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 <algorithm>
20 #include <iterator>
21 #include <memory>
22 #include <vector>
23 
24 #include "LegacyUtils.h"
25 #include "OperationResolver.h"
26 #include "Operations.h"
27 #include "OperationsUtils.h"
28 #include "Tracing.h"
29 
30 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
31 #include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
32 #include <tensorflow/lite/kernels/internal/reference/integer_ops/conv.h>
33 #include <tensorflow/lite/kernels/internal/types.h>
34 
35 #include "CpuOperationUtils.h"
36 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
37 
38 namespace android {
39 namespace nn {
40 namespace conv_2d {
41 
42 constexpr char kOperationName[] = "CONV_2D";
43 
44 constexpr uint32_t kNumInputsArray[] = {7, 8, 10, 11, 13};
45 constexpr uint32_t kInputTensor = 0;
46 constexpr uint32_t kFilterTensor = 1;
47 constexpr uint32_t kBiasTensor = 2;
48 
49 constexpr uint32_t kNumOutputs = 1;
50 constexpr uint32_t kOutputTensor = 0;
51 
52 namespace {
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::__anoncf6a863c0110::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 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
133 #define ANDROID_NN_CONV_PARAMETERS(Type)                                          \
134     uint32_t height = getSizeOfDimension(inputShape, 1);                          \
135     uint32_t width = getSizeOfDimension(inputShape, 2);                           \
136     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);                   \
137     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);                    \
138     uint32_t outHeight = getSizeOfDimension(outputShape, 1);                      \
139     uint32_t outWidth = getSizeOfDimension(outputShape, 2);                       \
140     uint32_t inDepth = getSizeOfDimension(inputShape, 3);                         \
141                                                                                   \
142     uint32_t paddingHeight = (uint32_t)padding_top;                               \
143     uint32_t paddingWidth = (uint32_t)padding_left;                               \
144                                                                                   \
145     tflite::Dims<4> im2colDim;                                                    \
146     im2colDim.sizes[3] = (int)getSizeOfDimension(outputShape, 0);                 \
147     im2colDim.sizes[2] = (int)getSizeOfDimension(outputShape, 1);                 \
148     im2colDim.sizes[1] = (int)getSizeOfDimension(outputShape, 2);                 \
149     im2colDim.sizes[0] = (int)inDepth * filterHeight * filterWidth;               \
150                                                                                   \
151     im2colDim.strides[0] = 1;                                                     \
152     for (int i = 1; i < 4; i++) {                                                 \
153         im2colDim.strides[i] = im2colDim.strides[i - 1] * im2colDim.sizes[i - 1]; \
154     }                                                                             \
155                                                                                   \
156     Type* im2colData = nullptr;                                                   \
157     uint64_t im2colByteSize = sizeof(Type);                                       \
158     std::unique_ptr<Type[]> im2colGuard;                                          \
159     for (int i = 0; i < 4; i++) {                                                 \
160         im2colByteSize *= im2colDim.sizes[i];                                     \
161     }                                                                             \
162     /* http://b/77982879, tflite::optimized_ops::Conv uses int for offsets */     \
163     if (im2colByteSize >= 0x7fffffff) {                                           \
164         LOG(ERROR) << "Conv size is too large, not enough memory";                \
165         return false;                                                             \
166     }                                                                             \
167     if (im2colByteSize <= kStaticBufferSize) {                                    \
168         im2colData = reinterpret_cast<Type*>(static_scratch_buffer);              \
169     } else {                                                                      \
170         im2colData = new (std::nothrow) Type[im2colByteSize / sizeof(Type)];      \
171         if (im2colData == nullptr) {                                              \
172             LOG(ERROR) << "Conv size is too large, not enough memory";            \
173             return false;                                                         \
174         }                                                                         \
175         im2colGuard.reset(im2colData);                                            \
176     }
177 
needim2colData(const Shape & filterShape,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor)178 bool needim2colData(const Shape& filterShape, int32_t stride_width, int32_t stride_height,
179                     int32_t dilation_width_factor, int32_t dilation_height_factor) {
180     // Within tflite::optimized_ops::Conv, the following tests are performed,
181     // and in the case (!need_dilated_im2col && !need_im2col), then the
182     // method doesn't expect to receive outputData. In debug mode this is
183     // asserted and fails tests, so we need to perform this check as the caller
184     // also. See:
185     // tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h:2655
186     const int filter_width = getSizeOfDimension(filterShape, 2);
187     const int filter_height = getSizeOfDimension(filterShape, 1);
188     const bool need_dilated_im2col = dilation_width_factor != 1 || dilation_height_factor != 1;
189     const bool need_im2col =
190             stride_width != 1 || stride_height != 1 || filter_width != 1 || filter_height != 1;
191     return need_dilated_im2col || need_im2col;
192 }
193 
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)194 bool convNhwc(const float* inputData, const Shape& inputShape, const float* filterData,
195               const Shape& filterShape, const float* biasData, const Shape& biasShape,
196               int32_t padding_left, int32_t padding_right, int32_t padding_top,
197               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
198               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
199               float* outputData, const Shape& outputShape) {
200     NNTRACE_TRANS("convFloat32");
201 
202     ANDROID_NN_CONV_PARAMETERS(float)
203 
204     float output_activation_min, output_activation_max;
205     CalculateActivationRangeFloat(activation, &output_activation_min, &output_activation_max);
206 
207     // Prevent concurrent executions that may access the scratch buffer.
