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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
|
#include "slang-ir-pytorch-cpp-binding.h"
#include "slang-ir.h"
#include "slang-ir-insts.h"
#include "slang-diagnostics.h"
namespace Slang
{
static bool getHostReturnTypeImpl(List<IRType*>& elementTypes, IRBuilder& builder, IRType* type)
{
bool isValid = true;
if (as<IRVoidType>(type))
return true;
if (as<IRBasicType>(type))
elementTypes.add(type);
else if (as<IRTorchTensorType>(type))
elementTypes.add(type);
else if (auto vectorType = as<IRVectorType>(type))
{
auto count = as<IRIntLit>(vectorType->getElementCount());
if (!count)
{
return false;
}
for (IRIntegerValue i = 0; i < count->getValue(); i++)
{
elementTypes.addRange(vectorType->getElementType());
}
}
else if (auto arrayType = as<IRArrayType>(type))
{
auto arraySize = as<IRIntLit>(arrayType->getElementCount());
if (!arraySize)
{
return false;
}
List<IRType*> subElementTypes;
isValid &= getHostReturnTypeImpl(subElementTypes, builder, arrayType->getElementType());
for (IRIntegerValue i = 0; i < arraySize->getValue(); i++)
{
elementTypes.addRange(subElementTypes);
}
}
else if (auto structType = as<IRStructType>(type))
{
for (auto field : structType->getFields())
{
isValid &= getHostReturnTypeImpl(elementTypes, builder, field->getFieldType());
}
}
else
{
return false;
}
return isValid;
}
static IRType* getHostReturnType(IRBuilder& builder, IRType* type)
{
List<IRType*> types;
bool isValid = getHostReturnTypeImpl(types, builder, type);
if (isValid)
return builder.getTargetTupleType((UInt)types.getCount(), types.getBuffer());
return nullptr;
}
static void flattenToTupleImpl(List<IRInst*>& result, IRBuilder& builder, IRInst* val)
{
auto type = val->getDataType();
if (as<IRVoidType>(type))
return;
if (as<IRBasicType>(type))
result.add(val);
else if (as<IRTorchTensorType>(type))
result.add(val);
else if (auto vectorType = as<IRVectorType>(type))
{
auto count = as<IRIntLit>(vectorType->getElementCount());
if (!count)
{
return;
}
for (IRIntegerValue i = 0; i < count->getValue(); i++)
{
result.add(builder.emitElementExtract(vectorType->getElementType(), builder.getIntValue(builder.getIntType(), i)));
}
}
else if (auto arrayType = as<IRArrayType>(type))
{
auto arraySize = as<IRIntLit>(arrayType->getElementCount());
if (!arraySize)
{
return;
}
for (IRIntegerValue i = 0; i < arraySize->getValue(); i++)
{
auto elementVal = builder.emitElementExtract(val, builder.getIntValue(builder.getIntType(), i));
flattenToTupleImpl(result, builder, elementVal);
}
}
else if (auto structType = as<IRStructType>(type))
{
for (auto field : structType->getFields())
{
auto elementVal = builder.emitFieldExtract(field->getFieldType(), val, field->getKey());
flattenToTupleImpl(result, builder, elementVal);
}
}
}
static IRInst* flattenToHostReturnTuple(IRBuilder& builder, IRType* type, IRInst* val)
{
List<IRInst*> vals;
flattenToTupleImpl(vals, builder, val);
return builder.emitMakeTargetTuple(type, (UInt)vals.getCount(), vals.getBuffer());
}
static void generateCppBindingForFunc(IRFunc* func, DiagnosticSink* sink)
{
IRBuilder builder(func);
builder.setInsertBefore(func);
auto hostReturnType = getHostReturnType(builder, func->getResultType());
if (!hostReturnType)
{
sink->diagnose(func->sourceLoc, Diagnostics::invalidTorchKernelReturnType, func->getResultType());
return;
}
List<IRType*> hostParamTypes;
