[RFC] Implement basic on disk caching

stack-info: PR: #336, branch: oulgen/stack/26

Author: oulgen

helion/_testing.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import collections
+import contextlib
 import importlib
 import inspect
 import operator
@@ -15,6 +16,7 @@
 import torch
 from triton.testing import do_bench
 
+from ._utils import counters
 from .runtime.config import Config
 from helion._compat import get_tensor_descriptor_fn_name
 
@@ -291,6 +293,20 @@ def tearDownClass(cls) -> None:
         super().tearDownClass()
         del cls._expected_journal
 
+    def setUp(self) -> None:
+        super().setUp()
+        self._test_stack = contextlib.ExitStack()
+
+        from torch._inductor.utils import fresh_cache
+
+        self._test_stack.enter_context(fresh_cache())
+
+        counters.clear()
+
+    def tearDown(self) -> None:
+        super().tearDown()
+        self._test_stack.close()
+
     def assertExpectedJournal(self, value: str) -> None:
         """
         Assert that the given value matches the expected output stored in <testfile>.expected.

helion/_utils.py

@@ -0,0 +1,7 @@
+from __future__ import annotations
+
+import collections
+
+counters: collections.defaultdict[str, collections.Counter[str]] = (
+    collections.defaultdict(collections.Counter)
+)

helion/autotuner/__init__.py

@@ -9,4 +9,6 @@
     DifferentialEvolutionSearch as DifferentialEvolutionSearch,
 )
 from .finite_search import FiniteSearch as FiniteSearch
+from .local_cache import LocalAutotuneCache as LocalAutotuneCache
+from .local_cache import StrictLocalAutotuneCache as StrictLocalAutotuneCache
 from .random_search import RandomSearch as RandomSearch

helion/autotuner/base_cache.py

@@ -0,0 +1,149 @@
+from __future__ import annotations
+
+import abc
+import dataclasses
+import functools
+import hashlib
+import logging
+import os
+from typing import TYPE_CHECKING
+from typing import Hashable
+from typing import Sequence
+
+from .._utils import counters
+
+if TYPE_CHECKING:
+    from ..runtime.config import Config
+    from ..runtime.kernel import BoundKernel
+    from .base_search import BaseSearch
+
+log: logging.Logger = logging.getLogger(__name__)
+
+
+@functools.cache
+def helion_key() -> str:
+    from torch._inductor.codecache import build_code_hash
+
+    here = os.path.abspath(__file__)
+    helion_path = os.path.dirname(os.path.dirname(here))
+
+    combined_hash = hashlib.sha256()
+    build_code_hash([helion_path], "", combined_hash)
+    return combined_hash.hexdigest()
+
+
+@functools.cache
+def torch_key_wrapper() -> str:
+    from torch._inductor.codecache import torch_key
+
+    return torch_key().hex()
+
+
+@functools.cache
+def triton_key_wrapper() -> str:
+    from torch._inductor.runtime.triton_compat import triton_key
+
+    return triton_key()
+
+
+class CacheKeyBase:
+    """
+    Base class to provide utility functions to all cache key dataclasses
+    """
+
+    def stable_hash(self) -> str:
+        return hashlib.sha256(repr(self).encode("utf-8")).hexdigest()
+
+
+@dataclasses.dataclass(frozen=True)
+class BoundKernelInMemoryCacheKey(CacheKeyBase):
+    """
+    Default in memory cache key.
+
+    This key includes:
+
+    specialization_key: Information about all kernel inputs.
+                        For tensors this means their device, shape, size etc.
+    extra_results: Information regarding `hl.specialize` decisions
+    """
+
+    specialization_key: tuple[Hashable, ...]
+    extra_results: tuple[Hashable, ...]
+
+
+@dataclasses.dataclass(frozen=True)
+class LooseAutotuneCacheKey(BoundKernelInMemoryCacheKey):
+    """
+    Autotune Cache key to use for most use cases.
+
+    This key includes (in addition to BoundKernelInMemoryCacheKey):
+
+    kernel_source_hash: Hash of source code of input Helion kernel
+    hardware: Hardware of the input device
+    runtime_name: Version of the cuda/rocm arch
+    """
+
+    kernel_source_hash: str
+    hardware: str
+    runtime_name: str
+
+    def stable_hash(self) -> str:
+        return hashlib.sha256(repr(self).encode("utf-8")).hexdigest()
+
+
+@dataclasses.dataclass(frozen=True)
+class StrictAutotuneCacheKey(LooseAutotuneCacheKey):
+    """
+    Autotune Cache key to use for utmost strictness in terms of re-autotuning
+    when library source code changes.
+
+    This key includes (in addition to StrictAutotuneCacheKey):
+
+    helion_key: Hash of source code of Helion
+    torch_key: Hash of source code of PyTorch
+    triton_key: Hash of source code of Triton
+    """
+
+    helion_key: str = dataclasses.field(default_factory=helion_key)
+    torch_key: str = dataclasses.field(default_factory=torch_key_wrapper)
+    triton_key: str = dataclasses.field(default_factory=triton_key_wrapper)
+
+
+class AutotuneCacheBase(abc.ABC):
+    """
+    Abstract base class that all autotune caches need to implement.
+    Any user defined cache will need to extend this class, and
+    provide implementations for get and put methods.
+    """
+
+    def __init__(
+        self, kernel: BoundKernel, args: Sequence[object], autotuner: BaseSearch
+    ) -> None:
+        self.autotuner = autotuner
+        self.kernel = kernel
+        self.args = args
+
+    @abc.abstractmethod
+    def get(self) -> Config | None:
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def put(self, config: Config) -> None:
+        raise NotImplementedError
+
+    def autotune(self) -> Config:
+        if (config := self.get()) is not None:
+            counters["autotune"]["cache_hit"] += 1
+            log.debug("cache hit: %s", str(config))
+            return config
+
+        counters["autotune"]["cache_miss"] += 1
+        log.debug("cache miss")
+
+        config = self.autotuner.autotune()
+
+        self.put(config)
+        counters["autotune"]["cache_put"] += 1
+        log.debug("cache put: %s", str(config))
+
+        return config

