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
|
#include <torch/csrc/jit/codegen/cuda/instrumentation.h>
#include <torch/csrc/jit/codegen/cuda/kernel_expr_evaluator.h>
#include <iostream>
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
namespace kir {
namespace {
template <typename T>
c10::optional<IntOrDouble> toOptionalIntOrDouble(c10::optional<T> i) {
if (!i) {
return c10::nullopt;
}
return IntOrDouble(i.value());
}
} // namespace
void ExpressionEvaluator::bind(const Val* value, IntOrDouble concrete_value) {
TORCH_CHECK(value->isScalar());
TORCH_CHECK(
value->dtype() == DataType::Int || value->dtype() == DataType::Double);
TORCH_CHECK(!value->isConstScalar(), "Tried to bind to a constant value");
TORCH_CHECK(
value->definition() == nullptr,
"Tried to bind to a value that is computed in the kernel IR: ",
value->toInlineString(),
" with ",
concrete_value);
known_values_[value] = concrete_value;
}
void ExpressionEvaluator::bind(
ParallelType pt,
Int::ScalarType concrete_value) {
TORCH_INTERNAL_ASSERT(isParallelTypeThread(pt));
if (precomputed_values_) {
// Need to bind the thread value to integer machine
// in pre-computed mode.
precomputed_values_->bindConcreteParallelTypeValue(pt, concrete_value);
} else {
known_parallel_dimensions_[pt] = concrete_value;
}
}
c10::optional<IntOrDouble> ExpressionEvaluator::evaluate(const Val* value) {
if (precomputed_values_ && precomputed_values_->ready()) {
if (precomputed_values_->getMaybeValueFor(value).has_value()) {
return toOptionalIntOrDouble(
precomputed_values_->getMaybeValueFor(value));
}
}
if (value->isScalar() && value->isConst()) {
if (value->isADouble()) {
return toOptionalIntOrDouble(value->as<Double>()->value());
}
return toOptionalIntOrDouble(value->as<Int>()->value());
} else {
FUSER_PERF_SCOPE("kir::ExpressionEvaluator::evaluate");
TORCH_CHECK(value->isScalar(), value->toString());
TORCH_CHECK(
value->dtype() == DataType::Int || value->dtype() == DataType::Double,
value->toString());
// Is the value known (either explicit binding or memoized)?
const auto pre_eval_it = known_values_.find(value);
if (pre_eval_it != known_values_.end()) {
return pre_eval_it->second;
}
OptOutConstDispatch::handle(value);
const auto post_eval_it = known_values_.find(value);
return post_eval_it != known_values_.end()
? c10::optional<IntOrDouble>(post_eval_it->second)
: c10::nullopt;
}
return c10::nullopt;
}
bool ExpressionEvaluator::isConst(const Val* value) {
return ExpressionEvaluator().evaluate(value).has_value();
}
void ExpressionEvaluator::print() const {
std::cout << "\nEvaluation context\n";
std::cout << "--------------------\n";
for (const auto& kv : known_values_) {
std::cout << kv.first->toString() << " = " << kv.second << "\n";
}
std::cout << "\nPre-computed Values\n";
if (precomputed_values_ != nullptr) {
precomputed_values_->print();
}
std::cout << "--------------------\n\n";
}
void ExpressionEvaluator::handle(const Int* value) {
TORCH_INTERNAL_ASSERT(!value->isConst());
if (auto def = value->definition()) {
OptOutConstDispatch::handle(def);
}
}
void ExpressionEvaluator::handle(const Double* value) {
TORCH_INTERNAL_ASSERT(!value->isConst());
if (auto def = value->definition()) {
OptOutConstDispatch::handle(def);
}
}
void ExpressionEvaluator::handle(const NamedScalar* named_scalar) {
const auto& name = named_scalar->name();
for (auto pt : kParallelTypeThreads) {
auto pt_val_it = known_parallel_dimensions_.find(pt);
if (pt_val_it == known_parallel_dimensions_.end()) {
continue;
}
if (name == stringifyThreadSize(pt)) {
known_values_[named_scalar] = pt_val_it->second;
return;
}
}
}
void ExpressionEvaluator::handle(const UnaryOp* unary_op) {
const auto in = evaluate(unary_op->in());
if (in.has_value()) {
switch (unary_op->getUnaryOpType()) {
case UnaryOpType::Neg:
known_values_[unary_op->out()] = -*in;
break;
case UnaryOpType::Set:
known_values_[unary_op->out()] = *in;
break;
case UnaryOpType::Cast:
if (unary_op->out()->getDataType() == DataType::Int) {
known_values_[unary_op->out()] = in->cast<int64_t>();
} else if (unary_op->out()->getDataType() == DataType::Double) {
known_values_[unary_op->out()] = in->cast<double>();
} else {
TORCH_INTERNAL_ASSERT(false, "dtype not supported in evaluator");
}
break;
default:
TORCH_CHECK(
false,
"Unexpected operator type ",
unary_op->getUnaryOpType(),
" in ",
unary_op->toString());
}
}
}
void ExpressionEvaluator::handle(const BinaryOp* binary_op) {
using namespace IntOrDouble_functions;
const auto lhs = evaluate(binary_op->lhs());
const auto rhs = evaluate(binary_op->rhs());
if (lhs.has_value() && rhs.has_value()) {
switch (binary_op->getBinaryOpType()) {
case BinaryOpType::Add:
known_values_[binary_op->out()] = *lhs + *rhs;
break;
case BinaryOpType::Sub:
known_values_[binary_op->out()] = *lhs - *rhs;
break;
case BinaryOpType::Mul:
known_values_[binary_op->out()] = *lhs * *rhs;
break;
case BinaryOpType::Div:
TORCH_CHECK(*rhs != 0);
known_values_[binary_op->out()] = *lhs / *rhs;
break;
case BinaryOpType::Mod:
TORCH_CHECK(*rhs != 0);
known_values_[binary_op->out()] = *lhs % *rhs;
break;
case BinaryOpType::CeilDiv:
TORCH_CHECK(*rhs != 0);
known_values_[binary_op->out()] = ceildiv(*lhs, *rhs);
break;
case BinaryOpType::And:
known_values_[binary_op->out()] = Int::ScalarType(*lhs && *rhs);
break;
default:
TORCH_CHECK(!"Unexpected operator type");
}
}
}
} // namespace kir
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch
|
