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 <algorithm>
20 #include <cfloat>
21 #include <cmath>
22 #include <vector>
23
24 #include "CpuOperationUtils.h"
25 #include "HalInterfaces.h"
26 #include "OperationResolver.h"
27 #include "OperationsUtils.h"
28 #include "Tracing.h"
29
30 namespace android {
31 namespace nn {
32 namespace roi_pooling {
33
34 constexpr char kOperationName[] = "ROI_POOLING";
35
36 constexpr uint32_t kNumInputs = 8;
37 constexpr uint32_t kInputTensor = 0;
38 constexpr uint32_t kRoiTensor = 1;
39 constexpr uint32_t kBatchSplitTensor = 2;
40 constexpr uint32_t kOutputHeightScalar = 3;
41 constexpr uint32_t kOutputWidthScalar = 4;
42 constexpr uint32_t kHeightStrideSalar = 5;
43 constexpr uint32_t kWidthStrideScalar = 6;
44 constexpr uint32_t kLayoutScalar = 7;
45
46 constexpr uint32_t kNumOutputs = 1;
47 constexpr uint32_t kOutputTensor = 0;
48
49 namespace {
50
51 using namespace hal;
52
53 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 & batchSplitShape,float heightStride,float widthStride,T_Input * outputData,const Shape & outputShape)54 inline bool roiPoolingNhwc(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData,
55 const Shape& roiShape, const int32_t* batchSplitData,
56 const Shape& batchSplitShape, float heightStride, float widthStride,
57 T_Input* outputData, const Shape& outputShape) {
58 NNTRACE_TRANS("RoiPooling");
59
60 const uint32_t kRoiDim = 4;
61 const T_Roi heightScale = 1.0f / heightStride;
62 const T_Roi widthScale = 1.0f / widthStride;
63
64 uint32_t numBatches = getSizeOfDimension(inputShape, 0);
65 uint32_t inHeight = getSizeOfDimension(inputShape, 1);
66 uint32_t inWidth = getSizeOfDimension(inputShape, 2);
67 uint32_t inDepth = getSizeOfDimension(inputShape, 3);
68 uint32_t outHeight = getSizeOfDimension(outputShape, 1);
69 uint32_t outWidth = getSizeOfDimension(outputShape, 2);
70 uint32_t numRois = getSizeOfDimension(roiShape, 0);
71 uint32_t roiInfoLength = getSizeOfDimension(roiShape, 1);
72
73 T_Input* outPtr = outputData;
74 const T_Roi* roiDataEnd = roiData + numRois * roiInfoLength;
75 uint32_t roiIndex = 0;
76 for (const T_Roi* roiInfo = roiData; roiInfo < roiDataEnd; roiInfo += kRoiDim, roiIndex++) {
77 uint32_t batchId = batchSplitData[roiIndex];
78 // Check for malformed data
79 // 1. invalid batch id
80 // 2. Region out of bound: x1|x2|y1|y2 < 0 || x1|x2 > inWidth || y1|y2 > inHeight
81 // 3. Invalid region: x2 < x1 || y2 < y1
82 NN_RET_CHECK_GE(batchId, 0);
83 NN_RET_CHECK_LT(batchId, numBatches);
84 NN_RET_CHECK(roiInfo[0] >= 0);
85 NN_RET_CHECK(roiInfo[1] >= 0);
86 NN_RET_CHECK(roiInfo[2] >= 0);
87 NN_RET_CHECK(roiInfo[3] >= 0);
88 NN_RET_CHECK(roiInfo[0] * widthScale <= inWidth);
89 NN_RET_CHECK(roiInfo[1] * heightScale <= inHeight);
90 NN_RET_CHECK(roiInfo[2] * widthScale <= inWidth);
91 NN_RET_CHECK(roiInfo[3] * heightScale <= inHeight);
92 NN_RET_CHECK(roiInfo[0] <= roiInfo[2]);
93 NN_RET_CHECK(roiInfo[1] <= roiInfo[3]);
94
95 int32_t wRoiStart = std::round(static_cast<float>(roiInfo[0] * widthScale));
96 int32_t hRoiStart = std::round(static_cast<float>(roiInfo[1] * heightScale));
97 int32_t wRoiEnd = std::round(static_cast<float>(roiInfo[2] * widthScale));
98 int32_t hRoiEnd = std::round(static_cast<float>(roiInfo[3] * heightScale));
99
100 // Rois with width/height < 1 are considered malformed and are forced to be 1
101 T_Roi roiWidth = static_cast<T_Roi>(std::max(wRoiEnd - wRoiStart + 1, 1));
102 T_Roi roiHeight = static_cast<T_Roi>(std::max(hRoiEnd - hRoiStart + 1, 1));