208     std::unique_lock<std::mutex> lock(executionMutex);
209     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
210 
211     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
212                                                 dilation_width_factor, dilation_height_factor);
213 
214     tflite::optimized_ops::Conv(
215             inputData, convertShapeToDims(inputShape), filterData, convertShapeToDims(filterShape),
216             biasData, convertShapeToDims(biasShape), stride_width, stride_height,
217             dilation_width_factor, dilation_height_factor, paddingWidth, paddingHeight,
218             output_activation_min, output_activation_max, outputData,
219             convertShapeToDims(outputShape), need_im2colData ? im2colData : nullptr, im2colDim);
220     return true;
221 }
222 
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)223 bool convNhwc(const uint8_t* inputData, const Shape& inputShape, const uint8_t* filterData,
224               const Shape& filterShape, const int32_t* biasData, const Shape& biasShape,
225               int32_t padding_left, int32_t padding_right, int32_t padding_top,
226               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
227               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
228               uint8_t* outputData, const Shape& outputShape) {
229     NNTRACE_TRANS("convQuant8");
230 
231     ANDROID_NN_CONV_PARAMETERS(uint8_t)
232 
233     int32_t inputOffset = -inputShape.offset;
234     int32_t filterOffset = -filterShape.offset;
235     int32_t outputOffset = outputShape.offset;
236 
237     double real_multiplier = 0.0;
238     int32_t output_multiplier = 0;
239     int32_t output_shift = 0;
240     int32_t output_activation_min = 0;
241     int32_t output_activation_max = 0;
242 
243     NN_RET_CHECK(GetQuantizedConvolutionMultipler(inputShape, filterShape, biasShape, outputShape,
244                                                   &real_multiplier));
245     int exponent;
246     NN_RET_CHECK(QuantizeMultiplier(real_multiplier, &output_multiplier, &exponent));
247     output_shift = -exponent;
248     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
249                                   &output_activation_max);
250 
251     static gemmlowp::GemmContext gemm_context;
252 
253     // Prevent concurrent executions that may access the scratch buffer and
254     // gemm_context.
255     std::unique_lock<std::mutex> lock(executionMutex);
256     // Alow gemmlowp automatically decide how many threads to use.
