• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2017 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #define LOG_TAG "Operations"
18 
19 #include "Conv2D.h"
20 
21 #include <algorithm>
22 #include <iterator>
23 #include <memory>
24 #include <vector>
25 
26 #include "LegacyUtils.h"
27 #include "OperationResolver.h"
28 #include "Operations.h"
29 #include "OperationsExecutionUtils.h"
30 #include "Tracing.h"
31 
32 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
33 #pragma clang diagnostic push
34 #pragma clang diagnostic ignored "-Wunused-parameter"
35 #pragma clang diagnostic ignored "-Wsign-compare"
36 #pragma clang diagnostic ignored "-Winvalid-partial-specialization"
37 #include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
38 #include <tensorflow/lite/kernels/internal/reference/integer_ops/conv.h>
39 #include <tensorflow/lite/kernels/internal/types.h>
40 #pragma clang diagnostic pop
41 
42 #include "CpuOperationUtils.h"
43 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
44 
45 namespace android {
46 namespace nn {
47 namespace conv_2d {
48 
49 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
50 namespace {
51 
52 // If possible we will use this static buffer for the tensor.
53 constexpr size_t kStaticBufferSize = 1605632;
54 [[maybe_unused]] char static_scratch_buffer[kStaticBufferSize];
55 
56 // executionMutex is used to protect concurrent access of the static_scratch_buffer
57 // and other non-threadsafe resources like gemmlowp::GemmContext.
58 // std::mutex is safe for pthreads on Android.
59 std::mutex executionMutex;
60 
61 struct Conv2dParam {
62     int32_t padding_left, padding_right;
63     int32_t padding_top, padding_bottom;
64     int32_t stride_width, stride_height;
65     int32_t dilation_width_factor = 1, dilation_height_factor = 1;
66     int32_t activation;
67     bool useNchw = false;
68 
initializeandroid::nn::conv_2d::__anon72a8f1430111::Conv2dParam69     bool initialize(const IOperationExecutionContext* context) {
70         uint32_t inCount = context->getNumInputs();
71         int32_t padding_implicit = 0;
72         bool useImplicitPadding = false;
73         if ((inCount >= 8 && context->getInputType(7) == OperandType::BOOL) || inCount == 7) {
74             padding_implicit = context->getInputValue<int32_t>(3);
75             stride_width = context->getInputValue<int32_t>(4);
76             stride_height = context->getInputValue<int32_t>(5);
77             activation = context->getInputValue<int32_t>(6);
78             if (inCount >= 8) {
79                 useNchw = context->getInputValue<bool>(7);
80             }
81             if (inCount == 10) {
82                 dilation_width_factor = context->getInputValue<int32_t>(8);
83                 dilation_height_factor = context->getInputValue<int32_t>(9);
84             }
85             useImplicitPadding = true;
86         } else if (inCount >= 10 && context->getInputType(7) == OperandType::INT32) {
87             padding_left = context->getInputValue<int32_t>(3);
88             padding_right = context->getInputValue<int32_t>(4);
89             padding_top = context->getInputValue<int32_t>(5);
90             padding_bottom = context->getInputValue<int32_t>(6);
91             stride_width = context->getInputValue<int32_t>(7);
92             stride_height = context->getInputValue<int32_t>(8);
93             activation = context->getInputValue<int32_t>(9);
94             if (inCount >= 11) {
95                 useNchw = context->getInputValue<bool>(10);
96             }
97             if (inCount == 13) {
98                 dilation_width_factor = context->getInputValue<int32_t>(11);
99                 dilation_height_factor = context->getInputValue<int32_t>(12);
100             }
101         } else {
102             NN_RET_CHECK_FAIL() << "Unsupported input spec for operation " << kOperationName;
103         }
104         if (useImplicitPadding) {
105             Shape inputShape = context->getInputShape(kInputTensor);
106             Shape filterShape = context->getInputShape(kFilterTensor);
107             int32_t input_width = getSizeOfDimension(inputShape, useNchw ? 3 : 2);
108             int32_t input_height = getSizeOfDimension(inputShape, useNchw ? 2 : 1);
109             int32_t filter_width = getSizeOfDimension(filterShape, 2);
110             int32_t filter_height = getSizeOfDimension(filterShape, 1);
111             calculateExplicitPadding(input_width, stride_width, dilation_width_factor, filter_width,
112                                      padding_implicit, &padding_left, &padding_right);
113             calculateExplicitPadding(input_height, stride_height, dilation_height_factor,
114                                      filter_height, padding_implicit, &padding_top,
115                                      &padding_bottom);
116         }
117         NN_RET_CHECK_GE(padding_left, 0);
118         NN_RET_CHECK_GE(padding_right, 0);
119         NN_RET_CHECK_GE(padding_top, 0);
120         NN_RET_CHECK_GE(padding_bottom, 0);
121         NN_RET_CHECK_GT(stride_width, 0);
122         NN_RET_CHECK_GT(stride_height, 0);
123         NN_RET_CHECK_GT(dilation_width_factor, 0);
124         NN_RET_CHECK_GT(dilation_height_factor, 0);
125         NN_RET_CHECK_GE(activation, 0);
126         return true;
127     }
128 };
129 
130 #define ANDROID_NN_CONV_PARAMETERS(Type)                                          \
131     [[maybe_unused]] uint32_t height = getSizeOfDimension(inputShape, 1);         \
132     [[maybe_unused]] uint32_t width = getSizeOfDimension(inputShape, 2);          \