103 T_Roi wStepSize = roiWidth / static_cast<T_Roi>(outWidth);
104 T_Roi hStepSize = roiHeight / static_cast<T_Roi>(outHeight);
105
106 const T_Input* batchBase = inputData + batchId * inHeight * inWidth * inDepth;
107 for (uint32_t i = 0; i < outHeight; i++) {
108 for (uint32_t j = 0; j < outWidth; j++) {
109 // Take floor on start, ceil on end, start included, end excluded, i.e. [start, end)
110 // end is guaranteed to larger than start by at least 1
111 uint32_t wStart = std::floor(static_cast<float>(wStepSize * j + wRoiStart));
112 uint32_t wEnd = std::ceil(static_cast<float>(wStepSize * (j + 1) + wRoiStart));
113 uint32_t hStart = std::floor(static_cast<float>(hStepSize * i + hRoiStart));
114 uint32_t hEnd = std::ceil(static_cast<float>(hStepSize * (i + 1) + hRoiStart));
115
116 wStart = std::min(wStart, inWidth);
117 wEnd = std::min(wEnd, inWidth);
118 hStart = std::min(hStart, inHeight);
119 hEnd = std::min(hEnd, inHeight);
120
121 for (uint32_t k = 0; k < inDepth; k++) {
122 T_Input maxValue = static_cast<T_Input>(inputShape.offset);
123 bool first = true;
124 for (uint32_t h = hStart; h < hEnd; h++) {
125 for (uint32_t w = wStart; w < wEnd; w++) {
126 T_Input inputValue = batchBase[h * inWidth * inDepth + w * inDepth + k];
127 if (first || inputValue > maxValue) {
128 maxValue = inputValue;
129 first = false;
130 }
131 }
132 }
133 outPtr[k] = maxValue;
134 }
135 outPtr += inDepth;
136 }
137 }
138 }
139 return true;
140 }
141
142 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)143 inline bool roiPooling(const T_Input* inputData, const Shape& inputShape, const T_Roi* roiData,
144 const Shape& roiShape, const int32_t* batchSplitData,
145 const Shape& batchSplitShape, float heightStride, float widthStride,
146 bool useNchw, T_Input* outputData, const Shape& outputShape) {
147 InputWithLayout<T_Input> input(useNchw);
148 OutputWithLayout<T_Input> output(useNchw);
149 NN_RET_CHECK(input.initialize(inputData, inputShape));
150 NN_RET_CHECK(output.initialize(outputData, outputShape));
151 NN_RET_CHECK(roiPoolingNhwc(input.getNhwcBuffer(), input.getNhwcShape(), roiData, roiShape,
152 batchSplitData, batchSplitShape, heightStride, widthStride,
153 output.getNhwcBuffer(), output.getNhwcShape()));
154 NN_RET_CHECK(output.commit());
155 return true;
156 }
157
158 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)159 inline bool roiPooling<uint8_t, uint16_t>(const uint8_t* inputData, const Shape& inputShape,
160 const uint16_t* roiData, const Shape& roiShape,
161 const int32_t* batchSplitData,
162 const Shape& batchSplitShape, float heightStride,
163 float widthStride, bool useNchw, uint8_t* outputData,
164 const Shape& outputShape) {
165 std::vector<float> roi_float32(getNumberOfElements(roiShape));
166 convertQuantToFloat32(roiData, roiShape.scale, roiShape.offset, &roi_float32);
167 NN_RET_CHECK(roiPooling(inputData, inputShape, roi_float32.data(), roiShape, batchSplitData,
168 batchSplitShape, heightStride, widthStride, useNchw, outputData,
169 outputShape));
170 return true;
171 }
172
173 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)174 inline bool roiPooling<int8_t, uint16_t>(const int8_t* inputData, const Shape& inputShape,
175 const uint16_t* roiData, const Shape& roiShape,
176 const int32_t* batchSplitData,
177 const Shape& batchSplitShape, float heightStride,
178 float widthStride, bool useNchw, int8_t* outputData,
179 const Shape& outputShape) {
180 std::vector<float> roi_float32(getNumberOfElements(roiShape));