257     gemm_context.set_max_num_threads(0);
258 
259     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
260 
261     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
262                                                 dilation_width_factor, dilation_height_factor);
263 
264     tflite::optimized_ops::Conv(inputData, convertShapeToDims(inputShape), inputOffset, filterData,
265                                 convertShapeToDims(filterShape), filterOffset, biasData,
266                                 convertShapeToDims(biasShape), stride_width, stride_height,
267                                 dilation_width_factor, dilation_height_factor, paddingWidth,
268                                 paddingHeight, outputOffset, output_multiplier, output_shift,
269                                 output_activation_min, output_activation_max, outputData,
270                                 convertShapeToDims(outputShape),
271                                 need_im2colData ? im2colData : nullptr, im2colDim, &gemm_context);
272     return true;
273 }
274 
275 // Passing input, filter and output shapes by value, so that we can change the
276 // 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)277 bool convNhwc(const int8_t* inputData, Shape inputShape, const int8_t* filterData,
278               Shape filterShape, const int32_t* biasData, const Shape& biasShape,
279               int32_t padding_left, int32_t padding_right, int32_t padding_top,
280               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
281               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
282               int8_t* outputData, Shape outputShape) {
283     NNTRACE_TRANS("convQuant8");
284 
285     std::vector<uint8_t> unsignedInput(getNumberOfElements(inputShape));
286     convertInt8ToUInt8(inputData, &unsignedInput);
287     inputShape.offset += 128;
288 
289     std::vector<uint8_t> unsignedFilter(getNumberOfElements(filterShape));
290     convertInt8ToUInt8(filterData, &unsignedFilter);
291     filterShape.offset += 128;
292 
293     std::vector<uint8_t> unsignedOutput(getNumberOfElements(outputShape));
294     outputShape.offset += 128;
295 
296     NN_RET_CHECK(convNhwc(unsignedInput.data(), inputShape, unsignedFilter.data(), filterShape,
297                           biasData, biasShape, padding_left, padding_right, padding_top,
298                           padding_bottom, stride_width, stride_height, dilation_width_factor,
299                           dilation_height_factor, activation, unsignedOutput.data(), outputShape));
300 
301     convertUInt8ToInt8(unsignedOutput, outputData);
302 
303     return true;
304 }
305 
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)306 bool convNhwc(const _Float16* inputData, const Shape& inputShape, const _Float16* filterData,
307               const Shape& filterShape, const _Float16* biasData, const Shape& biasShape,
308               int32_t padding_left, int32_t padding_right, int32_t padding_top,
309               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
310               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
311               _Float16* outputData, const Shape& outputShape) {
312     NNTRACE_TRANS("convFloat16");
313 
314     std::vector<float> inputData_float32(getNumberOfElements(inputShape));
315     std::vector<float> filterData_float32(getNumberOfElements(filterShape));
316     std::vector<float> biasData_float32(getNumberOfElements(biasShape));
317     std::vector<float> outputData_float32(getNumberOfElements(outputShape));
318 
319     convertFloat16ToFloat32(inputData, &inputData_float32);
320     convertFloat16ToFloat32(filterData, &filterData_float32);
321     convertFloat16ToFloat32(biasData, &biasData_float32);
322 
323     convNhwc(inputData_float32.data(), inputShape, filterData_float32.data(), filterShape,
324              biasData_float32.data(), biasShape, padding_left, padding_right, padding_top,
325              padding_bottom, stride_width, stride_height, dilation_width_factor,
326              dilation_height_factor, activation, outputData_float32.data(), outputShape);
327     convertFloat32ToFloat16(outputData_float32, outputData);
328 
329     return true;
330 }
331 
332 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)333 bool conv(const T_Input* inputData, const Shape& inputShape, const T_Filter* filterData,
334           const Shape& filterShape, const T_Bias* biasData, const Shape& biasShape,
335           int32_t padding_left, int32_t padding_right, int32_t padding_top, int32_t padding_bottom,
336           int32_t stride_width, int32_t stride_height, int32_t dilation_width_factor,
337           int32_t dilation_height_factor, int32_t activation, bool useNchw, T_Input* outputData,
338           const Shape& outputShape) {
339     InputWithLayout<T_Input> input(useNchw);
340     OutputWithLayout<T_Input> output(useNchw);
341     NN_RET_CHECK(input.initialize(inputData, inputShape));
342     NN_RET_CHECK(output.initialize(outputData, outputShape));
343     NN_RET_CHECK(convNhwc(input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape,
344                           biasData, biasShape, padding_left, padding_right, padding_top,
345                           padding_bottom, stride_width, stride_height, dilation_width_factor,
346                           dilation_height_factor, activation, output.getNhwcBuffer(),
347                           output.getNhwcShape()));
348     NN_RET_CHECK(output.commit());
349     return true;
350 }
351 
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)352 bool convQuant8PerChannelNhwc(const uint8_t* inputData, const Shape& inputShape,
353                               const int8_t* filterData, const Shape& filterShape,
354                               const float* filterScales, const int32_t* biasData,
