1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #define LOG_TAG "Operations"
18
19 #include "RoiPooling.h"
20
21 #include <algorithm>
22 #include <cfloat>
23 #include <cmath>
24 #include <vector>
25
26 #include "OperationResolver.h"
27 #include "OperationsExecutionUtils.h"
28 #include "Tracing.h"
29
30 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
31 #include "CpuOperationUtils.h"
32 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
33
34 namespace android {
35 namespace nn {
36 namespace roi_pooling {
37
38 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
39 namespace {
40
41 template <typename T_Input, typename T_Roi>
roiPoolingNhwc(const T_Input * inputData,const Shape & inputShape,const T_Roi * roiData,const Shape & roiShape,const int32_t * batchSplitData,const Shape &,float heightStride,float widthStride,T_Input * outputData,const Shape & outputShape)42 inline bool roiPoolingNhwc(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData,
43 const Shape& roiShape, const int32_t* batchSplitData,
44 const Shape& /*batchSplitShape*/, float heightStride, float widthStride,
45 T_Input* outputData, const Shape& outputShape) {
46 NNTRACE_TRANS("RoiPooling");
47
48 const uint32_t kRoiDim = 4;
49 const T_Roi heightScale = 1.0f / heightStride;
50 const T_Roi widthScale = 1.0f / widthStride;
51
52 uint32_t numBatches = getSizeOfDimension(inputShape, 0);
53 uint32_t inHeight = getSizeOfDimension(inputShape, 1);
54 uint32_t inWidth = getSizeOfDimension(inputShape, 2);
55 uint32_t inDepth = getSizeOfDimension(inputShape, 3);
56 uint32_t outHeight = getSizeOfDimension(outputShape, 1);
57 uint32_t outWidth = getSizeOfDimension(outputShape, 2);
58 uint32_t numRois = getSizeOfDimension(roiShape, 0);
59 uint32_t roiInfoLength = getSizeOfDimension(roiShape, 1);
60
61 T_Input* outPtr = outputData;
62 const T_Roi* roiDataEnd = roiData + numRois * roiInfoLength;
63 uint32_t roiIndex = 0;
64 for (const T_Roi* roiInfo = roiData; roiInfo < roiDataEnd; roiInfo += kRoiDim, roiIndex++) {
65 uint32_t batchId = batchSplitData[roiIndex];
66 // Check for malformed data
67 // 1. invalid batch id
68 // 2. Region out of bound: x1|x2|y1|y2 < 0 || x1|x2 > inWidth || y1|y2 > inHeight
69 // 3. Invalid region: x2 < x1 || y2 < y1
70 NN_RET_CHECK_GE(batchId, 0u);
71 NN_RET_CHECK_LT(batchId, numBatches);
72 NN_RET_CHECK(roiInfo[0] >= 0);
73 NN_RET_CHECK(roiInfo[1] >= 0);
74 NN_RET_CHECK(roiInfo[2] >= 0);
75 NN_RET_CHECK(roiInfo[3] >= 0);
76 NN_RET_CHECK(roiInfo[0] * widthScale <= inWidth);
77 NN_RET_CHECK(roiInfo[1] * heightScale <= inHeight);
78 NN_RET_CHECK(roiInfo[2] * widthScale <= inWidth);
79 NN_RET_CHECK(roiInfo[3] * heightScale <= inHeight);
80 NN_RET_CHECK(roiInfo[0] <= roiInfo[2]);
81 NN_RET_CHECK(roiInfo[1] <= roiInfo[3]);
82
83 int32_t wRoiStart = std::round(static_cast<float>(roiInfo[0] * widthScale));
84 int32_t hRoiStart = std::round(static_cast<float>(roiInfo[1] * heightScale));
85 int32_t wRoiEnd = std::round(static_cast<float>(roiInfo[2] * widthScale));
86 int32_t hRoiEnd = std::round(static_cast<float>(roiInfo[3] * heightScale));
87
88 // Rois with width/height < 1 are considered malformed and are forced to be 1
89 T_Roi roiWidth = static_cast<T_Roi>(std::max(wRoiEnd - wRoiStart + 1, 1));
