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::__anon14bee14a0111::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 
needim2colData(const Shape & filterShape,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor)177 bool needim2colData(const Shape& filterShape, int32_t stride_width, int32_t stride_height,
178                     int32_t dilation_width_factor, int32_t dilation_height_factor) {
179     // Within tflite::optimized_ops::Conv, the following tests are performed,
180     // and in the case (!need_dilated_im2col && !need_im2col), then the
181     // method doesn't expect to receive outputData. In debug mode this is
182     // asserted and fails tests, so we need to perform this check as the caller
183     // also. See:
184     // tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h:2655
185     const int filter_width = getSizeOfDimension(filterShape, 2);
186     const int filter_height = getSizeOfDimension(filterShape, 1);
187     const bool need_dilated_im2col = dilation_width_factor != 1 || dilation_height_factor != 1;
188     const bool need_im2col =
189             stride_width != 1 || stride_height != 1 || filter_width != 1 || filter_height != 1;
190     return need_dilated_im2col || need_im2col;
191 }
192 
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)193 bool convNhwc(const float* inputData, const Shape& inputShape, const float* filterData,
194               const Shape& filterShape, const float* biasData, const Shape& biasShape,
195               int32_t padding_left, int32_t padding_right, int32_t padding_top,
196               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
197               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
198               float* outputData, const Shape& outputShape) {
199     NNTRACE_TRANS("convFloat32");
200 
201     ANDROID_NN_CONV_PARAMETERS(float)
202 
203     float output_activation_min, output_activation_max;
204     CalculateActivationRangeFloat(activation, &output_activation_min, &output_activation_max);
205 
206     // Prevent concurrent executions that may access the scratch buffer.
207     std::unique_lock<std::mutex> lock(executionMutex);
208     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
209 
210     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
211                                                 dilation_width_factor, dilation_height_factor);
212 
213     tflite::optimized_ops::Conv(
214             inputData, convertShapeToDims(inputShape), filterData, convertShapeToDims(filterShape),
215             biasData, convertShapeToDims(biasShape), stride_width, stride_height,
216             dilation_width_factor, dilation_height_factor, paddingWidth, paddingHeight,
217             output_activation_min, output_activation_max, outputData,
218             convertShapeToDims(outputShape), need_im2colData ? im2colData : nullptr, im2colDim);
219     return true;
220 }
221 
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)222 bool convNhwc(const uint8_t* inputData, const Shape& inputShape, const uint8_t* filterData,
223               const Shape& filterShape, const int32_t* biasData, const Shape& biasShape,
224               int32_t padding_left, int32_t padding_right, int32_t padding_top,
225               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
226               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
227               uint8_t* outputData, const Shape& outputShape) {
228     NNTRACE_TRANS("convQuant8");
229 
230     ANDROID_NN_CONV_PARAMETERS(uint8_t)
231 
232     int32_t inputOffset = -inputShape.offset;
233     int32_t filterOffset = -filterShape.offset;
234     int32_t outputOffset = outputShape.offset;
235 
236     double real_multiplier = 0.0;
237     int32_t output_multiplier = 0;
238     int32_t output_shift = 0;
239     int32_t output_activation_min = 0;
240     int32_t output_activation_max = 0;
241 
242     NN_RET_CHECK(GetQuantizedConvolutionMultipler(inputShape, filterShape, biasShape, outputShape,
243                                                   &real_multiplier));
244     int exponent;
245     NN_RET_CHECK(QuantizeMultiplier(real_multiplier, &output_multiplier, &exponent));
246     output_shift = -exponent;
247     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
248                                   &output_activation_max);
249 
250     static gemmlowp::GemmContext gemm_context;
251 
252     // Prevent concurrent executions that may access the scratch buffer and
253     // gemm_context.
254     std::unique_lock<std::mutex> lock(executionMutex);
255     // Alow gemmlowp automatically decide how many threads to use.
