|
| 1 | +import numpy as np |
| 2 | +from inspect import getmembers, isfunction, isclass, isbuiltin |
| 3 | +from numbers import Number |
| 4 | +from types import FunctionType as ftype, BuiltinFunctionType as bftype |
| 5 | +import sys |
| 6 | +import inspect |
| 7 | +import dpctl |
| 8 | +from dpctl.memory import MemoryUSMShared |
| 9 | + |
| 10 | +debug = False |
| 11 | + |
| 12 | + |
| 13 | +def dprint(*args): |
| 14 | + if debug: |
| 15 | + print(*args) |
| 16 | + sys.stdout.flush() |
| 17 | + |
| 18 | + |
| 19 | +functions_list = [o[0] for o in getmembers(np) if isfunction(o[1]) or isbuiltin(o[1])] |
| 20 | +class_list = [o for o in getmembers(np) if isclass(o[1])] |
| 21 | + |
| 22 | +array_interface_property = "__sycl_usm_array_interface__" |
| 23 | + |
| 24 | + |
| 25 | +def has_array_interface(x): |
| 26 | + return hasattr(x, array_interface_property) |
| 27 | + |
| 28 | + |
| 29 | +class ndarray(np.ndarray): |
| 30 | + """ |
| 31 | + numpy.ndarray subclass whose underlying memory buffer is allocated |
| 32 | + with a foreign allocator. |
| 33 | + """ |
| 34 | + |
| 35 | + def __new__( |
| 36 | + subtype, shape, dtype=float, buffer=None, offset=0, strides=None, order=None |
| 37 | + ): |
| 38 | + # Create a new array. |
| 39 | + if buffer is None: |
| 40 | + dprint("dparray::ndarray __new__ buffer None") |
| 41 | + nelems = np.prod(shape) |
| 42 | + dt = np.dtype(dtype) |
| 43 | + isz = dt.itemsize |
| 44 | + nbytes = int(isz * max(1, nelems)) |
| 45 | + buf = MemoryUSMShared(nbytes) |
| 46 | + new_obj = np.ndarray.__new__( |
| 47 | + subtype, |
| 48 | + shape, |
| 49 | + dtype=dt, |
| 50 | + buffer=buf, |
| 51 | + offset=0, |
| 52 | + strides=strides, |
| 53 | + order=order, |
| 54 | + ) |
| 55 | + if hasattr(new_obj, array_interface_property): |
| 56 | + dprint("buffer None new_obj already has sycl_usm") |
| 57 | + else: |
| 58 | + dprint("buffer None new_obj will add sycl_usm") |
| 59 | + setattr(new_obj, array_interface_property, {}) |
| 60 | + return new_obj |
| 61 | + # zero copy if buffer is a usm backed array-like thing |
| 62 | + elif hasattr(buffer, array_interface_property): |
| 63 | + dprint("dparray::ndarray __new__ buffer", array_interface_property) |
| 64 | + # also check for array interface |
| 65 | + new_obj = np.ndarray.__new__( |
| 66 | + subtype, |
| 67 | + shape, |
| 68 | + dtype=dtype, |
| 69 | + buffer=buffer, |
| 70 | + offset=offset, |
| 71 | + strides=strides, |
| 72 | + order=order, |
| 73 | + ) |
| 74 | + if hasattr(new_obj, array_interface_property): |
| 75 | + dprint("buffer None new_obj already has sycl_usm") |
| 76 | + else: |
| 77 | + dprint("buffer None new_obj will add sycl_usm") |
| 78 | + setattr(new_obj, array_interface_property, {}) |
| 79 | + return new_obj |
| 80 | + else: |
| 81 | + dprint("dparray::ndarray __new__ buffer not None and not sycl_usm") |
| 82 | + nelems = np.prod(shape) |
| 83 | + # must copy |
| 84 | + ar = np.ndarray( |
| 85 | + shape, |
| 86 | + dtype=dtype, |
| 87 | + buffer=buffer, |
| 88 | + offset=offset, |
| 89 | + strides=strides, |
| 90 | + order=order, |
| 91 | + ) |
| 92 | + nbytes = int(ar.nbytes) |
| 93 | + buf = MemoryUSMShared(nbytes) |
| 94 | + new_obj = np.ndarray.__new__( |
| 95 | + subtype, |
| 96 | + shape, |
| 97 | + dtype=dtype, |
| 98 | + buffer=buf, |
| 99 | + offset=0, |
| 100 | + strides=strides, |
| 101 | + order=order, |
| 102 | + ) |
| 103 | + np.copyto(new_obj, ar, casting="no") |
| 104 | + if hasattr(new_obj, array_interface_property): |
| 105 | + dprint("buffer None new_obj already has sycl_usm") |
| 106 | + else: |
| 107 | + dprint("buffer None new_obj will add sycl_usm") |
| 108 | + setattr(new_obj, array_interface_property, {}) |
| 109 | + return new_obj |
| 110 | + |
| 111 | + def __array_finalize__(self, obj): |
| 112 | + dprint("__array_finalize__:", obj, hex(id(obj)), type(obj)) |
| 113 | + # When called from the explicit constructor, obj is None |
| 114 | + if obj is None: |
| 115 | + return |
| 116 | + # When called in new-from-template, `obj` is another instance of our own |
| 117 | + # subclass, that we might use to update the new `self` instance. |
| 118 | + # However, when called from view casting, `obj` can be an instance of any |
| 119 | + # subclass of ndarray, including our own. |
| 120 | + if hasattr(obj, array_interface_property): |
| 121 | + return |
| 122 | + if isinstance(obj, np.ndarray): |
| 123 | + ob = self |
