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
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2017-2021 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#pragma once
#include "common/LLVMWarningsPush.hpp"
#include <llvm/ADT/SetVector.h>
#include <llvm/Pass.h>
#include <llvmWrapper/IR/IRBuilder.h>
#include <llvm/IR/Dominators.h>
#include <llvm/Analysis/PostDominators.h>
#include "common/LLVMWarningsPop.hpp"
#include "Compiler/CodeGenPublic.h"
#include "Compiler/CISACodeGen/ShaderCodeGen.hpp"
void initializeThreadCombiningPass(llvm::PassRegistry&);
namespace IGC
{
enum dim
{
ThreadGroupSize_X,
ThreadGroupSize_Y,
ThreadGroupSize_Z
};
class ThreadCombining : public llvm::ModulePass
{
public:
ThreadCombining()
: ModulePass(ID)
, m_kernel(nullptr)
, m_SLMUsed(false)
{
initializeThreadCombiningPass(*llvm::PassRegistry::getPassRegistry());
}
~ThreadCombining() {}
virtual llvm::StringRef getPassName() const override
{
return "ThreadCombining";
}
bool runOnModule(llvm::Module& M) override;
virtual void getAnalysisUsage(llvm::AnalysisUsage& AU) const override
{
AU.addRequired<llvm::DominatorTreeWrapperPass>();
AU.addRequired<llvm::PostDominatorTreeWrapperPass>();
AU.addRequired<CodeGenContextWrapper>();
AU.addRequired<MetaDataUtilsWrapper>();
}
llvm::Function* m_kernel;
bool m_SLMUsed;
std::vector<llvm::Instruction*> m_barriers;
llvm::SetVector<llvm::Instruction*> m_aliveAcrossBarrier;
std::set<llvm::Instruction*> m_instructionsToMove;
std::map<llvm::Instruction*, std::set<llvm::Instruction*>> m_LiveRegistersPerBarrier;
static char ID;
bool isBarrier(llvm::Instruction& I) const;
bool isSLMUsed(llvm::Instruction* I) const;
unsigned int GetthreadGroupSize(llvm::Module& M, dim dimension);
void SetthreadGroupSize(llvm::Module& M, llvm::Constant* size, dim dimension);
void remapThreads(llvm::Module& M,
unsigned int newSizeX,
unsigned int newSizeY,
unsigned int threadGroupSize_X,
unsigned int threadGroupSize_Y,
IGCLLVM::IRBuilder<>& builder);
void CreateLoopKernel(
llvm::Module& M,
unsigned int newSizeX,
unsigned int newSizeY,
unsigned int threadGroupSize_X,
unsigned int threadGroupSize_Y,
llvm::Function* newFunc,
IGCLLVM::IRBuilder<>& builder);
void CreateNewKernel(llvm::Module& M,
IGCLLVM::IRBuilder<>& builder,
llvm::Function* newFunc);
bool canDoOptimization(llvm::Function* m_kernel, llvm::Module& M);
void PreAnalysis(llvm::Function* m_kernel, llvm::Module& M, std::vector<llvm::Instruction*>& barriers);
void FindRegistersAliveAcrossBarriers(llvm::Function* m_kernel, llvm::Module& M);
};
}
|
