File size: 9,124 Bytes
28c256d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
// Copyright 2019 Yan Yan
//
// Licensed 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.

#ifndef INDICE_CU_H_
#define INDICE_CU_H_
#include <utils/spconv/spconv/geometry.h>
#include <utils/spconv/tensorview/tensorview.h>

#include <utils/spconv/tensorview/helper_kernel.cuh>

template <typename Index, typename IndexGrid, unsigned NDim,
          int KernelMaxVolume = 256>
__global__ void prepareIndicePairsKernel(
    tv::TensorView<const Index> indicesIn, tv::TensorView<Index> indicesOut,
    tv::TensorView<IndexGrid> gridsOut, tv::TensorView<Index> indicePairs,
    tv::TensorView<Index> indiceNum, tv::TensorView<Index> indicePairUnique,
    const tv::SimpleVector<Index, NDim> kernelSize,
    const tv::SimpleVector<Index, NDim> stride,
    const tv::SimpleVector<Index, NDim> padding,
    const tv::SimpleVector<Index, NDim> dilation,
    const tv::SimpleVector<Index, NDim> outSpatialShape) {
  auto numActIn = indicesIn.dim(0);
  Index spatialVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    spatialVolume *= outSpatialShape[i];
  }
  Index kernelVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    kernelVolume *= kernelSize[i];
  }
  Index numValidPoints = 0;
  Index validPoints[KernelMaxVolume * (NDim + 1)];
  Index *pointPtr = nullptr;
  auto indicePairsDim2 = indicePairs.dim(2);
  Index index;
  for (int ix : tv::KernelLoopX<int>(numActIn)) {
    numValidPoints = getValidOutPos<Index, NDim>(
        indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
        stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
        validPoints);
    for (Index i = 0; i < numValidPoints; ++i) {
      pointPtr = validPoints + i * (NDim + 1);
      auto offset = pointPtr[NDim];
      auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
      indicePairs(offset, 0, oldNum) = ix;
      index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
              spatialVolume * indicesIn(ix, 0);
      indicePairs(offset, 1, oldNum) = index;
      indicePairUnique[offset * indicePairsDim2 + oldNum] = index;
    }
  }
}

template <typename Index, typename IndexGrid, unsigned NDim,
          int KernelMaxVolume = 256>
__global__ void prepareDeConvIndicePairsKernel(
    tv::TensorView<const Index> indicesIn, tv::TensorView<Index> indicesOut,
    tv::TensorView<IndexGrid> gridsOut, tv::TensorView<Index> indicePairs,
    tv::TensorView<Index> indiceNum, tv::TensorView<Index> indicePairUnique,
    const tv::SimpleVector<Index, NDim> kernelSize,
    const tv::SimpleVector<Index, NDim> stride,
    const tv::SimpleVector<Index, NDim> padding,
    const tv::SimpleVector<Index, NDim> dilation,
    const tv::SimpleVector<Index, NDim> outSpatialShape) {
  auto numActIn = indicesIn.dim(0);
  Index spatialVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    spatialVolume *= outSpatialShape[i];
  }
  Index kernelVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    kernelVolume *= kernelSize[i];
  }
  Index numValidPoints = 0;
  Index validPoints[KernelMaxVolume * (NDim + 1)];
  Index *pointPtr = nullptr;
  auto indicePairsDim2 = indicePairs.dim(2);
  Index index;
  for (int ix : tv::KernelLoopX<int>(numActIn)) {
    numValidPoints = getValidOutPosTranspose<Index, NDim>(
        indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
        stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
        validPoints);
    for (Index i = 0; i < numValidPoints; ++i) {
      pointPtr = validPoints + i * (NDim + 1);
      auto offset = pointPtr[NDim];
      auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
      indicePairs(offset, 0, oldNum) = ix;
      index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
              spatialVolume * indicesIn(ix, 0);
      indicePairs(offset, 1, oldNum) = index;
      indicePairUnique[offset * indicePairsDim2 + oldNum] = index;
    }
  }
}

template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void assignGridAndIndiceOutKernel(
    tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
    int numAct, tv::TensorView<Index> indicePairs,
    tv::TensorView<Index> indicePairUnique,
    const tv::SimpleVector<Index, NDim> outSpatialShape, int batchSize) {
  Index index;
  auto indicesOutPtr = indicesOut.data();
  for (int ix : tv::KernelLoopX<int>(numAct)) {
    index = indicePairUnique[ix];
    gridsOut[index] = ix;
    index = tv::rowArrayIdxInv<Index, NDim>(
        index, indicesOutPtr + ix * (NDim + 1) + 1, outSpatialShape.data());
    indicesOut[ix * (NDim + 1)] = index % batchSize;
  }
}

