• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2017 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 <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
20 #include <tensorflow/lite/kernels/internal/reference/integer_ops/conv.h>
21 #include <tensorflow/lite/kernels/internal/types.h>
22 
23 #include <algorithm>
24 #include <iterator>
25 #include <memory>
26 #include <vector>
27 
28 #include "CpuOperationUtils.h"
29 #include "HalInterfaces.h"
30 #include "OperationResolver.h"
31 #include "Operations.h"
32 #include "OperationsUtils.h"
33 #include "Tracing.h"
34 #include "Utils.h"
35 
36 namespace android {
37 namespace nn {
38 namespace conv_2d {
39 
40 constexpr char kOperationName[] = "CONV_2D";
41 
42 constexpr uint32_t kNumInputsArray[] = {7, 8, 10, 11, 13};
43 constexpr uint32_t kInputTensor = 0;
44 constexpr uint32_t kFilterTensor = 1;
45 constexpr uint32_t kBiasTensor = 2;
46 
47 constexpr uint32_t kNumOutputs = 1;
48 constexpr uint32_t kOutputTensor = 0;
49 
50 namespace {
51 
52 using namespace hal;
53 
54 // If possible we will use this static buffer for the tensor.
55 constexpr size_t kStaticBufferSize = 1605632;
56 char static_scratch_buffer[kStaticBufferSize];
57 
58 // executionMutex is used to protect concurrent access of the static_scratch_buffer
59 // and other non-threadsafe resources like gemmlowp::GemmContext.
60 // std::mutex is safe for pthreads on Android.
61 std::mutex executionMutex;
62 
63 struct Conv2dParam {
64     int32_t padding_left, padding_right;
65     int32_t padding_top, padding_bottom;
66     int32_t stride_width, stride_height;
67     int32_t dilation_width_factor = 1, dilation_height_factor = 1;
68     int32_t activation;
69     bool useNchw = false;
70 
initializeandroid::nn::conv_2d::__anon4d5fbfb50111::Conv2dParam71     bool initialize(const IOperationExecutionContext* context) {
72         uint32_t inCount = context->getNumInputs();
73         int32_t padding_implicit = 0;
74         bool useImplicitPadding = false;
75         if ((inCount >= 8 && context->getInputType(7) == OperandType::BOOL) || inCount == 7) {
76             padding_implicit = context->getInputValue<int32_t>(3);
77             stride_width = context->getInputValue<int32_t>(4);
78             stride_height = context->getInputValue<int32_t>(5);
79             activation = context->getInputValue<int32_t>(6);
80             if (inCount >= 8) {
81                 useNchw = context->getInputValue<bool>(7);
82             }
83             if (inCount == 10) {
84                 dilation_width_factor = context->getInputValue<int32_t>(8);
85                 dilation_height_factor = context->getInputValue<int32_t>(9);
86             }
87             useImplicitPadding = true;
88         } else if (inCount >= 10 && context->getInputType(7) == OperandType::INT32) {
89             padding_left = context->getInputValue<int32_t>(3);
90             padding_right = context->getInputValue<int32_t>(4);
91             padding_top = context->getInputValue<int32_t>(5);
92             padding_bottom = context->getInputValue<int32_t>(6);
93             stride_width = context->getInputValue<int32_t>(7);
94             stride_height = context->getInputValue<int32_t>(8);
95             activation = context->getInputValue<int32_t>(9);
96             if (inCount >= 11) {
97                 useNchw = context->getInputValue<bool>(10);
98             }
99             if (inCount == 13) {
100                 dilation_width_factor = context->getInputValue<int32_t>(11);
101                 dilation_height_factor = context->getInputValue<int32_t>(12);
102             }
103         } else {
104             NN_RET_CHECK_FAIL() << "Unsupported input spec for operation " << kOperationName;
105         }
106         if (useImplicitPadding) {
107             Shape inputShape = context->getInputShape(kInputTensor);
108             Shape filterShape = context->getInputShape(kFilterTensor);
109             int32_t input_width = getSizeOfDimension(inputShape, useNchw ? 3 : 2);
110             int32_t input_height = getSizeOfDimension(inputShape, useNchw ? 2 : 1);
111             int32_t filter_width = getSizeOfDimension(filterShape, 2);
112             int32_t filter_height = getSizeOfDimension(filterShape, 1);
113             calculateExplicitPadding(input_width, stride_width, dilation_width_factor, filter_width,
114                                      padding_implicit, &padding_left, &padding_right);
115             calculateExplicitPadding(input_height, stride_height, dilation_height_factor,
116                                      filter_height, padding_implicit, &padding_top,
117                                      &padding_bottom);
118         }
119         NN_RET_CHECK_GE(padding_left, 0);
120         NN_RET_CHECK_GE(padding_right, 0);
121         NN_RET_CHECK_GE(padding_top, 0);
122         NN_RET_CHECK_GE(padding_bottom, 0);
123         NN_RET_CHECK_GT(stride_width, 0);
124         NN_RET_CHECK_GT(stride_height, 0);
125         NN_RET_CHECK_GT(dilation_width_factor, 0);
