Merge branch 'main' of github.com:Accelergy-Project/timeloop-python

pytimeloop/fastfusion/layerdeduplication.py

@@ -1,5 +1,5 @@
 from collections import defaultdict
-from itertools import permutations
+from itertools import permutations, product
 
 from pytimeloop.looptree.mapping_utilities import get_intermediate_tensors
 
@@ -26,69 +26,72 @@ def is_equivalent(einsum_id1, einsum_id2, workload, analyzer):
     if einsum2_output_tensor is None:
         einsum2_output_tensor = set()
 
-    einsum1_tensors = einsum1_input_tensors | einsum1_output_tensor
-    einsum2_tensors = einsum2_input_tensors | einsum2_output_tensor
-
     intermediate_tensors = get_intermediate_tensors(workload)
 
-    tensor_properties = defaultdict(set)
-    for tensor in einsum1_input_tensors | einsum2_input_tensors:
-        tensor_properties[tensor].add('input')
-    for tensor in einsum1_output_tensor | einsum2_output_tensor:
-        tensor_properties[tensor].add('input')
-    for tensor in intermediate_tensors:
-        if tensor not in tensor_properties:
-            continue
-        tensor_properties[tensor].add('intermediate')
-    tensor_properties = {
-        tensor: frozenset(properties)
-        for tensor, properties in tensor_properties.items()
-    }
-    property_to_tensors = defaultdict(lambda: (set(), set()))
-    for tensor, property in tensor_properties:
-        tensor_sets = property_to_tensors[property]
-        if tensor in einsum1_tensors:
-            tensor_sets[0].add(tensor)
-        else:
-            tensor_sets[1].add(tensor)
-
-    for tensor_sets in property_to_tensors.values():
-        if len(tensor_sets[0]) != len(tensor_sets[1]):
-            return None, None
-
-
-
-    # Check if we can rename einsum1 ranks to create einsum2
-    for renamed_ranks in permutations(einsum2_ranks):
-        rank_renaming = {
-            r1: r2 for r1, r2 in zip(einsum1_ranks, renamed_ranks)
+    all_tensor_properties = []
+    all_tensors = [
+        (einsum1_input_tensors, einsum1_output_tensor),
+        (einsum2_input_tensors, einsum2_output_tensor)
+    ]
+    for input_tensors, output_tensors in all_tensors:
+        tensor_properties = defaultdict(set)
+        for tensor in input_tensors:
+            tensor_properties[tensor].add('input')
+        for tensor in output_tensors:
+            tensor_properties[tensor].add('output')
+        for tensor in tensor_properties:
+            if tensor in intermediate_tensors:
+                tensor_properties[tensor].add('intermediate')
+        tensor_properties = {
+            tensor: frozenset(properties)
+            for tensor, properties in tensor_properties.items()
         }
-        # for tensor_renaming in get_tensor_renamings(property_to_tensors):
-        for renamed_input_tensors in permutations(einsum2_input_tensors):
-            input_tensor_renaming = {
-                t1: t2 for t1, t2
-                in zip(einsum1_input_tensors, renamed_input_tensors)
+        all_tensor_properties.append(tensor_properties)
+
+    property_to_tensors = defaultdict(lambda: (set(), set()))
+    for i, tensor_properties in enumerate(all_tensor_properties):
+        for tensor, property in tensor_properties.items():
+            tensor_sets = property_to_tensors[property]
+            tensor_sets[i].add(tensor)
+
+    # Check if we can rename tensors in einsum1 to einsum2
+    for tensor_renaming in tensor_renamings(property_to_tensors):
+        # Check if we can rename einsum1 ranks to create einsum2
+        for renamed_ranks in permutations(einsum2_ranks):
+            rank_renaming = {
+                r1: r2 for r1, r2 in zip(einsum1_ranks, renamed_ranks)
             }
-            for renamed_output_tensors in permutations(einsum2_output_tensor):
-                output_tensor_renaming = {
-                    t1: t2 for t1, t2
-                    in zip(einsum1_output_tensor, renamed_output_tensors)
-                }
-                tensor_renaming = input_tensor_renaming | output_tensor_renaming
+            if not _shape_is_equivalent(rank_renaming, workload):
+                continue
 
