feat, perf: Refactor the PoC to support multiple dtypes (#757)

Co-authored-by: Mateusz Sokół <[email protected]>

Author: hameerabbasi

sparse/mlir_backend/_common.py

@@ -0,0 +1,14 @@
+import abc
+import functools
+
+from mlir import ir
+
+
+class MlirType(abc.ABC):
+    @classmethod
+    @abc.abstractmethod
+    def get_mlir_type(cls) -> ir.Type: ...
+
+
+def fn_cache(f, maxsize: int | None = None):
+    return functools.wraps(f)(functools.lru_cache(maxsize=maxsize)(f))

sparse/mlir_backend/_constructors.py

@@ -1,5 +1,7 @@
 import ctypes
 import ctypes.util
+import functools
+import weakref
 
 import mlir.execution_engine
 import mlir.passmanager
@@ -9,9 +11,26 @@
 import numpy as np
 import scipy.sparse as sps
 
-from ._core import DEBUG, MLIR_C_RUNNER_UTILS, SCRIPT_PATH, ctx
-from ._dtypes import DType, Float64, Index
-from ._memref import MemrefF64_1D, MemrefIdx_1D
+from ._common import fn_cache
+from ._core import CWD, DEBUG, MLIR_C_RUNNER_UTILS, ctx
+from ._dtypes import DType, Index, asdtype
+from ._memref import make_memref_ctype, ranked_memref_from_np
+
+
+def _hold_self_ref_in_ret(fn):
+    @functools.wraps(fn)
+    def wrapped(self, *a, **kw):
+        ptr = ctypes.py_object(self)
+        ctypes.pythonapi.Py_IncRef(ptr)
+        ret = fn(self, *a, **kw)
+
+        def finalizer(ptr):
+            ctypes.pythonapi.Py_DecRef(ptr)
+
+        weakref.finalize(ret, finalizer, ptr)
+        return ret
+
+    return wrapped
 
 
 class Tensor:
@@ -26,21 +45,21 @@ def __init__(self, obj, module, tensor_type, disassemble_fn, values_dtype, index
     def __del__(self):
         self.module.invoke("free_tensor", ctypes.pointer(self.obj))
 
+    @_hold_self_ref_in_ret
     def to_scipy_sparse(self):
         """
         Returns scipy.sparse or ndarray
         """
-        return self.disassemble_fn(self.module, self.obj)
+        return self.disassemble_fn(self.module, self.obj, self.values_dtype)
 
 
 class DenseFormat:
-    modules = {}
-
+    @fn_cache
     def get_module(shape: tuple[int], values_dtype: DType, index_dtype: DType):
         with ir.Location.unknown(ctx):
             module = ir.Module.create()
-            values_dtype = values_dtype.get()
-            index_dtype = index_dtype.get()
+            values_dtype = values_dtype.get_mlir_type()
+            index_dtype = index_dtype.get_mlir_type()
             index_width = getattr(index_dtype, "width", 0)
             levels = (sparse_tensor.LevelType.dense, sparse_tensor.LevelType.dense)
             ordering = ir.AffineMap.get_permutation([0, 1])
@@ -78,18 +97,19 @@ def free_tensor(tensor_shaped):
             disassemble.func_op.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
             free_tensor.func_op.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
             if DEBUG:
-                (SCRIPT_PATH / "dense_module.mlir").write_text(str(module))
+                (CWD / "dense_module.mlir").write_text(str(module))
             pm = mlir.passmanager.PassManager.parse("builtin.module(sparsifier{create-sparse-deallocs=1})")
             pm.run(module.operation)
             if DEBUG:
-                (SCRIPT_PATH / "dense_module_opt.mlir").write_text(str(module))
+                (CWD / "dense_module_opt.mlir").write_text(str(module))
 
         module = mlir.execution_engine.ExecutionEngine(module, opt_level=2, shared_libs=[MLIR_C_RUNNER_UTILS])
         return (module, dense_shaped)
 