126         NN_RET_CHECK_GT(dilation_height_factor, 0);
127         NN_RET_CHECK_GE(activation, 0);
128         return true;
129     }
130 };
131 
132 #define ANDROID_NN_CONV_PARAMETERS(Type)                                          \
133     uint32_t height = getSizeOfDimension(inputShape, 1);                          \
134     uint32_t width = getSizeOfDimension(inputShape, 2);                           \
135     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);                   \
136     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);                    \
137     uint32_t outHeight = getSizeOfDimension(outputShape, 1);                      \
138     uint32_t outWidth = getSizeOfDimension(outputShape, 2);                       \
139     uint32_t inDepth = getSizeOfDimension(inputShape, 3);                         \
140                                                                                   \
141     uint32_t paddingHeight = (uint32_t)padding_top;                               \
142     uint32_t paddingWidth = (uint32_t)padding_left;                               \
143                                                                                   \
144     tflite::Dims<4> im2colDim;                                                    \
145     im2colDim.sizes[3] = (int)getSizeOfDimension(outputShape, 0);                 \
146     im2colDim.sizes[2] = (int)getSizeOfDimension(outputShape, 1);                 \
147     im2colDim.sizes[1] = (int)getSizeOfDimension(outputShape, 2);                 \
148     im2colDim.sizes[0] = (int)inDepth * filterHeight * filterWidth;               \
149                                                                                   \
150     im2colDim.strides[0] = 1;                                                     \
151     for (int i = 1; i < 4; i++) {                                                 \
152         im2colDim.strides[i] = im2colDim.strides[i - 1] * im2colDim.sizes[i - 1]; \
153     }                                                                             \
154                                                                                   \
155     Type* im2colData = nullptr;                                                   \
156     uint64_t im2colByteSize = sizeof(Type);                                       \
157     std::unique_ptr<Type[]> im2colGuard;                                          \
158     for (int i = 0; i < 4; i++) {                                                 \
159         im2colByteSize *= im2colDim.sizes[i];                                     \
160     }                                                                             \
161     /* http://b/77982879, tflite::optimized_ops::Conv uses int for offsets */     \
162     if (im2colByteSize >= 0x7fffffff) {                                           \
163         LOG(ERROR) << "Conv size is too large, not enough memory";                \
164         return false;                                                             \
165     }                                                                             \
166     if (im2colByteSize <= kStaticBufferSize) {                                    \
167         im2colData = reinterpret_cast<Type*>(static_scratch_buffer);              \
168     } else {                                                                      \
169         im2colData = new (std::nothrow) Type[im2colByteSize / sizeof(Type)];      \
170         if (im2colData == nullptr) {                                              \
171             LOG(ERROR) << "Conv size is too large, not enough memory";            \
172             return false;                                                         \
173         }                                                                         \
174         im2colGuard.reset(im2colData);                                            \
175     }
176 
convNhwc(const float * inputData,const Shape & inputShape,const float * filterData,const Shape & filterShape,const float * biasData,const Shape & biasShape,int32_t padding_left,int32_t padding_right,int32_t padding_top,int32_t padding_bottom,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor,int32_t activation,float * outputData,const Shape & outputShape)177 bool convNhwc(const float* inputData, const Shape& inputShape, const float* filterData,
178               const Shape& filterShape, const float* biasData, const Shape& biasShape,
179               int32_t padding_left, int32_t padding_right, int32_t padding_top,
180               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
181               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
182               float* outputData, const Shape& outputShape) {
183     NNTRACE_TRANS("convFloat32");
184 
185     ANDROID_NN_CONV_PARAMETERS(float)
186 
187     float output_activation_min, output_activation_max;
188     CalculateActivationRangeFloat(activation, &output_activation_min, &output_activation_max);
189 
190     // Prevent concurrent executions that may access the scratch buffer.
