Saner defaults

Author: ricardoV94

pytensor/link/numba/cache.py

@@ -1,11 +1,17 @@
 import weakref
+from collections.abc import Callable
+from functools import singledispatch, wraps
 from hashlib import sha256
 from pathlib import Path
+from pickle import dumps
+from tempfile import NamedTemporaryFile
+from typing import Any
 
 from numba.core.caching import CacheImpl, _CacheLocator
 
 from pytensor import config
 from pytensor.graph.basic import Apply
+from pytensor.link.numba.compile import numba_funcify, numba_njit
 
 
 NUMBA_PYTENSOR_CACHE_ENABLED = True
@@ -74,3 +80,158 @@ def cache_node_key(node: Apply, extra_key="") -> str:
             ),
         ).encode()
     ).hexdigest()
+
+
+@singledispatch
+def numba_funcify_default_op_cache_key(
+    op, node=None, **kwargs
+) -> Callable | tuple[Callable, Any]:
+    """Funcify an Op and implement a default cache key.
+
+    The default cache key is based on the op class and its properties.
+    It does not take into account the node inputs or other context.
+    Note that numba will use the array dtypes, rank and layout as part of the cache key,
+    but not the static shape or constant values.
+    If the funcify implementation exploits this information, then this method should not be used.
+    Instead dispatch directly on `numba_funcify_and_cache_key` (or just numba_funcify)
+    which won't use any cache key.
+    """
+    # Default cache key of None which means "don't try to do directly cache this function"
+    raise NotImplementedError()
+
+
+def register_funcify_default_op_cache_key(op_type):
+    """Register a funcify implementation for both cache and non-cache versions."""
+
+    def decorator(dispatch_func):
+        # Register with the cache key dispatcher
+        numba_funcify_default_op_cache_key.register(op_type)(dispatch_func)
+
+        # Create a wrapper for the non-cache dispatcher
+        @wraps(dispatch_func)
+        def dispatch_func_wrapper(*args, **kwargs):
+            func, key = dispatch_func(*args, **kwargs)
+            # Discard the key for the non-cache version
+            return func
+
+        # Register the wrapper with the non-cache dispatcher
+        numba_funcify.register(op_type)(dispatch_func_wrapper)
+
+        return dispatch_func
+
+    return decorator
+
+
+@singledispatch
+def numba_funcify_and_cache_key(op, node=None, **kwargs) -> tuple[Callable, str | None]:
+    # Default cache key of None which means "don't try to do directly cache this function"
+    if hasattr(op, "_props"):
+        try:
+            func_and_salt = numba_funcify_default_op_cache_key(op, node=node, **kwargs)
+        except NotImplementedError:
+            pass
+        else:
+            if isinstance(func_and_salt, tuple):
+                func, salt = func_and_salt
+            else:
+                func, salt = func_and_salt, "0"
+            props_dict = op._props_dict()
+            if not props_dict:
+                # Simple op, just use the type string as key
+                key_bytes = str((type(op), salt)).encode()
+            else:
+                # Simple props, can use string representation of props as key
+                simple_types = (str, bool, int, type(None), float)
+                container_types = (tuple, frozenset)
+                if all(
+                    isinstance(v, simple_types)
+                    or (
+                        isinstance(v, container_types)
+                        and all(isinstance(i, simple_types) for i in v)
+                    )
+                    for v in props_dict.values()
+                ):
+                    key_bytes = str(
+                        (type(op), tuple(props_dict.items()), salt)
+                    ).encode()
+                else:
+                    # Complex props, use pickle to serialize them
+                    key_bytes = dumps((str(type(op)), tuple(props_dict.items()), salt))
+            return func, sha256(key_bytes).hexdigest()
+
+    # Fallback
+    return numba_funcify(op, node=node, **kwargs), None
+
+
+def register_funcify_and_cache_key(op_type):
+    """Register a funcify implementation for both cache and non-cache versions."""
+
+    def decorator(dispatch_func):
+        # Register with the cache key dispatcher
+        numba_funcify_and_cache_key.register(op_type)(dispatch_func)
+
+        # Create a wrapper for the non-cache dispatcher
+        @wraps(dispatch_func)
+        def dispatch_func_wrapper(*args, **kwargs):
+            func, key = dispatch_func(*args, **kwargs)
+            # Discard the key for the non-cache version
+            return func
+
+        # Register the wrapper with the non-cache dispatcher
+        numba_funcify.register(op_type)(dispatch_func_wrapper)
+
+        return dispatch_func_wrapper
+
+    return decorator
+
+
+def numba_njit_and_cache(op, node, **kwargs):
+    jitable_func, key = numba_funcify_and_cache_key(op, node=node, **kwargs)
+
+    if key is not None:
+        # To force numba to use our cache, we must compile the function so that any closure
+        # becomes a global variable...
+        op_name = op.__class__.__name__
+        cached_func = compile_and_cache_numba_function_src(
+            src=f"def {op_name}(*args): return jitable_func(*args)",
+            function_name=op_name,
+            global_env=globals() | dict(jitable_func=jitable_func),
+            cache_key=key,
+        )
+        return numba_njit(cached_func, final_function=True, cache=True)
+    else:
+        return numba_njit(
+            lambda *args: jitable_func(*args), final_function=True, cache=False
+        )
+
+
+def compile_and_cache_numba_function_src(
+    src: str,
+    function_name: str,
+    global_env: dict[Any, Any] | None = None,
+    local_env: dict[Any, Any] | None = None,
+    store_to_disk: bool = False,
+    cache_key: str | None = None,
+) -> Callable:
+    if store_to_disk:
+        with NamedTemporaryFile(delete=False) as f:
+            filename = f.name
+            f.write(src.encode())
+    else:
+        filename = "<string>"
+
+    if global_env is None:
+        global_env = {}
+
+    if local_env is None:
+        local_env = {}
+
+    mod_code = compile(src, filename, mode="exec")
+    exec(mod_code, global_env, local_env)
+
+    res = local_env[function_name]
+    res.__source__ = src  # type: ignore
+
+    if cache_key is not None:
+        CACHED_SRC_FUNCTIONS[res] = cache_key
+    return res

