scan_ops.cc revision b89251c6300b9941d06071543e5c4974d0db1984
1/* Copyright 2017 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 <vector> 17 18#include "tensorflow/compiler/tf2xla/shape_util.h" 19#include "tensorflow/compiler/tf2xla/type_util.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_util.h" 24#include "tensorflow/core/framework/op_kernel.h" 25#include "tensorflow/core/framework/partial_tensor_shape.h" 26#include "tensorflow/core/framework/register_types.h" 27#include "tensorflow/core/framework/tensor.h" 28#include "tensorflow/core/framework/tensor_types.h" 29#include "tensorflow/core/framework/types.h" 30#include "tensorflow/core/kernels/bounds_check.h" 31#include "tensorflow/core/kernels/concat_lib.h" 32#include "tensorflow/core/lib/core/status.h" 33#include "tensorflow/core/platform/types.h" 34 35namespace tensorflow { 36namespace { 37 38class ScanOp : public XlaOpKernel { 39 public: 40 ScanOp(OpKernelConstruction* ctx, bool sum) : XlaOpKernel(ctx), sum_(sum) { 41 OP_REQUIRES_OK(ctx, ctx->GetAttr("reverse", &reverse_)); 42 OP_REQUIRES_OK(ctx, ctx->GetAttr("exclusive", &exclusive_)); 43 } 44 45 void Compile(XlaOpKernelContext* ctx) override { 46 const TensorShape input_shape = ctx->InputShape(0); 47 const TensorShape tensor_axis_shape = ctx->InputShape(1); 48 49 OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(tensor_axis_shape), 50 errors::InvalidArgument("ScanOp: axis must be a scalar, not ", 51 tensor_axis_shape.DebugString())); 52 53 int64 axis; 54 OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntScalar(1, &axis)); 55 if (axis < 0) { 56 axis += input_shape.dims(); 57 } 58 OP_REQUIRES( 59 ctx, FastBoundsCheck(axis, input_shape.dims()), 60 errors::InvalidArgument("ScanOp: Expected scan axis in the range [", 61 -input_shape.dims(), ", ", input_shape.dims(), 62 "), but got ", axis)); 63 64 DataType dtype = ctx->input_type(0); 65 66 if (input_shape.num_elements() == 0) { 67 // Exit early if there is nothing to compute. 68 ctx->SetOutput(0, ctx->Input(0)); 69 return; 70 } 71 72 xla::ComputationBuilder* builder = ctx->builder(); 73 74 std::vector<int64> window_strides(input_shape.dims(), 1); 75 std::vector<int64> window_dims(input_shape.dims(), 1); 76 window_dims[axis] = input_shape.dim_size(axis); 77 78 std::vector<std::pair<int64, int64>> padding(input_shape.dims(), {0, 0}); 79 padding[axis].first = input_shape.dim_size(axis) - 1; 80 // In exclusive mode, add an extra padding element so there is a complete 81 // window of padding before the data starts. 82 if (exclusive_) { 83 ++padding[axis].first; 84 } 85 if (reverse_) { 86 std::swap(padding[axis].first, padding[axis].second); 87 } 88 89 xla::ComputationDataHandle input = ctx->Input(0); 90 xla::ComputationDataHandle init; 91 const xla::Computation* reducer; 92 if (sum_) { 93 init = XlaHelpers::Zero(builder, dtype); 94 reducer = ctx->GetOrCreateAdd(dtype); 95 } else { 96 init = XlaHelpers::One(builder, dtype); 97 reducer = ctx->GetOrCreateMul(dtype); 98 } 99 auto output = builder->ReduceWindowWithGeneralPadding( 100 ctx->Input(0), init, *reducer, window_dims, window_strides, padding); 101 102 // In exclusive mode, we have computed an extra element containing the sum 103 // of all the input elements. Slice off this extra "last" element. 104 if (exclusive_) { 105 if (reverse_) { 106 output = builder->SliceInDim(output, 1, input_shape.dim_size(axis) + 1, 107 1, axis); 108 109 } else { 110 output = 111 builder->SliceInDim(output, 0, input_shape.dim_size(axis), 1, axis); 112 } 113 } 114 ctx->SetOutput(0, output); 115 } 116 117 private: 118 const bool sum_; // True=cumulative sum. False=cumulative product. 119 bool reverse_; 120 bool exclusive_; 121}; 122 123class CumsumOp : public ScanOp { 124 public: 125 explicit CumsumOp(OpKernelConstruction* ctx) : ScanOp(ctx, /*sum=*/true) {} 126}; 127// TODO(phawkins): implement non-float windowed reductions in XLA and remove the 128// type constraint. 129REGISTER_XLA_OP(Name("Cumsum").TypeConstraint("T", DT_FLOAT), CumsumOp); 130 131class CumprodOp : public ScanOp { 132 public: 133 explicit CumprodOp(OpKernelConstruction* ctx) : ScanOp(ctx, /*sum=*/false) {} 134}; 135// TODO(phawkins): implement non-float windowed reductions in XLA and remove the 136// type constraint. 137REGISTER_XLA_OP(Name("Cumprod").TypeConstraint("T", DT_FLOAT), CumprodOp); 138 139} // anonymous namespace 140} // namespace tensorflow 141