191     std::unique_lock<std::mutex> lock(executionMutex);
192     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
193     tflite::optimized_ops::Conv(inputData, convertShapeToDims(inputShape), filterData,
194                                 convertShapeToDims(filterShape), biasData,
195                                 convertShapeToDims(biasShape), stride_width, stride_height,
196                                 dilation_width_factor, dilation_height_factor, paddingWidth,
197                                 paddingHeight, output_activation_min, output_activation_max,
198                                 outputData, convertShapeToDims(outputShape), im2colData, im2colDim);
199     return true;
200 }
201 
convNhwc(const uint8_t * inputData,const Shape & inputShape,const uint8_t * filterData,const Shape & filterShape,const int32_t * biasData,const Shape & biasShape,int32_t padding_left,int32_t padding_right,int32_t padding_top,int32_t padding_bottom,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor,int32_t activation,uint8_t * outputData,const Shape & outputShape)202 bool convNhwc(const uint8_t* inputData, const Shape& inputShape, const uint8_t* filterData,
203               const Shape& filterShape, const int32_t* biasData, const Shape& biasShape,
204               int32_t padding_left, int32_t padding_right, int32_t padding_top,
205               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
206               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
207               uint8_t* outputData, const Shape& outputShape) {
208     NNTRACE_TRANS("convQuant8");
209 
210     ANDROID_NN_CONV_PARAMETERS(uint8_t)
211 
212     int32_t inputOffset = -inputShape.offset;
213     int32_t filterOffset = -filterShape.offset;
214     int32_t outputOffset = outputShape.offset;
215 
216     double real_multiplier = 0.0;
217     int32_t output_multiplier = 0;
218     int32_t output_shift = 0;
219     int32_t output_activation_min = 0;
220     int32_t output_activation_max = 0;
221 
222     NN_RET_CHECK(GetQuantizedConvolutionMultipler(inputShape, filterShape, biasShape, outputShape,
223                                                   &real_multiplier));
224     int exponent;
225     NN_RET_CHECK(QuantizeMultiplier(real_multiplier, &output_multiplier, &exponent));
226     output_shift = -exponent;
227     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
228                                   &output_activation_max);
229 
230     static gemmlowp::GemmContext gemm_context;
231 
232     // Prevent concurrent executions that may access the scratch buffer and
233     // gemm_context.
234     std::unique_lock<std::mutex> lock(executionMutex);
235     // Alow gemmlowp automatically decide how many threads to use.
236     gemm_context.set_max_num_threads(0);
237 
238     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
239     tflite::optimized_ops::Conv(
240             inputData, convertShapeToDims(inputShape), inputOffset, filterData,
241             convertShapeToDims(filterShape), filterOffset, biasData, convertShapeToDims(biasShape),
242             stride_width, stride_height, dilation_width_factor, dilation_height_factor,
243             paddingWidth, paddingHeight, outputOffset, output_multiplier, output_shift,
244             output_activation_min, output_activation_max, outputData,
245             convertShapeToDims(outputShape), im2colData, im2colDim, &gemm_context);
246     return true;
247 }
248 
249 // Passing input, filter and output shapes by value, so that we can change the
250 // offsets without modifying the actual shapes.
convNhwc(const int8_t * inputData,Shape inputShape,const int8_t * filterData,Shape filterShape,const int32_t * biasData,const Shape & biasShape,int32_t padding_left,int32_t padding_right,int32_t padding_top,int32_t padding_bottom,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor,int32_t activation,int8_t * outputData,Shape outputShape)251 bool convNhwc(const int8_t* inputData, Shape inputShape, const int8_t* filterData,
252               Shape filterShape, const int32_t* biasData, const Shape& biasShape,
253               int32_t padding_left, int32_t padding_right, int32_t padding_top,
254               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
255               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
256               int8_t* outputData, Shape outputShape) {
257     NNTRACE_TRANS("convQuant8");
258 
259     std::vector<uint8_t> unsignedInput(getNumberOfElements(inputShape));
260     convertInt8ToUInt8(inputData, &unsignedInput);
261     inputShape.offset += 128;
262 
263     std::vector<uint8_t> unsignedFilter(getNumberOfElements(filterShape));
264     convertInt8ToUInt8(filterData, &unsignedFilter);
265     filterShape.offset += 128;
266 
267     std::vector<uint8_t> unsignedOutput(getNumberOfElements(outputShape));
268     outputShape.offset += 128;
269 
270     NN_RET_CHECK(convNhwc(unsignedInput.data(), inputShape, unsignedFilter.data(), filterShape,
271                           biasData, biasShape, padding_left, padding_right, padding_top,
272                           padding_bottom, stride_width, stride_height, dilation_width_factor,
273                           dilation_height_factor, activation, unsignedOutput.data(), outputShape));
274 
275     convertUInt8ToInt8(unsignedOutput, outputData);
276 
277     return true;
278 }
279 
convNhwc(const _Float16 * inputData,const Shape & inputShape,const _Float16 * filterData,const Shape & filterShape,const _Float16 * biasData,const Shape & biasShape,int32_t padding_left,int32_t padding_right,int32_t padding_top,int32_t padding_bottom,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor,int32_t activation,_Float16 * outputData,const Shape & outputShape)280 bool convNhwc(const _Float16* inputData, const Shape& inputShape, const _Float16* filterData,
