[Ansor][AutoTVM v2.0] Phase 1: XGBoost Cost Model (#6270)

* port xgb cost model

* add xgboost cost model

* fix lint

* address comments

* address comments

* Fix

include/tvm/auto_scheduler/feature.h

@@ -21,7 +21,7 @@
  * \file auto_scheduler/feature.h
  * \brief Feature extraction for the cost model.
  * We extract one feature vector per BufferStoreNode statement in a TIR Stmt,
- * so we call this feature as "Per Store" feature.
+ * so we call this feature as "per-store" feature.
  * The cost model also does prediction for each BufferStoreNode statement and aggregates
  * the predictions as the whole score for a TVM IR (Stmt).
  *

python/tvm/auto_scheduler/__init__.py

@@ -29,8 +29,8 @@
 from .auto_schedule import SearchTask, TuningOptions, HardwareParams, \
     auto_schedule, EmptyPolicy, SketchPolicy
 from .compute_dag import ComputeDAG
-from .cost_model import RandomModel
-from .measure import MeasureInput, LocalBuilder, LocalRunner, RPCRunner, \
+from .cost_model import RandomModel, XGBModel
+from .measure import MeasureInput, MeasureResult, LocalBuilder, LocalRunner, RPCRunner, \
     LocalRPCMeasureContext
 from .measure_record import RecordToFile, RecordReader, load_best, \
     load_records, save_records

python/tvm/auto_scheduler/auto_schedule.py

@@ -161,7 +161,9 @@ def __init__(self, task, schedule_cost_model=RandomModel(), params=None, seed=No
             seed or random.randint(1, 1 << 30), verbose, init_search_callbacks)
 
     def generate_sketches(self, print_for_debug=False):
-        """ Generate the sketches, this is mainly used for debug.
+        """ Generate the sketches.
+        This python interface is mainly used for debugging and testing.
+        The actual search is all doen in c++.
 
         Parameters
         ----------
@@ -180,6 +182,24 @@ def generate_sketches(self, print_for_debug=False):
                 print(s)
         return sketches
 
+    def sample_initial_population(self, pop_size):
+        """Sample initial population.
+        This python interface is mainly used for debugging and testing.
+        The actual search is all doen in c++.
+
+        Parameters
+        ----------
+        pop_size : int
+            The size of sampled population
+
+        Returns
+        -------
+        states: List[State]
+            The sampled states
+        """
+        states = _ffi_api.SketchPolicySampleInitialPopulation(self, pop_size)
+        return states
+
 @tvm._ffi.register_object("auto_scheduler.TuningOptions")
 class TuningOptions(Object):
     """ This controls the options of performance tuning.

python/tvm/auto_scheduler/cost_model/__init__.py

@@ -18,3 +18,4 @@
 """ Cost model that estimates the performance of programs """
 
 from .cost_model import RandomModel
+from .xgb_model import XGBModel

python/tvm/auto_scheduler/cost_model/cost_model.py

@@ -146,5 +146,25 @@ def predict_stages(self, task, states):
         -------
         scores: List[float]
             The predicted scores for all stages in all states in the packed format
+
+        Note
+        ----
+        For faster data copy between c++ and python, the python part returns scores in a
+        single flatten array using a packed format. The c++ part then unpacks the flatten array.
+
+        The packed format is:
+        {
+          float  scores[N];                 // scores[i] is the score for states[i].
+          int    n_stage_0;                 // the number of stages in states[0]
+          float  stage_scores_0[[n_stage_0] // the scores for all stages in states[0]
+          int    n_stage_1;                 // the number of stages in states[1]
+          float  stage_scores_1[n_stage_1]; // the scores for all stages in states[1]
+          ...
+          int    n_stage_i;                 // the number of stages in states[i]
+          float  stage_scores_1[n_stage_i]; // the scores for all stages in states[i]
+          ...  // untill i == N - 1
+        }
+        To implement this format, we also store int as float, so we can store all numbers
+        into a single float array.
         """
         raise NotImplementedError