     @classmethod
     def assemble(cls, module, arr: np.ndarray) -> ctypes.c_void_p:
-        data = MemrefF64_1D.from_numpy(arr.flatten())
+        assert arr.ndim == 2
+        data = ranked_memref_from_np(arr.flatten())
         out = ctypes.c_void_p()
         module.invoke(
             "assemble",
@@ -99,18 +119,18 @@ def assemble(cls, module, arr: np.ndarray) -> ctypes.c_void_p:
         return out
 
     @classmethod
-    def disassemble(cls, module: ir.Module, ptr: ctypes.c_void_p) -> np.ndarray:
+    def disassemble(cls, module: ir.Module, ptr: ctypes.c_void_p, dtype: type[DType]) -> np.ndarray:
         class Dense(ctypes.Structure):
             _fields_ = [
-                ("data", MemrefF64_1D),
+                ("data", make_memref_ctype(dtype, 1)),
                 ("data_len", np.ctypeslib.c_intp),
                 ("shape_x", np.ctypeslib.c_intp),
                 ("shape_y", np.ctypeslib.c_intp),
             ]
 
             def to_np(self) -> np.ndarray:
                 data = self.data.to_numpy()[: self.data_len]
-                return data.copy().reshape((self.shape_x, self.shape_y))
+                return data.reshape((self.shape_x, self.shape_y))
 
         arr = Dense()
         module.invoke(
@@ -122,18 +142,17 @@ def to_np(self) -> np.ndarray:
 
 
 class COOFormat:
-    modules = {}
     # TODO: implement
+    ...
 
 
 class CSRFormat:
-    modules = {}
-
-    def get_module(shape: tuple[int], values_dtype: DType, index_dtype: DType):
+    @fn_cache
+    def get_module(shape: tuple[int], values_dtype: type[DType], index_dtype: type[DType]):
         with ir.Location.unknown(ctx):
             module = ir.Module.create()
-            values_dtype = values_dtype.get()
-            index_dtype = index_dtype.get()
+            values_dtype = values_dtype.get_mlir_type()
+            index_dtype = index_dtype.get_mlir_type()
             index_width = getattr(index_dtype, "width", 0)
             levels = (sparse_tensor.LevelType.dense, sparse_tensor.LevelType.compressed)
             ordering = ir.AffineMap.get_permutation([0, 1])
@@ -175,11 +194,11 @@ def free_tensor(tensor_shaped):
             disassemble.func_op.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
             free_tensor.func_op.attributes["llvm.emit_c_interface"] = ir.UnitAttr.get()
             if DEBUG:
-                (SCRIPT_PATH / "scr_module.mlir").write_text(str(module))
+                (CWD / "csr_module.mlir").write_text(str(module))
             pm = mlir.passmanager.PassManager.parse("builtin.module(sparsifier{create-sparse-deallocs=1})")
             pm.run(module.operation)
             if DEBUG:
-                (SCRIPT_PATH / "csr_module_opt.mlir").write_text(str(module))
+                (CWD / "csr_module_opt.mlir").write_text(str(module))
 
         module = mlir.execution_engine.ExecutionEngine(module, opt_level=2, shared_libs=[MLIR_C_RUNNER_UTILS])
         return (module, csr_shaped)
@@ -189,20 +208,20 @@ def assemble(cls, module: ir.Module, arr: sps.csr_array) -> ctypes.c_void_p:
         out = ctypes.c_void_p()
         module.invoke(
             "assemble",
-            ctypes.pointer(ctypes.pointer(MemrefIdx_1D.from_numpy(arr.indptr))),
-            ctypes.pointer(ctypes.pointer(MemrefIdx_1D.from_numpy(arr.indices))),
-            ctypes.pointer(ctypes.pointer(MemrefF64_1D.from_numpy(arr.data))),
+            ctypes.pointer(ctypes.pointer(ranked_memref_from_np(arr.indptr))),
+            ctypes.pointer(ctypes.pointer(ranked_memref_from_np(arr.indices))),
+            ctypes.pointer(ctypes.pointer(ranked_memref_from_np(arr.data))),
             ctypes.pointer(out),
         )
         return out
 