90 T_Roi roiHeight = static_cast<T_Roi>(std::max(hRoiEnd - hRoiStart + 1, 1));
91 T_Roi wStepSize = roiWidth / static_cast<T_Roi>(outWidth);
92 T_Roi hStepSize = roiHeight / static_cast<T_Roi>(outHeight);
93
94 const T_Input* batchBase = inputData + batchId * inHeight * inWidth * inDepth;
95 for (uint32_t i = 0; i < outHeight; i++) {
96 for (uint32_t j = 0; j < outWidth; j++) {
97 // Take floor on start, ceil on end, start included, end excluded, i.e. [start, end)
98 // end is guaranteed to larger than start by at least 1
99 uint32_t wStart = std::floor(static_cast<float>(wStepSize * j + wRoiStart));
100 uint32_t wEnd = std::ceil(static_cast<float>(wStepSize * (j + 1) + wRoiStart));
101 uint32_t hStart = std::floor(static_cast<float>(hStepSize * i + hRoiStart));
102 uint32_t hEnd = std::ceil(static_cast<float>(hStepSize * (i + 1) + hRoiStart));
103
104 wStart = std::min(wStart, inWidth);
105 wEnd = std::min(wEnd, inWidth);
106 hStart = std::min(hStart, inHeight);
107 hEnd = std::min(hEnd, inHeight);
108
109 for (uint32_t k = 0; k < inDepth; k++) {
110 T_Input maxValue = static_cast<T_Input>(inputShape.offset);
111 bool first = true;
112 for (uint32_t h = hStart; h < hEnd; h++) {
113 for (uint32_t w = wStart; w < wEnd; w++) {
114 T_Input inputValue = batchBase[h * inWidth * inDepth + w * inDepth + k];
115 if (first || inputValue > maxValue) {
116 maxValue = inputValue;
117 first = false;
118 }
119 }
120 }
121 outPtr[k] = maxValue;
122 }
123 outPtr += inDepth;
124 }
125 }
126 }
127 return true;
128 }
129
130 template <typename T_Input, typename T_Roi>
roiPooling(const T_Input * inputData,const Shape & inputShape,const T_Roi * roiData,const Shape & roiShape,const int32_t * batchSplitData,const Shape & batchSplitShape,float heightStride,float widthStride,bool useNchw,T_Input * outputData,const Shape & outputShape)131 inline bool roiPooling(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData,
132 const Shape& roiShape, const int32_t* batchSplitData,
133 const Shape& batchSplitShape, float heightStride, float widthStride,
134 bool useNchw, T_Input* outputData, const Shape& outputShape) {
135 InputWithLayout<T_Input> input(useNchw);
136 OutputWithLayout<T_Input> output(useNchw);
137 NN_RET_CHECK(input.initialize(inputData, inputShape));
138 NN_RET_CHECK(output.initialize(outputData, outputShape));
139 NN_RET_CHECK(roiPoolingNhwc(input.getNhwcBuffer(), input.getNhwcShape(), roiData, roiShape,
140 batchSplitData, batchSplitShape, heightStride, widthStride,
141 output.getNhwcBuffer(), output.getNhwcShape()));
142 NN_RET_CHECK(output.commit());
143 return true;
144 }
145
146 template <>
roiPooling(const uint8_t * inputData,const Shape & inputShape,const uint16_t * roiData,const Shape & roiShape,const int32_t * batchSplitData,const Shape & batchSplitShape,float heightStride,float widthStride,bool useNchw,uint8_t * outputData,const Shape & outputShape)147 inline bool roiPooling<uint8_t, uint16_t>(const uint8_t* inputData, const Shape& inputShape,
148 const uint16_t* roiData, const Shape& roiShape,
149 const int32_t* batchSplitData,
150 const Shape& batchSplitShape, float heightStride,
151 float widthStride, bool useNchw, uint8_t* outputData,
152 const Shape& outputShape) {