181 convertQuantToFloat32(roiData, roiShape.scale, roiShape.offset, &roi_float32);
182 NN_RET_CHECK(roiPooling(inputData, inputShape, roi_float32.data(), roiShape, batchSplitData,
183 batchSplitShape, heightStride, widthStride, useNchw, outputData,
184 outputShape));
185 return true;
186 }
187
188 } // namespace
189
validate(const IOperationValidationContext * context)190 bool validate(const IOperationValidationContext* context) {
191 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
192 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
193 std::vector<OperandType> inExpectedTypes;
194 auto inputType = context->getInputType(kInputTensor);
195 if (inputType == OperandType::TENSOR_FLOAT32) {
196 inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
197 OperandType::TENSOR_INT32, OperandType::INT32,
198 OperandType::INT32, OperandType::FLOAT32,
199 OperandType::FLOAT32, OperandType::BOOL};
200 } else if (inputType == OperandType::TENSOR_FLOAT16) {
201 inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
202 OperandType::TENSOR_INT32, OperandType::INT32,
203 OperandType::INT32, OperandType::FLOAT16,
204 OperandType::FLOAT16, OperandType::BOOL};
205 } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
206 inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
207 inExpectedTypes = {inputType,
208 OperandType::TENSOR_QUANT16_ASYMM,
209 OperandType::TENSOR_INT32,
210 OperandType::INT32,
211 OperandType::INT32,
212 OperandType::FLOAT32,
213 OperandType::FLOAT32,
214 OperandType::BOOL};
215 } else {
216 LOG(ERROR) << "Unsupported input tensor type for operation " << kOperationName;
217 return false;
218 }
219 NN_RET_CHECK(validateInputTypes(context, inExpectedTypes));
220 NN_RET_CHECK(validateOutputTypes(context, {inputType}));
221 if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
222 return validateHalVersion(context, HalVersion::V1_3);
223 ;
224 } else {
225 return validateHalVersion(context, HalVersion::V1_2);
226 }
227 }
228
prepare(IOperationExecutionContext * context)229 bool prepare(IOperationExecutionContext* context) {
230 bool useNchw = context->getInputValue<bool>(kLayoutScalar);
231 Shape input = context->getInputShape(kInputTensor);
232 Shape roiShape = context->getInputShape(kRoiTensor);
233 Shape batchSplitShape = context->getInputShape(kBatchSplitTensor);
234 NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4);
235 NN_RET_CHECK_EQ(getNumberOfDimensions(roiShape), 2);
236
237 uint32_t numBatches = getSizeOfDimension(input, 0);
238 uint32_t inHeight = getSizeOfDimension(input, useNchw ? 2 : 1);
239 uint32_t inWidth = getSizeOfDimension(input, useNchw ? 3 : 2);
240 uint32_t inDepth = getSizeOfDimension(input, useNchw ? 1 : 3);
241 uint32_t numRois = getSizeOfDimension(roiShape, 0);
242 NN_RET_CHECK_EQ(getSizeOfDimension(roiShape, 1), 4);
243 NN_RET_CHECK_EQ(getSizeOfDimension(batchSplitShape, 0), numRois);
244
245 auto outputHeight = context->getInputValue<int32_t>(kOutputHeightScalar);
246 auto outputWidth = context->getInputValue<int32_t>(kOutputWidthScalar);
247 float heightStride, widthStride;
248 if (context->getInputType(kInputTensor) == OperandType::TENSOR_FLOAT16) {
249 heightStride = context->getInputValue<_Float16>(kHeightStrideSalar);
250 widthStride = context->getInputValue<_Float16>(kWidthStrideScalar);
251 } else {
252 heightStride = context->getInputValue<float>(kHeightStrideSalar);
253 widthStride = context->getInputValue<float>(kWidthStrideScalar);
254 }
255 NN_RET_CHECK_GT(outputHeight, 0);
256 NN_RET_CHECK_GT(outputWidth, 0);