355                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
356                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
357                               int32_t strideHeight, int32_t dilationWidthFactor,
358                               int32_t dilationHeightFactor, int32_t activation, uint8_t* outputData,
359                               const Shape& outputShape) {
360     NNTRACE_TRANS("convQuant8PerChannel");
361 
362     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
363     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
364     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
365     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
366     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
367     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
368     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
369     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
370     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
371     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
372 
373     int32_t inputOffset = -inputShape.offset;
374     int32_t outputOffset = outputShape.offset;
375 
376     auto realMultiplier = std::vector<double>(outputDepth, .0f);
377     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
378     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
379 
380     for (int i = 0; i < outputDepth; ++i) {
381         Shape filterChannelShape = filterShape;
382         filterChannelShape.scale = filterScales[i];
383         Shape biasChannelShape = biasShape;
384         biasChannelShape.scale = filterScales[i] * inputShape.scale;
385         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
386                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
387         int exponent;
388         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &exponent));
389         outputShift[i] = -exponent;
390     }
391 
392     int32_t output_activation_min = 0, output_activation_max = 0;
393     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
394                                   &output_activation_max);
395     const uint8_t* inputBase = inputData;
396     uint8_t* outPtr = outputData;
397     for (uint32_t b = 0; b < numBatches; b++) {
398         for (uint32_t h = 0; h < outputHeight; h++) {
399             for (uint32_t w = 0; w < outputWidth; w++) {
400                 const int8_t* filterBase = filterData;
401 
402                 for (uint32_t d = 0; d < outputDepth; d++) {
403                     int32_t wInputOrigin = static_cast<int32_t>(w) * strideWidth - paddingLeft;
404                     int32_t hInputOrigin = static_cast<int32_t>(h) * strideHeight - paddingTop;
405                     int32_t sum = 0.0f;
406 
407                     for (uint32_t i = 0; i < filterHeight; i++) {
408                         for (uint32_t j = 0; j < filterWidth; j++) {
409                             for (uint32_t k = 0; k < filterDepth; k++) {
410                                 int32_t hInput = hInputOrigin +
411                                                  dilationHeightFactor * static_cast<int32_t>(i);
412                                 int32_t wInput = wInputOrigin +
413                                                  dilationWidthFactor * static_cast<int32_t>(j);
414                                 uint32_t dInput = k;
415                                 if (hInput >= 0 && hInput < static_cast<int32_t>(inputHeight) &&
416                                     wInput >= 0 && wInput < static_cast<int32_t>(inputWidth)) {
417                                     uint32_t filterIndex =
418                                             i * filterWidth * filterDepth + j * filterDepth + k;
419                                     uint32_t inputIndex = hInput * inputWidth * inputDepth +
420                                                           wInput * inputDepth + dInput;
421                                     sum += (static_cast<int32_t>(filterBase[filterIndex])) *
422                                            (static_cast<int32_t>(inputBase[inputIndex]) +
423                                             inputOffset);
424                                 }
425                             }
426                         }
427                     }
428                     sum += biasData[d];
429                     sum = tflite::MultiplyByQuantizedMultiplier(sum, outputMultiplier[d],
430                                                                 -outputShift[d]);
431                     sum += outputOffset;
432                     sum = std::max(std::min(sum, output_activation_max), output_activation_min);
433                     outPtr[d] = static_cast<uint8_t>(sum);
434                     filterBase += filterHeight * filterWidth * filterDepth;
435                 }
436                 outPtr += outputDepth;
437             }
438         }
439         inputBase += inputHeight * inputWidth * inputDepth;
440     }
441 
442     return true;
443 }
444 
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)445 bool convQuant8PerChannelNhwc(const int8_t* inputData, const Shape& inputShape,
446                               const int8_t* filterData, const Shape& filterShape,
447                               const float* filterScales, const int32_t* biasData,
448                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
449                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
450                               int32_t strideHeight, int32_t dilationWidthFactor,
451                               int32_t dilationHeightFactor, int32_t activation, int8_t* outputData,
452                               const Shape& outputShape) {
453     NNTRACE_TRANS("convQuant8SignedPerChannel");
