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 #include "tensorflow/compiler/tf2xla/xla_compilation_device.h" 17 18 #include <functional> 19 #include <memory> 20 21 #include "tensorflow/compiler/tf2xla/shape_util.h" 22 #include "tensorflow/compiler/tf2xla/sharding_util.h" 23 #include "tensorflow/compiler/tf2xla/xla_context.h" 24 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 25 #include "tensorflow/core/common_runtime/local_device.h" 26 #include "tensorflow/core/framework/device_base.h" 27 #include "tensorflow/core/platform/mem.h" 28 29 namespace tensorflow { 30 31 // The XlaCompilationAllocator doesn't actually back any Tensors with storage 32 // buffers of values: instead for each Tensor it stores a 33 // XlaExpression which corresponds to the XLA computation 34 // represented by the Tensor. 35 class XlaCompilationAllocator : public Allocator { 36 public: 37 XlaCompilationAllocator() {} 38 ~XlaCompilationAllocator() override {} 39 40 string Name() override { return "xla_compilation"; } 41 42 void* AllocateRaw(size_t alignment, size_t num_bytes) override { 43 // Regardless of the size requested, always allocates an XlaExpression. 44 // Respects the aligment request because there is alignment checking even 45 // for Tensors whose data is never accessed. 46 void* p = port::AlignedMalloc(sizeof(XlaExpression), alignment); 47 XlaExpression* expression = reinterpret_cast<XlaExpression*>(p); 48 new (expression) XlaExpression(); 49 return expression; 50 } 51 52 void DeallocateRaw(void* ptr) override { 53 XlaExpression* expression = reinterpret_cast<XlaExpression*>(ptr); 54 expression->~XlaExpression(); 55 port::AlignedFree(ptr); 56 } 57 58 // Make sure that even tensors with 0 elements have allocated 59 // buffers, so they get ids to track. 60 bool ShouldAllocateEmptyTensors() override { return true; } 61 62 void GetStats(AllocatorStats* stats) override { stats->Clear(); } 63 64 private: 65 // Don't run any constructors or destructors for complex objects, 66 // since there is no backing store for the tensor to run them 67 // on. strings are the only complex objects currently stored in 68 // Tensors. If others are added, this set of overrides must be 69 // extended to include them. 70 void RunStringCtor(string* p, size_t n) override {} 71 void RunStringDtor(string* p, size_t n) override {} 72 void RunResourceCtor(ResourceHandle* p, size_t n) override {} 73 void RunResourceDtor(ResourceHandle* p, size_t n) override {} 74 }; 75 76 XlaCompilationDevice::XlaCompilationDevice(const SessionOptions& options, 77 DeviceType type) 78 : LocalDevice( 79 options, 80 Device::BuildDeviceAttributes( 81 strings::StrCat("/device:", type.type(), ":0"), type, 82 Bytes(256 << 20), DeviceLocality(), 83 strings::StrCat("device: XLA compilation device ", type.type()))), 84 allocator_(new XlaCompilationAllocator()) {} 85 86 XlaCompilationDevice::~XlaCompilationDevice() {} 87 88 Allocator* XlaCompilationDevice::GetAllocator(AllocatorAttributes attr) { 89 return allocator_.get(); 90 } 91 92 void XlaCompilationDevice::Compute(OpKernel* op_kernel, 93 OpKernelContext* context) { 94 VLOG(4) << "XlaCompilationDevice::Compute " 95 << SummarizeNodeDef(op_kernel->def()); 96 auto* b = XlaContext::Get(context).builder(); 97 xla::OpMetadata metadata; 98 metadata.set_op_type(op_kernel->type_string()); 99 metadata.set_op_name(op_kernel->name()); 100 b->SetOpMetadata(metadata); 101 102 auto sharding_parse_result = ParseShardingFromDevice( 103 op_kernel->def(), std::numeric_limits<int>::max()); 104 OP_REQUIRES_OK(context, sharding_parse_result.status()); 105 tensorflow::gtl::optional<xla::OpSharding> op_sharding = 106 sharding_parse_result.ValueOrDie(); 107 108 // If no sharding metadata is found, XLA is free to use whatever device it 109 // wants. In practice this usually has the effect of placing things on device 110 // 0. 111 xla::ScopedShardingAssignment assign_sharding(b, op_sharding); 112 op_kernel->Compute(context); 113 114 b->ClearOpMetadata(); 115 VLOG(4) << "Done"; 116 } 117 118 Status XlaCompilationDevice::Sync() { return Status::OK(); } 119 120 Status XlaCompilationDevice::MakeTensorFromProto( 121 const TensorProto& tensor_proto, const AllocatorAttributes alloc_attrs, 122 Tensor* tensor) { 123 return errors::InvalidArgument( 124 "XLACompilationDevice::MakeTensorFromProto should not be called"); 125 } 126 127 XlaExpression::XlaExpression() = default; 128 129 void XlaExpression::set_handle(const xla::ComputationDataHandle& h) { 130 handle_ = h; 131 } 132 133 void XlaExpression::set_constant_value(Tensor value) { 134 has_constant_value_ = true; 135 constant_value_ = std::move(value); 136 } 137 138 } // namespace tensorflow 139