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
|
//===- GenerateCubinAccessors.cpp - MLIR GPU lowering passes --------------===//
//
// Copyright 2019 The MLIR Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// =============================================================================
//
// This file implements a pass to generate LLVMIR functions that return the
// data stored in nvvm.cubin char* blob.
//
//===----------------------------------------------------------------------===//
#include "mlir/GPU/GPUDialect.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/Function.h"
#include "mlir/IR/Identifier.h"
#include "mlir/IR/Module.h"
#include "mlir/IR/StandardTypes.h"
#include "mlir/LLVMIR/LLVMDialect.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Pass/PassRegistry.h"
#include "llvm/ADT/STLExtras.h"
namespace mlir {
namespace {
// TODO(herhut): Move to shared location.
constexpr const char *kCubinAnnotation = "nvvm.cubin";
constexpr const char *kCubinGetterAnnotation = "nvvm.cubingetter";
constexpr const char *kCubinGetterSuffix = "_cubin";
constexpr const char *kMallocHelperName = "mcuMalloc";
/// A pass generating getter functions for all cubin blobs annotated on
/// functions via the nvvm.cubin attribute.
///
/// The functions allocate memory using a mcuMalloc helper function with the
/// signature void *mcuMalloc(int32_t size). This function has to be provided
/// by the actual runner that executes the generated code.
///
/// This is a stop-gap measure until MLIR supports global constants.
class GpuGenerateCubinAccessorsPass
: public ModulePass<GpuGenerateCubinAccessorsPass> {
private:
Function getMallocHelper(Location loc, Builder &builder) {
Function result = getModule().getNamedFunction(kMallocHelperName);
if (!result) {
result = Function::create(
loc, kMallocHelperName,
builder.getFunctionType(
ArrayRef<Type>{LLVM::LLVMType::getInt32Ty(llvmDialect)},
LLVM::LLVMType::getInt8PtrTy(llvmDialect)));
getModule().push_back(result);
}
return result;
}
// Generates a function that returns a char array at runtime that contains the
// data from blob. As there are currently no global constants, this uses a
// sequence of store operations.
// TODO(herhut): Use global constants instead.
Function generateCubinAccessor(Builder &builder, Function &orig,
StringAttr blob) {
Location loc = orig.getLoc();
SmallString<128> nameBuffer(orig.getName());
nameBuffer.append(kCubinGetterSuffix);
// Generate a function that returns void*.
Function result = Function::create(
loc, mlir::Identifier::get(nameBuffer, &getContext()),
builder.getFunctionType(ArrayRef<Type>{},
LLVM::LLVMType::getInt8PtrTy(llvmDialect)));
// Insert a body block that just returns the constant.
OpBuilder ob(result.getBody());
ob.createBlock();
auto sizeConstant = ob.create<LLVM::ConstantOp>(
loc, LLVM::LLVMType::getInt32Ty(llvmDialect),
builder.getI32IntegerAttr(blob.getValue().size()));
auto memory =
ob.create<LLVM::CallOp>(
loc, ArrayRef<Type>{LLVM::LLVMType::getInt8PtrTy(llvmDialect)},
builder.getFunctionAttr(getMallocHelper(loc, builder)),
ArrayRef<Value *>{sizeConstant})
.getResult(0);
for (auto byte : llvm::enumerate(blob.getValue().bytes())) {
auto index = ob.create<LLVM::ConstantOp>(
loc, LLVM::LLVMType::getInt32Ty(llvmDialect),
builder.getI32IntegerAttr(byte.index()));
auto gep =
ob.create<LLVM::GEPOp>(loc, LLVM::LLVMType::getInt8PtrTy(llvmDialect),
memory, ArrayRef<Value *>{index});
auto value = ob.create<LLVM::ConstantOp>(
loc, LLVM::LLVMType::getInt8Ty(llvmDialect),
builder.getIntegerAttr(builder.getIntegerType(8), byte.value()));
ob.create<LLVM::StoreOp>(loc, value, gep);
}
ob.create<LLVM::ReturnOp>(loc, ArrayRef<Value *>{memory});
// Store the name of the getter on the function for easier lookup.
orig.setAttr(kCubinGetterAnnotation, builder.getFunctionAttr(result));
return result;
}
public:
// Run the dialect converter on the module.
void runOnModule() override {
llvmDialect =
getModule().getContext()->getRegisteredDialect<LLVM::LLVMDialect>();
auto &module = getModule();
Builder builder(&getContext());
auto functions = module.getFunctions();
for (auto it = functions.begin(); it != functions.end();) {
// Move iterator to after the current function so that potential insertion
// of the accessor is after the kernel with cubin iself.
Function orig = *it++;
StringAttr cubinBlob = orig.getAttrOfType<StringAttr>(kCubinAnnotation);
if (!cubinBlob)
continue;
module.insert(it, generateCubinAccessor(builder, orig, cubinBlob));
}
}
private:
LLVM::LLVMDialect *llvmDialect;
};
} // anonymous namespace
ModulePassBase *createGenerateCubinAccessorPass() {
return new GpuGenerateCubinAccessorsPass();
}
static PassRegistration<GpuGenerateCubinAccessorsPass>
pass("generate-cubin-accessors",
"Generate LLVMIR functions that give access to cubin data");
} // namespace mlir
|
