1 /* Copyright 2018 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 "tensorflow/compiler/tf2xla/lib/scatter.h"
17
18 #include <memory>
19 #include <vector>
20
21 #include "absl/types/span.h"
22 #include "tensorflow/compiler/tf2xla/lib/util.h"
23 #include "tensorflow/compiler/xla/client/lib/arithmetic.h"
24 #include "tensorflow/compiler/xla/client/xla_builder.h"
25 #include "tensorflow/compiler/xla/literal.h"
26 #include "tensorflow/compiler/xla/shape_util.h"
27 #include "tensorflow/compiler/xla/status_macros.h"
28 #include "tensorflow/compiler/xla/util.h"
29 #include "tensorflow/core/lib/core/errors.h"
30
31 namespace tensorflow {
32
XlaScatter(const xla::XlaOp & buffer,const xla::XlaOp & updates,const xla::XlaOp & indices,bool indices_are_vectors,const std::function<xla::XlaOp (xla::XlaOp,xla::XlaOp,xla::XlaBuilder *)> & combiner,xla::XlaBuilder * builder)33 xla::StatusOr<xla::XlaOp> XlaScatter(
34 const xla::XlaOp& buffer, const xla::XlaOp& updates,
35 const xla::XlaOp& indices, bool indices_are_vectors,
36 const std::function<xla::XlaOp(xla::XlaOp, xla::XlaOp, xla::XlaBuilder*)>&
37 combiner,
38 xla::XlaBuilder* builder) {
39 TF_ASSIGN_OR_RETURN(xla::Shape buffer_shape, builder->GetShape(buffer));
40 TF_ASSIGN_OR_RETURN(xla::Shape updates_shape, builder->GetShape(updates));
41 TF_ASSIGN_OR_RETURN(xla::Shape indices_shape, builder->GetShape(indices));
42 absl::Span<const int64> indices_dims =
43 xla::AsInt64Slice(indices_shape.dimensions());
44
45 // If the indices are N-dimensional, the minor dimension of indices contains
46 // the indices to update. Otherwise the indices are all scalars.
47 int64 num_index_dims = 1;
48 if (indices_are_vectors) {
49 TF_RET_CHECK(!indices_dims.empty());
50 num_index_dims = indices_dims.back();
51 if (num_index_dims > buffer_shape.rank()) {
52 return errors::InvalidArgument(
53 "The size of the minor dimension of the indices (shape: ",
54 xla::ShapeUtil::HumanString(indices_shape),
55 ") must be <= the rank of the buffer (shape: ",
56 xla::ShapeUtil::HumanString(buffer_shape), ")");
57 }
58 indices_dims.remove_suffix(1);
59 }
60
61 int64 num_indices = 1;
62 for (int64 dim : indices_dims) {
63 num_indices *= dim;
64 }
65
66 // Degenerate case: nothing to update. Return the buffer unchanged.
67 if (num_indices == 0) {
68 return buffer;
69 }
70
71 // If any of the indexed dimensions are zero in the buffer, the update cannot
72 // succeed since it updates a slice of size 1.
73 for (int64 i = 0; i < num_index_dims; ++i) {
74 if (xla::ShapeUtil::GetDimension(buffer_shape, i) == 0) {
75 return errors::InvalidArgument("Scatter dimension ", i,
76 " is of size zero in tensor with shape ",
77 xla::ShapeUtil::HumanString(buffer_shape));
78 }
79 }
80
81 // Example of a 1-D scatter that updates two [3,1] tensors in a tensor of
82 // shape [3,3]:
83 // NOTE: ***This case will not be generated by any of the tf.scatter ops.***
84 //
85 // operand = s32[3,3] parameter(0)
86 // indices = s32[2] parameter(1)
87 // updates = s32[3,2] parameter(2)
88 // scatter = s32[3,3] scatter(operand, indices, updates),
89 // to_apply=update_computation,
90 // update_window_dims={0},
91 // inserted_window_dims={1},
92 // scatter_dims_to_operand_dims={1},
93 // index_vector_dim=1
94 //
95 //
96 // Example of a 1-D scatter that updates two [1,3] tensors in a tensor of
97 // shape [3,3]:
98 //
99 // operand = s32[3,3] parameter(0)
100 // indices = s32[2] parameter(1)
101 // updates = s32[2,3] parameter(2)
102 // scatter = s32[3,3] scatter(operand, indices, updates),
103 // to_apply=update_computation,
104 // update_window_dims={1},
105 // inserted_window_dims={0},
106 // scatter_dims_to_operand_dims={0},
107 // index_vector_dim=1
108 //
109 //
110 // Example of an N-D scatter updating slices of shape [1,1,2] in a tensor of
111 // shape [3,3,2]
112 //
113 // operand = s32[3,3,2] parameter(0)
114 // indices = s32[2,2] parameter(1)
115 // updates = s32[2,2] parameter(2)
116 // scatter = s32[3,3,2] scatter(operand, indices, updates),
117 // to_apply=update_computation,
118 // update_window_dims={1},
119 // inserted_window_dims={0,1},
120 // scatter_dims_to_operand_dims={0,1},
121 // index_vector_dim=1
122 //
123 //
124 // Example of a scatter updating slices of shape [] in a tensor of shape [1,1]
125 //
126 // operand = s32[1,1] parameter(0)
127 // indices = s32[1] parameter(1)
128 // updates = s32[1] parameter(2)
129 // scatter = s32[1,1] scatter(operand, indices, updates),
130 // to_apply=update_computation,
131 // update_window_dims={},
132 // inserted_window_dims={0,1},
133 // scatter_dims_to_operand_dims={0},
134 // index_vector_dim=1
135 // Note that updates operand would be broadcasted into [1] in this case.
