1/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2
3Licensed under the Apache License, Version 2.0 (the "License");
4you may not use this file except in compliance with the License.
5You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9Unless required by applicable law or agreed to in writing, software
10distributed under the License is distributed on an "AS IS" BASIS,
11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12See the License for the specific language governing permissions and
13limitations under the License.
14==============================================================================*/
15
16#include "tensorflow/core/kernels/cwise_ops_common.h"
17
18namespace tensorflow {
19
20BinaryOpShared::BinaryOpShared(OpKernelConstruction* ctx, DataType out,
21 DataType in)
22 : OpKernel(ctx) {
23#if !defined(INTEL_MKL) || !defined(ENABLE_MKL)
24 OP_REQUIRES_OK(ctx, ctx->MatchSignature({in, in}, {out}));
25#endif // !INTEL_MKL || !ENABLE_MKL
26}
27
28void BinaryOpShared::SetUnimplementedError(OpKernelContext* ctx) {
29 ctx->SetStatus(errors::Unimplemented(
30 "Broadcast between ", ctx->input(0).shape().DebugString(), " and ",
31 ctx->input(1).shape().DebugString(), " is not supported yet."));
32}
33
34void BinaryOpShared::SetComputeError(OpKernelContext* ctx) {
35 // For speed, errors during compute are caught only via boolean flag, with no
36 // associated information. This is sufficient for now, since the only binary
37 // ops that have compute errors are integer division and mod, and the only
38 // error they produce is zero division.
39 const string& op = ctx->op_kernel().type_string();
40 if ((op == "Div" || op == "Mod" || op == "FloorMod" || op == "FloorDiv") &&
41 DataTypeIsInteger(ctx->op_kernel().input_type(0))) {
42 ctx->CtxFailure(errors::InvalidArgument("Integer division by zero"));
43 } else if ((op == "Pow") &&
44 DataTypeIsInteger(ctx->op_kernel().input_type(0)) &&
45 DataTypeIsSigned(ctx->op_kernel().input_type(1))) {
46 ctx->CtxFailure(errors::InvalidArgument(
47 "Integers to negative integer powers are not allowed"));
48 } else {
49 ctx->CtxFailure(
50 errors::Internal("Unexpected error in binary operator "
51 "(only integer div and mod should have errors)"));
52 }
53}
54
55BinaryOpShared::BinaryOpState::BinaryOpState(OpKernelContext* ctx)
56 : in0(ctx->input(0)),
57 in1(ctx->input(1)),
58 bcast(BCast::FromShape(in0.shape()), BCast::FromShape(in1.shape())) {
59 if (!bcast.IsValid()) {
60 bool incompatible_shape_error;
61 bool has_attr =
62 TryGetNodeAttr(ctx->op_kernel().def(), "incompatible_shape_error",
63 &(incompatible_shape_error));
64 if (has_attr && !incompatible_shape_error) {
65 const string& op = ctx->op_kernel().type_string();
66 OP_REQUIRES_OK(ctx, ctx->allocate_output(0, TensorShape({}), &out));
67 result = (op == "NotEqual");
68 return;
69 }
70
71 ctx->SetStatus(errors::InvalidArgument(
72 "Incompatible shapes: ", in0.shape().DebugString(), " vs. ",
73 in1.shape().DebugString()));
74 return;
75 }
76
77 const TensorShape output_shape = BCast::ToShape(bcast.output_shape());
78 out_num_elements = output_shape.num_elements();
79 in0_num_elements = in0.NumElements();
80 in1_num_elements = in1.NumElements();
81 OP_REQUIRES_OK(ctx, ctx->forward_input_or_allocate_output(
82 {0, 1}, 0, output_shape, &out));
83
84 ndims = static_cast<int>(bcast.x_reshape().size());
85}
86
87} // namespace tensorflow
88