256     gemm_context.set_max_num_threads(0);
257 
258     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
259 
260     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
261                                                 dilation_width_factor, dilation_height_factor);
262 
263     tflite::optimized_ops::Conv(inputData, convertShapeToDims(inputShape), inputOffset, filterData,
264                                 convertShapeToDims(filterShape), filterOffset, biasData,
265                                 convertShapeToDims(biasShape), stride_width, stride_height,
266                                 dilation_width_factor, dilation_height_factor, paddingWidth,
267                                 paddingHeight, outputOffset, output_multiplier, output_shift,
268                                 output_activation_min, output_activation_max, outputData,
269                                 convertShapeToDims(outputShape),
270                                 need_im2colData ? im2colData : nullptr, im2colDim, &gemm_context);
271     return true;
272 }
273 
274 // Passing input, filter and output shapes by value, so that we can change the
275 // 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)276 bool convNhwc(const int8_t* inputData, Shape inputShape, const int8_t* filterData,
277               Shape filterShape, const int32_t* biasData, const Shape& biasShape,
278               int32_t padding_left, int32_t padding_right, int32_t padding_top,
279               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
280               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
281               int8_t* outputData, Shape outputShape) {
282     NNTRACE_TRANS("convQuant8");
283 
284     std::vector<uint8_t> unsignedInput(getNumberOfElements(inputShape));
285     convertInt8ToUInt8(inputData, &unsignedInput);
286     inputShape.offset += 128;
287 
288     std::vector<uint8_t> unsignedFilter(getNumberOfElements(filterShape));
289     convertInt8ToUInt8(filterData, &unsignedFilter);
290     filterShape.offset += 128;
291 
292     std::vector<uint8_t> unsignedOutput(getNumberOfElements(outputShape));
293     outputShape.offset += 128;
294 
295     NN_RET_CHECK(convNhwc(unsignedInput.data(), inputShape, unsignedFilter.data(), filterShape,
296                           biasData, biasShape, padding_left, padding_right, padding_top,
297                           padding_bottom, stride_width, stride_height, dilation_width_factor,
298                           dilation_height_factor, activation, unsignedOutput.data(), outputShape));
299 
300     convertUInt8ToInt8(unsignedOutput, outputData);
301 
302     return true;
303 }
304 
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)305 bool convNhwc(const _Float16* inputData, const Shape& inputShape, const _Float16* filterData,
306               const Shape& filterShape, const _Float16* biasData, const Shape& biasShape,
307               int32_t padding_left, int32_t padding_right, int32_t padding_top,
308               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
309               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
310               _Float16* outputData, const Shape& outputShape) {
311     NNTRACE_TRANS("convFloat16");
312 
313     std::vector<float> inputData_float32(getNumberOfElements(inputShape));
314     std::vector<float> filterData_float32(getNumberOfElements(filterShape));
315     std::vector<float> biasData_float32(getNumberOfElements(biasShape));
316     std::vector<float> outputData_float32(getNumberOfElements(outputShape));
317 
318     convertFloat16ToFloat32(inputData, &inputData_float32);
319     convertFloat16ToFloat32(filterData, &filterData_float32);
320     convertFloat16ToFloat32(biasData, &biasData_float32);
321 
322     convNhwc(inputData_float32.data(), inputShape, filterData_float32.data(), filterShape,
323              biasData_float32.data(), biasShape, padding_left, padding_right, padding_top,
324              padding_bottom, stride_width, stride_height, dilation_width_factor,
325              dilation_height_factor, activation, outputData_float32.data(), outputShape);
326     convertFloat32ToFloat16(outputData_float32, outputData);
327 
328     return true;
329 }
330 
331 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)332 bool conv(const T_Input* inputData, const Shape& inputShape, const T_Filter* filterData,
333           const Shape& filterShape, const T_Bias* biasData, const Shape& biasShape,
334           int32_t padding_left, int32_t padding_right, int32_t padding_top, int32_t padding_bottom,
335           int32_t stride_width, int32_t stride_height, int32_t dilation_width_factor,
336           int32_t dilation_height_factor, int32_t activation, bool useNchw, T_Input* outputData,
337           const Shape& outputShape) {
338     InputWithLayout<T_Input> input(useNchw);
339     OutputWithLayout<T_Input> output(useNchw);
340     NN_RET_CHECK(input.initialize(inputData, inputShape));
