-
Notifications
You must be signed in to change notification settings - Fork 3.5k
/
thrust.cu
267 lines (243 loc) · 11.4 KB
/
thrust.cu
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
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/
/*!
* \file Use external Thrust library call
*/
#include <thrust/device_ptr.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <thrust/gather.h>
#include <tvm/runtime/registry.h>
#include <dlpack/dlpack.h>
#include <algorithm>
#include <vector>
#include <functional>
namespace tvm {
namespace contrib {
using namespace runtime;
// Performs sorting along axis -1 and returns both sorted values and indices.
template<typename DataType, typename IndicesType>
void thrust_sort(DLTensor* input,
DLTensor* out_values,
DLTensor* out_indices,
bool is_ascend,
int n_values) {
thrust::device_ptr<DataType> data_ptr(static_cast<DataType *>(input->data));
thrust::device_ptr<DataType> values_ptr(static_cast<DataType *>(out_values->data));
thrust::device_ptr<IndicesType> indices_ptr(static_cast<IndicesType *>(out_indices->data));
size_t size = 1;
for (int i = 0; i < input->ndim; ++i) {
size *= input->shape[i];
}
thrust::copy(data_ptr, data_ptr + size, values_ptr);
if (input->ndim == 1 || (input->ndim == 2 && input->shape[0] == 1)) {
thrust::sequence(indices_ptr, indices_ptr + n_values);
if (is_ascend) {
thrust::sort_by_key(values_ptr, values_ptr + n_values, indices_ptr);
} else {
thrust::sort_by_key(values_ptr, values_ptr + n_values, indices_ptr,
thrust::greater<DataType>());
}
} else {
// segmented sort by key
// Follow the back-to-back stable_sort_by_key strategy explained below
// https://groups.google.com/g/thrust-users/c/BoLsxO6b4FY
thrust::device_vector<int64_t> argsort_order(size);
thrust::sequence(argsort_order.begin(), argsort_order.end());
// First, sort values and store the sorted order in argsort_order.
if (is_ascend) {
thrust::stable_sort_by_key(values_ptr, values_ptr + size, argsort_order.begin());
} else {
thrust::stable_sort_by_key(values_ptr, values_ptr + size, argsort_order.begin(),
thrust::greater<DataType>());
}
// The following is to create the indices array 0, 1, 2, 0, 1, 2 ... 0, 1, 2
// without materializing it
auto counting_iter = thrust::counting_iterator<int64_t>(0);
auto linear_index_to_sort_axis_index = [n_values] __host__ __device__(int64_t i) {
return i % n_values;
}; // NOLINT(*)
auto init_indices_iter = thrust::make_transform_iterator(counting_iter,
linear_index_to_sort_axis_index);
// This will reorder indices 0, 1, 2 ... in the sorted order of values_ptr
thrust::gather(argsort_order.begin(), argsort_order.end(), init_indices_iter, indices_ptr);
thrust::device_vector<int> segment_ids(size);
auto linear_index_to_segment_id = [n_values] __host__ __device__(int64_t i) {
return i / n_values;
}; // NOLINT(*)
// We also reorder segment indices 0, 0, 0, 1, 1, 1 ... in the order of values_ptr
thrust::transform(argsort_order.begin(), argsort_order.end(), segment_ids.begin(),
linear_index_to_segment_id);
// The second sort key-ed by segment_ids would bring segment_ids back to 0, 0, 0, 1, 1, 1 ...
// values_ptr and indices_ptr will also be sorted in the order of segmend_ids above
// Since sorting has been done in a stable way, relative orderings of values and indices
// in the segment do not change and hence they remain sorted.