     @classmethod
-    def disassemble(cls, module: ir.Module, ptr: ctypes.c_void_p) -> sps.csr_array:
+    def disassemble(cls, module: ir.Module, ptr: ctypes.c_void_p, dtype: type[DType]) -> sps.csr_array:
         class Csr(ctypes.Structure):
             _fields_ = [
-                ("data", MemrefF64_1D),
-                ("pos", MemrefIdx_1D),
-                ("crd", MemrefIdx_1D),
+                ("data", make_memref_ctype(dtype, 1)),
+                ("pos", make_memref_ctype(Index, 1)),
+                ("crd", make_memref_ctype(Index, 1)),
                 ("data_len", np.ctypeslib.c_intp),
                 ("pos_len", np.ctypeslib.c_intp),
                 ("crd_len", np.ctypeslib.c_intp),
@@ -214,7 +233,7 @@ def to_sps(self) -> sps.csr_array:
                 pos = self.pos.to_numpy()[: self.pos_len]
                 crd = self.crd.to_numpy()[: self.crd_len]
                 data = self.data.to_numpy()[: self.data_len]
-                return sps.csr_array((data.copy(), crd.copy(), pos.copy()), shape=(self.shape_x, self.shape_y))
+                return sps.csr_array((data, crd, pos), shape=(self.shape_x, self.shape_y))
 
         arr = Csr()
         module.invoke(
@@ -235,23 +254,21 @@ def _is_numpy_obj(x) -> bool:
 
 def asarray(obj) -> Tensor:
     # TODO: discover obj's dtype
-    values_dtype = Float64
-    index_dtype = Index
+    values_dtype = asdtype(obj.dtype)
 
     # TODO: support other scipy formats
     if _is_scipy_sparse_obj(obj):
         format_class = CSRFormat
+        # This can be int32 or int64
+        index_dtype = asdtype(obj.indptr.dtype)
     elif _is_numpy_obj(obj):
         format_class = DenseFormat
+        index_dtype = Index
     else:
         raise Exception(f"{type(obj)} not supported.")
 
     # TODO: support proper caching
-    if hash(obj.shape) in format_class.modules:
-        module, tensor_type = format_class.modules[hash(obj.shape)]
-    else:
-        module, tensor_type = format_class.get_module(obj.shape, values_dtype, index_dtype)
-        format_class.modules[hash(obj.shape)] = module, tensor_type
+    module, tensor_type = format_class.get_module(obj.shape, values_dtype, index_dtype)
 
     assembled_obj = format_class.assemble(module, obj)
     return Tensor(assembled_obj, module, tensor_type, format_class.disassemble, values_dtype, index_dtype)

sparse/mlir_backend/_core.py

@@ -5,7 +5,7 @@
 from mlir.ir import Context
 
 DEBUG = bool(int(os.environ.get("DEBUG", "0")))
-SCRIPT_PATH = pathlib.Path(__file__).parent
+CWD = pathlib.Path(".")
 
 MLIR_C_RUNNER_UTILS = ctypes.util.find_library("mlir_c_runner_utils")
 libc = ctypes.CDLL(ctypes.util.find_library("c")) if os.name != "nt" else ctypes.cdll.msvcrt

sparse/mlir_backend/_dtypes.py

@@ -1,71 +1,119 @@
+import inspect
+import math
+import sys
+import typing
+
 from mlir import ir
 