-                if not _shape_is_equivalent(rank_renaming, workload):
-                    continue
+            if not _dependency_is_equivalent(einsum_id1,
+                                            einsum_id2,
+                                            rank_renaming,
+                                            tensor_renaming,
+                                            analyzer):
+                continue
 
-                if not _dependency_is_equivalent(einsum_id1,
-                                                einsum_id2,
-                                                rank_renaming,
-                                                tensor_renaming,
-                                                analyzer):
-                    continue
+            return rank_renaming, tensor_renaming
+    return None, None
 
-                return rank_renaming, tensor_renaming
 
-    return None, None
+def tensor_renamings(property_to_tensors):
+    for tensors_of_1, tensors_of_2 in property_to_tensors.values():
+        if len(tensors_of_1) != len(tensors_of_2):
+            return
+
+    all_tensors_of_1 = [
+        t
+        for tensors_of_1, _ in property_to_tensors.values()
+        for t in tensors_of_1
+    ]
+    permutations_of_tensor_2_by_property = []
+    for _, tensors_of_2 in property_to_tensors.values():
+        permutations_of_tensor_2_by_property.append(permutations(tensors_of_2))
+    for permutation_of_2 in product(*permutations_of_tensor_2_by_property):
+        permutation_of_2 = tuple(t for tupl in permutation_of_2 for t in tupl)
+        renaming = dict(zip(all_tensors_of_1, permutation_of_2))
+        yield renaming
 
 
 def _shape_is_equivalent(rank_renaming, workload):

pytimeloop/fastfusion/mapper/mapper2.py

@@ -1,9 +1,12 @@
+from collections import defaultdict
+from copy import deepcopy
 import logging.handlers
 from pathlib import Path
 import logging
 logger = logging.getLogger(__name__)
 
 from ruamel.yaml import YAML
+from joblib import Parallel, delayed
 
 yaml = YAML(typ="safe")
 
@@ -14,7 +17,8 @@
 from pytimeloop.fastfusion.mapper.constraints import *
 from pytimeloop.fastfusion.layerdeduplication import is_equivalent
 from pytimeloop.fastfusion.mapper.logging import make_queue_and_listener
-from pytimeloop.fastfusion.mapper.per_einsum_mapper import mapper_one_einsum
+from pytimeloop.fastfusion.mapper.per_einsum_mapper import get_top_loop_jobs, mapper_place_fusion_level
+from pytimeloop.fastfusion.sim import Tiling, Loop, TensorStorage
 
 from pytimeloop.timeloopfe.v4 import Ert
 from pytimeloop.timeloopfe.common.backend_calls import call_accelergy_verbose
@@ -47,36 +51,130 @@ def mapper(
     ert = Ert(ert_dict["ERT"])
     energy_dict = ert.to_dict()
 
