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 // Native XLA implementations of XLA Relu Ops 17 18 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 19 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" 20 #include "tensorflow/compiler/tf2xla/xla_op_registry.h" 21 #include "tensorflow/compiler/xla/client/lib/constants.h" 22 #include "tensorflow/compiler/xla/client/xla_builder.h" 23 #include "tensorflow/compiler/xla/literal.h" 24 25 namespace tensorflow { 26 namespace { 27 28 class ReluOp : public XlaOpKernel { 29 public: ReluOp(OpKernelConstruction * ctx)30 explicit ReluOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 31 // Computes the max of the scalar input x and 0. Compile(XlaOpKernelContext * ctx)32 void Compile(XlaOpKernelContext* ctx) override { 33 xla::XlaBuilder* builder = ctx->builder(); 34 auto zero = XlaHelpers::Zero(builder, input_type(0)); 35 ctx->SetOutput(0, xla::Max(zero, ctx->Input(0))); 36 } 37 }; 38 REGISTER_XLA_OP(Name("Relu"), ReluOp); 39 40 class Relu6Op : public XlaOpKernel { 41 public: Relu6Op(OpKernelConstruction * ctx)42 explicit Relu6Op(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 43 // Clamp the scalar input between 0 and 6. Compile(XlaOpKernelContext * ctx)44 void Compile(XlaOpKernelContext* ctx) override { 45 xla::XlaBuilder* builder = ctx->builder(); 46 auto zero = XlaHelpers::Zero(builder, input_type(0)); 47 auto six = XlaHelpers::IntegerLiteral(builder, input_type(0), 6); 48 ctx->SetOutput(0, xla::Clamp(zero, ctx->Input(0), six)); 49 } 50 }; 51 REGISTER_XLA_OP(Name("Relu6"), Relu6Op); 52 53 class LeakyReluOp : public XlaOpKernel { 54 public: LeakyReluOp(OpKernelConstruction * ctx)55 explicit LeakyReluOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { 56 OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &alpha_)); 57 } Compile(XlaOpKernelContext * ctx)58 void Compile(XlaOpKernelContext* ctx) override { 59 auto features = ctx->Input("features"); 60 auto output = 61 xla::Max(features, features * xla::ScalarLike(features, alpha_)); 62 ctx->SetOutput(0, output); 63 } 64 float alpha_; 65 }; 66 REGISTER_XLA_OP(Name("LeakyRelu"), LeakyReluOp); 67 68 class ReluGradOp : public XlaOpKernel { 69 public: ReluGradOp(OpKernelConstruction * ctx)70 explicit ReluGradOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 71 // Return the lhs (incoming gradient) if the rhs (input feature) > 0, 72 // otherwise return 0. Compile(XlaOpKernelContext * ctx)73 void Compile(XlaOpKernelContext* ctx) override { 74 xla::XlaBuilder* b = ctx->builder(); 75 const TensorShape shape = ctx->InputShape(0); 76 const auto zero = 77 xla::Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); 78 const auto pred = xla::Gt(ctx->Input(1), zero); 79 ctx->SetOutput(0, xla::Select(pred, ctx->Input(0), zero)); 80 } 81 }; 82 REGISTER_XLA_OP(Name("ReluGrad"), ReluGradOp); 83 84 class Relu6GradOp : public XlaOpKernel { 85 public: Relu6GradOp(OpKernelConstruction * ctx)86 explicit Relu6GradOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 87 // Return the lhs (incoming gradient) if the rhs (input feature) > 0, 88 // otherwise return 0. Compile(XlaOpKernelContext * ctx)89 void Compile(XlaOpKernelContext* ctx) override { 90 xla::XlaBuilder* b = ctx->builder(); 91 const TensorShape shape = ctx->InputShape(0); 92 const auto zero = 93 xla::Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); 94 const auto six = xla::Broadcast( 95 XlaHelpers::IntegerLiteral(b, input_type(0), 6), shape.dim_sizes()); 96 auto out = xla::Select( 97 xla::And(xla::Lt(ctx->Input(1), six), xla::Gt(ctx->Input(1), zero)), 98 ctx->Input(0), zero); 99 ctx->SetOutput(0, out); 100 } 101 }; 102 REGISTER_XLA_OP(Name("Relu6Grad"), Relu6GradOp); 103 104 class LeakyReluGradOp : public XlaOpKernel { 105 public: LeakyReluGradOp(OpKernelConstruction * ctx)106 explicit LeakyReluGradOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { 107 OP_REQUIRES_OK(ctx, ctx->GetAttr("alpha", &alpha_)); 108 } Compile(XlaOpKernelContext * ctx)109 void Compile(XlaOpKernelContext* ctx) override { 110 auto gradients = ctx->Input("gradients"); 111 auto features = ctx->Input("features"); 112 auto output = 113 xla::Select(xla::Gt(features, xla::ScalarLike(features, 0)), gradients, 114 gradients * xla::ScalarLike(gradients, alpha_)); 115 ctx->SetOutput(0, output); 116 } 117 float alpha_; 118 }; 119 REGISTER_XLA_OP(Name("LeakyReluGrad"), LeakyReluGradOp); 120 121 } // namespace 122 } // namespace tensorflow 123