Merge branch 'main' of https://github.com/Accelergy-Project/accelforge into main

Author: tanner-andrulis

accelforge/mapper/FFM/_make_pmappings/make_pmappings_from_templates/make_tile_shapes.py

@@ -761,6 +761,11 @@ def _try_admit(n: int) -> bool:
         else:
             n += 1
 
+        # Force n to be a multiple of the inner size. ctrl-f for
+        # IMPERFECT_OUTER_ONLY_ASSUMPTION. Note we also require that the enumeration
+        # order is inner -> outer if we're doing imperfect.
+        n = math.ceil(n / inner_size) * inner_size
+
     # One more in case coarseness jumped the max value
     _try_admit(outer_size)
 
@@ -906,8 +911,9 @@ def get_padded_choices(
         if subs in choices_padded:
             choices_padded[s] = choices_padded[subs]
         else:
-            choices_padded[s] = ones * subs
-
+            choices_padded[s] = np.full(
+                choices_enumerated.shape[0], subs, dtype=choices_enumerated.dtype
+            )
     for k, v in choices_padded.items():
         assert isinstance(v, np.ndarray), f"{k} is not a numeric array: {v}"
 
@@ -1219,10 +1225,11 @@ def hybrid(s):
         if isinstance(tiled_by, Symbol) and tiled_by not in enumerated_set:
             score_unknown -= 1
 
-        # Score for bringing in an entirely-new formula
+        # Score for bringing in an entirely-new formula. NOTE: This caused memory to
+        # blow up for some workloads when the punishment was set higher.
         for o in objectives:
             if s in o._free_symbols and not o._free_symbols & enumerated_set:
-                score_unknown -= 10
+                score_unknown -= .1
 
         return (
             -keep,  # Punish enumerating keep symbols because they're always diff
@@ -1781,10 +1788,10 @@ def update_symbol2goal(
                         # Haven't done any pruning, so valid is the # of total choices
                         valid = [choices_enumerated.shape[0]]
 
-                    porp = sum(valid) / max(1, choices_enumerated.shape[0])
+                    porp = int(np.sum(valid)) / max(1, choices_enumerated.shape[0])
                     job.log_porp_pmappings_kept(
                         f"{objective.name}",
-                        sum(valid) / max(1, prev_size),
+                        int(np.sum(valid)) / max(1, prev_size),
                     )
                     DEBUG and log_message(
                         f"Valid check", f"{objective.name}", f"porp={porp:.2%}"
@@ -1832,10 +1839,10 @@ def update_symbol2goal(
                         # Everyone valid (for counting purposes)
                         valid = [choices_enumerated.shape[0]]
 
-                    porp = sum(valid) / max(1, choices_enumerated.shape[0])
+                    porp = int(np.sum(valid)) / max(1, choices_enumerated.shape[0])
                     job.log_porp_pmappings_kept(
                         f"{objective.name}",
-                        sum(valid) / max(1, prev_size),
+                        int(np.sum(valid)) / max(1, prev_size),
                     )
                     DEBUG and log_message(
                         f"Valid check", f"{objective.name}", f"porp={porp:.2%}"
@@ -1980,7 +1987,7 @@ def update_symbol2goal(
                 prev_size = choices_enumerated.shape[0]
                 choices_enumerated = choices_enumerated[keep]
                 job.log_porp_pmappings_kept(
-                    f"Pareto", sum(keep) / choices_enumerated.shape[0]
+                    f"Pareto", int(np.sum(keep)) / choices_enumerated.shape[0]
                 )
                 DEBUG and log_message(
                     "pareto", f"size {prev_size} -> {choices_enumerated.shape[0]}"

tests/hwcomponents_expected.json

@@ -15,6 +15,8 @@
       }
     },
     "GlobalBuffer": {
+      "area": 6.8574526567307686e-06,
+      "leak_power": 0.00025508039081958506,
       "area": 6.8574526567307686e-06,
       "leak_power": 0.00025508039081958506,
       "actions": {
@@ -29,6 +31,8 @@
       }
     },
     "InputScratchpad": {
+      "area": 3.8736612980769225e-10,
+      "leak_power": 4.3815839033400895e-07,
       "area": 3.8736612980769225e-10,
       "leak_power": 4.3815839033400895e-07,
       "actions": {
@@ -43,6 +47,8 @@
       }
     },
     "WeightScratchpad": {
+      "area": 3.534394038461538e-09,
+      "leak_power": 7.663684917510134e-07,
       "area": 3.534394038461538e-09,
       "leak_power": 7.663684917510134e-07,
       "actions": {
@@ -57,6 +63,8 @@
       }
     },
     "OutputScratchpad": {
+      "area": 4.5613394230769223e-10,
+      "leak_power": 4.4143907783400897e-07,
       "area": 4.5613394230769223e-10,
       "leak_power": 4.4143907783400897e-07,
       "actions": {
@@ -71,6 +79,8 @@
       }
     },
     "MAC": {
+      "area": 1.1748028846153846e-09,
+      "leak_power": 8.851870775236185e-06,
       "area": 1.1748028846153846e-09,
       "leak_power": 8.851870775236185e-06,
       "actions": {
@@ -97,6 +107,8 @@
       }
     },
     "GlobalBuffer": {
+      "area": 2.3951167124150845e-07,
+      "leak_power": 9.50865964351506e-06,
       "area": 2.3951167124150845e-07,
       "leak_power": 9.50865964351506e-06,
       "actions": {
@@ -111,6 +123,8 @@
       }
     },
     "InputBuffer": {
+      "area": 5.391208241958042e-09,
+      "leak_power": 1.2069544874830208e-06,
       "area": 5.391208241958042e-09,
       "leak_power": 1.2069544874830208e-06,
       "actions": {
@@ -125,6 +139,8 @@
       }
     },
     "WeightBuffer": {
+      "area": 1.0703430381765736e-07,
+      "leak_power": 1.2611510646849734e-05,
       "area": 1.0703430381765736e-07,
       "leak_power": 1.2611510646849734e-05,
       "actions": {
@@ -139,6 +155,8 @@
       }
     },
     "AccumulationBuffer": {
+      "area": 2.2003072874125873e-09,
+      "leak_power": 6.034779422382222e-07,
       "area": 2.2003072874125873e-09,
       "leak_power": 6.034779422382222e-07,
       "actions": {
@@ -167,6 +185,8 @@
       }
     },
     "MAC": {
+      "area": 1.4715951923076922e-10,
+      "leak_power": 1.2382323818087997e-06,
       "area": 1.4715951923076922e-10,
       "leak_power": 1.2382323818087997e-06,
       "actions": {