281               const Shape& filterShape, const _Float16* biasData, const Shape& biasShape,
282               int32_t padding_left, int32_t padding_right, int32_t padding_top,
283               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
284               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
285               _Float16* outputData, const Shape& outputShape) {
286     NNTRACE_TRANS("convFloat16");
287 
288     std::vector<float> inputData_float32(getNumberOfElements(inputShape));
289     std::vector<float> filterData_float32(getNumberOfElements(filterShape));
290     std::vector<float> biasData_float32(getNumberOfElements(biasShape));
291     std::vector<float> outputData_float32(getNumberOfElements(outputShape));
292 
293     convertFloat16ToFloat32(inputData, &inputData_float32);
294     convertFloat16ToFloat32(filterData, &filterData_float32);
295     convertFloat16ToFloat32(biasData, &biasData_float32);
296 
297     convNhwc(inputData_float32.data(), inputShape, filterData_float32.data(), filterShape,
298              biasData_float32.data(), biasShape, padding_left, padding_right, padding_top,
299              padding_bottom, stride_width, stride_height, dilation_width_factor,
300              dilation_height_factor, activation, outputData_float32.data(), outputShape);
301     convertFloat32ToFloat16(outputData_float32, outputData);
302 
303     return true;
304 }
305 
306 template <typename T_Input, typename T_Filter, typename T_Bias>
conv(const T_Input * inputData,const Shape & inputShape,const T_Filter * filterData,const Shape & filterShape,const T_Bias * biasData,const Shape & biasShape,int32_t padding_left,int32_t padding_right,int32_t padding_top,int32_t padding_bottom,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor,int32_t activation,bool useNchw,T_Input * outputData,const Shape & outputShape)307 bool conv(const T_Input* inputData, const Shape& inputShape, const T_Filter* filterData,
308           const Shape& filterShape, const T_Bias* biasData, const Shape& biasShape,
309           int32_t padding_left, int32_t padding_right, int32_t padding_top, int32_t padding_bottom,
310           int32_t stride_width, int32_t stride_height, int32_t dilation_width_factor,
311           int32_t dilation_height_factor, int32_t activation, bool useNchw, T_Input* outputData,
312           const Shape& outputShape) {
313     InputWithLayout<T_Input> input(useNchw);
314     OutputWithLayout<T_Input> output(useNchw);
315     NN_RET_CHECK(input.initialize(inputData, inputShape));
316     NN_RET_CHECK(output.initialize(outputData, outputShape));
317     NN_RET_CHECK(convNhwc(input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape,
318                           biasData, biasShape, padding_left, padding_right, padding_top,
319                           padding_bottom, stride_width, stride_height, dilation_width_factor,
320                           dilation_height_factor, activation, output.getNhwcBuffer(),
321                           output.getNhwcShape()));
322     NN_RET_CHECK(output.commit());
323     return true;
324 }
325 
convQuant8PerChannelNhwc(const uint8_t * inputData,const Shape & inputShape,const int8_t * filterData,const Shape & filterShape,const float * filterScales,const int32_t * biasData,const Shape & biasShape,int32_t paddingLeft,int32_t paddingRight,int32_t paddingTop,int32_t paddingBottom,int32_t strideWidth,int32_t strideHeight,int32_t dilationWidthFactor,int32_t dilationHeightFactor,int32_t activation,uint8_t * outputData,const Shape & outputShape)326 bool convQuant8PerChannelNhwc(const uint8_t* inputData, const Shape& inputShape,
327                               const int8_t* filterData, const Shape& filterShape,
328                               const float* filterScales, const int32_t* biasData,
329                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
330                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
331                               int32_t strideHeight, int32_t dilationWidthFactor,
332                               int32_t dilationHeightFactor, int32_t activation, uint8_t* outputData,
333                               const Shape& outputShape) {
334     NNTRACE_TRANS("convQuant8PerChannel");
335 
336     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
337     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
338     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
339     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
340     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
341     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
342     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
343     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
344     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
345     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
346 
347     int32_t inputOffset = -inputShape.offset;
348     int32_t outputOffset = outputShape.offset;
349 
350     auto realMultiplier = std::vector<double>(outputDepth, .0f);
351     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
352     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
353 
354     for (int i = 0; i < outputDepth; ++i) {
355         Shape filterChannelShape = filterShape;
356         filterChannelShape.scale = filterScales[i];
357         Shape biasChannelShape = biasShape;
358         biasChannelShape.scale = filterScales[i] * inputShape.scale;
359         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
360                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
361         int exponent;
362         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &exponent));
363         outputShift[i] = -exponent;
364     }
365 
366     int32_t output_activation_min = 0, output_activation_max = 0;
367     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
368                                   &output_activation_max);