454 
455     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
456     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
457     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
458     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
459     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
460     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
461     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
462     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
463     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
464     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
465 
466     int32_t inputOffset = -inputShape.offset;
467     int32_t outputOffset = outputShape.offset;
468 
469     auto realMultiplier = std::vector<double>(outputDepth, .0f);
470     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
471     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
472 
473     for (int i = 0; i < outputDepth; ++i) {
474         Shape filterChannelShape = filterShape;
475         filterChannelShape.scale = filterScales[i];
476         Shape biasChannelShape = biasShape;
477         biasChannelShape.scale = filterScales[i] * inputShape.scale;
478         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
479                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
480         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &outputShift[i]));
481     }
482 
483     int32_t output_activation_min = 0, output_activation_max = 0;
484     CalculateActivationRangeInt8(activation, outputShape, &output_activation_min,
485                                  &output_activation_max);
486 
487     tflite::ConvParams convParams;
488     convParams.input_offset = -inputShape.offset;
489     convParams.output_offset = outputShape.offset;
490     convParams.stride_height = strideHeight;
491     convParams.stride_width = strideWidth;
492     convParams.dilation_height_factor = dilationHeightFactor;
493     convParams.dilation_width_factor = dilationWidthFactor;
494     convParams.padding_values.height = paddingTop;
495     convParams.padding_values.width = paddingLeft;
496     convParams.quantized_activation_min = output_activation_min;
497     convParams.quantized_activation_max = output_activation_max;
498 
499     NNTRACE_COMP_SWITCH("reference_integer_ops::ConvPerChannel");
500     tflite::reference_integer_ops::ConvPerChannel(
501             convParams, outputMultiplier.data(), outputShift.data(),
502             convertShapeToTflshape(inputShape), inputData, convertShapeToTflshape(filterShape),
503             filterData, convertShapeToTflshape(biasShape), biasData,
504             convertShapeToTflshape(outputShape), outputData);
505     return true;
506 }
507 
508 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)509 bool convQuant8PerChannel(const T* inputData, const Shape& inputShape, const int8_t* filterData,
510                           const Shape& filterShape, const float* filterScales,
511                           const int32_t* biasData, const Shape& biasShape, int32_t paddingLeft,
512                           int32_t paddingRight, int32_t paddingTop, int32_t paddingBottom,
513                           int32_t strideWidth, int32_t strideHeight, int32_t dilationWidthFactor,
514                           int32_t dilationHeightFactor, int32_t activation, bool useNchw,
515                           T* outputData, const Shape& outputShape) {
516     InputWithLayout<T> input(useNchw);
517     OutputWithLayout<T> output(useNchw);
518     NN_RET_CHECK(input.initialize(inputData, inputShape));
519     NN_RET_CHECK(output.initialize(outputData, outputShape));
520     NN_RET_CHECK(convQuant8PerChannelNhwc(
521             input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape, filterScales,
522             biasData, biasShape, paddingLeft, paddingRight, paddingTop, paddingBottom, strideWidth,
523             strideHeight, dilationWidthFactor, dilationHeightFactor, activation,
524             output.getNhwcBuffer(), output.getNhwcShape()));
525     NN_RET_CHECK(output.commit());
526     return true;
527 }
528 
529 #undef ANDROID_NN_CONV_PARAMETERS
530 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
531 
532 }  // namespace
533 
validate(const IOperationValidationContext * context)534 Result<Version> validate(const IOperationValidationContext* context) {
535     const uint32_t numInputs = context->getNumInputs();
536     NN_RET_CHECK(
537             std::binary_search(std::begin(kNumInputsArray), std::end(kNumInputsArray), numInputs));
538     NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
539     const auto inputRank = getNumberOfDimensions(context->getInputShape(kInputTensor));
540     const auto filterRank = getNumberOfDimensions(context->getInputShape(kFilterTensor));
541     if (inputRank != 0) {
542         NN_RET_CHECK_EQ(inputRank, 4);
543     }
544     if (filterRank != 0) {
545         NN_RET_CHECK_EQ(filterRank, 4);
546     }
547     auto inputCount = context->getNumInputs();
548     auto inputType = context->getInputType(kInputTensor);
549     auto filterType = context->getInputType(kFilterTensor);
550     std::vector<OperandType> inExpectedTypes;
551     if (inputType == OperandType::TENSOR_FLOAT32) {
552         inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
553                            OperandType::TENSOR_FLOAT32, OperandType::INT32,
554                            OperandType::INT32,          OperandType::INT32,
555                            OperandType::INT32};
556     } else if (inputType == OperandType::TENSOR_FLOAT16) {
557         inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
558                            OperandType::TENSOR_FLOAT16, OperandType::INT32,
