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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
|
/*========================== begin_copyright_notice ============================
Copyright (C) 2017-2021 Intel Corporation
SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/
#include "GenISAIntrinsics/GenIntrinsicInst.h"
#include "ConvertMSAAPayloadTo16Bit.hpp"
#include "Compiler/IGCPassSupport.h"
#include "common/LLVMUtils.h"
#include "Compiler/CodeGenPublic.h"
#include "common/LLVMWarningsPush.hpp"
#include "llvmWrapper/IR/DerivedTypes.h"
#include <llvm/Support/CommandLine.h>
#include <llvm/IR/Function.h>
#include <llvm/ADT/SmallVector.h>
#include "common/LLVMWarningsPop.hpp"
#include "Probe/Assertion.h"
using namespace llvm;
using namespace IGC;
// Register pass to igc-opt
#define PASS_FLAG "igc-ConverMSAAPayloadTo16Bit"
#define PASS_DESCRIPTION "Comvert normal MSAA intrinsics to 16 bit payload intrinsics"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS true
IGC_INITIALIZE_PASS_BEGIN(ConvertMSAAPayloadTo16Bit, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_DEPENDENCY(CodeGenContextWrapper)
IGC_INITIALIZE_PASS_END(ConvertMSAAPayloadTo16Bit, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
char ConvertMSAAPayloadTo16Bit::ID = 0;
ConvertMSAAPayloadTo16Bit::ConvertMSAAPayloadTo16Bit() : FunctionPass(ID)
{
initializeWorkaroundAnalysisPass(*PassRegistry::getPassRegistry());
};
/*
Returns vector of LdmsIntrinsics using inst directly and indirectly
input: inst - instruction that ldms intrinsics we search for are using
output: ldmsUsing - vector of ldmsIntrinsics using inst
*/
void ConvertMSAAPayloadTo16Bit::findLdmsUsingInstDownInTree(Value* inst, std::vector<GenIntrinsicInst*>& ldmsUsing)
{
if (LdmsInstrinsic* ldms = dyn_cast<LdmsInstrinsic>(inst))
{
ldmsUsing.push_back(ldms);
return;
}
else
{
for (User* user : inst->users())
{
findLdmsUsingInstDownInTree(user, ldmsUsing);
}
}
}
void ConvertMSAAPayloadTo16Bit::visitCallInst(CallInst& I)
{
if (GenIntrinsicInst* intr = dyn_cast<GenIntrinsicInst>(&I))
{
GenISAIntrinsic::ID intrID = intr->getIntrinsicID();
switch (intrID)
{
case GenISAIntrinsic::GenISA_ldmcsptr:
{
GenIntrinsicInst* ldmcs = intr;
Type* coordType = m_builder->getInt32Ty();
if (m_pCtxWrapper->getCodeGenContext()->platform.supports16BitLdMcs())
{
coordType = m_builder->getInt16Ty();
}
Type* types_ldmcs[] = {
IGCLLVM::FixedVectorType::get(m_builder->getInt16Ty(), 4),
coordType,
ldmcs->getOperand(4)->getType() };
Function* func_ldmcs =
GenISAIntrinsic::getDeclaration(
I.getParent()->getParent()->getParent(),
GenISAIntrinsic::GenISA_ldmcsptr,
types_ldmcs);
m_builder->SetInsertPoint(ldmcs);
Value* packed_tex_params_ldmcs[] = {
m_builder->CreateTrunc(ldmcs->getOperand(0), coordType),
m_builder->CreateTrunc(ldmcs->getOperand(1), coordType),
m_builder->CreateTrunc(ldmcs->getOperand(2), coordType),
m_builder->CreateTrunc(ldmcs->getOperand(3), coordType),
ldmcs->getOperand(4),
ldmcs->getOperand(5),
ldmcs->getOperand(6),
ldmcs->getOperand(7)
};
llvm::CallInst* new_mcs_call = m_builder->CreateCall(func_ldmcs, packed_tex_params_ldmcs);
llvm::Value* mcs0 = m_builder->CreateExtractElement(new_mcs_call, m_builder->getInt32(0));