341     NN_RET_CHECK(output.initialize(outputData, outputShape));
342     NN_RET_CHECK(convNhwc(input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape,
343                           biasData, biasShape, padding_left, padding_right, padding_top,
344                           padding_bottom, stride_width, stride_height, dilation_width_factor,
345                           dilation_height_factor, activation, output.getNhwcBuffer(),
346                           output.getNhwcShape()));
347     NN_RET_CHECK(output.commit());
348     return true;
349 }
350 
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)351 bool convQuant8PerChannelNhwc(const uint8_t* inputData, const Shape& inputShape,
352                               const int8_t* filterData, const Shape& filterShape,
353                               const float* filterScales, const int32_t* biasData,
354                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
355                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
356                               int32_t strideHeight, int32_t dilationWidthFactor,
357                               int32_t dilationHeightFactor, int32_t activation, uint8_t* outputData,
358                               const Shape& outputShape) {
359     NNTRACE_TRANS("convQuant8PerChannel");
360 
361     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
362     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
363     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
364     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
365     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
366     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
367     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
368     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
369     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
370     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
371 
372     int32_t inputOffset = -inputShape.offset;
373     int32_t outputOffset = outputShape.offset;
374 
375     auto realMultiplier = std::vector<double>(outputDepth, .0f);
376     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
377     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
378 
379     for (int i = 0; i < outputDepth; ++i) {
380         Shape filterChannelShape = filterShape;
381         filterChannelShape.scale = filterScales[i];
382         Shape biasChannelShape = biasShape;
383         biasChannelShape.scale = filterScales[i] * inputShape.scale;
384         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
385                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
386         int exponent;
387         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &exponent));
388         outputShift[i] = -exponent;
389     }
390 
391     int32_t output_activation_min = 0, output_activation_max = 0;
392     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
393                                   &output_activation_max);
394     const uint8_t* inputBase = inputData;
395     uint8_t* outPtr = outputData;
396     for (uint32_t b = 0; b < numBatches; b++) {
397         for (uint32_t h = 0; h < outputHeight; h++) {
398             for (uint32_t w = 0; w < outputWidth; w++) {
399                 const int8_t* filterBase = filterData;
400 
401                 for (uint32_t d = 0; d < outputDepth; d++) {
402                     int32_t wInputOrigin = static_cast<int32_t>(w) * strideWidth - paddingLeft;
403                     int32_t hInputOrigin = static_cast<int32_t>(h) * strideHeight - paddingTop;
404                     int32_t sum = 0.0f;
405 
406                     for (uint32_t i = 0; i < filterHeight; i++) {
407                         for (uint32_t j = 0; j < filterWidth; j++) {
408                             for (uint32_t k = 0; k < filterDepth; k++) {
409                                 int32_t hInput = hInputOrigin +
410                                                  dilationHeightFactor * static_cast<int32_t>(i);
411                                 int32_t wInput = wInputOrigin +
412                                                  dilationWidthFactor * static_cast<int32_t>(j);
413                                 uint32_t dInput = k;
414                                 if (hInput >= 0 && hInput < static_cast<int32_t>(inputHeight) &&
415                                     wInput >= 0 && wInput < static_cast<int32_t>(inputWidth)) {
416                                     uint32_t filterIndex =
417                                             i * filterWidth * filterDepth + j * filterDepth + k;
418                                     uint32_t inputIndex = hInput * inputWidth * inputDepth +
419                                                           wInput * inputDepth + dInput;
420                                     sum += (static_cast<int32_t>(filterBase[filterIndex])) *
