblast.KERNEL_ATTRIBUTES + ["kernel"] + clblast.ARGUMENT_ATTRIBUTES)
+    groups = database.groupby(clblast.KERNEL_ATTRIBUTES + ["kernel"] + clblast.ARGUMENT_ATTRIBUTES)
     for group_name, database_group in groups:
-        default_values = database_group.min(axis=0)
+        default_values = get_smallest_best(database_group)
         default_values["device_vendor"] = clblast.VENDOR_DEFAULT
         default_values["device_type"] = clblast.DEVICE_TYPE_DEFAULT
         default_values = set_default_device(default_values)
@@ -56,3 +60,53 @@ def calculate_defaults(df):
 
     # Database with both types of defaults only
     return database_defaults
+
+
+def get_smallest_best(database):
+    """Sets defaults based on the smallest values of all known entries. The average might be better for performance but
+    some parameters might not be supported on other devices."""
+    database_best_results = bests.get_best_results(database)
+    return database_best_results.min(axis=0)
+
+
+def get_common_best(database, group_name):
+    """Sets defaults based on the best values of entries supported by all devices. This might cause a problem in case
+    not every device was tuned with the same parameters."""
+    # TODO: Quite a bit slower than the above `get_smallest_best` method
+
+    # Counts the number of devices in this group
+    num_devices = len(database.groupby(clblast.DEVICE_ATTRIBUTES))
+
+    # Removes columns without any values
+    database = database.dropna(axis=1, how='all')
+
+    # Retrieves the parameter names for this kernel
+    all_column_names = list(database.columns.values)
+    parameter_column_names = [c for c in all_column_names if "parameters." in c]
+
+    # Removes entries which are not available for all devices
+    database_common = pd.DataFrame()
+    database_by_parameters = database.groupby(parameter_column_names)
+    for parameter_values, database_parameters in database_by_parameters:
+        num_entries = database_parameters.shape[0]
+        if num_entries == num_devices:
+            database_common = database_common.append(database_parameters)
+
+    # Fall back to another method in case there are no shared entries at all across devices
+    if database_common.shape[0] == 0:
+        # print("Skipping: " + str(group_name) + " with devices: " + str(num_devices) + " " + str(database.shape[0]))
+        return get_smallest_best(database)
+
+    # Computes the sum of the execution times over the different devices
+    database_common['time'] = database_common.groupby(parameter_column_names)['time'].transform(sum)
+
+    # Retrieves the entries with the best execution time
+    best_time = database_common["time"].min()
+    database_bests = database_common[database_common["time"] == best_time]
+
+    # Retrieves one example only (the parameters are the same anyway)
+    database_bests = database_bests.drop_duplicates(["time"])
+    # print(str(group_name) + " with num devices: " + str(num_devices) + " " + str(database_bests.shape))
+    assert database_bests.shape[0] == 1
+
+    return database_bests