559                            OperandType::INT32,          OperandType::INT32,
560                            OperandType::INT32};
561     } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
562                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
563         NN_RET_CHECK(filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL ||
564                      filterType == inputType)
565                 << "Unsupported filter tensor type for operation " << kOperationName;
566         inExpectedTypes = {inputType,          filterType,         OperandType::TENSOR_INT32,
567                            OperandType::INT32, OperandType::INT32, OperandType::INT32,
568                            OperandType::INT32};
569 
570         if (filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
571             NN_RET_CHECK_EQ(std::get<Operand::SymmPerChannelQuantParams>(
572                                     context->getInputExtraParams(kFilterTensor))
573                                     .channelDim,
574                             0)
575                     << "Unsupported filter tensor channel dimension for operation "
576                     << kOperationName;
577         }
578     } else {
579         NN_RET_CHECK_FAIL() << "Unsupported input tensor type for operation " << kOperationName;
580     }
581 
582     // NeuralNetworks.h specifies that ANEURALNETWORKS_CONV_2D's output must
583     // meet "outputScale > inputScale * filterScale" for the operand type
584     // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM before API level 29. For other
585     // operand types (e.g., ANEURALNETWORKS_TENSOR_FLOAT32), this constraint
586     // does not apply, so by default the constraint is met.
587     bool meetsQuantizedScaleConstraintBeforeV1_2 = true;
588     if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
589         const float inputScale = context->getInputShape(kInputTensor).scale;
590         const float filterScale = context->getInputShape(kFilterTensor).scale;
591         const float outputScale = context->getInputShape(kOutputTensor).scale;
592         meetsQuantizedScaleConstraintBeforeV1_2 = (outputScale > inputScale * filterScale);
593     }
594 
595     bool withExplicitPadding = false;
596     bool withLayout = false;
597     bool withDilation = false;
598     if (inputCount >= 8) {
599         if (context->getInputType(7) == OperandType::INT32 && inputCount >= 10) {
600             std::vector<OperandType> explicitScalarTypes(3, OperandType::INT32);
601             inExpectedTypes.insert(inExpectedTypes.end(), explicitScalarTypes.begin(),
602                                    explicitScalarTypes.end());
603             withExplicitPadding = true;
604         }
605         int inputOffset = withExplicitPadding ? 3 : 0;
606         if (inputCount >= 8 + inputOffset) {
607             inExpectedTypes.push_back(OperandType::BOOL);
608             withLayout = true;
609         }
610         NN_RET_CHECK_NE(inputCount, 9 + inputOffset)
611                 << "Provided only one dilation factor value, two values are requred for operation "
612                 << kOperationName;
613         if (inputCount == 10 + inputOffset) {
614             inExpectedTypes.push_back(OperandType::INT32);
615             inExpectedTypes.push_back(OperandType::INT32);
616             withDilation = true;
617         }
618     }
619 
620     auto minSupportedVersion = Version::ANDROID_OC_MR1;
621     if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
622         minSupportedVersion = Version::ANDROID_R;
623     } else if (inputType == OperandType::TENSOR_FLOAT16 ||
624                filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL || withLayout ||
625                withDilation || !meetsQuantizedScaleConstraintBeforeV1_2) {
626         minSupportedVersion = Version::ANDROID_Q;
627     } else {
628         minSupportedVersion = Version::ANDROID_OC_MR1;
629     }
630     NN_RET_CHECK(validateInputTypes(context, inExpectedTypes));
631     NN_RET_CHECK(validateOutputTypes(context, {inputType}));
632     return minSupportedVersion;
633 }
634 
635 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
prepare(IOperationExecutionContext * context)636 bool prepare(IOperationExecutionContext* context) {
637     Shape input = context->getInputShape(kInputTensor);
638     Shape filter = context->getInputShape(kFilterTensor);
639     Shape bias = context->getInputShape(kBiasTensor);
640 
641     if (filter.type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
642         NN_RET_CHECK(input.type == OperandType::TENSOR_QUANT8_ASYMM ||
643                      input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED);
644     } else {
645         NN_RET_CHECK(input.type == filter.type);
646     }
647     if (input.type == OperandType::TENSOR_QUANT8_ASYMM ||
648         input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
649         NN_RET_CHECK(bias.type == OperandType::TENSOR_INT32);
650     } else {
651         NN_RET_CHECK(input.type == bias.type);
652     }
653     NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4);
654     NN_RET_CHECK_EQ(getNumberOfDimensions(filter), 4);
655     NN_RET_CHECK_EQ(getNumberOfDimensions(bias), 1);
656 
657     Conv2dParam param;
658     NN_RET_CHECK(param.initialize(context));
659 
660     uint32_t batches = getSizeOfDimension(input, 0);
661     uint32_t height = getSizeOfDimension(input, param.useNchw ? 2 : 1);
662     uint32_t width = getSizeOfDimension(input, param.useNchw ? 3 : 2);
663     uint32_t channels_in = getSizeOfDimension(input, param.useNchw ? 1 : 3);
664     uint32_t channels_out = getSizeOfDimension(filter, 0);
665     uint32_t filterHeight = getSizeOfDimension(filter, 1);
666     uint32_t filterWidth = getSizeOfDimension(filter, 2);
667     // Only batches can be zero.
