• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 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 // XLA-specific Concat Ops.
17 
18 #include <limits>
19 #include <vector>
20 
21 #include "tensorflow/compiler/tf2xla/type_util.h"
22 #include "tensorflow/compiler/tf2xla/xla_helpers.h"
23 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
24 #include "tensorflow/compiler/tf2xla/xla_op_registry.h"
25 #include "tensorflow/compiler/xla/client/xla_builder.h"
26 #include "tensorflow/compiler/xla/literal_util.h"
27 #include "tensorflow/core/framework/bounds_check.h"
28 #include "tensorflow/core/framework/op_kernel.h"
29 #include "tensorflow/core/framework/register_types.h"
30 #include "tensorflow/core/framework/tensor.h"
31 #include "tensorflow/core/framework/tensor_shape.h"
32 #include "tensorflow/core/framework/tensor_types.h"
33 #include "tensorflow/core/framework/types.h"
34 #include "tensorflow/core/lib/core/status.h"
35 #include "tensorflow/core/platform/types.h"
36 
37 namespace tensorflow {
38 namespace {
39 
40 // --------------------------------------------------------------------------
41 class ConcatBaseOp : public XlaOpKernel {
42  public:
ConcatBaseOp(OpKernelConstruction * c,int axis_index)43   ConcatBaseOp(OpKernelConstruction* c, int axis_index)
44       : XlaOpKernel(c), axis_index_(axis_index) {}
45 
Compile(XlaOpKernelContext * ctx)46   void Compile(XlaOpKernelContext* ctx) override {
47     const TensorShape concat_dim_tensor_shape = ctx->InputShape(axis_index_);
48     OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(concat_dim_tensor_shape),
49                 errors::InvalidArgument(
50                     "Concat dim tensor should be a scalar, but got shape ",
51                     concat_dim_tensor_shape.DebugString()));
52     int64 concat_dim;
53     OP_REQUIRES_OK(ctx,
54                    ctx->ConstantInputAsIntScalar(axis_index_, &concat_dim));
55 
56     std::vector<xla::XlaOp> values;
57     std::vector<TensorShape> shapes;
58     OP_REQUIRES_OK(ctx, ctx->InputList("values", &values, &shapes));
59     const int N = values.size();
60     const int input_dims = shapes[0].dims();
61     const TensorShape& input_shape = shapes[0];
62 
63     int32 axis = concat_dim < 0 ? concat_dim + input_dims : concat_dim;
64     OP_REQUIRES(ctx, 0 <= axis && axis < input_dims,
65                 errors::InvalidArgument(
66                     "ConcatOp : Expected concatenating dimensions in the range "
67                     "[",
68                     -input_dims, ", ", input_dims, "), but got ", concat_dim));
69 
70     // Make a vector holding the XlaOp for each of the inputs that has non-zero
71     // elements.
72     std::vector<xla::XlaOp> input_data;
73     int output_concat_dim = 0;
74     for (int i = 0; i < N; ++i) {
75       xla::XlaOp handle = values[i];
76       const TensorShape& in_shape = shapes[i];
77       OP_REQUIRES(
78           ctx, in_shape.dims() == input_dims,
79           errors::InvalidArgument(
80               "ConcatOp : Ranks of all input tensors should match: shape[0] = ",
81               input_shape.DebugString(), " vs. shape[", i,
82               "] = ", in_shape.DebugString()));
83       if (in_shape.dims() == 0) {
84         // Inputs that come in as scalars must be reshaped to 1-vectors.
85         input_data.push_back(xla::Reshape(handle, {1}));
86       } else {
87         input_data.push_back(handle);
88       }
89       output_concat_dim += in_shape.dims() > 0 ? in_shape.dim_size(axis) : 1;
90     }
91 
92     VLOG(1) << "Concat dim " << concat_dim << " equivalent to " << axis;
93     ctx->SetOutput(0, xla::ConcatInDim(ctx->builder(), input_data, axis));
94   }
95 
96  private:
97   int axis_index_;
98 };
99 
100 class ConcatOp : public ConcatBaseOp {
101  public:
ConcatOp(OpKernelConstruction * c)102   explicit ConcatOp(OpKernelConstruction* c)
103       : ConcatBaseOp(c, /* axis_index */ 0) {}
104 };
105 
106 // ConcatV2 operation is the same as Concat except 'concat_dim'
107 // is the last input instead of the first and renamed to 'axis'.