136 //
137
138 xla::ScatterDimensionNumbers dim_numbers;
139 dim_numbers.set_index_vector_dim(indices_are_vectors
140 ? indices_shape.dimensions_size() - 1
141 : indices_shape.dimensions_size());
142
143 int64 updates_rank = updates_shape.rank();
144 int64 buffer_rank = buffer_shape.rank();
145 int64 num_window_dims_in_updates = buffer_rank - num_index_dims;
146
147 // If the rank of `updates` is 0 and does not match the expected rank of
148 // updates, broadcast `updates` to the expected shape of updates.
149 auto new_updates = updates;
150 std::vector<int64> expected_updates_dims(indices_dims.begin(),
151 indices_dims.end());
152 for (int64 dim = num_index_dims; dim < buffer_rank; ++dim) {
153 expected_updates_dims.push_back(buffer_shape.dimensions(dim));
154 }
155 int64 expected_updates_rank = expected_updates_dims.size();
156 if (updates_rank == 0 && expected_updates_rank != 0) {
157 new_updates = xla::Broadcast(updates, expected_updates_dims);
158 TF_ASSIGN_OR_RETURN(updates_shape, builder->GetShape(new_updates));
159 updates_rank = updates_shape.rank();
160 }
161
162 if (updates_rank > 0) {
163 for (int64 i = (updates_rank - num_window_dims_in_updates);
164 i < updates_rank; ++i) {
165 dim_numbers.add_update_window_dims(i);
166 }
167 }
168
169 for (int64 i = 0; i < num_index_dims; ++i) {
170 dim_numbers.add_inserted_window_dims(i);
171 dim_numbers.add_scatter_dims_to_operand_dims(i);
172 }
173
174 // Build the combiner computation.
175 xla::XlaComputation combiner_computation;
176 {
177 xla::XlaBuilder cb("scatter-combiner");
178 auto xla_scalar_shape =
179 xla::ShapeUtil::MakeShape(buffer_shape.element_type(), {});
180 auto p0 = xla::Parameter(&cb, 0, xla_scalar_shape, "p0");
181 auto p1 = xla::Parameter(&cb, 1, xla_scalar_shape, "p1");
182 if (combiner) {
183 combiner(p0, p1, &cb);
184 }
185 combiner_computation = cb.Build().ConsumeValueOrDie();
186 }
187
188 VLOG(3) << "Scatter op:";
189 VLOG(3) << " Input: " << xla::ShapeUtil::HumanString(buffer_shape);
190 VLOG(3) << " Indices: " << xla::ShapeUtil::HumanString(indices_shape);
191 VLOG(3) << " Updates: " << xla::ShapeUtil::HumanString(updates_shape);
192 VLOG(3) << " Scatter Dimension Numbers: ";
193 VLOG(3) << " index_vector_dim: " << dim_numbers.index_vector_dim();
194 VLOG(3) << " update_window_dims: ["
195 << absl::StrJoin(dim_numbers.update_window_dims(), ",") << "]";
196 VLOG(3) << " inserted_window_dims: ["
197 << absl::StrJoin(dim_numbers.inserted_window_dims(), ",") << "]";
198 VLOG(3) << " scatter_dims_to_operand_dims: ["
199 << absl::StrJoin(dim_numbers.scatter_dims_to_operand_dims(), ",")
200 << "]";
201
202 return xla::Scatter(buffer, indices, new_updates, combiner_computation,
203 dim_numbers);
204 }
205
206 } // namespace tensorflow
207