668     NN_RET_CHECK_EQ(channels_in, getSizeOfDimension(filter, 3));
669     NN_RET_CHECK_EQ(channels_out, getSizeOfDimension(bias, 0));
670     NN_RET_CHECK_GT(height, 0);
671     NN_RET_CHECK_GT(width, 0);
672     NN_RET_CHECK_GT(channels_in, 0);
673     NN_RET_CHECK_GT(channels_out, 0);
674 
675     int32_t effectiveFilterWidth = (filterWidth - 1) * param.dilation_width_factor + 1;
676     int32_t effectiveFilterHeight = (filterHeight - 1) * param.dilation_height_factor + 1;
677     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_left);
678     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_right);
679     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_top);
680     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_bottom);
681 
682     uint32_t outWidth =
683             computeOutSize(width, filterWidth, param.stride_width, param.dilation_width_factor,
684                            param.padding_left, param.padding_right);
685     uint32_t outHeight =
686             computeOutSize(height, filterHeight, param.stride_height, param.dilation_height_factor,
687                            param.padding_top, param.padding_bottom);
688 
689     Shape output = context->getOutputShape(kOutputTensor);
690     output.type = input.type;
691     if (param.useNchw) {
692         output.dimensions = {batches, channels_out, outHeight, outWidth};
693     } else {
694         output.dimensions = {batches, outHeight, outWidth, channels_out};
695     }
696     return context->setOutputShape(kOutputTensor, output);
697 }
698 
execute(IOperationExecutionContext * context)699 bool execute(IOperationExecutionContext* context) {
700     // Bypass execution in the case of zero-sized input.
701     if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
702     Conv2dParam param;
703     NN_RET_CHECK(param.initialize(context));
704     switch (context->getInputType(kInputTensor)) {
705         case OperandType::TENSOR_FLOAT32:
706             return conv(context->getInputBuffer<float>(kInputTensor),
707                         context->getInputShape(kInputTensor),
708                         context->getInputBuffer<float>(kFilterTensor),
709                         context->getInputShape(kFilterTensor),
710                         context->getInputBuffer<float>(kBiasTensor),
711                         context->getInputShape(kBiasTensor), param.padding_left,
712                         param.padding_right, param.padding_top, param.padding_bottom,
713                         param.stride_width, param.stride_height, param.dilation_width_factor,
714                         param.dilation_height_factor, param.activation, param.useNchw,
715                         context->getOutputBuffer<float>(kOutputTensor),
716                         context->getOutputShape(kOutputTensor));
717         case OperandType::TENSOR_FLOAT16:
718             return conv(context->getInputBuffer<_Float16>(kInputTensor),
719                         context->getInputShape(kInputTensor),
720                         context->getInputBuffer<_Float16>(kFilterTensor),
721                         context->getInputShape(kFilterTensor),
722                         context->getInputBuffer<_Float16>(kBiasTensor),
723                         context->getInputShape(kBiasTensor), param.padding_left,
724                         param.padding_right, param.padding_top, param.padding_bottom,
725                         param.stride_width, param.stride_height, param.dilation_width_factor,
726                         param.dilation_height_factor, param.activation, param.useNchw,
727                         context->getOutputBuffer<_Float16>(kOutputTensor),
728                         context->getOutputShape(kOutputTensor));
729         case OperandType::TENSOR_QUANT8_ASYMM:
730             if (context->getInputType(kFilterTensor) ==
731                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
732                 return convQuant8PerChannel(
733                         context->getInputBuffer<uint8_t>(kInputTensor),
734                         context->getInputShape(kInputTensor),
735                         context->getInputBuffer<int8_t>(kFilterTensor),
736                         context->getInputShape(kFilterTensor),
737                         std::get<Operand::SymmPerChannelQuantParams>(
738                                 context->getInputExtraParams(kFilterTensor))
739                                 .scales.data(),
740                         context->getInputBuffer<int32_t>(kBiasTensor),