pytensor/link/numba/compile.py

@@ -1,7 +1,6 @@
 import warnings
 from collections.abc import Callable
-from tempfile import NamedTemporaryFile
-from typing import Any
+from functools import singledispatch
 
 import numba
 import numpy as np
@@ -11,8 +10,6 @@
 
 from pytensor import config
 from pytensor.graph import Apply, FunctionGraph, Type
-from pytensor.link.numba.cache import CACHED_SRC_FUNCTIONS
-from pytensor.link.numba.dispatch.sparse import CSCMatrixType, CSRMatrixType
 from pytensor.scalar import ScalarType
 from pytensor.sparse import SparseTensorType
 from pytensor.tensor import TensorType
@@ -59,36 +56,17 @@ def numba_njit(*args, fastmath=None, final_function: bool = False, **kwargs):
         return func(*args, fastmath=fastmath, **kwargs)
 
 
-def compile_and_cache_numba_function_src(
-    src: str,
-    function_name: str,
-    global_env: dict[Any, Any] | None = None,
-    local_env: dict[Any, Any] | None = None,
-    store_to_disk: bool = False,
-    cache_key: str | None = None,
-) -> Callable:
-    if store_to_disk:
-        with NamedTemporaryFile(delete=False) as f:
-            filename = f.name
-            f.write(src.encode())
-    else:
-        filename = "<string>"
-
-    if global_env is None:
-        global_env = {}
-
-    if local_env is None:
-        local_env = {}
-
-    mod_code = compile(src, filename, mode="exec")
-    exec(mod_code, global_env, local_env)
+@singledispatch
+def numba_funcify(
+    typ, node=None, storage_map=None, **kwargs
+) -> Callable | tuple[Callable, str | int | None]:
+    """Generate a numba function for a given op and apply node (or Fgraph).
 
-    res = local_env[function_name]
-    res.__source__ = src  # type: ignore
-
-    if cache_key is not None:
-        CACHED_SRC_FUNCTIONS[res] = cache_key
-    return res
+    The resulting function will usually use the `no_cpython_wrapper`
+    argument in numba, so it can not be called directly from python,
+    but only from other jit functions.
+    """
+    raise NotImplementedError(f"Numba funcify not implemented for type {typ}")
 
 
 def get_numba_type(
@@ -124,6 +102,8 @@ def get_numba_type(
         numba_dtype = numba.from_dtype(dtype)
         return numba_dtype
     elif isinstance(pytensor_type, SparseTensorType):
+        from pytensor.link.numba.dispatch.sparse import CSCMatrixType, CSRMatrixType
+
         dtype = pytensor_type.numpy_dtype
         numba_dtype = numba.from_dtype(dtype)
         if pytensor_type.format == "csr":