133     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);                   \
134     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);                    \
135     [[maybe_unused]] uint32_t outHeight = getSizeOfDimension(outputShape, 1);     \
136     [[maybe_unused]] uint32_t outWidth = getSizeOfDimension(outputShape, 2);      \
137     uint32_t inDepth = getSizeOfDimension(inputShape, 3);                         \
138                                                                                   \
139     uint32_t paddingHeight = (uint32_t)padding_top;                               \
140     uint32_t paddingWidth = (uint32_t)padding_left;                               \
141                                                                                   \
142     tflite::Dims<4> im2colDim;                                                    \
143     im2colDim.sizes[3] = (int)getSizeOfDimension(outputShape, 0);                 \
144     im2colDim.sizes[2] = (int)getSizeOfDimension(outputShape, 1);                 \
145     im2colDim.sizes[1] = (int)getSizeOfDimension(outputShape, 2);                 \
146     im2colDim.sizes[0] = (int)inDepth * filterHeight * filterWidth;               \
147                                                                                   \
148     im2colDim.strides[0] = 1;                                                     \
149     for (int i = 1; i < 4; i++) {                                                 \
150         im2colDim.strides[i] = im2colDim.strides[i - 1] * im2colDim.sizes[i - 1]; \
151     }                                                                             \
152                                                                                   \
153     Type* im2colData = nullptr;                                                   \
154     uint64_t im2colByteSize = sizeof(Type);                                       \
155     std::unique_ptr<Type[]> im2colGuard;                                          \
156     for (int i = 0; i < 4; i++) {                                                 \
157         im2colByteSize *= im2colDim.sizes[i];                                     \
158     }                                                                             \
159     /* http://b/77982879, tflite::optimized_ops::Conv uses int for offsets */     \
160     if (im2colByteSize >= 0x7fffffff) {                                           \
161         LOG(ERROR) << "Conv size is too large, not enough memory";                \
162         return false;                                                             \
163     }                                                                             \
164     if (im2colByteSize <= kStaticBufferSize) {                                    \
165         im2colData = reinterpret_cast<Type*>(static_scratch_buffer);              \
166     } else {                                                                      \
167         im2colData = new (std::nothrow) Type[im2colByteSize / sizeof(Type)];      \
168         if (im2colData == nullptr) {                                              \
169             LOG(ERROR) << "Conv size is too large, not enough memory";            \
170             return false;                                                         \
171         }                                                                         \
172         im2colGuard.reset(im2colData);                                            \
173     }
174 
needim2colData(const Shape & filterShape,int32_t stride_width,int32_t stride_height,int32_t dilation_width_factor,int32_t dilation_height_factor)175 bool needim2colData(const Shape& filterShape, int32_t stride_width, int32_t stride_height,
176                     int32_t dilation_width_factor, int32_t dilation_height_factor) {
177     // Within tflite::optimized_ops::Conv, the following tests are performed,
178     // and in the case (!need_dilated_im2col && !need_im2col), then the
179     // method doesn't expect to receive outputData. In debug mode this is
180     // asserted and fails tests, so we need to perform this check as the caller
181     // also. See:
182     // tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h:2655
183     const int filter_width = getSizeOfDimension(filterShape, 2);
184     const int filter_height = getSizeOfDimension(filterShape, 1);
185     const bool need_dilated_im2col = dilation_width_factor != 1 || dilation_height_factor != 1;
186     const bool need_im2col =
187             stride_width != 1 || stride_height != 1 || filter_width != 1 || filter_height != 1;
188     return need_dilated_im2col || need_im2col;
189 }
190 
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,int32_t padding_top,int32_t,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)191 bool convNhwc(const float* inputData, const Shape& inputShape, const float* filterData,
192               const Shape& filterShape, const float* biasData, const Shape& biasShape,
193               int32_t padding_left, int32_t /*padding_right*/, int32_t padding_top,
194               int32_t /*padding_bottom*/, int32_t stride_width, int32_t stride_height,
195               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
196               float* outputData, const Shape& outputShape) {
197     NNTRACE_TRANS("convFloat32");
198 
199     ANDROID_NN_CONV_PARAMETERS(float)
200 
201     float output_activation_min, output_activation_max;
202     CalculateActivationRangeFloat(activation, &output_activation_min, &output_activation_max);
203 
204     // Prevent concurrent executions that may access the scratch buffer.