auto funcType = as<IRFuncType>(func->getDataType());
for (UInt i = 0; i < funcType->getParamCount(); i++)
{
hostParamTypes.add(funcType->getParamType(i));
}
auto bindingFuncType = builder.getFuncType(hostParamTypes, hostReturnType);
func->setFullType(bindingFuncType);
builder.setInsertBefore(func->getFirstBlock()->getFirstOrdinaryInst());
auto allocator = builder.emitVar(builder.getType(kIROp_TorchKernelMemoryAllocatorType));
List<IRInst*> instsToRemove;
for (auto block : func->getBlocks())
{
for (auto inst : block->getChildren())
{
if (auto kernelDispatch = as<IRDispatchKernel>(inst))
{
builder.setInsertBefore(kernelDispatch);
List<IRInst*> kernelArgs;
auto kernelArgCount = kernelDispatch->getArgCount();
auto argArrayType = builder.getArrayType(builder.getPtrType(builder.getVoidType()),
builder.getIntValue(builder.getIntType(), kernelArgCount));
auto argArrayVar = builder.emitVar(argArrayType);
for (UInt i = 0; i < kernelArgCount; i++)
{
auto arg = kernelDispatch->getArg(i);
auto argVar = builder.emitVar(arg->getFullType());
builder.emitStore(argVar, arg);
auto addr = builder.emitElementAddress(argArrayVar, builder.getIntValue(builder.getIntType(), i));
builder.emitStore(addr, argVar);
}
auto argArrayPtr = builder.emitElementAddress(argArrayVar, builder.getIntValue(builder.getIntType(), 0));
builder.emitCudaKernelLaunch(
kernelDispatch->getBaseFn(),
kernelDispatch->getDispatchSize(),
kernelDispatch->getThreadGroupSize(),
argArrayPtr,
builder.emitGetTorchCudaStream());
instsToRemove.add(inst);
}
else if (auto getView = as<IRTorchTensorGetView>(inst))
{
builder.setInsertBefore(getView);
auto makeView = builder.emitMakeTensorView(getView->getFullType(), allocator, inst->getOperand(0));
getView->replaceUsesWith(makeView);
instsToRemove.add(getView);
}
else if (auto ret = as<IRReturn>(inst))
{
builder.setInsertBefore(ret);
auto retVal = flattenToHostReturnTuple(builder, hostReturnType, ret->getVal());
ret->setOperand(0, retVal);
}
}
}
for (auto inst : instsToRemove)
inst->removeAndDeallocate();
}
void generatePyTorchCppBinding(IRModule* module, DiagnosticSink* sink)
{
List<IRFunc*> workList;
List<IRFunc*> cudaKernels;
for (auto globalInst : module->getGlobalInsts())
{
auto func = as<IRFunc>(globalInst);
if (!func)
continue;
if (func->findDecoration<IRTorchEntryPointDecoration>())
{
workList.add(func);
}
else if (func->findDecoration<IRCudaKernelDecoration>())
{
cudaKernels.add(func);
}
else
{
// Remove all other export decorations if this is not a cuda host func.
if (auto decor = func->findDecoration<IRPublicDecoration>())
decor->removeAndDeallocate();
if (auto decor = func->findDecoration<IRHLSLExportDecoration>())
decor->removeAndDeallocate();
if (auto decor = func->findDecoration<IRKeepAliveDecoration>())
decor->removeAndDeallocate();
if (auto decor = func->findDecoration<IRDllExportDecoration>())
decor->removeAndDeallocate();
}
}
for (auto func : workList)
generateCppBindingForFunc(func, sink);
for (auto func : cudaKernels)
{
for (auto block = func->getFirstBlock(); block;)
{
auto nextBlock = block->getNextBlock();
block->removeAndDeallocate();
block = nextBlock;
}
}
}
// Remove all [TorchEntryPoint] functions when emitting CUDA source.
void removeTorchKernels(IRModule* module)
{
for (auto globalInst : module->getGlobalInsts())
{
if (!as<IRFunc>(globalInst))
continue;
if (globalInst->findDecoration<IRTorchEntryPointDecoration>())
globalInst->removeAndDeallocate();
}
}
}
|