template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void assignIndicePairsKernel(
    tv::TensorView<Index> indicesOut, tv::TensorView<IndexGrid> gridsOut,
    int numActIn, tv::TensorView<Index> indicePairs,
    tv::TensorView<Index> indicePairUnique,
    const tv::SimpleVector<Index, NDim> outSpatialShape) {
  Index index;
  int kernelVolume = indicePairs.dim(0);
  for (int ix : tv::KernelLoopX<int>(numActIn)) {
    for (int i = 0; i < kernelVolume; ++i) {
      index = indicePairs(i, 1, ix);
      if (index > -1) {
        indicePairs(i, 1, ix) = gridsOut[index];
      }
    }
  }
}

template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void prepareSubMGridKernel(
    tv::TensorView<const Index> indicesIn, tv::TensorView<IndexGrid> gridsOut,
    const tv::SimpleVector<Index, NDim> outSpatialShape) {
  auto numActIn = indicesIn.dim(0);
  Index spatialVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    spatialVolume *= outSpatialShape[i];
  }
  Index index = 0;
  for (int ix : tv::KernelLoopX<int>(numActIn)) {
    index = tv::rowArrayIdx<Index, NDim>(indicesIn.data() + ix * (NDim + 1) + 1,
                                         outSpatialShape.data()) +
            spatialVolume * indicesIn(ix, 0);
    gridsOut[index] = ix;
  }
}

template <typename Index, typename IndexGrid, unsigned NDim,
          int KernelMaxVolume = 256>
__global__ void getSubMIndicePairsKernel(
    tv::TensorView<const Index> indicesIn, tv::TensorView<IndexGrid> gridsOut,
    tv::TensorView<Index> indicePairs, tv::TensorView<Index> indiceNum,
    const tv::SimpleVector<Index, NDim> kernelSize,
    const tv::SimpleVector<Index, NDim> stride,
    const tv::SimpleVector<Index, NDim> padding,
    const tv::SimpleVector<Index, NDim> dilation,
    const tv::SimpleVector<Index, NDim> outSpatialShape) {
  auto numActIn = indicesIn.dim(0);
  Index spatialVolume = 1;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    spatialVolume *= outSpatialShape[i];
  }
  Index numValidPoints = 0;
  Index validPoints[KernelMaxVolume * (NDim + 1)];
  Index *pointPtr = nullptr;
  Index index = 0;
  for (int ix : tv::KernelLoopX<int>(numActIn)) {
    numValidPoints = getValidOutPos<Index, NDim>(
        indicesIn.data() + ix * (NDim + 1) + 1, kernelSize.data(),
        stride.data(), padding.data(), dilation.data(), outSpatialShape.data(),
        validPoints);
    for (int i = 0; i < numValidPoints; ++i) {
      pointPtr = validPoints + i * (NDim + 1);
      auto offset = pointPtr[NDim];
      index = tv::rowArrayIdx<Index, NDim>(pointPtr, outSpatialShape.data()) +
              spatialVolume * indicesIn(ix, 0);
      if (gridsOut[index] > -1) {
        auto oldNum = atomicAdd(indiceNum.data() + offset, Index(1));
        indicePairs(offset, 1, oldNum) = gridsOut[index];
        indicePairs(offset, 0, oldNum) = ix;
      }
    }
  }
}

template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void resetGridKernel(const Index *indicePairUnique,
                                tv::TensorView<IndexGrid> gridsOut,
                                int numAct) {
  for (int ix : tv::KernelLoopX<int>(numAct)) {
    gridsOut[indicePairUnique[ix]] = -1;
  }
}

template <typename Index, typename IndexGrid, unsigned NDim>
__global__ void resetGridSubMKernel(
    const Index *indices, tv::TensorView<IndexGrid> gridsOut,
    const tv::SimpleVector<Index, NDim> outSpatialShape, int numAct) {
  int outSpatialShapeReg[NDim];
  for (int i = 0; i < NDim; ++i) {
    outSpatialShapeReg[i] = outSpatialShape[i];
  }
  Index spatialVolume = 1;
  auto indsPtr = indices;
#pragma unroll
  for (int i = 0; i < NDim; ++i) {
    spatialVolume *= outSpatialShape[i];
  }
  Index index;
  for (int ix : tv::KernelLoopX<int>(numAct)) {
    indsPtr = indices + ix * (NDim + 1);
    index = tv::rowArrayIdx<Index, NDim>(indsPtr + 1, outSpatialShapeReg);
    gridsOut[index + spatialVolume * indsPtr[0]] = -1;
  }
}

#endif