-
Notifications
You must be signed in to change notification settings - Fork 29
/
Copy pathdispatch_stub.h
337 lines (279 loc) · 18.8 KB
/
dispatch_stub.h
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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
/*
* Copyright (c) 2020-2021, Intel Corporation
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* 3. Neither the name of the Intel Corporation nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once
#include "utils.h"
#include "ProcessGroupCCL.hpp"
namespace oneccl_bindings_for_pytorch {
using namespace c10d;
constexpr typename std::underlying_type<c10::DeviceType>::type to_int(c10::DeviceType dev_type) noexcept {
return static_cast<typename std::underlying_type<c10::DeviceType>::type>(dev_type);
}
class DispatchStub {
public:
DispatchStub() {}
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allreduce(std::vector<at::Tensor>& tensors,
const AllreduceOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allreduce_coalesced(std::vector<at::Tensor>& tensors,
const AllreduceOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce(std::vector<at::Tensor>& tensors,
const ReduceOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> broadcast(std::vector<at::Tensor>& tensors,
const BroadcastOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allgather(std::vector<std::vector<at::Tensor>>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> _allgather_base(
at::Tensor& outputBuffer,
at::Tensor& inputBuffer,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allgather_into_tensor_coalesced(
std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> gather(std::vector<std::vector<at::Tensor>>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const GatherOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> scatter(std::vector<at::Tensor>& outputTensors,
std::vector<std::vector<at::Tensor>>& inputTensors,
const ScatterOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> _reduce_scatter_base(at::Tensor & outputTensor,
at::Tensor& inputTensor,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce_scatter(std::vector<at::Tensor>& outputTensors,
std::vector<std::vector<at::Tensor>>& inputTensors,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce_scatter_tensor_coalesced(
std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> alltoall_base(at::Tensor& outputTensor,
at::Tensor& inputTensor,
std::vector<int64_t>& outputSplitSizes,
std::vector<int64_t>& inputSplitSizes,
const AllToAllOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> alltoall(std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllToAllOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> send(std::vector<at::Tensor>& tensors,
int dstRank,
int tag,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> recv(std::vector<at::Tensor>& tensors,
int srcRank,
int tag,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> barrier(const BarrierOptions& opts,
ProcessGroupCCL& pg_ccl);
static c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> end_coalescing(ProcessGroupCCL& pg_ccl);
static void reset_all();
virtual ~DispatchStub() {};
static void register_ccl_stub(c10::DeviceType devoce_type, DispatchStub* stub);
static DispatchStub* get_ccl_stub(c10::DeviceType dev_type);
virtual void reset() {};
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce_(std::vector<at::Tensor>& tensors,
const ReduceOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(), "reduce");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allgather_(std::vector<std::vector<at::Tensor>>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensors[0].device().type(), "allgather");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> _allgather_base_(at::Tensor& outputTensor,
at::Tensor& inputTensor,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensor.device().type(), "_allgather_base");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allgather_into_tensor_coalesced_(std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllgatherOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensors[0].device().type(), "allgather_into_tensor_coalesced_");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> gather_(std::vector<std::vector<at::Tensor>>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const GatherOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensors[0].device().type(), "gather");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> scatter_(std::vector<at::Tensor>& outputTensors,
std::vector<std::vector<at::Tensor>>& inputTensors,
const ScatterOptions& opts,
ProcessGroupCCL& pg_ccl){
fail(outputTensors[0].device().type(), "scatter");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce_scatter_(std::vector<at::Tensor>& outputTensors,
std::vector<std::vector<at::Tensor>>& inputTensors,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(outputTensors[0].device().type(), "reduce_scatter");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> _reduce_scatter_base_(at::Tensor& outputTensor,
at::Tensor& inputTensor,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensor.device().type(), "_reduce_scatter_base");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> reduce_scatter_tensor_coalesced_(std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const ReduceScatterOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensors[0].device().type(), "reduce_scatter_tensor_coalesced_");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> broadcast_(std::vector<at::Tensor>& tensors,
const BroadcastOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(), "broadcast");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> alltoall_base_(at::Tensor& outputTensor,
at::Tensor& inputTensor,
std::vector<int64_t>& outputSplitSizes,
std::vector<int64_t>& inputSplitSizes,
const AllToAllOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensor[0].device().type(), "alltoall_base");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> alltoall_(std::vector<at::Tensor>& outputTensors,
std::vector<at::Tensor>& inputTensors,
const AllToAllOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(inputTensors[0].device().type(), "alltoall");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> barrier_(const BarrierOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(c10::DeviceType::XPU, "barrier");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allreduce_(std::vector<at::Tensor>& tensors,
const AllreduceOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(), "allreduce");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> allreduce_coalesced_(std::vector<at::Tensor>& tensors,
const AllreduceOptions& opts,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(), "allreduce_coalesced");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> send_(std::vector<at::Tensor>& tensors,
int dstRank,
int tag,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(),"send");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> recv_(std::vector<at::Tensor>& tensors,
int srcRank,
int tag,
ProcessGroupCCL& pg_ccl) {
fail(tensors[0].device().type(),"recv");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
virtual c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> end_coalescing_(ProcessGroupCCL& pg_ccl) {
TORCH_CHECK(false, "oneccl_bindings_for_pytorch: end_coalescing isn't implementd on backend [", c10::DeviceType::CPU, "].");
return c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL>();
}
private:
static void fail(c10::DeviceType dev_type, const std::string method) {
TORCH_CHECK(false, "oneccl_bindings_for_pytorch: ", method, " isn't implementd on backend [", dev_type, "].");
}
};
} // namespace oneccl_bindings_for_pytorch
namespace {
std::string reduce_op_to_string(c10d::ReduceOp op) {
switch (op) {
case c10d::ReduceOp::SUM:
return "SUM";
case c10d::ReduceOp::PRODUCT:
return "PRODUCT";
case c10d::ReduceOp::MIN:
return "MIN";
case c10d::ReduceOp::MAX:
return "MAX";
case c10d::ReduceOp::BAND:
return "BAND";
case c10d::ReduceOp::BOR:
return "BOR";
case c10d::ReduceOp::BXOR:
return "BXOR";
case c10d::ReduceOp::AVG:
return "AVG";
default:
return "UNKNOWN";
}
}
void check_supported_reduce_op(c10::DeviceType dev_type, c10d::ReduceOp op) {
if (dev_type == c10::DeviceType::XPU) {
switch (op) {
case c10d::ReduceOp::BAND:
case c10d::ReduceOp::BOR:
case c10d::ReduceOp::BXOR:
case c10d::ReduceOp::AVG:
case c10d::ReduceOp::PREMUL_SUM:
case c10d::ReduceOp::UNUSED:
TORCH_CHECK(false, ("Cannot use ReduceOp." + reduce_op_to_string(op) + " with XPU"));
default:
// No action needed for supported operations
break;
}
}
}
} // namespace