205     std::unique_lock<std::mutex> lock(executionMutex);
206     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
207 
208     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
209                                                 dilation_width_factor, dilation_height_factor);
210 
211     tflite::optimized_ops::Conv(
212             inputData, convertShapeToDims(inputShape), filterData, convertShapeToDims(filterShape),
213             biasData, convertShapeToDims(biasShape), stride_width, stride_height,
214             dilation_width_factor, dilation_height_factor, paddingWidth, paddingHeight,
215             output_activation_min, output_activation_max, outputData,
216             convertShapeToDims(outputShape), need_im2colData ? im2colData : nullptr, im2colDim);
217     return true;
218 }
219 
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,int32_t padding_top,int32_t,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)220 bool convNhwc(const uint8_t* inputData, const Shape& inputShape, const uint8_t* filterData,
221               const Shape& filterShape, const int32_t* biasData, const Shape& biasShape,
222               int32_t padding_left, int32_t /*padding_right*/, int32_t padding_top,
223               int32_t /*padding_bottom*/, int32_t stride_width, int32_t stride_height,
224               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
225               uint8_t* outputData, const Shape& outputShape) {
226     NNTRACE_TRANS("convQuant8");
227 
228     ANDROID_NN_CONV_PARAMETERS(uint8_t)
229 
230     int32_t inputOffset = -inputShape.offset;
231     int32_t filterOffset = -filterShape.offset;
232     int32_t outputOffset = outputShape.offset;
233 
234     double real_multiplier = 0.0;
235     int32_t output_multiplier = 0;
236     int32_t output_shift = 0;
237     int32_t output_activation_min = 0;
238     int32_t output_activation_max = 0;
239 
240     NN_RET_CHECK(GetQuantizedConvolutionMultiplier(inputShape, filterShape, biasShape, outputShape,
241                                                    &real_multiplier));
242     int exponent;
243     NN_RET_CHECK(QuantizeMultiplier(real_multiplier, &output_multiplier, &exponent));
244     output_shift = -exponent;
245     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
246                                   &output_activation_max);
247 
248     static gemmlowp::GemmContext gemm_context;
249 
250     // Prevent concurrent executions that may access the scratch buffer and
251     // gemm_context.
252     std::unique_lock<std::mutex> lock(executionMutex);
253     // Alow gemmlowp automatically decide how many threads to use.