421                                            (static_cast<int32_t>(inputBase[inputIndex]) +
422                                             inputOffset);
423                                 }
424                             }
425                         }
426                     }
427                     sum += biasData[d];
428                     sum = tflite::MultiplyByQuantizedMultiplier(sum, outputMultiplier[d],
429                                                                 -outputShift[d]);
430                     sum += outputOffset;
431                     sum = std::max(std::min(sum, output_activation_max), output_activation_min);
432                     outPtr[d] = static_cast<uint8_t>(sum);
433                     filterBase += filterHeight * filterWidth * filterDepth;
434                 }
435                 outPtr += outputDepth;
436             }
437         }
438         inputBase += inputHeight * inputWidth * inputDepth;
439     }
440 
441     return true;
442 }
443 
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)444 bool convQuant8PerChannelNhwc(const int8_t* inputData, const Shape& inputShape,
445                               const int8_t* filterData, const Shape& filterShape,
446                               const float* filterScales, const int32_t* biasData,
447                               const Shape& biasShape, int32_t paddingLeft, int32_t paddingRight,
448                               int32_t paddingTop, int32_t paddingBottom, int32_t strideWidth,
449                               int32_t strideHeight, int32_t dilationWidthFactor,
450                               int32_t dilationHeightFactor, int32_t activation, int8_t* outputData,
451                               const Shape& outputShape) {
452     NNTRACE_TRANS("convQuant8SignedPerChannel");
453 
454     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
455     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
456     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
457     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
458     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
459     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
460     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
461     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
462     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
463     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
464 
465     int32_t inputOffset = -inputShape.offset;
466     int32_t outputOffset = outputShape.offset;
467 
468     auto realMultiplier = std::vector<double>(outputDepth, .0f);
469     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
470     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
471 
472     for (int i = 0; i < outputDepth; ++i) {
473         Shape filterChannelShape = filterShape;
474         filterChannelShape.scale = filterScales[i];
475         Shape biasChannelShape = biasShape;
476         biasChannelShape.scale = filterScales[i] * inputShape.scale;
477         NN_RET_CHECK(GetQuantizedConvolutionMultipler(
478                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
479         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &outputShift[i]));
480     }
481 
482     int32_t output_activation_min = 0, output_activation_max = 0;
483     CalculateActivationRangeInt8(activation, outputShape, &output_activation_min,
484                                  &output_activation_max);
485 
486     tflite::ConvParams convParams;
487     convParams.input_offset = -inputShape.offset;
488     convParams.output_offset = outputShape.offset;
489     convParams.stride_height = strideHeight;
490     convParams.stride_width = strideWidth;
491     convParams.dilation_height_factor = dilationHeightFactor;
492     convParams.dilation_width_factor = dilationWidthFactor;
493     convParams.padding_values.height = paddingTop;
494     convParams.padding_values.width = paddingLeft;
495     convParams.quantized_activation_min = output_activation_min;
496     convParams.quantized_activation_max = output_activation_max;
497 
498     NNTRACE_COMP_SWITCH("reference_integer_ops::ConvPerChannel");
499     tflite::reference_integer_ops::ConvPerChannel(
500             convParams, outputMultiplier.data(), outputShift.data(),
501             convertShapeToTflshape(inputShape), inputData, convertShapeToTflshape(filterShape),
502             filterData, convertShapeToTflshape(biasShape), biasData,
503             convertShapeToTflshape(outputShape), outputData);
504     return true;
505 }
506 
507 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)508 bool convQuant8PerChannel(const T* inputData, const Shape& inputShape, const int8_t* filterData,
509                           const Shape& filterShape, const float* filterScales,
510                           const int32_t* biasData, const Shape& biasShape, int32_t paddingLeft,
511                           int32_t paddingRight, int32_t paddingTop, int32_t paddingBottom,
