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
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
|
#include <torch/csrc/jit/tensorexpr/tensor.h>
#include <c10/util/Logging.h>
#include <c10/util/irange.h>
#include <torch/csrc/jit/tensorexpr/reduction.h>
namespace torch {
namespace jit {
namespace tensorexpr {
StmtPtr Tensor::constructStmt(
const std::vector<VarPtr>& args,
ExprPtr body,
const std::vector<ExprPtr>& reduce_dims,
const std::vector<VarPtr>& reduce_args) const {
std::vector<ExprPtr> indices(args.begin(), args.end());
size_t ndim = buf()->ndim();
size_t reduce_ndim = reduce_dims.size();
auto reduce_op = to<ReduceOp>(body);
auto acc_buf = reduce_ndim > 0 ? reduce_op->getAccBuf() : nullptr;
StmtPtr s = alloc<Store>(buf_, indices, body);
if (reduce_ndim > 0) {
TORCH_INTERNAL_ASSERT(reduce_op != nullptr);
if (acc_buf != nullptr) {
auto reducer = reduce_op->reducer();
std::vector<ExprPtr> output_args(args.begin(), args.end());
ExprPtr new_reduce_op = reducer(
to<Buf>(acc_buf),
alloc<Cast>(acc_buf->dtype(), reduce_op->getRiOperand()),
output_args,
reduce_args);
new_reduce_op->set_dtype(acc_buf->dtype());
s = alloc<Store>(to<Buf>(acc_buf), indices, new_reduce_op);
}
}
if (ndim == 0 && reduce_ndim == 0) {
return s;
}
if (reduce_ndim > 0) {
TORCH_INTERNAL_ASSERT(reduce_op != nullptr);
for (const auto i : c10::irange(reduce_ndim)) {
// Going in reverse order: from innermost loop to the outermost
size_t dim_index = reduce_ndim - i - 1;
auto const& dim = reduce_dims[dim_index];
s = alloc<For>(reduce_args[dim_index], immLike(dim, 0), dim, s);
}
s = alloc<Block>(std::vector<StmtPtr>({s}));
BufPtr init_buf = acc_buf ? to<Buf>(acc_buf) : buf();
ExprPtr init_expr =
acc_buf ? to<Buf>(acc_buf)->initializer() : buf()->initializer();
if (init_expr) {
StorePtr init_stmt = alloc<Store>(init_buf, indices, init_expr);
to<Block>(s)->prepend_stmt(init_stmt);
}
if (acc_buf != nullptr) {
LoadPtr load_acc = alloc<Load>(acc_buf, indices);
auto cast = alloc<Cast>(buf()->dtype(), load_acc);
StorePtr post_stmt = alloc<Store>(buf(), indices, cast);
to<Block>(s)->append_stmt(post_stmt);
}
}
TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
buf_->is_contiguous() ||
buf_->is_contiguous(at::MemoryFormat::ChannelsLast) ||
buf_->is_contiguous(at::MemoryFormat::ChannelsLast3d) ||
buf_->is_channels_last_1d_contiguous());
auto loop_order_fn = [&]() {
std::vector<int32_t> loop_order;
if (buf_->is_contiguous()) {
for (int32_t i = args.size() - 1; i >= 0; i--) {
loop_order.push_back(i);
}
} else if (buf_->is_contiguous(c10::MemoryFormat::ChannelsLast)) {
loop_order = {1, 3, 2, 0};
} else if (buf_->is_contiguous(c10::MemoryFormat::ChannelsLast3d)) {
loop_order = {1, 4, 3, 2, 0};
} else {
loop_order = {1, 2, 0};
}
return loop_order;
};
auto loop_order = loop_order_fn();
for (auto dim_index : loop_order) {
auto const& dim = buf()->dim(dim_index);
s = alloc<For>(args[dim_index], immLike(dim, 0), dim, s);
}
return s;
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const std::vector<VarHandle>&)>& body_func) {
std::vector<VarHandle> args = create_index_vars(dims);
ExprHandle body = body_func(args);
BufHandle buf = Buf::make(name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(buf, args, body);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const std::vector<VarHandle>&)>& body_func) {
return Compute(name, dims, c10::nullopt, body_func);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const VarHandle&)>& body_func) {
if (dims.size() != 1) {
throw malformed_input("mismatch between body and arg size (1)");
}
std::vector<VarHandle> args = create_index_vars(dims);
ExprHandle body = body_func(args[0]);
BufHandle buf = Buf::make(name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(buf, args, body);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const VarHandle&)>& body_func) {
return Compute(name, dims, c10::nullopt, body_func);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(const VarHandle&, const VarHandle&)>&
body_func) {
if (dims.size() != 2) {
throw malformed_input("mismatch between body and arg size (2)");
}
std::vector<VarHandle> args = create_index_vars(dims);
ExprHandle body = body_func(args[0], args[1]);
BufHandle buf = Buf::make(name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(buf, args, body);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(const VarHandle&, const VarHandle&)>&
body_func) {
return Compute(name, dims, c10::nullopt, body_func);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<
ExprHandle(const VarHandle&, const VarHandle&, const VarHandle&)>&
body_func) {
if (dims.size() != 3) {
throw malformed_input("mismatch between body and arg size (3)");
}
std::vector<VarHandle> args = create_index_vars(dims);
ExprHandle body = body_func(args[0], args[1], args[2]);
BufHandle buf = Buf::make(name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(buf, args, body);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
const std::function<
ExprHandle(const VarHandle&, const VarHandle&, const VarHandle&)>&
body_func) {
return Compute(name, dims, c10::nullopt, body_func);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const std::function<ExprHandle(
const VarHandle&,
const VarHandle&,
const VarHandle&,
const VarHandle&)>& body_func) {
if (dims.size() != 4) {
throw malformed_input("mismatch between body and arg size (4)");
}
std::vector<VarHandle> args = create_index_vars(dims);
ExprHandle body = body_func(args[0], args[1], args[2], args[3]);
BufHandle buf = Buf::make(name, dims, body.dtype(), c10::nullopt, strides);
return Tensor(buf, args, body);
}
Tensor Compute(
const std::string& name,
const std::vector<ExprHandle>& dims,
const std::function<ExprHandle(
const VarHandle&,
const VarHandle&,
const VarHandle&,
const VarHandle&)>& body_func) {
return Compute(name, dims, c10::nullopt, body_func);
}
Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
const BufHandle& buffer,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(
name,
dims,
strides,
reducer,
[&](ParameterList& p) { return buffer.load(p); },
reduce_dims);
}
Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
const BufHandle& buffer,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(name, dims, c10::nullopt, reducer, buffer, reduce_dims);
}
Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
c10::optional<std::vector<ExprHandle>> strides,
const Reducer& reducer,
Tensor tensor,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(
name,
dims,
strides,
reducer,
[&](ParameterList& p) { return tensor.load(p); },
reduce_dims);
}
Tensor Reduce(
const std::string& name,
const std::vector<ExprHandle>& dims,
const Reducer& reducer,
Tensor tensor,
const std::vector<ExprHandle>& reduce_dims) {
return Reduce(name, dims, c10::nullopt, reducer, tensor, reduce_dims);
}
} // namespace tensorexpr
} // namespace jit
} // namespace torch
|