| 124 | + while isinstance(ob, np.ndarray): |
| 125 | + if hasattr(ob, array_interface_property): |
| 126 | + return |
| 127 | + ob = ob.base |
| 128 | + |
| 129 | + # Just raise an exception since __array_ufunc__ makes all reasonable cases not |
| 130 | + # need the code below. |
| 131 | + raise ValueError( |
| 132 | + "Non-USM allocated ndarray can not viewed as a USM-allocated one without a copy" |
| 133 | + ) |
| 134 | + |
| 135 | + # Tell Numba to not treat this type just like a NumPy ndarray but to propagate its type. |
| 136 | + # This way it will use the custom dparray allocator. |
| 137 | + __numba_no_subtype_ndarray__ = True |
| 138 | + |
| 139 | + # Convert to a NumPy ndarray. |
| 140 | + def as_ndarray(self): |
| 141 | + return np.copy(self) |
| 142 | + |
| 143 | + def __array__(self): |
| 144 | + return self |
| 145 | + |
| 146 | + def __array_ufunc__(self, ufunc, method, *inputs, **kwargs): |
| 147 | + if method == "__call__": |
| 148 | + N = None |
| 149 | + scalars = [] |
| 150 | + typing = [] |
| 151 | + for inp in inputs: |
| 152 | + if isinstance(inp, Number): |
| 153 | + scalars.append(inp) |
| 154 | + typing.append(inp) |
| 155 | + elif isinstance(inp, (self.__class__, np.ndarray)): |
| 156 | + if isinstance(inp, self.__class__): |
| 157 | + scalars.append(np.ndarray(inp.shape, inp.dtype, inp)) |
| 158 | + typing.append(np.ndarray(inp.shape, inp.dtype)) |
| 159 | + else: |
| 160 | + scalars.append(inp) |
| 161 | + typing.append(inp) |
| 162 | + if N is not None: |
| 163 | + if N != inp.shape: |
| 164 | + raise TypeError("inconsistent sizes") |
| 165 | + else: |
| 166 | + N = inp.shape |
| 167 | + else: |
| 168 | + return NotImplemented |
| 169 | + # Have to avoid recursive calls to array_ufunc here. |
| 170 | + # If no out kwarg then we create a dparray out so that we get |
| 171 | + # USM memory. However, if kwarg has dparray-typed out then |
| 172 | + # array_ufunc is called recursively so we cast out as regular |
| 173 | + # NumPy ndarray (having a USM data pointer). |
| 174 | + if kwargs.get("out", None) is None: |
| 175 | + # maybe copy? |
| 176 | + # deal with multiple returned arrays, so kwargs['out'] can be tuple |
| 177 | + res_type = np.result_type(*typing) |
| 178 | + out = empty(inputs[0].shape, dtype=res_type) |
| 179 | + out_as_np = np.ndarray(out.shape, out.dtype, out) |
| 180 | + kwargs["out"] = out_as_np |
| 181 | + else: |
| 182 | + # If they manually gave dparray as out kwarg then we have to also |
| 183 | + # cast as regular NumPy ndarray to avoid recursion. |
| 184 | + if isinstance(kwargs["out"], ndarray): |
| 185 | + out = kwargs["out"] |
| 186 | + kwargs["out"] = np.ndarray(out.shape, out.dtype, out) |
| 187 | + else: |
| 188 | + out = kwargs["out"] |
| 189 | + ret = ufunc(*scalars, **kwargs) |
| 190 | + return out |
| 191 | + else: |
| 192 | + return NotImplemented |
| 193 | + |
| 194 | + |
| 195 | +def isdef(x): |
| 196 | + try: |
| 197 | + eval(x) |
| 198 | + return True |
| 199 | + except NameError: |
| 200 | + return False |
| 201 | + |
| 202 | + |
| 203 | +for c in class_list: |
| 204 | + cname = c[0] |
| 205 | + if isdef(cname): |
| 206 | + continue |
| 207 | + # For now we do the simple thing and copy the types from NumPy module into dparray module. |
| 208 | + new_func = "%s = np.%s" % (cname, cname) |
| 209 | + try: |
| 210 | + the_code = compile(new_func, "__init__", "exec") |
| 211 | + exec(the_code) |
| 212 | + except: |
| 213 | + print("Failed to exec type propagation", cname) |
| 214 | + pass |
| 215 | + |
| 216 | +# Redefine all Numpy functions in this module and if they |
| 217 | +# return a Numpy array, transform that to a USM-backed array |
| 218 | +# instead. This is a stop-gap. We should eventually find a |
| 219 | +# way to do the allocation correct to start with. |
| 220 | +for fname in functions_list: |
| 221 | + if isdef(fname): |
| 222 | + continue |
| 223 | + new_func = "def %s(*args, **kwargs):\n" % fname |
| 224 | + new_func += " ret = np.%s(*args, **kwargs)\n" % fname |
| 225 | + new_func += " if type(ret) == np.ndarray:\n" |
| 226 | + new_func += " ret = ndarray(ret.shape, ret.dtype, ret)\n" |
| 227 | + new_func += " return ret\n" |
| 228 | + the_code = compile(new_func, "__init__", "exec") |
| 229 | + exec(the_code) |
| 230 | + |
| 231 | + |
| 232 | +def from_ndarray(x): |
| 233 | + return copy(x) |
| 234 | + |
| 235 | + |
| 236 | +def as_ndarray(x): |
| 237 | + return np.copy(x) |
0 commit comments