369     const uint8_t* inputBase = inputData;
370     uint8_t* outPtr = outputData;
371     for (uint32_t b = 0; b < numBatches; b++) {
372         for (uint32_t h = 0; h < outputHeight; h++) {
373             for (uint32_t w = 0; w < outputWidth; w++) {
374                 const int8_t* filterBase = filterData;
375 
376                 for (uint32_t d = 0; d < outputDepth; d++) {
377                     int32_t wInputOrigin = static_cast<int32_t>(w) * strideWidth - paddingLeft;
378                     int32_t hInputOrigin = static_cast<int32_t>(h) * strideHeight - paddingTop;
379                     int32_t sum = 0.0f;
380 
381                     for (uint32_t i = 0; i < filterHeight; i++) {
382                         for (uint32_t j = 0; j < filterWidth; j++) {
383                             for (uint32_t k = 0; k < filterDepth; k++) {
384                                 int32_t hInput = hInputOrigin +
385                                                  dilationHeightFactor * static_cast<int32_t>(i);
386                                 int32_t wInput = wInputOrigin +
387                                                  dilationWidthFactor * static_cast<int32_t>(j);
388                                 uint32_t dInput = k;
389                                 if (hInput >= 0 && hInput < static_cast<int32_t>(inputHeight) &&
390                                     wInput >= 0 && wInput < static_cast<int32_t>(inputWidth)) {
391                                     uint32_t filterIndex =
392                                             i * filterWidth * filterDepth + j * filterDepth + k;
393                                     uint32_t inputIndex = hInput * inputWidth * inputDepth +
394                                                           wInput * inputDepth + dInput;
395                                     sum += (static_cast<int32_t>(filterBase[filterIndex])) *
396                                            (static_cast<int32_t>(inputBase[inputIndex]) +
397                                             inputOffset);
398                                 }
399                             }
400                         }
401                     }
402                     sum += biasData[d];
403                     sum = tflite::MultiplyByQuantizedMultiplier(sum, outputMultiplier[d],
404                                                                 -outputShift[d]);
405                     sum += outputOffset;
406                     sum = std::max(std::min(sum, output_activation_max), output_activation_min);
407                     outPtr[d] = static_cast<uint8_t>(sum);
408                     filterBase += filterHeight * filterWidth * filterDepth;
409                 }
410                 outPtr += outputDepth;
411             }
412         }
413         inputBase += inputHeight * inputWidth * inputDepth;
414     }
415 
416     return true;
417 }
418 
convQuant8PerChannelNhwc(const int8_t * inputData,const Shape & inputShape,const int8_t * filterData,const Shape & filterShape,const float * filterScales,const int32_t * biasData,const Shape & biasShape,int32_t paddingLeft,int32_t paddingRight,int32_t paddingTop,int32_t paddingBottom,int32_t strideWidth,int32_t strideHeight,int32_t dilationWidthFactor,int32_t dilationHeightFactor,int32_t activation,int8_t * outputData,const Shape & outputShape)419 bool convQuant8PerChannelNhwc(const int8_t* inputData, const Shape& inputShape,
420                               const int8_t* filterData, const Shape& filterShape,
421                               const float* filterScales, const int32_t* biasData,
422                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
423                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
424                               int32_t strideHeight, int32_t dilationWidthFactor,
425                               int32_t dilationHeightFactor, int32_t activation, int8_t* outputData,
426                               const Shape& outputShape) {
427     NNTRACE_TRANS("convQuant8SignedPerChannel");
428 
429     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
430     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
431     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
432     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
433     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
434     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
435     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
436     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
437     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
438     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
439 
440     int32_t inputOffset = -inputShape.offset;
441     int32_t outputOffset = outputShape.offset;
442 
443     auto realMultiplier = std::vector<double>(outputDepth, .0f);
444     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
445     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
446 
447     for (int i = 0; i < outputDepth; ++i) {
448         Shape filterChannelShape = filterShape;
449         filterChannelShape.scale = filterScales[i];
450         Shape biasChannelShape = biasShape;
451         biasChannelShape.scale = filterScales[i] * inputShape.scale;
452         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
453                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
454         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &outputShift[i]));
455     }
456 
457     int32_t output_activation_min = 0, output_activation_max = 0;
458     CalculateActivationRangeInt8(activation, outputShape, &output_activation_min,
459                                  &output_activation_max);
460 
461     tflite::ConvParams convParams;
462     convParams.input_offset = -inputShape.offset;
463     convParams.output_offset = outputShape.offset;