153 std::vector<float> roi_float32(getNumberOfElements(roiShape));
154 convertQuantToFloat32(roiData, roiShape.scale, roiShape.offset, &roi_float32);
155 NN_RET_CHECK(roiPooling(inputData, inputShape, roi_float32.data(), roiShape, batchSplitData,
156 batchSplitShape, heightStride, widthStride, useNchw, outputData,
157 outputShape));
158 return true;
159 }
160
161 template <>
roiPooling(const int8_t * inputData,const Shape & inputShape,const uint16_t * roiData,const Shape & roiShape,const int32_t * batchSplitData,const Shape & batchSplitShape,float heightStride,float widthStride,bool useNchw,int8_t * outputData,const Shape & outputShape)162 inline bool roiPooling<int8_t, uint16_t>(const int8_t* inputData, const Shape& inputShape,
163 const uint16_t* roiData, const Shape& roiShape,
164 const int32_t* batchSplitData,
165 const Shape& batchSplitShape, float heightStride,
166 float widthStride, bool useNchw, int8_t* outputData,
167 const Shape& outputShape) {
168 std::vector<float> roi_float32(getNumberOfElements(roiShape));
169 convertQuantToFloat32(roiData, roiShape.scale, roiShape.offset, &roi_float32);
170 NN_RET_CHECK(roiPooling(inputData, inputShape, roi_float32.data(), roiShape, batchSplitData,
171 batchSplitShape, heightStride, widthStride, useNchw, outputData,
172 outputShape));
173 return true;
174 }
175
176 } // namespace
177
prepare(IOperationExecutionContext * context)178 bool prepare(IOperationExecutionContext* context) {
179 bool useNchw = context->getInputValue<bool>(kLayoutScalar);
180 Shape input = context->getInputShape(kInputTensor);
181 Shape roiShape = context->getInputShape(kRoiTensor);
182 Shape batchSplitShape = context->getInputShape(kBatchSplitTensor);
183 NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4u);
184 NN_RET_CHECK_EQ(getNumberOfDimensions(roiShape), 2u);
185
186 [[maybe_unused]] uint32_t numBatches = getSizeOfDimension(input, 0);
187 [[maybe_unused]] uint32_t inHeight = getSizeOfDimension(input, useNchw ? 2 : 1);
188 [[maybe_unused]] uint32_t inWidth = getSizeOfDimension(input, useNchw ? 3 : 2);
189 uint32_t inDepth = getSizeOfDimension(input, useNchw ? 1 : 3);
190 uint32_t numRois = getSizeOfDimension(roiShape, 0);
191 NN_RET_CHECK_EQ(getSizeOfDimension(roiShape, 1), 4u);
192 NN_RET_CHECK_EQ(getSizeOfDimension(batchSplitShape, 0), numRois);
193
194 auto outputHeight = context->getInputValue<int32_t>(kOutputHeightScalar);
195 auto outputWidth = context->getInputValue<int32_t>(kOutputWidthScalar);
196 float heightStride, widthStride;
197 if (context->getInputType(kInputTensor) == OperandType::TENSOR_FLOAT16) {
198 heightStride = context->getInputValue<_Float16>(kHeightStrideSalar);
199 widthStride = context->getInputValue<_Float16>(kWidthStrideScalar);
200 } else {
201 heightStride = context->getInputValue<float>(kHeightStrideSalar);
202 widthStride = context->getInputValue<float>(kWidthStrideScalar);
203 }
204 NN_RET_CHECK_GT(outputHeight, 0);
205 NN_RET_CHECK_GT(outputWidth, 0);
206 NN_RET_CHECK_GT(heightStride, 0);
207 NN_RET_CHECK_GT(widthStride, 0);
208
209 if (roiShape.type == OperandType::TENSOR_QUANT16_ASYMM) {
210 NN_RET_CHECK_EQ(roiShape.scale, 0.125f);
211 NN_RET_CHECK_EQ(roiShape.offset, 0);
212 }
213
214 Shape output = input;
215 if (useNchw) {
