|
| 1 | +# Copyright 2024 The AI Edge Torch Authors. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | +# ============================================================================== |
| 15 | +import uuid |
| 16 | + |
| 17 | +from ai_edge_torch.odml_torch import export_utils |
| 18 | +from ai_edge_torch.odml_torch.lowerings import context |
| 19 | +from ai_edge_torch.odml_torch.lowerings import registry |
| 20 | +from jax._src.lib.mlir import ir |
| 21 | +from jax._src.lib.mlir.dialects import func |
| 22 | +from jax._src.lib.mlir.dialects import hlo as stablehlo |
| 23 | +import numpy as np |
| 24 | +import torch |
| 25 | +import torch.utils._pytree as pytree |
| 26 | + |
| 27 | +LoweringContext = context.LoweringContext |
| 28 | +lower = registry.lower |
| 29 | + |
| 30 | + |
| 31 | +def _random_lowering( |
| 32 | + lctx: LoweringContext, |
| 33 | + size: list[int], |
| 34 | + generator, |
| 35 | + dtype: torch.dtype, |
| 36 | + rand_tensor, |
| 37 | + composite_name: str, |
| 38 | +): |
| 39 | + if dtype is None: |
| 40 | + dtype = torch.float32 |
| 41 | + |
| 42 | + rand_tensor = rand_tensor.type(dtype) |
| 43 | + data = rand_tensor.detach().numpy() |
| 44 | + |
| 45 | + shape, _ = pytree.tree_flatten(size) |
| 46 | + elty = export_utils.torch_dtype_to_ir_element_type(dtype) |
| 47 | + |
| 48 | + decomp_name = f"{composite_name}.impl_{uuid.uuid4().hex[:8]}" |
| 49 | + |
| 50 | + with ir.InsertionPoint(lctx.ir_module.body): |
| 51 | + |
| 52 | + @func.FuncOp.from_py_func( |
| 53 | + ir.RankedTensorType.get( |
| 54 | + [len(shape)], |
| 55 | + ir.IntegerType.get_signless(32), |
| 56 | + ), |
| 57 | + name=decomp_name, |
| 58 | + ) |
| 59 | + def _rand_impl(_): |
| 60 | + return [stablehlo.constant(ir.DenseElementsAttr.get(data))] |
| 61 | + |
| 62 | + seed, seed2 = ( |
| 63 | + torch.randint( |
| 64 | + torch.iinfo(torch.int64).min, |
| 65 | + torch.iinfo(torch.int64).max, |
| 66 | + (2,), |
| 67 | + dtype=torch.int64, |
| 68 | + generator=generator, |
| 69 | + ) |
| 70 | + .detach() |
| 71 | + .numpy() |
| 72 | + ) |
| 73 | + |
| 74 | + shape_ = stablehlo.constant( |
| 75 | + ir.DenseElementsAttr.get(np.array(shape, dtype=np.int32)) |
| 76 | + ) |
| 77 | + return stablehlo.CompositeOp( |
| 78 | + result=[ir.RankedTensorType.get(shape, elty)], |
| 79 | + inputs=[shape_], |
| 80 | + name=composite_name, |
| 81 | + composite_attributes=ir.DictAttr.get({ |
| 82 | + "seed": ir.IntegerAttr.get(ir.IntegerType.get_signless(64), seed), |
| 83 | + "seed2": ir.IntegerAttr.get(ir.IntegerType.get_signless(64), seed2), |
| 84 | + }), |
| 85 | + decomposition=decomp_name, |
| 86 | + ).results[0] |
| 87 | + |
| 88 | + |
| 89 | +# Schema: |
| 90 | +# - aten::rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, |
| 91 | +# Device? device=None, bool? pin_memory=None) -> Tensor |
| 92 | +# - aten::rand.generator(SymInt[] size, *, Generator? generator, |
| 93 | +# ScalarType? dtype=None, Layout? layout=None, Device? device=None, |
| 94 | +# bool? pin_memory=None) -> Tensor |
| 95 | +@registry.lower(torch.ops.aten.rand) |
| 96 | +def _aten_rand( |
| 97 | + lctx: LoweringContext, |
| 98 | + size, |
| 99 | + generator=None, |
| 100 | + dtype=None, |
| 101 | + layout=torch.strided, |
| 102 | + device=None, |
| 103 | + pin_memory=False, |
| 104 | +): |
| 105 | + return _random_lowering( |
| 106 | + lctx, |
| 107 | + size, |
| 108 | + generator, |
| 109 | + dtype, |
| 110 | + rand_tensor=torch.ops.aten.rand.generator( |
| 111 | + size, generator=generator, dtype=dtype |
| 112 | + ), |
| 113 | + composite_name="odml.random_uniform", |
| 114 | + ) |
| 115 | + |
| 116 | + |
| 117 | +# Schema: |
| 118 | +# - aten::randn(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, |
| 119 | +# Device? device=None, bool? pin_memory=None) -> Tensor |
| 120 | +# - aten::randn.generator(SymInt[] size, *, Generator? generator, |
| 121 | +# ScalarType? dtype=None, Layout? layout=None, Device? device=None, |
| 122 | +# bool? pin_memory=None) -> Tensor |
| 123 | +@registry.lower(torch.ops.aten.randn) |
| 124 | +def _aten_randn( |
| 125 | + lctx: LoweringContext, |
| 126 | + size, |
| 127 | + generator=None, |
| 128 | + dtype=None, |
| 129 | + layout=torch.strided, |
| 130 | + device=None, |
| 131 | + pin_memory=False, |
| 132 | +): |
| 133 | + return _random_lowering( |
| 134 | + lctx, |
| 135 | + size, |
| 136 | + generator, |
| 137 | + dtype, |
| 138 | + rand_tensor=torch.ops.aten.randn.generator( |
| 139 | + size, generator=generator, dtype=dtype |
| 140 | + ), |
| 141 | + composite_name="odml.random_standard_normal", |
| 142 | + ) |
0 commit comments