108 class ConcatV2Op : public ConcatBaseOp {
109  public:
ConcatV2Op(OpKernelConstruction * c)110   explicit ConcatV2Op(OpKernelConstruction* c)
111       : ConcatBaseOp(c, /* axis_index */ c->num_inputs() - 1) {}
112 };
113 
114 REGISTER_XLA_OP(Name("Concat").CompileTimeConstantInput("concat_dim"),
115                 ConcatOp);
116 REGISTER_XLA_OP(Name("ConcatV2")
117                     .TypeConstraint("Tidx", DT_INT32)
118                     .CompileTimeConstantInput("axis"),
119                 ConcatV2Op);
120 
121 class ConcatOffsetOp : public XlaOpKernel {
122  public:
ConcatOffsetOp(OpKernelConstruction * ctx)123   explicit ConcatOffsetOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {}
124 
Compile(XlaOpKernelContext * ctx)125   void Compile(XlaOpKernelContext* ctx) override {
126     const TensorShape concat_dim_shape = ctx->InputShape(0);
127     OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(concat_dim_shape),
128                 errors::InvalidArgument(
129                     "Concat dim tensor should be a scalar, but got shape ",
130                     concat_dim_shape.DebugString()));
131     for (int i = 1; i < ctx->num_inputs(); ++i) {
132       OP_REQUIRES(ctx, TensorShapeUtils::IsVector(ctx->InputShape(i)),
133                   errors::InvalidArgument("input ", i,
134                                           " should be a vector, but got shape ",
135                                           ctx->InputShape(i).DebugString()));
136     }
137     // Suppose a Concat() op needs to Concatenate N tensors, each of
138     // which has the same number of dimensions.  Their shapes match
139     // except the concat dimension.
140     //
141     // E.g., say, we want to concatenate 3 tensors in the 2nd
142     // dimension, and their shapes are:
143     //
144     //  [2, 2, 5, 7]
145     //  [2, 3, 5, 7]
146     //  [2, 4, 5, 7]
147     //
148     // Here, N=3, cdim=1, dims=4. The concatenated tensor has shape
149     // [2,9,5,7]. We will compute the cumulative sum along the 2nd
150     // dimension to figure out each input's offset in the concatenated
151     // output:
152     //  [0, 0, 0, 0]
153     //  [0, 2, 0, 0]
154     //  [0, 5, 0, 0]
155     const int32 N = ctx->num_inputs() - 1;
156     const TensorShape inp0_shape = ctx->InputShape(1);
157     std::vector<int64> inp0_dims;
158     OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(1, &inp0_dims));
159     const int64 inp0_rank = inp0_shape.num_elements();
160 
161     int64 cdim;
162     OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntScalar(0, &cdim));
163 
164     VLOG(1) << "ConcatOffset " << cdim << "," << inp0_rank;
165     int32 axis = cdim < 0 ? cdim + inp0_rank : cdim;
166     OP_REQUIRES(ctx, FastBoundsCheck(axis, inp0_rank),
167                 errors::InvalidArgument("Concat dim is out of range: ", axis,
168                                         " vs. ", inp0_rank));
169     int32 offset = 0;
170     for (int i = 0; i < N; ++i) {
171       const TensorShape inp_shape = ctx->InputShape(1 + i);
172       OP_REQUIRES(ctx, inp0_rank == inp_shape.num_elements(),
173                   errors::InvalidArgument("input ", i, " should contain ",
174                                           inp0_rank, " elements, but got ",
175                                           inp_shape.num_elements()));
176       std::vector<int64> inp_dims;
177       OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(1 + i, &inp_dims));
178 
179       Tensor out_constant(DT_INT32, TensorShape({inp0_rank}));
180       auto out_vec = out_constant.vec<int32>();
181       for (int64 j = 0; j < inp0_rank; ++j) {
182         if (j == axis) {
183           out_vec(j) = offset;
184           offset += inp_dims[j];
185         } else {
186           const int32 inp0_element = inp0_dims[j];
187           const int32 inp_element = inp_dims[j];
188           OP_REQUIRES(ctx, inp0_element == inp_element,
189                       errors::InvalidArgument("input[", i, ",", j,
190                                               "] mismatch: ", inp0_element,
191                                               " vs. ", inp_element));
192           out_vec(j) = 0;
193         }
194       }
195 
196       ctx->SetConstantOutput(i, out_constant);
197     }
198   }
199 };
200 
201 REGISTER_XLA_OP(Name("ConcatOffset")
202                     .CompileTimeConstantInput("concat_dim")
203                     .CompileTimeConstantInput("shape"),
204                 ConcatOffsetOp);
205 
206 }  // namespace
207 }  // namespace tensorflow
208