216 output.dimensions = {numRois, inDepth, static_cast<uint32_t>(outputHeight),
217 static_cast<uint32_t>(outputWidth)};
218 } else {
219 output.dimensions = {numRois, static_cast<uint32_t>(outputHeight),
220 static_cast<uint32_t>(outputWidth), inDepth};
221 }
222 return context->setOutputShape(kOutputTensor, output);
223 }
224
execute(IOperationExecutionContext * context)225 bool execute(IOperationExecutionContext* context) {
226 switch (context->getInputType(kInputTensor)) {
227 case OperandType::TENSOR_FLOAT16:
228 return roiPooling(context->getInputBuffer<_Float16>(kInputTensor),
229 context->getInputShape(kInputTensor),
230 context->getInputBuffer<_Float16>(kRoiTensor),
231 context->getInputShape(kRoiTensor),
232 context->getInputBuffer<int32_t>(kBatchSplitTensor),
233 context->getInputShape(kBatchSplitTensor),
234 context->getInputValue<_Float16>(kHeightStrideSalar),
235 context->getInputValue<_Float16>(kWidthStrideScalar),
236 context->getInputValue<bool>(kLayoutScalar),
237 context->getOutputBuffer<_Float16>(kOutputTensor),
238 context->getOutputShape(kOutputTensor));
239 case OperandType::TENSOR_FLOAT32:
240 return roiPooling(context->getInputBuffer<float>(kInputTensor),
241 context->getInputShape(kInputTensor),
242 context->getInputBuffer<float>(kRoiTensor),
243 context->getInputShape(kRoiTensor),
244 context->getInputBuffer<int32_t>(kBatchSplitTensor),
245 context->getInputShape(kBatchSplitTensor),
246 context->getInputValue<float>(kHeightStrideSalar),
247 context->getInputValue<float>(kWidthStrideScalar),
248 context->getInputValue<bool>(kLayoutScalar),
249 context->getOutputBuffer<float>(kOutputTensor),
250 context->getOutputShape(kOutputTensor));
251 case OperandType::TENSOR_QUANT8_ASYMM:
252 return roiPooling(context->getInputBuffer<uint8_t>(kInputTensor),
253 context->getInputShape(kInputTensor),
254 context->getInputBuffer<uint16_t>(kRoiTensor),
255 context->getInputShape(kRoiTensor),
256 context->getInputBuffer<int32_t>(kBatchSplitTensor),
257 context->getInputShape(kBatchSplitTensor),
258 context->getInputValue<float>(kHeightStrideSalar),
259 context->getInputValue<float>(kWidthStrideScalar),
260 context->getInputValue<bool>(kLayoutScalar),
261 context->getOutputBuffer<uint8_t>(kOutputTensor),
262 context->getOutputShape(kOutputTensor));
263 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
264 return roiPooling(context->getInputBuffer<int8_t>(kInputTensor),
265 context->getInputShape(kInputTensor),
266 context->getInputBuffer<uint16_t>(kRoiTensor),
267 context->getInputShape(kRoiTensor),
268 context->getInputBuffer<int32_t>(kBatchSplitTensor),
269 context->getInputShape(kBatchSplitTensor),
270 context->getInputValue<float>(kHeightStrideSalar),
271 context->getInputValue<float>(kWidthStrideScalar),
272 context->getInputValue<bool>(kLayoutScalar),
273 context->getOutputBuffer<int8_t>(kOutputTensor),
274 context->getOutputShape(kOutputTensor));
275 default:
276 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
277 }
278 }
279 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
280
281 } // namespace roi_pooling
282
283 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(ROI_POOLING, roi_pooling::prepare, roi_pooling::execute);
284
285 } // namespace nn
286 } // namespace android
287