llvm::Value* mcs1 = m_builder->CreateExtractElement(new_mcs_call, m_builder->getInt32(1));
llvm::Value* mcs2 = m_builder->CreateExtractElement(new_mcs_call, m_builder->getInt32(2));
llvm::Value* mcs3 = m_builder->CreateExtractElement(new_mcs_call, m_builder->getInt32(3));
// find ldms using this ldmcs call and convert them to 16 bit
std::vector<GenIntrinsicInst*> ldmsUsingldmcs;
findLdmsUsingInstDownInTree(ldmcs, ldmsUsingldmcs);
for (GenIntrinsicInst* ldms : ldmsUsingldmcs)
{
Type* types_ldms[] = { ldms->getType(), ldms->getOperand(7)->getType() };
Function* func_ldms = GenISAIntrinsic::getDeclaration(
ldms->getParent()->getParent()->getParent(),
GenISAIntrinsic::GenISA_ldmsptr16bit,
types_ldms);
m_builder->SetInsertPoint(ldms);
llvm::Value* packed_tex_params_ldms[] = {
m_builder->CreateTrunc(ldms->getOperand(0), m_builder->getInt16Ty(), ""),
mcs0,
mcs1,
mcs2,
mcs3,
m_builder->CreateTrunc(ldms->getOperand(3), m_builder->getInt16Ty()),
m_builder->CreateTrunc(ldms->getOperand(4), m_builder->getInt16Ty()),
m_builder->CreateTrunc(ldms->getOperand(5), m_builder->getInt16Ty()),
m_builder->CreateTrunc(ldms->getOperand(6), m_builder->getInt16Ty()),
ldms->getOperand(7),
ldms->getOperand(8),
ldms->getOperand(9),
ldms->getOperand(10)
};
llvm::CallInst* new_ldms = m_builder->CreateCall(func_ldms, packed_tex_params_ldms);
ldms->replaceAllUsesWith(new_ldms);
}
// There are uses of ldmcs other then ldms, using vector of int32 type.
// Fix them to use newly created 16bit ldmcs.
if (ldmcs->getType()->isVectorTy() &&
cast<IGCLLVM::FixedVectorType>(ldmcs->getType())->getElementType() == m_builder->getInt32Ty())
{
m_builder->SetInsertPoint(ldmcs);
uint32_t ldmcsNumOfElements = (uint32_t)cast<IGCLLVM::FixedVectorType>(ldmcs->getType())->getNumElements();
uint32_t newLdmcsNumOfElements = (uint32_t)cast<IGCLLVM::FixedVectorType>(new_mcs_call->getType())->getNumElements();
// vec of 16bit ints to vec of 32bit ints
Type* newLdmcsVecType = IGCLLVM::FixedVectorType::get(m_builder->getInt32Ty(), newLdmcsNumOfElements);
Value* ldmcsExtendedToInt32 = m_builder->CreateSExt(new_mcs_call, newLdmcsVecType);
// if ldmcs has fewer elements than new ldmcs, extend vector
Value* ldmcsInt32CorrectlySized;
if (newLdmcsNumOfElements != ldmcsNumOfElements)
{
IGC_ASSERT(newLdmcsNumOfElements * 2 >= ldmcsNumOfElements);
SmallVector<uint32_t, 4> maskVals;
for (uint i = 0; i < ldmcsNumOfElements; i++)
{
maskVals.push_back(i);
}
auto* pMask = ConstantDataVector::get(I.getContext(), maskVals);
ldmcsInt32CorrectlySized = m_builder->CreateShuffleVector(ldmcsExtendedToInt32, UndefValue::get(ldmcsExtendedToInt32->getType()), pMask);
}
else
{
ldmcsInt32CorrectlySized = ldmcsExtendedToInt32;
}
IGC_ASSERT(ldmcsInt32CorrectlySized);
ldmcs->replaceAllUsesWith(ldmcsInt32CorrectlySized);
}
break;
}
default:
break;
}
}
}
bool ConvertMSAAPayloadTo16Bit::runOnFunction(Function& F)
{
m_pCtxWrapper = &getAnalysis<CodeGenContextWrapper>();
CodeGenContext* cgCtx = m_pCtxWrapper->getCodeGenContext();
IRBuilder<> builder(F.getContext());
m_builder = &builder;
visit(F);
DumpLLVMIR(cgCtx, "AfterMSAA16bitPayload");
return true;
}
|
