1 | /* Copyright 2022 The TensorFlow Authors. All Rights Reserved. |
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 | |
16 | #include <algorithm> |
17 | #include <array> |
18 | #include <memory> |
19 | |
20 | #include "absl/strings/str_format.h" |
21 | #include "absl/strings/string_view.h" |
22 | #include "tensorflow/core/framework/op.h" |
23 | #include "tensorflow/core/framework/op_kernel.h" |
24 | #include "tensorflow/core/framework/op_requires.h" |
25 | #include "tensorflow/core/framework/register_types.h" |
26 | #include "tensorflow/core/framework/shape_inference.h" |
27 | #include "tensorflow/core/framework/tensor_shape.h" |
28 | #include "tensorflow/core/framework/tensor_util.h" |
29 | #include "tensorflow/core/framework/types.pb.h" |
30 | #include "tensorflow/core/kernels/random_index_shuffle.h" |
31 | #include "tensorflow/core/platform/errors.h" |
32 | #include "tensorflow/core/platform/types.h" |
33 | #include "tensorflow/core/profiler/lib/traceme.h" |
34 | #include "tensorflow/core/protobuf/error_codes.pb.h" |
35 | |
36 | namespace tensorflow { |
37 | namespace { |
38 | |
39 | constexpr absl::string_view kRounds = "rounds" ; |
40 | |
41 | template <typename DType> |
42 | std::array<uint32_t, 3> CastSeedFrom(const Tensor& seed_t, const int row) { |
43 | const auto seed_vals = seed_t.flat<DType>(); |
44 | return {static_cast<uint32_t>(seed_vals(3 * row)), |
45 | static_cast<uint32_t>(seed_vals(3 * row + 1)), |
46 | static_cast<uint32_t>(seed_vals(3 * row + 2))}; |
47 | } |
48 | |
49 | Status GetSeed(const Tensor& seed_t, const int row, |
50 | std::array<uint32_t, 3>* seed) { |
51 | if (seed_t.dtype() == DT_INT32) { |
52 | *seed = CastSeedFrom<int32_t>(seed_t, row); |
53 | } else if (seed_t.dtype() == DT_UINT32) { |
54 | *seed = CastSeedFrom<uint32_t>(seed_t, row); |
55 | } else if (seed_t.dtype() == DT_INT64) { |
56 | *seed = CastSeedFrom<int64_t>(seed_t, row); |
57 | } else if (seed_t.dtype() == DT_UINT64) { |
58 | *seed = CastSeedFrom<uint64_t>(seed_t, row); |
59 | } else { |
60 | return errors::InvalidArgument("Invalid seed type: " , |
61 | DataTypeString(seed_t.dtype())); |
62 | } |
63 | return OkStatus(); |
64 | } |
65 | |
66 | template <typename IntType> |
67 | class RandomIndexShuffleOp : public OpKernel { |
68 | public: |
69 | explicit RandomIndexShuffleOp(OpKernelConstruction* context) |
70 | : OpKernel(context) { |
71 | OP_REQUIRES_OK(context, context->GetAttr(kRounds, &rounds_)); |
72 | } |
73 | |
74 | void Compute(OpKernelContext* context) override { |
75 | const Tensor& index_t = context->input(0); |
76 | const Tensor& seed_t = context->input(1); |
77 | const Tensor& max_index_t = context->input(2); |
78 | |
79 | const bool all_scalar = |
80 | index_t.dims() == 0 && seed_t.dims() == 1 && max_index_t.dims() == 0; |
81 | const int64_t num_outputs = |
82 | std::max(std::max(index_t.NumElements(), max_index_t.NumElements()), |
83 | seed_t.NumElements() / 3); |
84 | |
85 | // Check shapes. |
86 | OP_REQUIRES(context, |
87 | index_t.dims() == 0 || |
88 | (index_t.dims() == 1 && index_t.dim_size(0) == num_outputs), |
89 | errors::InvalidArgument("Index bust be a scalar or vector." )); |
90 | OP_REQUIRES(context, |
91 | (seed_t.dims() == 1 && seed_t.dim_size(0) == 3) || |
92 | (seed_t.dims() == 2 && seed_t.dim_size(0) == num_outputs && |
93 | seed_t.dim_size(1) == 3), |
94 | errors::InvalidArgument(absl::StrFormat( |
95 | "Seed must be a vector of size [3] " |
96 | "or a matrix of size [%d, 3] but got %s." , |
97 | num_outputs, seed_t.shape().DebugString()))); |
98 | OP_REQUIRES( |
99 | context, |
100 | max_index_t.dims() == 0 || |
101 | (max_index_t.dims() == 1 && max_index_t.dim_size(0) == num_outputs), |
102 | errors::InvalidArgument( |
103 | absl::StrFormat("Maxval must be a scalar or a vector of " |
104 | "the same size as index but got %s" , |
105 | max_index_t.shape().DebugString()))); |
106 | |
107 | // Create output tensor. |
108 | Tensor* new_index_t; |
109 | if (all_scalar) { |
110 | OP_REQUIRES_OK( |
111 | context, context->allocate_output(0, index_t.shape(), &new_index_t)); |
112 | } else { |
113 | TensorShape new_index_shape({num_outputs}); |
114 | OP_REQUIRES_OK( |
115 | context, context->allocate_output(0, new_index_shape, &new_index_t)); |
116 | } |
117 | |
118 | for (int64_t i = 0; i < num_outputs; ++i) { |
119 | const auto index = |
120 | static_cast<uint64_t>(index_t.dims() ? index_t.vec<IntType>()(i) |
121 | : index_t.scalar<IntType>()()); |
122 | const auto max_index = static_cast<uint64_t>( |
123 | max_index_t.dims() ? max_index_t.vec<IntType>()(i) |
124 | : max_index_t.scalar<IntType>()()); |
125 | std::array<uint32_t, 3> seed; |
126 | OP_REQUIRES_OK(context, |
127 | GetSeed(seed_t, seed_t.dims() == 1 ? 0 : i, &seed)); |
128 | const auto new_index = |
129 | tensorflow::random::index_shuffle(index, seed, max_index, rounds_); |
130 | new_index_t->flat<IntType>()(i) = static_cast<IntType>(new_index); |
131 | } |
132 | } |
133 | |
134 | private: |
135 | int32_t rounds_; // Number of rounds for the block cipher. |
136 | |
137 | TF_DISALLOW_COPY_AND_ASSIGN(RandomIndexShuffleOp); |
138 | }; |
139 | |
140 | #define REGISTER(TYPE) \ |
141 | REGISTER_KERNEL_BUILDER(Name("RandomIndexShuffle") \ |
142 | .Device(DEVICE_CPU) \ |
143 | .TypeConstraint<TYPE>("dtype"), \ |
144 | RandomIndexShuffleOp<TYPE>); |
145 | |
146 | TF_CALL_int32(REGISTER); |
147 | TF_CALL_int64(REGISTER); |
148 | TF_CALL_uint32(REGISTER); |
149 | TF_CALL_uint64(REGISTER); |
150 | |
151 | } // namespace |
152 | } // namespace tensorflow |
153 | |