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 "tensorflow/compiler/xla/service/cpu/elemental_ir_emitter.h"
17
18#include <string>
19
20#include "llvm/IR/Instructions.h"
21#include "llvm/IR/Module.h"
22#include "tensorflow/compiler/xla/service/hlo_opcode.h"
23#include "tensorflow/compiler/xla/service/llvm_ir/llvm_util.h"
24#include "tensorflow/compiler/xla/types.h"
25#include "tensorflow/compiler/xla/util.h"
26#include "tensorflow/compiler/xla/xla_data.pb.h"
27
28namespace xla {
29namespace cpu {
30
31StatusOr<llvm::Value*> CpuElementalIrEmitter::EmitFloatUnaryOp(
32    const HloInstruction* op, llvm::Value* operand_value) const {
33  switch (op->opcode()) {
34    case HloOpcode::kTanh: {
35      PrimitiveType element_type = op->shape().element_type();
36      bool cast_result_to_fp16 = false;
37      string function_name;
38      switch (element_type) {
39        case F16:
40          cast_result_to_fp16 = true;
41          operand_value = ir_builder_->CreateFPCast(operand_value,
42                                                    ir_builder_->getFloatTy());
43          TF_FALLTHROUGH_INTENDED;
44        case F32:
45          function_name = "tanhf";
46          break;
47        case F64:
48          function_name = "tanh";
49          break;
50        default:
51          return Unimplemented("tanh");
52      }
53      // Create a function declaration.
54      llvm::Function* function =
55          llvm::cast<llvm::Function>(module_->getOrInsertFunction(
56              llvm_ir::AsStringRef(function_name), operand_value->getType(),
57              operand_value->getType()));
58      function->setCallingConv(llvm::CallingConv::C);
59      function->setDoesNotThrow();
60      function->setDoesNotAccessMemory();
61      // Create an instruction to call the function.
62      llvm::Value* result = ir_builder_->CreateCall(function, operand_value);
63      if (cast_result_to_fp16) {
64        result = ir_builder_->CreateFPCast(result, ir_builder_->getHalfTy());
65      }
66      return result;
67    }
68    default:
69      return ElementalIrEmitter::EmitFloatUnaryOp(op, operand_value);
70  }
71}
72
73StatusOr<llvm::Value*> CpuElementalIrEmitter::EmitAtan2(
74    PrimitiveType prim_type, llvm::Value* lhs, llvm::Value* rhs) const {
75  string function_name;
76  bool cast_result_to_fp16 = false;
77  switch (prim_type) {
78    case F16:
79      cast_result_to_fp16 = true;
80      lhs = ir_builder_->CreateFPCast(lhs, ir_builder_->getFloatTy());
81      rhs = ir_builder_->CreateFPCast(rhs, ir_builder_->getFloatTy());
82      TF_FALLTHROUGH_INTENDED;
83    case F32:
84      function_name = "atan2f";
85      break;
86    case F64:
87      function_name = "atan2";
88      break;
89    default:
90      return Unimplemented("atan2");
91  }
92  // Create a function declaration.
93  llvm::Function* function =
94      llvm::cast<llvm::Function>(module_->getOrInsertFunction(
95          llvm_ir::AsStringRef(function_name), lhs->getType(), lhs->getType(),
96          rhs->getType()));
97  function->setCallingConv(llvm::CallingConv::C);
98  function->setDoesNotThrow();
99  function->setDoesNotAccessMemory();
100  // Create an instruction to call the function.
101  llvm::Value* result = ir_builder_->CreateCall(function, {lhs, rhs});
102  if (cast_result_to_fp16) {
103    result = ir_builder_->CreateFPCast(result, ir_builder_->getHalfTy());
104  }
105  return result;
106}
107
108llvm_ir::ElementGenerator CpuElementalIrEmitter::MakeElementGenerator(
109    const HloInstruction* hlo,
110    const HloToElementGeneratorMap& operand_to_generator) const {
111  if (hlo->opcode() == HloOpcode::kMap) {
112    return [this, hlo, &operand_to_generator](
113               const llvm_ir::IrArray::Index& index) -> StatusOr<llvm::Value*> {
114      std::vector<llvm::Value*> operands;
115      for (int i = 0; i < hlo->operand_count(); i++) {
116        TF_ASSIGN_OR_RETURN(llvm::Value * operand_value,
117                            operand_to_generator.at(hlo->operand(i))(
118                                ElementwiseSourceIndex(index, *hlo, 0)));
119        operands.push_back(operand_value);
120      }
121      return ir_emitter_->EmitScalarCall(hlo->shape().element_type(),
122                                         hlo->to_apply(), operands,
123                                         llvm_ir::IrName(hlo));
124    };
125  }
126  return ElementalIrEmitter::MakeElementGenerator(hlo, operand_to_generator);
127}
128}  // namespace cpu
129}  // namespace xla
130