Add initial cgra compiler

compiler_gym/datasets/datasets.py

@@ -152,7 +152,6 @@ def dataset_from_parsed_uri(self, uri: BenchmarkUri) -> Dataset:
         key = self._dataset_key_from_uri(uri)
 
         if key not in self._datasets:
-            print("datasets are ", str(self._datasets))
             raise LookupError(f"Dataset not found: {key}")
 
         return self._datasets[key]

compiler_gym/envs/cgra/BUILD

@@ -2,14 +2,15 @@
 #
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
-load("@rules_python//python:defs.bzl", "py_library", "py_test")
+load("@rules_python//python:defs.bzl", "py_library")
 
 py_library(
     name = "cgra",
     srcs = [
         "__init__.py",
         "cgra_rewards.py",
         "Operations.py",
+        "compile_settings.py",
         "DFG.py"
     ],
     data = [
@@ -18,6 +19,7 @@ py_library(
     visibility = ["//visibility:public"],
     deps = [
         "//compiler_gym/envs/cgra/datasets",
+        "//compiler_gym/envs/cgra/architectures",
         "//compiler_gym/errors",
         "//compiler_gym/service:client_service_compiler_env",
         "//compiler_gym/service/runtime",  # Implicit dependency of service.

compiler_gym/envs/cgra/CMakeLists.txt

@@ -7,17 +7,19 @@ cg_add_all_subdirs()
 
 cg_py_library(
   NAME
-    gcc
+    cgra
   SRCS
     "__init__.py"
-    "gcc.py"
-    "gcc_env.py"
-    "gcc_rewards.py"
+    "cgra_rewards.py"
+    "Operations.py"
+    "compile_settings.py"
+    "DFG.py"
   DATA
-    compiler_gym::envs::gcc::service::service
+    compiler_gym::envs::cgra::service::service
   DEPS
     compiler_gym::service::client_service_compiler_env
-    compiler_gym::envs::gcc::datasets::datasets
+    compiler_gym::envs::cgra::datasets::datasets
+    compiler_gym::engs::cgra::architectures::architectures
     compiler_gym::errors::errors
     compiler_gym::service::runtime::runtime
     compiler_gym::util::util

compiler_gym/envs/cgra/DFG.py

@@ -1,3 +1,8 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
 import json
 from pathlib import Path
 import random
@@ -24,41 +29,26 @@ def __str__(self):
         return "Node with name " + self.name + " and op " + str(self.operation)
 
 class DFG(object):
-    def __init__(self, working_directory: Optional[Path] = None, benchmark: Optional[Benchmark] = None, from_json: Optional[Path] = None, from_text: Optional[str] = None):
+    def __init__(self, working_directory: Optional[Path] = None, from_json: Optional[Path] = None, from_text: Optional[str] = None):
         # Copied from here: https://github.com/facebookresearch/CompilerGym/blob/development/examples/loop_optimizations_service/service_py/loops_opt_service.py
         # self.inst2vec = _INST2VEC_ENCODER
 
         if from_json is not None:
             self.load_dfg_from_json(from_json)
         elif from_text is not None:
             self.load_dfg_from_text(from_text)
-        elif benchmark is not None:
-            # Only re-create the JSON file if we aren't providing an existing one.
-            # The existing ones are mostly a debugging functionality.
-            with open(self.working_directory / "benchmark.c", "wb") as f:
-                f.write(benchmark.program.contents)
-
-            # We use CGRA-Mapper to produce a DFG in JSON.
-            run_command(
-                ["cgra-mapper", self.src_path, self.dfg_path]
-            )
-
-            # Now, load in the DFG.
-            self.load_dfg_from_json(self.dfg_path)
 
     def __str__(self):
         res = "nodes are: " + str(self.nodes) + " and edges are " + str(self.adj)
         return res
 
     def load_dfg_from_json(self, path):
-        import json
         with open(path, 'r') as p:
             # This isnt' text, but I think the json.loads
             # that this calls just works?
             self.load_dfg_from_text(p)
 
     def load_dfg_from_text(self, text):
-        import json
         f = json.loads(text)
         self.nodes = {}
         self.node_names = []
@@ -97,7 +87,13 @@ def get_succs(self, node):
             succs.append(self.nodes[n])
         return succs
 