464     convParams.stride_height = strideHeight;
465     convParams.stride_width = strideWidth;
466     convParams.dilation_height_factor = dilationHeightFactor;
467     convParams.dilation_width_factor = dilationWidthFactor;
468     convParams.padding_values.height = paddingTop;
469     convParams.padding_values.width = paddingLeft;
470     convParams.quantized_activation_min = output_activation_min;
471     convParams.quantized_activation_max = output_activation_max;
472 
473     NNTRACE_COMP_SWITCH("reference_integer_ops::ConvPerChannel");
474     tflite::reference_integer_ops::ConvPerChannel(
475             convParams, outputMultiplier.data(), outputShift.data(),
476             convertShapeToTflshape(inputShape), inputData, convertShapeToTflshape(filterShape),
477             filterData, convertShapeToTflshape(biasShape), biasData,
478             convertShapeToTflshape(outputShape), outputData);
479     return true;
480 }
481 
482 template <typename T>
convQuant8PerChannel(const T * inputData,const Shape & inputShape,const int8_t * filterData,const Shape & filterShape,const float * filterScales,const int32_t * biasData,const Shape & biasShape,int32_t paddingLeft,int32_t paddingRight,int32_t paddingTop,int32_t paddingBottom,int32_t strideWidth,int32_t strideHeight,int32_t dilationWidthFactor,int32_t dilationHeightFactor,int32_t activation,bool useNchw,T * outputData,const Shape & outputShape)483 bool convQuant8PerChannel(const T* inputData, const Shape& inputShape, const int8_t* filterData,
484                           const Shape& filterShape, const float* filterScales,
485                           const int32_t* biasData, const Shape& biasShape, int32_t paddingLeft,
486                           int32_t paddingRight, int32_t paddingTop, int32_t paddingBottom,
487                           int32_t strideWidth, int32_t strideHeight, int32_t dilationWidthFactor,
488                           int32_t dilationHeightFactor, int32_t activation, bool useNchw,
489                           T* outputData, const Shape& outputShape) {
490     InputWithLayout<T> input(useNchw);
491     OutputWithLayout<T> output(useNchw);
492     NN_RET_CHECK(input.initialize(inputData, inputShape));
493     NN_RET_CHECK(output.initialize(outputData, outputShape));
494     NN_RET_CHECK(convQuant8PerChannelNhwc(
495             input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape, filterScales,
496             biasData, biasShape, paddingLeft, paddingRight, paddingTop, paddingBottom, strideWidth,
497             strideHeight, dilationWidthFactor, dilationHeightFactor, activation,
498             output.getNhwcBuffer(), output.getNhwcShape()));
499     NN_RET_CHECK(output.commit());
500     return true;
501 }
502 
503 #undef ANDROID_NN_CONV_PARAMETERS
504 
505 }  // namespace
506 
validate(const IOperationValidationContext * context)507 bool validate(const IOperationValidationContext* context) {
508     const uint32_t numInputs = context->getNumInputs();
509     NN_RET_CHECK(
510             std::binary_search(std::begin(kNumInputsArray), std::end(kNumInputsArray), numInputs));
511     NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
512     const auto inputRank = getNumberOfDimensions(context->getInputShape(kInputTensor));
513     const auto filterRank = getNumberOfDimensions(context->getInputShape(kFilterTensor));
514     if (inputRank != 0) {
515         NN_RET_CHECK_EQ(inputRank, 4);
516     }
517     if (filterRank != 0) {
518         NN_RET_CHECK_EQ(filterRank, 4);
519     }
520     auto inputCount = context->getNumInputs();
521     auto inputType = context->getInputType(kInputTensor);
522     auto filterType = context->getInputType(kFilterTensor);
523     std::vector<OperandType> inExpectedTypes;
524     if (inputType == OperandType::TENSOR_FLOAT32) {
525         inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
526                            OperandType::TENSOR_FLOAT32, OperandType::INT32,
527                            OperandType::INT32,          OperandType::INT32,
528                            OperandType::INT32};
529     } else if (inputType == OperandType::TENSOR_FLOAT16) {
530         inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
531                            OperandType::TENSOR_FLOAT16, OperandType::INT32,
532                            OperandType::INT32,          OperandType::INT32,
533                            OperandType::INT32};
534     } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
535                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
536         NN_RET_CHECK(filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL ||
537                      filterType == inputType)
538                 << "Unsupported filter tensor type for operation " << kOperationName;
539         inExpectedTypes = {inputType,          filterType,         OperandType::TENSOR_INT32,
540                            OperandType::INT32, OperandType::INT32, OperandType::INT32,
541                            OperandType::INT32};
542 
543         if (filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
544             NN_RET_CHECK_EQ(context->getInputExtraParams(kFilterTensor).channelQuant().channelDim,
545                             0)
546                     << "Unsupported filter tensor channel dimension for operation "
547                     << kOperationName;
548         }
549     } else {
550         NN_RET_CHECK_FAIL() << "Unsupported input tensor type for operation " << kOperationName;
551     }
552 
553     // NeuralNetworks.h specifies that ANEURALNETWORKS_CONV_2D's output must
554     // meet "outputScale > inputScale * filterScale" for the operand type
555     // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM before API level 29. For other
556     // operand types (e.g., ANEURALNETWORKS_TENSOR_FLOAT32), this constraint
557     // does not apply, so by default the constraint is met.