512                           int32_t strideWidth, int32_t strideHeight, int32_t dilationWidthFactor,
513                           int32_t dilationHeightFactor, int32_t activation, bool useNchw,
514                           T* outputData, const Shape& outputShape) {
515     InputWithLayout<T> input(useNchw);
516     OutputWithLayout<T> output(useNchw);
517     NN_RET_CHECK(input.initialize(inputData, inputShape));
518     NN_RET_CHECK(output.initialize(outputData, outputShape));
519     NN_RET_CHECK(convQuant8PerChannelNhwc(
520             input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape, filterScales,
521             biasData, biasShape, paddingLeft, paddingRight, paddingTop, paddingBottom, strideWidth,
522             strideHeight, dilationWidthFactor, dilationHeightFactor, activation,
523             output.getNhwcBuffer(), output.getNhwcShape()));
524     NN_RET_CHECK(output.commit());
525     return true;
526 }
527 
528 #undef ANDROID_NN_CONV_PARAMETERS
529 
530 }  // namespace
531 
validate(const IOperationValidationContext * context)532 bool validate(const IOperationValidationContext* context) {
533     const uint32_t numInputs = context->getNumInputs();
534     NN_RET_CHECK(
535             std::binary_search(std::begin(kNumInputsArray), std::end(kNumInputsArray), numInputs));
536     NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
537     const auto inputRank = getNumberOfDimensions(context->getInputShape(kInputTensor));
538     const auto filterRank = getNumberOfDimensions(context->getInputShape(kFilterTensor));
539     if (inputRank != 0) {
540         NN_RET_CHECK_EQ(inputRank, 4);
541     }
542     if (filterRank != 0) {
543         NN_RET_CHECK_EQ(filterRank, 4);
544     }
545     auto inputCount = context->getNumInputs();
546     auto inputType = context->getInputType(kInputTensor);
547     auto filterType = context->getInputType(kFilterTensor);
548     std::vector<OperandType> inExpectedTypes;
549     if (inputType == OperandType::TENSOR_FLOAT32) {
550         inExpectedTypes = {OperandType::TENSOR_FLOAT32, OperandType::TENSOR_FLOAT32,
551                            OperandType::TENSOR_FLOAT32, OperandType::INT32,
552                            OperandType::INT32,          OperandType::INT32,
553                            OperandType::INT32};
554     } else if (inputType == OperandType::TENSOR_FLOAT16) {
555         inExpectedTypes = {OperandType::TENSOR_FLOAT16, OperandType::TENSOR_FLOAT16,
556                            OperandType::TENSOR_FLOAT16, OperandType::INT32,
557                            OperandType::INT32,          OperandType::INT32,
558                            OperandType::INT32};
559     } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM ||
560                inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
561         NN_RET_CHECK(filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL ||
562                      filterType == inputType)
563                 << "Unsupported filter tensor type for operation " << kOperationName;
564         inExpectedTypes = {inputType,          filterType,         OperandType::TENSOR_INT32,
565                            OperandType::INT32, OperandType::INT32, OperandType::INT32,
566                            OperandType::INT32};
567 
568         if (filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
569             NN_RET_CHECK_EQ(context->getInputExtraParams(kFilterTensor).channelQuant().channelDim,
570                             0)
571                     << "Unsupported filter tensor channel dimension for operation "
572                     << kOperationName;
573         }
574     } else {
575         NN_RET_CHECK_FAIL() << "Unsupported input tensor type for operation " << kOperationName;
576     }
577 
578     // NeuralNetworks.h specifies that ANEURALNETWORKS_CONV_2D's output must
579     // meet "outputScale > inputScale * filterScale" for the operand type
580     // ANEURALNETWORKS_TENSOR_QUANT8_ASYMM before API level 29. For other
581     // operand types (e.g., ANEURALNETWORKS_TENSOR_FLOAT32), this constraint
582     // does not apply, so by default the constraint is met.
583     bool meetsQuantizedScaleConstraintBeforeV1_2 = true;
584     if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
585         const float inputScale = context->getInputShape(kInputTensor).scale;
586         const float filterScale = context->getInputShape(kFilterTensor).scale;
587         const float outputScale = context->getInputShape(kOutputTensor).scale;
588         meetsQuantizedScaleConstraintBeforeV1_2 = (outputScale > inputScale * filterScale);
589     }
590 
591     bool withExplicitPadding = false;
592     bool withLayout = false;
593     bool withDilation = false;
594     if (inputCount >= 8) {
595         if (context->getInputType(7) == OperandType::INT32 && inputCount >= 10) {
596             std::vector<OperandType> explicitScalarTypes(3, OperandType::INT32);