257 NN_RET_CHECK_GT(heightStride, 0);
258 NN_RET_CHECK_GT(widthStride, 0);
259
260 if (roiShape.type == OperandType::TENSOR_QUANT16_ASYMM) {
261 NN_RET_CHECK_EQ(roiShape.scale, 0.125f);
262 NN_RET_CHECK_EQ(roiShape.offset, 0);
263 }
264
265 Shape output = input;
266 if (useNchw) {
267 output.dimensions = {numRois, inDepth, static_cast<uint32_t>(outputHeight),
268 static_cast<uint32_t>(outputWidth)};
269 } else {
270 output.dimensions = {numRois, static_cast<uint32_t>(outputHeight),
271 static_cast<uint32_t>(outputWidth), inDepth};
272 }
273 return context->setOutputShape(kOutputTensor, output);
274 }
275
execute(IOperationExecutionContext * context)276 bool execute(IOperationExecutionContext* context) {
277 switch (context->getInputType(kInputTensor)) {
278 case OperandType::TENSOR_FLOAT16:
279 return roiPooling(context->getInputBuffer<_Float16>(kInputTensor),
280 context->getInputShape(kInputTensor),
281 context->getInputBuffer<_Float16>(kRoiTensor),
282 context->getInputShape(kRoiTensor),
283 context->getInputBuffer<int32_t>(kBatchSplitTensor),
284 context->getInputShape(kBatchSplitTensor),
285 context->getInputValue<_Float16>(kHeightStrideSalar),
286 context->getInputValue<_Float16>(kWidthStrideScalar),
287 context->getInputValue<bool>(kLayoutScalar),
288 context->getOutputBuffer<_Float16>(kOutputTensor),
289 context->getOutputShape(kOutputTensor));
290 case OperandType::TENSOR_FLOAT32:
291 return roiPooling(context->getInputBuffer<float>(kInputTensor),
292 context->getInputShape(kInputTensor),
293 context->getInputBuffer<float>(kRoiTensor),
294 context->getInputShape(kRoiTensor),
295 context->getInputBuffer<int32_t>(kBatchSplitTensor),
296 context->getInputShape(kBatchSplitTensor),
297 context->getInputValue<float>(kHeightStrideSalar),
298 context->getInputValue<float>(kWidthStrideScalar),
299 context->getInputValue<bool>(kLayoutScalar),
300 context->getOutputBuffer<float>(kOutputTensor),
301 context->getOutputShape(kOutputTensor));
302 case OperandType::TENSOR_QUANT8_ASYMM:
303 return roiPooling(context->getInputBuffer<uint8_t>(kInputTensor),
304 context->getInputShape(kInputTensor),
305 context->getInputBuffer<uint16_t>(kRoiTensor),
306 context->getInputShape(kRoiTensor),
307 context->getInputBuffer<int32_t>(kBatchSplitTensor),
308 context->getInputShape(kBatchSplitTensor),
309 context->getInputValue<float>(kHeightStrideSalar),
310 context->getInputValue<float>(kWidthStrideScalar),
311 context->getInputValue<bool>(kLayoutScalar),
312 context->getOutputBuffer<uint8_t>(kOutputTensor),
313 context->getOutputShape(kOutputTensor));
314 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
315 return roiPooling(context->getInputBuffer<int8_t>(kInputTensor),
316 context->getInputShape(kInputTensor),
317 context->getInputBuffer<uint16_t>(kRoiTensor),
318 context->getInputShape(kRoiTensor),
319 context->getInputBuffer<int32_t>(kBatchSplitTensor),
320 context->getInputShape(kBatchSplitTensor),
321 context->getInputValue<float>(kHeightStrideSalar),
322 context->getInputValue<float>(kWidthStrideScalar),
323 context->getInputValue<bool>(kLayoutScalar),
324 context->getOutputBuffer<int8_t>(kOutputTensor),
325 context->getOutputShape(kOutputTensor));
326 default:
327 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
328 }
329 }
330
331 } // namespace roi_pooling
332
333 NN_REGISTER_OPERATION(ROI_POOLING, roi_pooling::kOperationName, roi_pooling::validate,
334 roi_pooling::prepare, roi_pooling::execute);
335
336 } // namespace nn
337 } // namespace android
338