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 Ops for broadcasting used in gradient 17 // code. 18 19 #include "absl/strings/str_join.h" 20 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 21 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" 22 #include "tensorflow/compiler/tf2xla/xla_op_registry.h" 23 #include "tensorflow/compiler/xla/literal.h" 24 #include "tensorflow/core/platform/macros.h" 25 #include "tensorflow/core/platform/types.h" 26 #include "tensorflow/core/util/bcast.h" 27 28 namespace tensorflow { 29 namespace { 30 31 // Given shapes of two tensors, computes the broadcast shape. 32 class BCastArgsOp : public XlaOpKernel { 33 public: BCastArgsOp(OpKernelConstruction * ctx)34 explicit BCastArgsOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { 35 OP_REQUIRES_OK(ctx, ctx->MatchSignature({DT_INT32, DT_INT32}, {DT_INT32})); 36 } 37 Compile(XlaOpKernelContext * ctx)38 void Compile(XlaOpKernelContext* ctx) override { 39 OP_REQUIRES( 40 ctx, ctx->num_inputs() == 2, 41 errors::Unimplemented("Broadcast for n-ary operations (n > 2)")); 42 absl::InlinedVector<BCast::Vec, 2> shapes; 43 for (int i = 0; i < ctx->num_inputs(); ++i) { 44 const TensorShape in_shape = ctx->InputShape(i); 45 OP_REQUIRES(ctx, TensorShapeUtils::IsVector(in_shape), 46 errors::InvalidArgument("In[", i, "] must be a vector.", 47 in_shape.DebugString())); 48 std::vector<int64> shape; 49 OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(i, &shape)); 50 shapes.push_back(BCast::Vec(shape.begin(), shape.end())); 51 } 52 BCast bcast(shapes[0], shapes[1]); 53 OP_REQUIRES(ctx, bcast.IsValid(), 54 errors::InvalidArgument( 55 "Incompatible shapes: [", absl::StrJoin(shapes[0], ","), 56 "] vs. [", absl::StrJoin(shapes[1], ","), "]")); 57 58 const int64 len = bcast.output_shape().size(); 59 Tensor output(DT_INT32, TensorShape({len})); 60 for (int64 i = 0; i < len; ++i) { 61 output.flat<int32>()(i) = static_cast<int32>(bcast.output_shape()[i]); 62 } 63 ctx->SetConstantOutput(0, output); 64 } 65 66 private: 67 TF_DISALLOW_COPY_AND_ASSIGN(BCastArgsOp); 68 }; 69 REGISTER_XLA_OP(Name("BroadcastArgs") 70 .CompileTimeConstantInput("s0") 71 .CompileTimeConstantInput("s1"), 72 BCastArgsOp); 73 74 // Given shapes of two tensors, computes the reduction indices for the 75 // gradient computation. 76 // 77 // TODO(zhifengc): 78 // 1. Adds support for n-ary (n >= 2). 79 class BCastGradArgsOp : public XlaOpKernel { 80 public: BCastGradArgsOp(OpKernelConstruction * ctx)81 explicit BCastGradArgsOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { 82 OP_REQUIRES_OK( 83 ctx, ctx->MatchSignature({DT_INT32, DT_INT32}, {DT_INT32, DT_INT32})); 84 } 85 Compile(XlaOpKernelContext * ctx)86 void Compile(XlaOpKernelContext* ctx) override { 87 OP_REQUIRES( 88 ctx, ctx->num_inputs() == 2, 89 errors::Unimplemented("Broadcast for n-ary operations (n > 2)")); 90 91 absl::InlinedVector<BCast::Vec, 4> shapes; 92 for (int i = 0; i < ctx->num_inputs(); ++i) { 93 const TensorShape in_shape = ctx->InputShape(i); 94 OP_REQUIRES(ctx, TensorShapeUtils::IsVector(in_shape), 95 errors::InvalidArgument("In[", i, "] must be a vector.", 96 in_shape.DebugString())); 97 std::vector<int64> vec; 98 OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector(i, &vec)); 99 100 shapes.push_back(BCast::Vec(vec.begin(), vec.end())); 101 } 102 BCast bcast(shapes[0], shapes[1]); 103 OP_REQUIRES(ctx, bcast.IsValid(), 104 errors::InvalidArgument( 105 "Incompatible shapes: [", absl::StrJoin(shapes[0], ","), 106 "] vs. [", absl::StrJoin(shapes[1], ","), "]")); 107 Output(ctx, 0, bcast.grad_x_reduce_idx()); 108 Output(ctx, 1, bcast.grad_y_reduce_idx()); 109 } 110 111 private: Output(XlaOpKernelContext * ctx,int idx,const BCast::Vec & v)112 void Output(XlaOpKernelContext* ctx, int idx, const BCast::Vec& v) { 113 const int64 len = v.size(); 114 Tensor constant(DT_INT32, TensorShape({len})); 115 for (int64 i = 0; i < len; ++i) { 116 constant.flat<int32>()(i) = static_cast<int32>(v[i]); 117 } 118 ctx->SetConstantOutput(idx, constant); 119 } 120 121 TF_DISALLOW_COPY_AND_ASSIGN(BCastGradArgsOp); 122 }; 123 124 REGISTER_XLA_OP(Name("BroadcastGradientArgs") 125 .CompileTimeConstantInput("s0") 126 .CompileTimeConstantInput("s1"), 127 BCastGradArgsOp); 128 129 } // namespace 130 } // namespace tensorflow 131