1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2017-2021 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#include "Compiler/Optimizer/OpenCLPasses/CorrectlyRoundedDivSqrt/CorrectlyRoundedDivSqrt.hpp"
#include "Compiler/MetaDataApi/IGCMetaDataHelper.h"
#include "Compiler/IGCPassSupport.h"
#include "GenISAIntrinsics/GenIntrinsicInst.h"
#include "llvmWrapper/IR/DerivedTypes.h"
#include "common/LLVMWarningsPush.hpp"
#include <llvm/IR/Function.h>
#include <llvm/IR/Instructions.h>
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/IRBuilder.h>
#include "common/LLVMWarningsPop.hpp"
using namespace llvm;
using namespace IGC;
using namespace IGC::IGCMD;
// Register pass to igc-opt
#define PASS_FLAG "igc-correctly-rounded-div-sqrt"
#define PASS_DESCRIPTION "Ensures single precision divide and sqrt are correctly rounded"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS false
IGC_INITIALIZE_PASS_BEGIN(CorrectlyRoundedDivSqrt, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_DEPENDENCY(MetaDataUtilsWrapper)
IGC_INITIALIZE_PASS_END(CorrectlyRoundedDivSqrt, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
char CorrectlyRoundedDivSqrt::ID = 0;
CorrectlyRoundedDivSqrt::CorrectlyRoundedDivSqrt() : ModulePass(ID),
m_forceCR(false), m_hasHalfTy(false), m_IsCorrectlyRounded(false)
{
initializeCorrectlyRoundedDivSqrtPass(*PassRegistry::getPassRegistry());
}
CorrectlyRoundedDivSqrt::CorrectlyRoundedDivSqrt(bool forceCR, bool HasHalf) : ModulePass(ID),
m_forceCR(forceCR), m_hasHalfTy(HasHalf), m_IsCorrectlyRounded(false)
{
initializeCorrectlyRoundedDivSqrtPass(*PassRegistry::getPassRegistry());
}
bool CorrectlyRoundedDivSqrt::runOnModule(Module& M)
{
// Was the module compiled with the CR flag on?
m_IsCorrectlyRounded = getAnalysis<MetaDataUtilsWrapper>().getModuleMetaData()->compOpt.CorrectlyRoundedDivSqrt;
// Even if it wasn't, it's possible that CR was requested through a build-time option
// (This is relevant at least for SPIR)
if (!m_IsCorrectlyRounded && !m_forceCR)
{
return false;
}
m_changed = false;
m_module = &M;
for (Function& F : M)
{
if (F.isDeclaration())
{
if (!m_hasHalfTy)
m_changed |= processDeclaration(F);
}
else
{
visit(F);
}
}
m_module = nullptr;
return m_changed;
}
bool CorrectlyRoundedDivSqrt::processDeclaration(Function& F)
{
StringRef name = F.getName();
if (name.startswith("_Z4sqrt"))
{
std::string newName = name.str();
newName[2] = '7';
newName.insert(7, "_cr");
F.setName(newName);
return true;
}
#if defined(IGC_SCALAR_USE_KHRONOS_SPIRV_TRANSLATOR)
else if (name.startswith("_Z16__spirv_ocl_sqrt"))
{
std::string newName = name.str();
newName[3] = '9';
newName.insert(20, "_cr");
F.setName(newName);
return true;
}
#else
else if (name.startswith("__builtin_spirv_OpenCL_sqrt_") &&
!name.startswith("__builtin_spirv_OpenCL_sqrt_cr"))
{
std::string newName = name.str();
newName.insert(28, "cr_");
F.setName(newName);
return true;
}
#endif
// not sqrt function
return false;
}
Value* CorrectlyRoundedDivSqrt::emitIEEEDivide(BinaryOperator* I, Value* Op0, Value* Op1)
{
Type* Ty = Op0->getType();
IRBuilder<> IRB(I);
std::string FuncName = "__builtin_spirv_divide_cr_f32_f32";
SmallVector<Type*, 2> ArgsTypes{ Ty->getScalarType(), Ty->getScalarType() };
auto FT = FunctionType::get(Ty->getScalarType(), ArgsTypes, false);
auto IEEEDivide = m_module->getOrInsertFunction(FuncName, FT);
Value* Divide = nullptr;
if (!isa<VectorType>(Ty))
{
Value* Args[] = { Op0, Op1 };
Divide = IRB.CreateCall(IEEEDivide, Args);
}
else
{
auto vType = dyn_cast<IGCLLVM::FixedVectorType>(Ty);
unsigned VecLen = (uint32_t)vType->getNumElements();
Divide = UndefValue::get(Ty);
for (unsigned i = 0; i < VecLen; i++)
{
auto* SOp0 = IRB.CreateExtractElement(Op0, i);
auto* SOp1 = IRB.CreateExtractElement(Op1, i);
Value* Args[] = { SOp0, SOp1 };
auto* ScalarDivide = IRB.CreateCall(IEEEDivide, Args);
Divide = IRB.CreateInsertElement(Divide, ScalarDivide, i);
}
}
return Divide;
}
void CorrectlyRoundedDivSqrt::visitFDiv(BinaryOperator& I)
{
Type* Ty = I.getType();
if (Ty->getScalarType()->isFloatTy())
{
auto* Divide = emitIEEEDivide(&I, I.getOperand(0), I.getOperand(1));
I.replaceAllUsesWith(Divide);
I.eraseFromParent();
m_changed = true;
}
}
|
