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
|
#include <torch/csrc/jit/codegen/cuda/ir_utils.h>
#include <torch/csrc/jit/codegen/cuda/lower2device.h>
#include <torch/csrc/jit/codegen/cuda/contiguity.h>
namespace torch {
namespace jit {
namespace fuser {
namespace cuda {
ContigIDs::ContigIDs(
const std::vector<IterDomain*>& ids,
const std::vector<IterDomain*>& root_domain,
const std::vector<bool>& root_contiguity,
std::unordered_map<IterDomain*, IterDomain*> concrete_to_ref,
std::unordered_map<IterDomain*, IterDomain*> p2c_id_map,
bool ignore_halo_constraint,
bool ignore_indexability)
: root_domain_(root_domain),
root_contiguity_(root_contiguity),
concrete_to_ref_(std::move(concrete_to_ref)),
p2c_id_map_(std::move(p2c_id_map)),
ignore_indexability_(ignore_indexability) {
if (ids.empty()) {
return;
}
TORCH_INTERNAL_ASSERT(
root_domain_.size() == root_contiguity_.size(),
"Arguments don't match ",
root_domain_.size(),
" != ",
root_contiguity_.size());
// GpuLower is required to honor halo constraints
if (!ignore_halo_constraint) {
TORCH_INTERNAL_ASSERT(GpuLower::hasCurrent(), "GpuLower not found");
}
for (const auto i : c10::irange(root_domain_.size())) {
auto root_domain_i = root_domain_[i]->as<IterDomain>();
root_to_indexed_id_[root_domain_i] = root_domain_i;
// Initialize to false
is_contig_root_[root_domain_i] = false;
// If a root domain has halo, can't use merged domain even if
// both inputs are contiguous. HaloInfo is also initialized for
// rfactor root domains, which should just return "zero"
// RootAxisInfo. This should be safe as no rfactor tensor should
// need halo.
if (root_contiguity_[i] &&
(ignore_halo_constraint ||
!GpuLower::current()
->haloInfo()
.getRootAxisInfo(root_domain_i)
.hasHalo())) {
contig_ids_.emplace(root_domain_i);
is_contig_root_[root_domain_i] = true;
within_contig_ids_[root_domain_i] = std::unordered_set<IterDomain*>();
}
}
if (!contig_ids_.empty()) {
auto exprs = StmtSort::getExprs(ids[0]->fusion(), {ids.begin(), ids.end()});
for (auto expr : exprs) {
handle(expr);
}
}
}
void ContigIDs::handle(Merge* merge) {
// If either input is non-contiguous so is output.
const auto inner = merge->inner();
const auto outer = merge->outer();
const auto out = merge->out();
if (!isContig(inner) || !isContig(outer)) {
return;
}
// Stop contig merging if the merge output is not indexable.
if (!ignore_indexability_ && !isIndexable(out)) {
return;
}
// Grab inputs, make sure they're in root domain, check if they're
// contiguous.
auto lhs_inputs =
ir_utils::iterDomainInputsOfOrderedAs({outer}, root_domain_);
auto rhs_inputs =
ir_utils::iterDomainInputsOfOrderedAs({inner}, root_domain_);
TORCH_INTERNAL_ASSERT(
inRoot(lhs_inputs) && inRoot(rhs_inputs),
"Found an invalid merge operation, inputs of its arguments are not in the root domain.");
std::deque<IterDomain*> ordered_inputs(lhs_inputs.begin(), lhs_inputs.end());
ordered_inputs.insert(
ordered_inputs.end(), rhs_inputs.begin(), rhs_inputs.end());
// If any root input is not contig, output is not contig
if (!(std::all_of(
ordered_inputs.begin(), ordered_inputs.end(), [this](IterDomain* id) {
// Allow reduction tensors in contiguity check since we're using
// this to check contiguous vectors of reference tensors in
// schedulers (to set vectorization sizes), those reference tensors
// may have reduction dims, don't bail on contiguity just because
// it's a reduction dimension.
return is_contig_root_.at(id);
}))) {
return;
}
std::deque<IterDomain*> root_copy(root_domain_.begin(), root_domain_.end());
// Forward to first matching argument
while (!root_copy.empty() && !ordered_inputs.empty()) {
if (root_copy.front() != ordered_inputs.front()) {
root_copy.pop_front();
} else {
break;
}
}
// Forward through all matching arguments
while (!root_copy.empty() && !ordered_inputs.empty()) {
if (root_copy.front() == ordered_inputs.front()) {
root_copy.pop_front();
ordered_inputs.pop_front();
} else if (
root_copy.front()->isReduction() || root_copy.front()->isBroadcast()) {
// This was a cause of an error with
// ReductionSchedulerMultiDimNonFastest. The test no longer
// fails.
root_copy.pop_front();
} else {
break;
}
}
// If we matched all inputs, the output is contiguous. Only want to keep the
// top contig ID, lower ids should be placed in the "within_contig_ids" map
// of top id.
if (ordered_inputs.empty()) {
if (contig_ids_.find(inner) != contig_ids_.end()) {
contig_ids_.erase(inner);
}
if (contig_ids_.find(outer) != contig_ids_.end()) {
contig_ids_.erase(outer);
}
contig_ids_.emplace(out);
std::unordered_set<IterDomain*> within_out;
within_out.emplace(inner);
if (within_contig_ids_.find(inner) != within_contig_ids_.end()) {
auto in_inner = within_contig_ids_.at(inner);
within_out.insert(in_inner.begin(), in_inner.end());
within_contig_ids_.erase(inner);
}
within_out.emplace(outer);
if (within_contig_ids_.find(outer) != within_contig_ids_.end()) {
auto in_outer = within_contig_ids_.at(outer);
within_out.insert(in_outer.begin(), in_outer.end());
within_contig_ids_.erase(outer);
}
within_contig_ids_[out] = within_out;
for (auto root : lhs_inputs) {
root_to_indexed_id_[root] = out;
}
for (auto root : rhs_inputs) {
root_to_indexed_id_[root] = out;
}
}
}
IterDomain* ContigIDs::getMappedId(IterDomain* id) const {
auto it = p2c_id_map_.find(id);
if (it != p2c_id_map_.end()) {
return it->second;
} else {
return id;
}
}
IterDomain* ContigIDs::getCAIndexConcreteId(IterDomain* id) const {
TORCH_INTERNAL_ASSERT(
GpuLower::current() != nullptr, "GpuLower is not found");
auto c_id = GpuLower::current()->caMap()->getConcreteMappedID(
getMappedId(id), IdMappingMode::EXACT);
return c_id;
}
bool ContigIDs::isIndexable(IterDomain* id) const {
// If ID is mapped to consumer through persmissive map but not exact map it
// will not be mapped through to the exact map through the p2c map. Therefore
// reject because it involves broadcast resolution.
if (!GpuLower::current()->caMap()->idExistsInMap(getMappedId(id))) {
return false;
}
auto c_id = getCAIndexConcreteId(id);
return concrete_to_ref_.find(c_id) != concrete_to_ref_.end();
}
} // namespace cuda
} // namespace fuser
} // namespace jit
} // namespace torch
|