741                         context->getInputShape(kBiasTensor), param.padding_left,
742                         param.padding_right, param.padding_top, param.padding_bottom,
743                         param.stride_width, param.stride_height, param.dilation_width_factor,
744                         param.dilation_height_factor, param.activation, param.useNchw,
745                         context->getOutputBuffer<uint8_t>(kOutputTensor),
746                         context->getOutputShape(kOutputTensor));
747             } else if (context->getInputType(kFilterTensor) == OperandType::TENSOR_QUANT8_ASYMM) {
748                 return conv(context->getInputBuffer<uint8_t>(kInputTensor),
749                             context->getInputShape(kInputTensor),
750                             context->getInputBuffer<uint8_t>(kFilterTensor),
751                             context->getInputShape(kFilterTensor),
752                             context->getInputBuffer<int32_t>(kBiasTensor),
753                             context->getInputShape(kBiasTensor), param.padding_left,
754                             param.padding_right, param.padding_top, param.padding_bottom,
755                             param.stride_width, param.stride_height, param.dilation_width_factor,
756                             param.dilation_height_factor, param.activation, param.useNchw,
757                             context->getOutputBuffer<uint8_t>(kOutputTensor),
758                             context->getOutputShape(kOutputTensor));
759             } else {
760                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
761             }
762         case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
763             if (context->getInputType(kFilterTensor) ==
764                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
765                 return convQuant8PerChannel(
766                         context->getInputBuffer<int8_t>(kInputTensor),
767                         context->getInputShape(kInputTensor),
768                         context->getInputBuffer<int8_t>(kFilterTensor),
769                         context->getInputShape(kFilterTensor),
770                         std::get<Operand::SymmPerChannelQuantParams>(
771                                 context->getInputExtraParams(kFilterTensor))
772                                 .scales.data(),
773                         context->getInputBuffer<int32_t>(kBiasTensor),
774                         context->getInputShape(kBiasTensor), param.padding_left,
775                         param.padding_right, param.padding_top, param.padding_bottom,
776                         param.stride_width, param.stride_height, param.dilation_width_factor,
777                         param.dilation_height_factor, param.activation, param.useNchw,
778                         context->getOutputBuffer<int8_t>(kOutputTensor),
779                         context->getOutputShape(kOutputTensor));
780             } else if (context->getInputType(kFilterTensor) ==
781                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
782                 return conv(context->getInputBuffer<int8_t>(kInputTensor),
783                             context->getInputShape(kInputTensor),
784                             context->getInputBuffer<int8_t>(kFilterTensor),
785                             context->getInputShape(kFilterTensor),
786                             context->getInputBuffer<int32_t>(kBiasTensor),
787                             context->getInputShape(kBiasTensor), param.padding_left,
788                             param.padding_right, param.padding_top, param.padding_bottom,
789                             param.stride_width, param.stride_height, param.dilation_width_factor,
790                             param.dilation_height_factor, param.activation, param.useNchw,
791                             context->getOutputBuffer<int8_t>(kOutputTensor),
792                             context->getOutputShape(kOutputTensor));
793             } else {
794                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
795             }
796         default:
797             NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
798     }
799 }
800 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
801 
802 }  // namespace conv_2d
803 
804 NN_REGISTER_OPERATION(CONV_2D, conv_2d::kOperationName, conv_2d::validate, conv_2d::prepare,
805                       conv_2d::execute, .allowZeroSizedInput = true);
806 
807 }  // namespace nn
808 }  // namespace android
809