558     bool meetsQuantizedScaleConstraintBeforeV1_2 = true;
559     if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
560         const float inputScale = context->getInputShape(kInputTensor).scale;
561         const float filterScale = context->getInputShape(kFilterTensor).scale;
562         const float outputScale = context->getInputShape(kOutputTensor).scale;
563         meetsQuantizedScaleConstraintBeforeV1_2 = (outputScale > inputScale * filterScale);
564     }
565 
566     bool withExplicitPadding = false;
567     bool withLayout = false;
568     bool withDilation = false;
569     if (inputCount >= 8) {
570         if (context->getInputType(7) == OperandType::INT32 && inputCount >= 10) {
571             std::vector<OperandType> explicitScalarTypes(3, OperandType::INT32);
572             inExpectedTypes.insert(inExpectedTypes.end(), explicitScalarTypes.begin(),
573                                    explicitScalarTypes.end());
574             withExplicitPadding = true;
575         }
576         int inputOffset = withExplicitPadding ? 3 : 0;
577         if (inputCount >= 8 + inputOffset) {
578             inExpectedTypes.push_back(OperandType::BOOL);
579             withLayout = true;
580         }
581         NN_RET_CHECK_NE(inputCount, 9 + inputOffset)
582                 << "Provided only one dilation factor value, two values are requred for operation "
583                 << kOperationName;
584         if (inputCount == 10 + inputOffset) {
585             inExpectedTypes.push_back(OperandType::INT32);
586             inExpectedTypes.push_back(OperandType::INT32);
587             withDilation = true;
588         }
589     }
590 
591     if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
592         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_3));
593     } else if (inputType == OperandType::TENSOR_FLOAT16 ||
594                filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL || withLayout ||
595                withDilation || !meetsQuantizedScaleConstraintBeforeV1_2) {
596         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_2));
597     } else {
598         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_0));
599     }
600     return validateInputTypes(context, inExpectedTypes) &&
601            validateOutputTypes(context, {inputType});
602 }
603 
prepare(IOperationExecutionContext * context)604 bool prepare(IOperationExecutionContext* context) {
605     Shape input = context->getInputShape(kInputTensor);
606     Shape filter = context->getInputShape(kFilterTensor);
607     Shape bias = context->getInputShape(kBiasTensor);
608 
609     if (filter.type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
610         NN_RET_CHECK(input.type == OperandType::TENSOR_QUANT8_ASYMM ||
611                      input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED);
612     } else {
613         NN_RET_CHECK(input.type == filter.type);
614     }
615     if (input.type == OperandType::TENSOR_QUANT8_ASYMM ||
616         input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
617         NN_RET_CHECK(bias.type == OperandType::TENSOR_INT32);
618     } else {
619         NN_RET_CHECK(input.type == bias.type);
620     }
621     NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4);
622     NN_RET_CHECK_EQ(getNumberOfDimensions(filter), 4);
623     NN_RET_CHECK_EQ(getNumberOfDimensions(bias), 1);
624 
625     Conv2dParam param;
626     NN_RET_CHECK(param.initialize(context));
627 
628     uint32_t batches = getSizeOfDimension(input, 0);
629     uint32_t height = getSizeOfDimension(input, param.useNchw ? 2 : 1);
630     uint32_t width = getSizeOfDimension(input, param.useNchw ? 3 : 2);
631     uint32_t channels_in = getSizeOfDimension(input, param.useNchw ? 1 : 3);
632     uint32_t channels_out = getSizeOfDimension(filter, 0);
633     uint32_t filterHeight = getSizeOfDimension(filter, 1);
634     uint32_t filterWidth = getSizeOfDimension(filter, 2);
635     // Only batches can be zero.
636     NN_RET_CHECK_EQ(channels_in, getSizeOfDimension(filter, 3));
637     NN_RET_CHECK_EQ(channels_out, getSizeOfDimension(bias, 0));
638     NN_RET_CHECK_GT(height, 0);
639     NN_RET_CHECK_GT(width, 0);
640     NN_RET_CHECK_GT(channels_in, 0);
641     NN_RET_CHECK_GT(channels_out, 0);
642 
643     int32_t effectiveFilterWidth = (filterWidth - 1) * param.dilation_width_factor + 1;
644     int32_t effectiveFilterHeight = (filterHeight - 1) * param.dilation_height_factor + 1;
645     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_left);
646     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_right);
647     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_top);
648     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_bottom);
649 
650     uint32_t outWidth =
651             computeOutSize(width, filterWidth, param.stride_width, param.dilation_width_factor,
652                            param.padding_left, param.padding_right);
653     uint32_t outHeight =
654             computeOutSize(height, filterHeight, param.stride_height, param.dilation_height_factor,
655                            param.padding_top, param.padding_bottom);
656 
657     Shape output = context->getOutputShape(kOutputTensor);
658     output.type = input.type;
659     if (param.useNchw) {
660         output.dimensions = {batches, channels_out, outHeight, outWidth};
661     } else {
662         output.dimensions = {batches, outHeight, outWidth, channels_out};
663     }
664     return context->setOutputShape(kOutputTensor, output);
665 }
666 
execute(IOperationExecutionContext * context)667 bool execute(IOperationExecutionContext* context) {
668     // Bypass execution in the case of zero-sized input.
