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
|
//===-- CUFAddConstructor.cpp ---------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#include "flang/Optimizer/Builder/BoxValue.h"
#include "flang/Optimizer/Builder/CUFCommon.h"
#include "flang/Optimizer/Builder/FIRBuilder.h"
#include "flang/Optimizer/Builder/Runtime/RTBuilder.h"
#include "flang/Optimizer/Builder/Todo.h"
#include "flang/Optimizer/CodeGen/Target.h"
#include "flang/Optimizer/CodeGen/TypeConverter.h"
#include "flang/Optimizer/Dialect/CUF/CUFOps.h"
#include "flang/Optimizer/Dialect/FIRAttr.h"
#include "flang/Optimizer/Dialect/FIRDialect.h"
#include "flang/Optimizer/Dialect/FIROps.h"
#include "flang/Optimizer/Dialect/FIROpsSupport.h"
#include "flang/Optimizer/Dialect/FIRType.h"
#include "flang/Optimizer/Support/DataLayout.h"
#include "flang/Runtime/CUDA/registration.h"
#include "flang/Runtime/entry-names.h"
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
#include "mlir/IR/Value.h"
#include "mlir/Pass/Pass.h"
#include "llvm/ADT/SmallVector.h"
namespace fir {
#define GEN_PASS_DEF_CUFADDCONSTRUCTOR
#include "flang/Optimizer/Transforms/Passes.h.inc"
} // namespace fir
using namespace Fortran::runtime::cuda;
namespace {
static constexpr llvm::StringRef cudaFortranCtorName{
"__cudaFortranConstructor"};
struct CUFAddConstructor
: public fir::impl::CUFAddConstructorBase<CUFAddConstructor> {
void runOnOperation() override {
mlir::ModuleOp mod = getOperation();
mlir::SymbolTable symTab(mod);
mlir::OpBuilder opBuilder{mod.getBodyRegion()};
fir::FirOpBuilder builder(opBuilder, mod);
fir::KindMapping kindMap{fir::getKindMapping(mod)};
builder.setInsertionPointToEnd(mod.getBody());
mlir::Location loc = mod.getLoc();
auto *ctx = mod.getContext();
auto voidTy = mlir::LLVM::LLVMVoidType::get(ctx);
auto idxTy = builder.getIndexType();
auto funcTy =
mlir::LLVM::LLVMFunctionType::get(voidTy, {}, /*isVarArg=*/false);
std::optional<mlir::DataLayout> dl =
fir::support::getOrSetDataLayout(mod, /*allowDefaultLayout=*/false);
if (!dl) {
mlir::emitError(mod.getLoc(),
"data layout attribute is required to perform " +
getName() + "pass");
}
// Symbol reference to CUFRegisterAllocator.
builder.setInsertionPointToEnd(mod.getBody());
auto registerFuncOp = builder.create<mlir::LLVM::LLVMFuncOp>(
loc, RTNAME_STRING(CUFRegisterAllocator), funcTy);
registerFuncOp.setVisibility(mlir::SymbolTable::Visibility::Private);
auto cufRegisterAllocatorRef = mlir::SymbolRefAttr::get(
mod.getContext(), RTNAME_STRING(CUFRegisterAllocator));
builder.setInsertionPointToEnd(mod.getBody());
// Create the constructor function that call CUFRegisterAllocator.
auto func = builder.create<mlir::LLVM::LLVMFuncOp>(loc, cudaFortranCtorName,
funcTy);
func.setLinkage(mlir::LLVM::Linkage::Internal);
builder.setInsertionPointToStart(func.addEntryBlock(builder));
builder.create<mlir::LLVM::CallOp>(loc, funcTy, cufRegisterAllocatorRef);
auto gpuMod = symTab.lookup<mlir::gpu::GPUModuleOp>(cudaDeviceModuleName);
if (gpuMod) {
auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(ctx);
auto registeredMod = builder.create<cuf::RegisterModuleOp>(
loc, llvmPtrTy, mlir::SymbolRefAttr::get(ctx, gpuMod.getName()));
fir::LLVMTypeConverter typeConverter(mod, /*applyTBAA=*/false,
/*forceUnifiedTBAATree=*/false, *dl);
// Register kernels
for (auto func : gpuMod.getOps<mlir::gpu::GPUFuncOp>()) {
if (func.isKernel()) {
auto kernelName = mlir::SymbolRefAttr::get(
builder.getStringAttr(cudaDeviceModuleName),
{mlir::SymbolRefAttr::get(builder.getContext(), func.getName())});
builder.create<cuf::RegisterKernelOp>(loc, kernelName, registeredMod);
}
}
// Register variables
for (fir::GlobalOp globalOp : mod.getOps<fir::GlobalOp>()) {
auto attr = globalOp.getDataAttrAttr();
if (!attr)
continue;
mlir::func::FuncOp func;
switch (attr.getValue()) {
case cuf::DataAttribute::Device:
case cuf::DataAttribute::Constant: {
func = fir::runtime::getRuntimeFunc<mkRTKey(CUFRegisterVariable)>(
loc, builder);
auto fTy = func.getFunctionType();
// Global variable name
std::string gblNameStr = globalOp.getSymbol().getValue().str();
gblNameStr += '\0';
mlir::Value gblName = fir::getBase(
fir::factory::createStringLiteral(builder, loc, gblNameStr));
// Global variable size
std::optional<uint64_t> size;
if (auto boxTy =
mlir::dyn_cast<fir::BaseBoxType>(globalOp.getType())) {
mlir::Type structTy = typeConverter.convertBoxTypeAsStruct(boxTy);
size = dl->getTypeSizeInBits(structTy) / 8;
}
if (!size) {
size = fir::getTypeSizeAndAlignmentOrCrash(loc, globalOp.getType(),
*dl, kindMap)
.first;
}
auto sizeVal = builder.createIntegerConstant(loc, idxTy, *size);
// Global variable address
mlir::Value addr = builder.create<fir::AddrOfOp>(
loc, globalOp.resultType(), globalOp.getSymbol());
llvm::SmallVector<mlir::Value> args{fir::runtime::createArguments(
builder, loc, fTy, registeredMod, addr, gblName, sizeVal)};
builder.create<fir::CallOp>(loc, func, args);
} break;
case cuf::DataAttribute::Managed:
TODO(loc, "registration of managed variables");
default:
break;
}
}
}
builder.create<mlir::LLVM::ReturnOp>(loc, mlir::ValueRange{});
// Create the llvm.global_ctor with the function.
// TODO: We might want to have a utility that retrieve it if already
// created and adds new functions.
builder.setInsertionPointToEnd(mod.getBody());
llvm::SmallVector<mlir::Attribute> funcs;
funcs.push_back(
mlir::FlatSymbolRefAttr::get(mod.getContext(), func.getSymName()));
llvm::SmallVector<int> priorities;
priorities.push_back(0);
builder.create<mlir::LLVM::GlobalCtorsOp>(
mod.getLoc(), builder.getArrayAttr(funcs),
builder.getI32ArrayAttr(priorities));
}
};
} // end anonymous namespace
|
