Implemented the pyatf_strategies, which enables using the pyATF strategies in Kernel Tuner

Author: fjwillemsen

kernel_tuner/interface.py

@@ -47,7 +47,6 @@
 except ImportError:
     torch = util.TorchPlaceHolder()
 
-from kernel_tuner.strategies.wrapper import OptAlgWrapper
 from kernel_tuner.strategies import (
     basinhopping,
     bayes_opt,
@@ -62,9 +61,11 @@
     mls,
     ordered_greedy_mls,
     pso,
+    pyatf_strategies,
     random_sample,
-    simulated_annealing
+    simulated_annealing,
 )
+from kernel_tuner.strategies.wrapper import OptAlgWrapper
 
 strategy_map = {
     "brute_force": brute_force,
@@ -81,7 +82,8 @@
     "pso": pso,
     "simulated_annealing": simulated_annealing,
     "firefly_algorithm": firefly_algorithm,
-    "bayes_opt": bayes_opt
+    "bayes_opt": bayes_opt,
+    "pyatf_strategies": pyatf_strategies,
 }
 
 
@@ -629,6 +631,7 @@ def tune_kernel(
     logging.debug("tuning_options: %s", util.get_config_string(tuning_options))
     logging.debug("device_options: %s", util.get_config_string(device_options))
 
+    strategy_string = strategy
     if strategy:
         if strategy in strategy_map:
             strategy = strategy_map[strategy]
@@ -861,10 +864,9 @@ def tune_kernel_T1(
     strategy: str=None,
     strategy_options: dict={},
 ) -> tuple:
-    """
-    Call the tune function with a T1 input file.
+    """Call the tune function with a T1 input file.
     
-        The device, strategy and strategy_options can be overridden by passing a strategy name and options, otherwise the input file specification is used.
+    The device, strategy and strategy_options can be overridden by passing a strategy name and options, otherwise the input file specification is used.
     """
     inputs = get_input_file(input_filepath)
     kernelspec: dict = inputs["KernelSpecification"]

kernel_tuner/searchspace.py

@@ -76,6 +76,7 @@ def __init__(
         framework_l = framework.lower()
         restrictions = restrictions if restrictions is not None else []
         self.tune_params = tune_params
+        self.tune_params_pyatf = None
         self._tensorspace = None
         self.tensor_dtype = torch.float32 if torch_available else None
         self.tensor_device = torch.device("cpu") if torch_available else None
@@ -376,10 +377,13 @@ def get_params():
                     constraint = res
                 params.append(TP(key, vals, constraint, constraint_source))
             return params
+        
+        # set data
+        self.tune_params_pyatf = get_params()
 
         # tune
         _, _, tuning_data = (
-            Tuner().verbosity(0).tuning_parameters(*get_params()).search_technique(Exhaustive()).tune(costfunc)
+            Tuner().verbosity(0).tuning_parameters(*self.tune_params_pyatf).search_technique(Exhaustive()).tune(costfunc)
         )
 
         # transform the result into a list of parameter configurations for validation

kernel_tuner/strategies/common.py

@@ -356,7 +356,6 @@ def scale_from_params(params, tune_params, eps):
 
 def unscale_and_snap_to_nearest_valid(x, params, searchspace, eps):
     """Helper func to snap to the nearest valid configuration"""
-
     # params is nearest unscaled point, but is not valid
     neighbors = get_neighbors(params, searchspace)