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
|
#pragma once
#include <torch/csrc/jit/tensorexpr/codegen.h>
#include <torch/csrc/jit/tensorexpr/ir.h>
#include <torch/csrc/jit/tensorexpr/ir_visitor.h>
#include <torch/csrc/jit/tensorexpr/tensor.h>
namespace torch {
namespace jit {
namespace tensorexpr {
// Walk the Statment looking for Half size loads/stores.
class HalfChecker : public IRVisitor {
public:
HalfChecker(const std::vector<CodeGen::BufferArg>& args) {
for (const auto& BA : args) {
hasHalf_ |= BA.dtype().scalar_type() == ScalarType::Half;
}
}
bool hasHalf() const {
return hasHalf_;
}
bool hasBFloat16() const {
return hasBFloat16_;
}
void visit(LoadPtr v) override {
hasHalf_ |= v->dtype().scalar_type() == ScalarType::Half;
hasBFloat16_ |= v->dtype().scalar_type() == ScalarType::BFloat16;
IRVisitor::visit(v);
}
void visit(StorePtr v) override {
hasHalf_ |= v->buf()->dtype().scalar_type() == ScalarType::Half;
hasBFloat16_ |= v->buf()->dtype().scalar_type() == ScalarType::BFloat16;
IRVisitor::visit(v);
}
void visit(HalfImmPtr v) override {
hasHalf_ = true;
}
void visit(BFloat16ImmPtr v) override {
hasBFloat16_ = true;
}
void visit(CastPtr v) override {
hasHalf_ |= v->dtype().scalar_type() == ScalarType::Half;
hasBFloat16_ |= v->dtype().scalar_type() == ScalarType::BFloat16;
IRVisitor::visit(v);
}
private:
bool hasHalf_{false};
bool hasBFloat16_{false};
};
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
class HalfRewriter : public IRMutator {
ExprPtr mutate(LoadPtr v) override {
ExprPtr child = IRMutator::mutate(v);
if (!isHalf(child)) {
return child;
}
ExprPtr ret = alloc<Cast>(
child->dtype().cloneWithScalarType(ScalarType::Float), child);
inserted_half_casts_.insert(ret);
return ret;
}
StmtPtr mutate(StorePtr v) override {
// Since mutation changes the `value()` expression in-place, we need to
// get the dtype of the `value()` before that is mutated.
auto newType = v->value()->dtype();
ExprPtr new_val = v->value()->accept_mutator(this);
if (isHalf(newType.scalar_type())) {
new_val = alloc<Cast>(newType, new_val);
inserted_half_casts_.insert(new_val);
}
v->set_value(new_val);
return v;
}
ExprPtr mutate(HalfImmPtr v) override {
return alloc<Cast>(kFloat, v);
}
ExprPtr mutate(BFloat16ImmPtr v) override {
return alloc<Cast>(kFloat, v);
}
ExprPtr mutate(CastPtr v) override {
ExprPtr child = v->src_value()->accept_mutator(this);
// just don't allow half casts we didn't insert.
if (isHalf(v)) {
if (inserted_half_casts_.count(v) < 1) {
v->set_src_value(child);
v->set_dtype(v->dtype().cloneWithScalarType(c10::kFloat));
return v;
}
}
// Remove Half(Float()) and friends.
CastPtr cast_child = to<Cast>(child);
if (cast_child) {
auto cast_to_double = v->dtype().scalar_type() == ScalarType::Double;
auto from_half = isHalf(cast_child->src_value());
// Cannot simplify the double(float(half)) to double(half) as NNC does
// not support cast BF16 to double directly.
auto not_cast_half_to_doulbe = !(cast_to_double && from_half);
if (v->dtype().is_floating_point() &&
cast_child->dtype().is_floating_point() && not_cast_half_to_doulbe) {
return alloc<Cast>(v->dtype(), cast_child->src_value());
}
}
if (child == v->src_value()) {
return v;
}
return alloc<Cast>(v->dtype(), child);
}
StmtPtr mutate(LetPtr v) override {
if (isHalf(v->var()->dtype().scalar_type())) {
VarPtr load_new_var = alloc<Var>(v->var()->name_hint(), kFloat);
ExprPtr new_value = alloc<Cast>(
v->var()->dtype().cloneWithScalarType(ScalarType::Float),
v->value()->accept_mutator(this));
var_map[v->var()] = load_new_var;
return alloc<Let>(load_new_var, new_value);
}
return IRMutator::mutate(v);
}
ExprPtr mutate(VarPtr v) override {
auto it = var_map.find(v);
if (it != var_map.end()) {
return it->second;
}
return v;
}
template <typename T>
ExprPtr mutateArithmetic(T v) {
IRMutator::mutate(v);
if (isHalf(v)) {
v->set_dtype(v->dtype().cloneWithScalarType(c10::kFloat));
}
return v;
}
ExprPtr mutate(AddPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(SubPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(MulPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(DivPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(MaxPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(MinPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(CompareSelectPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(BroadcastPtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(IfThenElsePtr v) override {
return mutateArithmetic(v);
}
ExprPtr mutate(IntrinsicsPtr v) override {
return mutateArithmetic(v);
}
private:
static bool isHalf(ScalarType st) {
return st == ScalarType::Half || st == ScalarType::BFloat16;
}
static bool isHalf(ExprPtr v) {
return isHalf(v->dtype().scalar_type());
}
std::unordered_set<ExprPtr> inserted_half_casts_;
std::unordered_map<VarPtr, VarPtr> var_map;
};
} // namespace tensorexpr
} // namespace jit
} // namespace torch
|