helion/autotuner/local_cache.py

@@ -0,0 +1,107 @@
+from __future__ import annotations
+
+import hashlib
+import inspect
+import logging
+import os
+from pathlib import Path
+import textwrap
+from typing import TYPE_CHECKING
+from typing import Sequence
+
+import torch
+
+from ..runtime.config import Config
+from .base_cache import AutotuneCacheBase
+from .base_cache import LooseAutotuneCacheKey
+from .base_cache import StrictAutotuneCacheKey
+
+if TYPE_CHECKING:
+    from ..runtime.kernel import BoundKernel
+    from .base_search import BaseSearch
+
+log: logging.Logger = logging.getLogger(__name__)
+
+
+class LocalAutotuneCache(AutotuneCacheBase):
+    """
+    This class implements the local autotune cache, storing the
+    best config artifact on the local file system either by default
+    on torch's cache directory, or at a user specified HELION_CACHE_DIR
+    directory.
+    It uses the LooseAutotuneCacheKey implementation for the cache key
+    which takes into account device and source code properties, but does
+    not account for library level code changes such as Triton, Helion or
+    PyTorch. Use StrictLocalAutotuneCache to consider these properties.
+    """
+
+    def __init__(
+        self, kernel: BoundKernel, args: Sequence[object], autotuner: BaseSearch
+    ) -> None:
+        super().__init__(kernel, args, autotuner)
+        self.key = self._generate_key()
+
+    def _generate_key(self) -> LooseAutotuneCacheKey:
+        in_memory_cache_key = self.kernel.kernel._create_bound_kernel_cache_key(
+            self.kernel,
+            tuple(self.args),
+            self.kernel.kernel.specialization_key(self.args),
+        )
+        kernel_source = textwrap.dedent(inspect.getsource(self.kernel.kernel.fn))
+        kernel_source_hash = hashlib.sha256(kernel_source.encode("utf-8")).hexdigest()
+
+        hardware = None
+        runtime_name = None
+
+        for arg in self.args:
+            if isinstance(arg, torch.Tensor):
+                device_properties = torch.cuda.get_device_properties(arg.device)
+                if torch.version.cuda is not None:  # pyright: ignore[reportAttributeAccessIssue]
+                    hardware = device_properties.name
+                    runtime_name = torch.version.cuda  # pyright: ignore[reportAttributeAccessIssue]
+                else:
+                    hardware = device_properties.gcnArchName
+                    runtime_name = torch.version.hip  # pyright: ignore[reportAttributeAccessIssue]
+
+        assert hardware is not None and runtime_name is not None
+        return LooseAutotuneCacheKey(
+            specialization_key=in_memory_cache_key.specialization_key,
+            extra_results=in_memory_cache_key.extra_results,
+            kernel_source_hash=kernel_source_hash,
+            hardware=hardware,
+            runtime_name=runtime_name,
+        )
+
+    def _get_local_cache_path(self) -> Path:
+        if (user_path := os.environ.get("HELION_CACHE_DIR", None)) is not None:
+            cache_path = Path(user_path)
+        else:
+            from torch._inductor.runtime.cache_dir_utils import (
+                cache_dir,  # pyright: ignore[reportPrivateImportUsage]
+            )
+
+            cache_path = Path(cache_dir()) / "helion"
+
+        return cache_path / f"{self.key.stable_hash()}.best_config"
+
+    def get(self) -> Config | None:
+        path = self._get_local_cache_path()
+        try:
+            return Config.load(path)
+        except Exception:
+            return None
+
+    def put(self, config: Config) -> None:
+        path = self._get_local_cache_path()
+        config.save(path)
+
+
+class StrictLocalAutotuneCache(LocalAutotuneCache):
+    """
+    Stricter implementation of the local autotune cache, which takes into
+    account library level code changes such as Triton, Helion or PyTorch.
+    """
+
+    def _generate_key(self) -> StrictAutotuneCacheKey:
+        loose_key = super()._generate_key()
+        return StrictAutotuneCacheKey(**vars(loose_key))