597             inExpectedTypes.insert(inExpectedTypes.end(), explicitScalarTypes.begin(),
598                                    explicitScalarTypes.end());
599             withExplicitPadding = true;
600         }
601         int inputOffset = withExplicitPadding ? 3 : 0;
602         if (inputCount >= 8 + inputOffset) {
603             inExpectedTypes.push_back(OperandType::BOOL);
604             withLayout = true;
605         }
606         NN_RET_CHECK_NE(inputCount, 9 + inputOffset)
607                 << "Provided only one dilation factor value, two values are requred for operation "
608                 << kOperationName;
609         if (inputCount == 10 + inputOffset) {
610             inExpectedTypes.push_back(OperandType::INT32);
611             inExpectedTypes.push_back(OperandType::INT32);
612             withDilation = true;
613         }
614     }
615 
616     if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
617         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_3));
618     } else if (inputType == OperandType::TENSOR_FLOAT16 ||
619                filterType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL || withLayout ||
620                withDilation || !meetsQuantizedScaleConstraintBeforeV1_2) {
621         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_2));
622     } else {
623         NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_0));
624     }
625     return validateInputTypes(context, inExpectedTypes) &&
626            validateOutputTypes(context, {inputType});
627 }
628 
prepare(IOperationExecutionContext * context)629 bool prepare(IOperationExecutionContext* context) {
630     Shape input = context->getInputShape(kInputTensor);
631     Shape filter = context->getInputShape(kFilterTensor);
632     Shape bias = context->getInputShape(kBiasTensor);
633 
634     if (filter.type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
635         NN_RET_CHECK(input.type == OperandType::TENSOR_QUANT8_ASYMM ||
636                      input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED);
637     } else {
638         NN_RET_CHECK(input.type == filter.type);
639     }
640     if (input.type == OperandType::TENSOR_QUANT8_ASYMM ||
641         input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
642         NN_RET_CHECK(bias.type == OperandType::TENSOR_INT32);
643     } else {
644         NN_RET_CHECK(input.type == bias.type);
645     }
646     NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4);
647     NN_RET_CHECK_EQ(getNumberOfDimensions(filter), 4);
648     NN_RET_CHECK_EQ(getNumberOfDimensions(bias), 1);
649 
650     Conv2dParam param;
651     NN_RET_CHECK(param.initialize(context));
652 
653     uint32_t batches = getSizeOfDimension(input, 0);
654     uint32_t height = getSizeOfDimension(input, param.useNchw ? 2 : 1);
655     uint32_t width = getSizeOfDimension(input, param.useNchw ? 3 : 2);
656     uint32_t channels_in = getSizeOfDimension(input, param.useNchw ? 1 : 3);
657     uint32_t channels_out = getSizeOfDimension(filter, 0);
658     uint32_t filterHeight = getSizeOfDimension(filter, 1);
659     uint32_t filterWidth = getSizeOfDimension(filter, 2);
660     // Only batches can be zero.
661     NN_RET_CHECK_EQ(channels_in, getSizeOfDimension(filter, 3));
662     NN_RET_CHECK_EQ(channels_out, getSizeOfDimension(bias, 0));
663     NN_RET_CHECK_GT(height, 0);
664     NN_RET_CHECK_GT(width, 0);
665     NN_RET_CHECK_GT(channels_in, 0);
666     NN_RET_CHECK_GT(channels_out, 0);
667 
668     int32_t effectiveFilterWidth = (filterWidth - 1) * param.dilation_width_factor + 1;
669     int32_t effectiveFilterHeight = (filterHeight - 1) * param.dilation_height_factor + 1;
670     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_left);
671     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_right);
672     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_top);
673     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_bottom);
674 
675     uint32_t outWidth =
676             computeOutSize(width, filterWidth, param.stride_width, param.dilation_width_factor,
677                            param.padding_left, param.padding_right);
678     uint32_t outHeight =
679             computeOutSize(height, filterHeight, param.stride_height, param.dilation_height_factor,
680                            param.padding_top, param.padding_bottom);
681 
682     Shape output = context->getOutputShape(kOutputTensor);
683     output.type = input.type;
684     if (param.useNchw) {
685         output.dimensions = {batches, channels_out, outHeight, outWidth};
686     } else {
687         output.dimensions = {batches, outHeight, outWidth, channels_out};
688     }
689     return context->setOutputShape(kOutputTensor, output);
690 }
691 
execute(IOperationExecutionContext * context)692 bool execute(IOperationExecutionContext* context) {
693     // Bypass execution in the case of zero-sized input.