-    # TODO -- fix this, because for a graph with multiple entry nodes,
+    def build_preds_lookup(self):
+        preds_lookup = {}
+        for n in self.node_names:
+            preds_lookup[n] = self.get_preds(self.nodes[n])
+        return preds_lookup
+
+    # TODO(jcw) -- fix this, because for a graph with multiple entry nodes,
     # this doesn't actually give the right answer :)
     # (should do in most cases)
     def bfs(self):
@@ -106,6 +102,10 @@ def bfs(self):
         print(self.entry_points)
         seen = set()
 
+        # build a lookup based on the predecessors
+        # for each node.
+        preds_lookup = self.build_preds_lookup()
+
         while len(to_explore) > 0:
             head = to_explore[0]
             to_explore = to_explore[1:]
@@ -114,9 +114,30 @@ def bfs(self):
             seen.add(head)
             yield self.nodes[head]
 
-            # Get the following nodes.
-            following_nodes = self.adj[head]
-            to_explore += following_nodes
+            # Add the next batch of nodes that we have
+            # visited all the preds for if there are more
+            # nodes to explore.
+            if len(to_explore) == 0 and len(seen) != len(self.node_names):
+                for node_name in self.node_names:
+                    if node_name in seen:
+                        continue
+                    else:
+                        # Unseen --- have we seen all th preds?
+                        failed = False
+                        for p in preds_lookup[node_name]:
+                            if p.name not in seen:
+                                failed = True
+                        if not failed:
+                            to_explore.append(node_name)
+                if len(to_explore) == 0: # We added nothing despite trying
+                    # to.
+
+                    # TODO(jcw) -- Fix this, as support for cyclical DFGs
+                    # is important to be able to support loops with 
+                    # cross-loop dependencies.
+                    print("Cyclical DFG --- Impossible to do a true BFS")
+                    print("DFG is ", str(self))
+                    assert False
 
 # Generate a test DFG using the operations in
 # 'operations'.
@@ -175,7 +196,7 @@ def generate_DFG(operations: List[Operation], size, seed=0):
             else:
                 inputs.append(random.choice(nodes_list))
         # If the node has no arguments, then we should add it
-        # as an entry point.  --- todo --- should we just skip
+        # as an entry point.  --- todo(jcw) --- should we just skip
         # this avoid creating graphs with too many constant loads?
         if operation.inputs == 0:
             entry_points.append(name)

compiler_gym/envs/cgra/Operations.py

@@ -1,3 +1,7 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
 
 class Operation(object):
     def __init__(self, name, inputs, outputs, latency):
@@ -10,7 +14,7 @@ def __str__(self):
         return self.name
 
 Operations = [
-    # TODO --- should we support more operations as heterogeneous?
+    # TODO(jcw) --- should we support more operations as heterogeneous?
     # IMO most of the other things that are scheduled are
     # pretty vacuous, although we could explore supporting those.
     # Operation is: name, inputs, outputs, cycles.
@@ -44,7 +48,7 @@ def operation_index_of(op):
     return -1
 
 def operation_latency(op):
-    # TODO --- model latency --- or at least expost this
+    # TODO(jcw) --- model latency --- or at least expost this
     # to a configuration.
     return op.latency
 

compiler_gym/envs/cgra/architectures/BUILD

@@ -0,0 +1,15 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+load("@rules_python//python:defs.bzl", "py_library")
+
+py_library(
+    name = "architectures",
+    srcs = [
+        "__init__.py",
+        "CGRA.py"
+    ],
+    visibility = ["//visibility:public"]
+)