helion/runtime/config.py

@@ -118,6 +118,7 @@ def from_json(cls, json_str: str) -> Config:
 
     def save(self, path: str | Path) -> None:
         """Save the config to a JSON file."""
+        Path(path).parent.mkdir(parents=True, exist_ok=True)
         Path(path).write_text(self.to_json())
 
     @classmethod

helion/runtime/kernel.py

@@ -45,6 +45,7 @@
     from torch._guards import Source
 
     from ..autotuner import ConfigSpec
+    from ..autotuner.base_cache import BoundKernelInMemoryCacheKey
 
     ConfigLike = Config | dict[str, object]
 
@@ -53,12 +54,6 @@
 CompiledConfig = Callable[..., _R]
 
 
-@dataclasses.dataclass(frozen=True)
-class BoundKernelInMemoryCacheKey:
-    specialization_key: tuple[Hashable, ...]
-    extra_results: tuple[Hashable, ...]
-
-
 class Kernel(Generic[_R]):
     def __init__(
         self,
@@ -114,6 +109,8 @@ def __init__(
     def _get_bound_kernel_cache_key(
         self, args: tuple[object, ...], signature: tuple[Hashable, ...]
     ) -> BoundKernelInMemoryCacheKey | None:
+        from ..autotuner.base_cache import BoundKernelInMemoryCacheKey
+
         extra_fns = self._specialize_extra.get(signature)
         if extra_fns is not None:
             extra_results: tuple[Hashable, ...] = tuple([s(args) for s in extra_fns])
@@ -126,6 +123,8 @@ def _create_bound_kernel_cache_key(
         args: tuple[object, ...],
         signature: tuple[Hashable, ...],
     ) -> BoundKernelInMemoryCacheKey:
+        from ..autotuner.base_cache import BoundKernelInMemoryCacheKey
+
         self._specialize_extra[signature] = extra_fns = bound_kernel._specialize_extra()
         extra_results: tuple[Hashable, ...] = tuple([s(args) for s in extra_fns])
         return BoundKernelInMemoryCacheKey(signature, extra_results)
@@ -458,12 +457,18 @@ def autotune(
             self.settings.check_autotuning_disabled()
 
             from ..autotuner import DifferentialEvolutionSearch
+            from ..autotuner import LocalAutotuneCache
 
-            config = DifferentialEvolutionSearch(
+            config = LocalAutotuneCache(
                 self,
                 args,
-                **kwargs,  # pyright: ignore[reportArgumentType]
+                DifferentialEvolutionSearch(
+                    self,
+                    args,
+                    **kwargs,  # pyright: ignore[reportArgumentType]
+                ),
             ).autotune()
+
         self.set_config(config)
         return config