694     if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
695     Conv2dParam param;
696     NN_RET_CHECK(param.initialize(context));
697     switch (context->getInputType(kInputTensor)) {
698         case OperandType::TENSOR_FLOAT32:
699             return conv(context->getInputBuffer<float>(kInputTensor),
700                         context->getInputShape(kInputTensor),
701                         context->getInputBuffer<float>(kFilterTensor),
702                         context->getInputShape(kFilterTensor),
703                         context->getInputBuffer<float>(kBiasTensor),
704                         context->getInputShape(kBiasTensor), param.padding_left,
705                         param.padding_right, param.padding_top, param.padding_bottom,
706                         param.stride_width, param.stride_height, param.dilation_width_factor,
707                         param.dilation_height_factor, param.activation, param.useNchw,
708                         context->getOutputBuffer<float>(kOutputTensor),
709                         context->getOutputShape(kOutputTensor));
710         case OperandType::TENSOR_FLOAT16:
711             return conv(context->getInputBuffer<_Float16>(kInputTensor),
712                         context->getInputShape(kInputTensor),
713                         context->getInputBuffer<_Float16>(kFilterTensor),
714                         context->getInputShape(kFilterTensor),
715                         context->getInputBuffer<_Float16>(kBiasTensor),
716                         context->getInputShape(kBiasTensor), param.padding_left,
717                         param.padding_right, param.padding_top, param.padding_bottom,
718                         param.stride_width, param.stride_height, param.dilation_width_factor,
719                         param.dilation_height_factor, param.activation, param.useNchw,
720                         context->getOutputBuffer<_Float16>(kOutputTensor),
721                         context->getOutputShape(kOutputTensor));
722         case OperandType::TENSOR_QUANT8_ASYMM:
723             if (context->getInputType(kFilterTensor) ==
724                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
725                 return convQuant8PerChannel(
726                         context->getInputBuffer<uint8_t>(kInputTensor),
727                         context->getInputShape(kInputTensor),
728                         context->getInputBuffer<int8_t>(kFilterTensor),
729                         context->getInputShape(kFilterTensor),
730                         context->getInputExtraParams(kFilterTensor).channelQuant().scales.data(),
731                         context->getInputBuffer<int32_t>(kBiasTensor),
732                         context->getInputShape(kBiasTensor), param.padding_left,
733                         param.padding_right, param.padding_top, param.padding_bottom,
734                         param.stride_width, param.stride_height, param.dilation_width_factor,
735                         param.dilation_height_factor, param.activation, param.useNchw,
736                         context->getOutputBuffer<uint8_t>(kOutputTensor),
737                         context->getOutputShape(kOutputTensor));
738             } else if (context->getInputType(kFilterTensor) == OperandType::TENSOR_QUANT8_ASYMM) {
739                 return conv(context->getInputBuffer<uint8_t>(kInputTensor),
740                             context->getInputShape(kInputTensor),
741                             context->getInputBuffer<uint8_t>(kFilterTensor),
742                             context->getInputShape(kFilterTensor),
743                             context->getInputBuffer<int32_t>(kBiasTensor),
744                             context->getInputShape(kBiasTensor), param.padding_left,
745                             param.padding_right, param.padding_top, param.padding_bottom,
746                             param.stride_width, param.stride_height, param.dilation_width_factor,
747                             param.dilation_height_factor, param.activation, param.useNchw,
748                             context->getOutputBuffer<uint8_t>(kOutputTensor),
749                             context->getOutputShape(kOutputTensor));
750             } else {
751                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
752             }
753         case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
754             if (context->getInputType(kFilterTensor) ==
755                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
756                 return convQuant8PerChannel(
757                         context->getInputBuffer<int8_t>(kInputTensor),
758                         context->getInputShape(kInputTensor),
759                         context->getInputBuffer<int8_t>(kFilterTensor),
760                         context->getInputShape(kFilterTensor),
761                         context->getInputExtraParams(kFilterTensor).channelQuant().scales.data(),
762                         context->getInputBuffer<int32_t>(kBiasTensor),
763                         context->getInputShape(kBiasTensor), param.padding_left,
764                         param.padding_right, param.padding_top, param.padding_bottom,
765                         param.stride_width, param.stride_height, param.dilation_width_factor,
766                         param.dilation_height_factor, param.activation, param.useNchw,
767                         context->getOutputBuffer<int8_t>(kOutputTensor),
768                         context->getOutputShape(kOutputTensor));
769             } else if (context->getInputType(kFilterTensor) ==
770                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
771                 return conv(context->getInputBuffer<int8_t>(kInputTensor),
772                             context->getInputShape(kInputTensor),
773                             context->getInputBuffer<int8_t>(kFilterTensor),
774                             context->getInputShape(kFilterTensor),
775                             context->getInputBuffer<int32_t>(kBiasTensor),
776                             context->getInputShape(kBiasTensor), param.padding_left,
777                             param.padding_right, param.padding_top, param.padding_bottom,
778                             param.stride_width, param.stride_height, param.dilation_width_factor,
779                             param.dilation_height_factor, param.activation, param.useNchw,
780                             context->getOutputBuffer<int8_t>(kOutputTensor),
781                             context->getOutputShape(kOutputTensor));
782             } else {
783                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
784             }
785         default:
786             NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
787     }
788 }
789 
790 }  // namespace conv_2d
791 
792 NN_REGISTER_OPERATION(CONV_2D, conv_2d::kOperationName, conv_2d::validate, conv_2d::prepare,
793                       conv_2d::execute, .allowZeroSizedInput = true);
794 
795 }  // namespace nn
796 }  // namespace android
797