-    data = {}
-    per_einsum_args = [
-        dict(
-            einsum_id=einsum_id,
-            config=config,
-            pe_array_constraint=pe_array_constraint,
-            mac_array_constraint=mac_array_constraint,
-            explore_glb_uneven=explore_glb_uneven,
-            explore_pe_uneven=explore_pe_uneven,
-            spec=spec,
-            energy_dict=energy_dict,
-            log_queue=log_queue,
-            verbose_stream=verbose_stream,
-        )
-        for einsum_id in einsum_name_to_id.values()
-    ]
-
-    from joblib import Parallel, delayed
-
-    logger.debug("Starting workers")
+    grouped_similar_einsums = convert_rank_to_group_renaming(
+        detect_similar_einsums(workload, analyzer),
+        equivalent_groups
+    )
+    logger.info(f"Found {len(grouped_similar_einsums)} unique Einsums\n"
+                + f"\tConverter: {grouped_similar_einsums}")
+
+    args = get_top_loop_jobs(
+        einsums_to_explore=list(grouped_similar_einsums.keys()),
+        config=config,
+        pe_array_constraint=pe_array_constraint,
+        mac_array_constraint=mac_array_constraint,
+        explore_glb_uneven=explore_glb_uneven,
+        explore_pe_uneven=explore_pe_uneven,
+        spec=spec,
+        energy_dict=energy_dict,
+        log_queue=log_queue,
+        verbose_stream=verbose_stream,
+    )
+
+    print(f'Number of jobs: {len(args)}')
+    n_workers = 128
+    logger.debug(f"Starting {n_workers} workers")
     log_queue_listener.start()
-    data = Parallel(n_jobs=32)(
-        delayed(mapper_one_einsum)(**args) for args in per_einsum_args
+
+    result = Parallel(n_jobs=n_workers)(
+        delayed(mapper_place_fusion_level)(**a) for a in args
     )
+    data = defaultdict(dict)
+    for einsum_id, mappings in result:
+        for k, v in mappings.items():
+            if k in data[einsum_id]:
+                data[einsum_id][k] += v
+            else:
+                data[einsum_id][k] = v
+
     log_queue_listener.stop()
-
-    data = {
-        einsum_id: mapping
-        for einsum_id, mapping in zip(einsum_name_to_id.values(), data)
-    }
     logger.info(f"Mapper finished for {spec}")
+
+    generated_data = {}
+    logger.info(f"Generating data for non-unique Einsums")
+    for ref_einsum, others in grouped_similar_einsums.items():
+        for other_einsum, (rank_renaming, tensor_renaming) in others.items():
+            logger.info(f"Generating data for {other_einsum}. "
+                        + f"Rank renaming={rank_renaming}. "
+                        + f"Tensor renaming={tensor_renaming}")
+            generated_data[other_einsum] = generate_data(data[ref_einsum],
+                                                         rank_renaming,
+                                                         tensor_renaming)
+
+    for einsum, mapping in generated_data.items():
+        data[einsum] = mapping
+
+    logger.info(f"Final set of Einsums: {set(data.keys())}")
+
+    # data has to come out in sorted Einsum-id order
+    data = {k: v for k, v in sorted(data.items(), key=lambda item: item[0])}
+
     return data
 
+
+def generate_data(data, rank_renaming, tensor_renaming):
+    return {
+        _convert_tiling(tiling, rank_renaming, tensor_renaming)
+        :
+        _convert_stats(stats, rank_renaming, tensor_renaming)
+        for tiling, stats in data.items()
+    }
+
+
+def _convert_tiling(tiling: Tiling, rank_renaming, tensor_renaming):
+    return Tiling(
+        loops=tuple(Loop(rank_renaming[l.rank_id], l.bound, l.is_spatial)
+                    for l in tiling.loops),
+        tensors=frozenset(TensorStorage(tensor_renaming[ts.tensor_id],
+                                        ts.backer_id,
+                                        ts.above_loop_index,
+                                        ts.tile_size)
+                          for ts in tiling.tensors)
+    )
+
+
+def _convert_stats(stats, rank_renaming, tensor_renaming):
+    return deepcopy(stats)
+
+
+def detect_similar_einsums(workload, analyzer, return_all_as_unique=False):
+    if return_all_as_unique:
+        return {ref: {} for ref in workload.einsum_id_to_name()}
+
+    ref_to_other_einsums = {}
+    for einsum in workload.einsum_id_to_name():
+        found = False
+        for ref_einsum in ref_to_other_einsums:
+            rank_renaming, tensor_renaming = is_equivalent(ref_einsum,
+                                                           einsum,
+                                                           workload,
+                                                           analyzer)
+            if rank_renaming is not None:
+                ref_to_other_einsums[ref_einsum][einsum] = (rank_renaming,
+                                                            tensor_renaming)
+                found = True
+                break
+        if not found:
+            ref_to_other_einsums[einsum] = {}
+    return ref_to_other_einsums
+
+
+def convert_rank_to_group_renaming(ref_to_other_einsums, equiv_ranks):
+    return {
+        ref: {
+            other: (_convert_rank_renaming(rank_renaming, equiv_ranks),
+                    tensor_renaming)
+            for other, (rank_renaming, tensor_renaming) in others.items()
+        }
+        for ref, others in ref_to_other_einsums.items()
+    }
+
+
+def _convert_rank_renaming(rank_renaming, equiv_ranks):
+    # The Tiling class uses string ids
+    return {
+        str(equiv_ranks.rank_to_group_id[r1])
+        :
+        str(equiv_ranks.rank_to_group_id[r2])
+        for r1, r2 in rank_renaming.items()
+    }