auto key_val_zip = thrust::make_zip_iterator(thrust::make_tuple(values_ptr, indices_ptr));
thrust::stable_sort_by_key(segment_ids.begin(), segment_ids.end(), key_val_zip);
}
}
void thrust_sort_common(DLTensor* input,
DLTensor* values_out,
DLTensor* indices_out,
bool is_ascend,
int sort_len,
std::string data_dtype,
std::string out_dtype) {
if (data_dtype == "float32") {
if (out_dtype == "int32") {
thrust_sort<float, int32_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "int64") {
thrust_sort<float, int64_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float32") {
thrust_sort<float, float>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float64") {
thrust_sort<float, double>(input, values_out, indices_out, is_ascend, sort_len);
} else {
LOG(FATAL) << "Unsupported output dtype: " << out_dtype;
}
} else if (data_dtype == "float64") {
if (out_dtype == "int32") {
thrust_sort<double, int32_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "int64") {
thrust_sort<double, int64_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float32") {
thrust_sort<double, float>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float64") {
thrust_sort<double, double>(input, values_out, indices_out, is_ascend, sort_len);
} else {
LOG(FATAL) << "Unsupported output dtype: " << out_dtype;
}
} else if (data_dtype == "int32") {
if (out_dtype == "int32") {
thrust_sort<int32_t, int32_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "int64") {
thrust_sort<int32_t, int64_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float32") {
thrust_sort<int32_t, float>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float64") {
thrust_sort<int32_t, double>(input, values_out, indices_out, is_ascend, sort_len);
} else {
LOG(FATAL) << "Unsupported output dtype: " << out_dtype;
}
} else if (data_dtype == "int64") {
if (out_dtype == "int32") {
thrust_sort<int64_t, int32_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "int64") {
thrust_sort<int64_t, int64_t>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float32") {
thrust_sort<int64_t, float>(input, values_out, indices_out, is_ascend, sort_len);
} else if (out_dtype == "float64") {
thrust_sort<int64_t, double>(input, values_out, indices_out, is_ascend, sort_len);
} else {
LOG(FATAL) << "Unsupported output dtype: " << out_dtype;
}
} else {
LOG(FATAL) << "Unsupported input dtype: " << data_dtype;
}
}
TVM_REGISTER_GLOBAL("tvm.contrib.thrust.sort")
.set_body([](TVMArgs args, TVMRetValue* ret) {
ICHECK_GE(args.num_args, 4);
DLTensor* input = args[0];
DLTensor* values_out = args[1];
DLTensor* indices_out = args[2];
bool is_ascend = args[3];
auto data_dtype = DLDataType2String(input->dtype);
auto out_dtype = DLDataType2String(indices_out->dtype);
int n_values = input->shape[input->ndim - 1];
thrust_sort_common(input, values_out, indices_out, is_ascend, n_values,
data_dtype, out_dtype);
});
template<typename KeyType, typename ValueType>
void thrust_stable_sort_by_key(DLTensor* keys_in,
DLTensor* values_in,
DLTensor* keys_out,
DLTensor* values_out,
bool for_scatter) {
const auto size = keys_in->shape[0];
thrust::device_ptr<KeyType> keys_in_ptr(static_cast<KeyType *>(keys_in->data));
thrust::device_ptr<ValueType> values_in_ptr(static_cast<ValueType *>(values_in->data));
thrust::device_ptr<KeyType> keys_out_ptr(static_cast<KeyType *>(keys_out->data));
thrust::device_ptr<ValueType> values_out_ptr(static_cast<ValueType *>(values_out->data));
if (for_scatter) {
thrust::transform(keys_in_ptr, keys_in_ptr + size, keys_out_ptr, [size] __device__(KeyType k) {
if (k < 0) return k + static_cast<KeyType>(size);
return k;
});
} else {
thrust::copy(keys_in_ptr, keys_in_ptr + size, keys_out_ptr);
}
thrust::copy(values_in_ptr, values_in_ptr + size, values_out_ptr);
thrust::stable_sort_by_key(keys_out_ptr, keys_out_ptr + size, values_out_ptr);
}
TVM_REGISTER_GLOBAL("tvm.contrib.thrust.stable_sort_by_key")
.set_body([](TVMArgs args, TVMRetValue* ret) {
ICHECK_GE(args.num_args, 5);
DLTensor* keys_in = args[0];
DLTensor* values_in = args[1];
DLTensor* keys_out = args[2];
DLTensor* values_out = args[3];
bool for_scatter = args[4];
auto key_dtype = DLDataType2String(keys_in->dtype);
auto value_dtype = DLDataType2String(values_in->dtype);
if (key_dtype == "int32") {
if (value_dtype == "int32") {
thrust_stable_sort_by_key<int, int>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "int64") {
thrust_stable_sort_by_key<int, int64_t>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "float32") {
thrust_stable_sort_by_key<int, float>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else {
LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
}
} else if (key_dtype == "int64") {
if (value_dtype == "int32") {
thrust_stable_sort_by_key<int64_t, int>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "int64") {
thrust_stable_sort_by_key<int64_t, int64_t>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "float32") {
thrust_stable_sort_by_key<int64_t, float>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else {
LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
}
} else if (key_dtype == "float32") {
if (value_dtype == "int32") {
thrust_stable_sort_by_key<float, int>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "int64") {
thrust_stable_sort_by_key<float, int64_t>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else if (value_dtype == "float32") {
thrust_stable_sort_by_key<float, float>(keys_in, values_in, keys_out, values_out,
for_scatter);
} else {
LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
}
} else {
LOG(FATAL) << "Unsupported key dtype: " << key_dtype;
}
});
} // namespace contrib
} // namespace tvm