669     if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
670     Conv2dParam param;
671     NN_RET_CHECK(param.initialize(context));
672     switch (context->getInputType(kInputTensor)) {
673         case OperandType::TENSOR_FLOAT32:
674             return conv(context->getInputBuffer<float>(kInputTensor),
675                         context->getInputShape(kInputTensor),
676                         context->getInputBuffer<float>(kFilterTensor),
677                         context->getInputShape(kFilterTensor),
678                         context->getInputBuffer<float>(kBiasTensor),
679                         context->getInputShape(kBiasTensor), param.padding_left,
680                         param.padding_right, param.padding_top, param.padding_bottom,
681                         param.stride_width, param.stride_height, param.dilation_width_factor,
682                         param.dilation_height_factor, param.activation, param.useNchw,
683                         context->getOutputBuffer<float>(kOutputTensor),
684                         context->getOutputShape(kOutputTensor));
685         case OperandType::TENSOR_FLOAT16:
686             return conv(context->getInputBuffer<_Float16>(kInputTensor),
687                         context->getInputShape(kInputTensor),
688                         context->getInputBuffer<_Float16>(kFilterTensor),
689                         context->getInputShape(kFilterTensor),
690                         context->getInputBuffer<_Float16>(kBiasTensor),
691                         context->getInputShape(kBiasTensor), param.padding_left,
692                         param.padding_right, param.padding_top, param.padding_bottom,
693                         param.stride_width, param.stride_height, param.dilation_width_factor,
694                         param.dilation_height_factor, param.activation, param.useNchw,
695                         context->getOutputBuffer<_Float16>(kOutputTensor),
696                         context->getOutputShape(kOutputTensor));
697         case OperandType::TENSOR_QUANT8_ASYMM:
698             if (context->getInputType(kFilterTensor) ==
699                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
700                 return convQuant8PerChannel(
701                         context->getInputBuffer<uint8_t>(kInputTensor),
702                         context->getInputShape(kInputTensor),
703                         context->getInputBuffer<int8_t>(kFilterTensor),
704                         context->getInputShape(kFilterTensor),
705                         context->getInputExtraParams(kFilterTensor).channelQuant().scales.data(),
706                         context->getInputBuffer<int32_t>(kBiasTensor),
707                         context->getInputShape(kBiasTensor), param.padding_left,
708                         param.padding_right, param.padding_top, param.padding_bottom,
709                         param.stride_width, param.stride_height, param.dilation_width_factor,
710                         param.dilation_height_factor, param.activation, param.useNchw,
711                         context->getOutputBuffer<uint8_t>(kOutputTensor),
712                         context->getOutputShape(kOutputTensor));
713             } else if (context->getInputType(kFilterTensor) == OperandType::TENSOR_QUANT8_ASYMM) {
714                 return conv(context->getInputBuffer<uint8_t>(kInputTensor),
715                             context->getInputShape(kInputTensor),
716                             context->getInputBuffer<uint8_t>(kFilterTensor),
717                             context->getInputShape(kFilterTensor),
718                             context->getInputBuffer<int32_t>(kBiasTensor),
719                             context->getInputShape(kBiasTensor), param.padding_left,
720                             param.padding_right, param.padding_top, param.padding_bottom,
721                             param.stride_width, param.stride_height, param.dilation_width_factor,
722                             param.dilation_height_factor, param.activation, param.useNchw,
723                             context->getOutputBuffer<uint8_t>(kOutputTensor),
724                             context->getOutputShape(kOutputTensor));
725             } else {
726                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
727             }
728         case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
729             if (context->getInputType(kFilterTensor) ==
730                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
731                 return convQuant8PerChannel(
732                         context->getInputBuffer<int8_t>(kInputTensor),
733                         context->getInputShape(kInputTensor),
734                         context->getInputBuffer<int8_t>(kFilterTensor),
735                         context->getInputShape(kFilterTensor),
736                         context->getInputExtraParams(kFilterTensor).channelQuant().scales.data(),
737                         context->getInputBuffer<int32_t>(kBiasTensor),
738                         context->getInputShape(kBiasTensor), param.padding_left,
739                         param.padding_right, param.padding_top, param.padding_bottom,
740                         param.stride_width, param.stride_height, param.dilation_width_factor,
741                         param.dilation_height_factor, param.activation, param.useNchw,
742                         context->getOutputBuffer<int8_t>(kOutputTensor),
743                         context->getOutputShape(kOutputTensor));
744             } else if (context->getInputType(kFilterTensor) ==
745                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
746                 return conv(context->getInputBuffer<int8_t>(kInputTensor),
747                             context->getInputShape(kInputTensor),
748                             context->getInputBuffer<int8_t>(kFilterTensor),
749                             context->getInputShape(kFilterTensor),
750                             context->getInputBuffer<int32_t>(kBiasTensor),
751                             context->getInputShape(kBiasTensor), param.padding_left,
752                             param.padding_right, param.padding_top, param.padding_bottom,
753                             param.stride_width, param.stride_height, param.dilation_width_factor,
754                             param.dilation_height_factor, param.activation, param.useNchw,
755                             context->getOutputBuffer<int8_t>(kOutputTensor),
756                             context->getOutputShape(kOutputTensor));
757             } else {
758                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
759             }
760         default:
761             NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
762     }
763 }
764 
765 }  // namespace conv_2d
766 
767 NN_REGISTER_OPERATION(CONV_2D, conv_2d::kOperationName, conv_2d::validate, conv_2d::prepare,
768                       conv_2d::execute, .allowZeroSizedInput = true);
769 
770 }  // namespace nn
771 }  // namespace android
772