 import numpy as np
 
+from ._common import MlirType
+
+
+def _get_pointer_width() -> int:
+    return round(math.log2(sys.maxsize + 1.0)) + 1
+
+
+_PTR_WIDTH = _get_pointer_width()
+
+
+def _make_int_classes(namespace: dict[str, object], bit_widths: typing.Iterable[int]) -> None:
+    for bw in bit_widths:
+
+        class SignedBW(SignedIntegerDType):
+            np_dtype = getattr(np, f"int{bw}")
+            bit_width = bw
+
+            @classmethod
+            def get_mlir_type(cls):
+                return ir.IntegerType.get_signless(cls.bit_width)
+
+        SignedBW.__name__ = f"Int{bw}"
+        SignedBW.__module__ = __name__
+
+        class UnsignedBW(UnsignedIntegerDType):
+            np_dtype = getattr(np, f"uint{bw}")
+            bit_width = bw
+
+            @classmethod
+            def get_mlir_type(cls):
+                return ir.IntegerType.get_signless(cls.bit_width)
+
+        UnsignedBW.__name__ = f"UInt{bw}"
+        UnsignedBW.__module__ = __name__
+
+        namespace[SignedBW.__name__] = SignedBW
+        namespace[UnsignedBW.__name__] = UnsignedBW
 
-class DType:
-    pass
 
+class DType(MlirType):
+    np_dtype: np.dtype
+    bit_width: int
 
-class Float64(DType):
+
+class FloatingDType(DType): ...
+
+
+class Float64(FloatingDType):
     np_dtype = np.float64
+    bit_width = 64
 
     @classmethod
-    def get(cls):
+    def get_mlir_type(cls):
         return ir.F64Type.get()
 
 
-class Float32(DType):
+class Float32(FloatingDType):
     np_dtype = np.float32
+    bit_width = 32
 
     @classmethod
-    def get(cls):
+    def get_mlir_type(cls):
         return ir.F32Type.get()
 
 
-class Int64(DType):
-    np_dtype = np.int64
+class Float16(FloatingDType):
+    np_dtype = np.float16
+    bit_width = 16
 
     @classmethod
-    def get(cls):
-        return ir.IntegerType.get_signed(64)
+    def get_mlir_type(cls):
+        return ir.F16Type.get()
 
 
-class UInt64(DType):
-    np_dtype = np.uint64
+class IntegerDType(DType): ...
 
-    @classmethod
-    def get(cls):
-        return ir.IntegerType.get_unsigned(64)
 
+class UnsignedIntegerDType(IntegerDType): ...
 
-class Int32(DType):
-    np_dtype = np.int32
-
-    @classmethod
-    def get(cls):
-        return ir.IntegerType.get_signed(32)
 
+class SignedIntegerDType(IntegerDType): ...
 
-class UInt32(DType):
-    np_dtype = np.uint32
 
-    @classmethod
-    def get(cls):
-        return ir.IntegerType.get_unsigned(32)
+_make_int_classes(locals(), [8, 16, 32, 64])
 
 
 class Index(DType):
     np_dtype = np.intp
 
     @classmethod
-    def get(cls):
+    def get_mlir_type(cls):
         return ir.IndexType.get()
 
 
-class SignlessInt64(DType):
-    np_dtype = np.int64
+IntP: type[SignedIntegerDType] = locals()[f"Int{_PTR_WIDTH}"]
+UIntP: type[UnsignedIntegerDType] = locals()[f"UInt{_PTR_WIDTH}"]
 
-    @classmethod
-    def get(cls):
-        return ir.IntegerType.get_signless(64)
+
+def isdtype(dt, /) -> bool:
+    return isinstance(dt, type) and issubclass(dt, DType) and not inspect.isabstract(dt)
+
+
+NUMPY_DTYPE_MAP = {np.dtype(dt.np_dtype): dt for dt in locals().values() if isdtype(dt)}
+
+
+def asdtype(dt, /) -> type[DType]:
+    if isdtype(dt):
+        return dt
+
+    return NUMPY_DTYPE_MAP[np.dtype(dt)]