[AutoScheduler] Fix distill record (apache#7439)

* [AutoScheduler] Fix distill record

* update comments

python/tvm/auto_scheduler/measure_record.py

@@ -286,26 +286,42 @@ def distill_record_file(in_file, out_file):
     if os.path.isfile(out_file):
         out_context = load_records(out_file)
         context = itertools.chain(context, out_context)
-    context, context_clone = itertools.tee(context)
-    best_context = ApplyHistoryBest(context)
-    best_set = set()
 
     def measure_input_str_key(inp):
         return _ffi_api.SerializeMeasureInput(inp)
 
-    for v in best_context.best_by_model.values():
-        best_set.add(measure_input_str_key(v[0]))
+    # Dict[target key,
+    #   Dict[workload hash,
+    #     Dict[workload args, (cost, (MeasureInput, MeasureResult))]]]
+    # Full type: Dict[str, Dict[str, Dict[Tuple, Tuple[float, Tuple[Measureinput, MeasureResult]]]]]
+    best_records = {}
 
-    for v in best_context.best_by_targetkey.values():
-        best_set.add(measure_input_str_key(v[0]))
+    for inp, res in context:
+        if res.error_no != 0:
+            continue
+
+        # Keep the best record for each target and workload.
+        costs = [x.value for x in res.costs if isinstance(x, tvm.tir.expr.FloatImm)]
+        cost = np.mean(costs)
+        for k in inp.task.target.keys:
+            entry, _, workload_args = ApplyHistoryBest.get_workload_entry(
+                best_records, k, inp.task.workload_key
+            )
+            if workload_args not in entry or cost < entry[workload_args][0]:
+                entry[workload_args] = (cost, (inp, res))
+
+    # Remove duplications by multiple target keys.
+    out_records = {}
+    for target_entry in best_records.values():
+        for workload_entry in target_entry.values():
+            for _, (inp, res) in workload_entry.values():
+                out_records[measure_input_str_key(inp)] = (inp, res)
 
     inputs = []
     results = []
-    for inp, res in context_clone:
-        if measure_input_str_key(inp) in best_set:
-            inputs.append(inp)
-            results.append(res)
-            best_set.remove(measure_input_str_key(inp))
+    for inp, res in out_records.values():
+        inputs.append(inp)
+        results.append(res)
 
     # create a new file and save the best records
     open(out_file, "w")
@@ -316,23 +332,23 @@ def measure_input_str_key(inp):
 def main():
     """The main function for CLI."""
     parser = argparse.ArgumentParser()
-    parser.add_argument("--mode", choices=["distill"], required=True)
-    parser.add_argument("--i", type=str, help="input file")
-    parser.add_argument("--o", type=str, default=None, help="output file")
+    parser.add_argument("--mode", choices=["distill"], default="distill")
+    parser.add_argument("-i", "--input", type=str, help="input file")
+    parser.add_argument("-o", "--output", type=str, default=None, help="output file")
 
     args = parser.parse_args()
     logging.basicConfig()
     logger.setLevel(logging.INFO)
 
     if args.mode == "distill":
-        args.o = args.o or args.i + ".best.json"
-        distill_record_file(args.i, args.o)
+        args.output = args.output or args.input + ".best.json"
+        distill_record_file(args.input, args.output)
 
 
 """
 Usage:
 * Distill the best entries from a large log file
-e.g. python -m tvm.auto_scheduler.measure_record --mode distill --i input.json
+e.g. python -m tvm.auto_scheduler.measure_record --mode distill -i input.json
 """
 if __name__ == "__main__":
     main()

tutorials/auto_scheduler/tune_network_arm.py

@@ -408,7 +408,7 @@ def tune_and_evaluate():
 # 1. During the tuning, the auto-scheduler needs to compile many programs and
 #    extract feature from them. This part is CPU-intensive,
 #    so a high-performance CPU with many cores is recommended for faster search.
-# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill --i log.json`
+# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json`
 #    to distill the large log file and only save the best useful records.
 # 3. You can resume a search from the previous log file. You just need to
 #    add a new argument :code:`load_log_file` when creating the task scheduler

tutorials/auto_scheduler/tune_network_cuda.py

@@ -299,7 +299,7 @@ def run_tuning():
 # 1. During the tuning, the auto-scheduler needs to compile many programs and
 #    extract feature from them. This part is CPU-intensive,
 #    so a high-performance CPU with many cores is recommended for faster search.
-# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill --i log.json`
+# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json`
 #    to distill the large log file and only save the best useful records.
 # 3. You can resume a search from the previous log file. You just need to
 #    add a new argument :code:`load_log_file` when creating the task scheduler

tutorials/auto_scheduler/tune_network_mali.py

@@ -349,7 +349,7 @@ def tune_and_evaluate():
 # 1. During the tuning, the auto-scheduler needs to compile many programs and
 #    extract feature from them. This part is CPU-intensive,
 #    so a high-performance CPU with many cores is recommended for faster search.
-# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill --i log.json`
+# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json`
 #    to distill the large log file and only save the best useful records.
 # 3. You can resume a search from the previous log file. You just need to
 #    add a new argument :code:`load_log_file` when creating the task scheduler

tutorials/auto_scheduler/tune_network_x86.py

@@ -298,7 +298,7 @@ def run_tuning():
 # 1. During the tuning, the auto-scheduler needs to compile many programs and
 #    extract feature from them. This part is CPU-intensive,
 #    so a high-performance CPU with many cores is recommended for faster search.
-# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill --i log.json`
+# 2. You can use :code:`python3 -m tvm.auto_scheduler.measure_record --mode distill -i log.json`
 #    to distill the large log file and only save the best useful records.
 # 3. You can resume a search from the previous log file. You just need to
 #    add a new argument :code:`load_log_file` when creating the task scheduler