254     gemm_context.set_max_num_threads(0);
255 
256     NNTRACE_COMP_SWITCH("optimized_ops::Conv");
257 
258     const bool need_im2colData = needim2colData(filterShape, stride_width, stride_height,
259                                                 dilation_width_factor, dilation_height_factor);
260 
261     tflite::optimized_ops::Conv(inputData, convertShapeToDims(inputShape), inputOffset, filterData,
262                                 convertShapeToDims(filterShape), filterOffset, biasData,
263                                 convertShapeToDims(biasShape), stride_width, stride_height,
264                                 dilation_width_factor, dilation_height_factor, paddingWidth,
265                                 paddingHeight, outputOffset, output_multiplier, output_shift,
266                                 output_activation_min, output_activation_max, outputData,
267                                 convertShapeToDims(outputShape),
268                                 need_im2colData ? im2colData : nullptr, im2colDim, &gemm_context);
269     return true;
270 }
271 
272 // Passing input, filter and output shapes by value, so that we can change the
273 // 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)274 bool convNhwc(const int8_t* inputData, Shape inputShape, const int8_t* filterData,
275               Shape filterShape, const int32_t* biasData, const Shape& biasShape,
276               int32_t padding_left, int32_t padding_right, int32_t padding_top,
277               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
278               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
279               int8_t* outputData, Shape outputShape) {
280     NNTRACE_TRANS("convQuant8");
281 
282     std::vector<uint8_t> unsignedInput(getNumberOfElements(inputShape));
283     convertInt8ToUInt8(inputData, &unsignedInput);
284     inputShape.offset += 128;
285 
286     std::vector<uint8_t> unsignedFilter(getNumberOfElements(filterShape));
287     convertInt8ToUInt8(filterData, &unsignedFilter);
288     filterShape.offset += 128;
289 
290     std::vector<uint8_t> unsignedOutput(getNumberOfElements(outputShape));
291     outputShape.offset += 128;
292 
293     NN_RET_CHECK(convNhwc(unsignedInput.data(), inputShape, unsignedFilter.data(), filterShape,
294                           biasData, biasShape, padding_left, padding_right, padding_top,
295                           padding_bottom, stride_width, stride_height, dilation_width_factor,
296                           dilation_height_factor, activation, unsignedOutput.data(), outputShape));
297 
298     convertUInt8ToInt8(unsignedOutput, outputData);
299 
300     return true;
301 }
302 
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)303 bool convNhwc(const _Float16* inputData, const Shape& inputShape, const _Float16* filterData,
304               const Shape& filterShape, const _Float16* biasData, const Shape& biasShape,
305               int32_t padding_left, int32_t padding_right, int32_t padding_top,
306               int32_t padding_bottom, int32_t stride_width, int32_t stride_height,
307               int32_t dilation_width_factor, int32_t dilation_height_factor, int32_t activation,
308               _Float16* outputData, const Shape& outputShape) {
309     NNTRACE_TRANS("convFloat16");
310 
311     std::vector<float> inputData_float32(getNumberOfElements(inputShape));
312     std::vector<float> filterData_float32(getNumberOfElements(filterShape));
313     std::vector<float> biasData_float32(getNumberOfElements(biasShape));
314     std::vector<float> outputData_float32(getNumberOfElements(outputShape));
315 
316     convertFloat16ToFloat32(inputData, &inputData_float32);
317     convertFloat16ToFloat32(filterData, &filterData_float32);
318     convertFloat16ToFloat32(biasData, &biasData_float32);
319 
320     convNhwc(inputData_float32.data(), inputShape, filterData_float32.data(), filterShape,
321              biasData_float32.data(), biasShape, padding_left, padding_right, padding_top,
322              padding_bottom, stride_width, stride_height, dilation_width_factor,
323              dilation_height_factor, activation, outputData_float32.data(), outputShape);
324     convertFloat32ToFloat16(outputData_float32, outputData);
325 
326     return true;
327 }
328 
329 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)330 bool conv(const T_Input* inputData, const Shape& inputShape, const T_Filter* filterData,
331           const Shape& filterShape, const T_Bias* biasData, const Shape& biasShape,
332           int32_t padding_left, int32_t padding_right, int32_t padding_top, int32_t padding_bottom,
333           int32_t stride_width, int32_t stride_height, int32_t dilation_width_factor,
334           int32_t dilation_height_factor, int32_t activation, bool useNchw, T_Input* outputData,
335           const Shape& outputShape) {
336     InputWithLayout<T_Input> input(useNchw);
337     OutputWithLayout<T_Input> output(useNchw);
338     NN_RET_CHECK(input.initialize(inputData, inputShape));
339     NN_RET_CHECK(output.initialize(outputData, outputShape));
340     NN_RET_CHECK(convNhwc(input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape,
341                           biasData, biasShape, padding_left, padding_right, padding_top,
342                           padding_bottom, stride_width, stride_height, dilation_width_factor,
343                           dilation_height_factor, activation, output.getNhwcBuffer(),
344                           output.getNhwcShape()));
345     NN_RET_CHECK(output.commit());
346     return true;
347 }
348 
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,int32_t paddingTop,int32_t,int32_t strideWidth,int32_t strideHeight,int32_t dilationWidthFactor,int32_t dilationHeightFactor,int32_t activation,uint8_t * outputData,const Shape & outputShape)349 bool convQuant8PerChannelNhwc(const uint8_t* inputData, const Shape& inputShape,
350                               const int8_t* filterData, const Shape& filterShape,
351                               const float* filterScales, const int32_t* biasData,
352                               const Shape& biasShape, int32_t paddingLeft, int32_t /*paddingRight*/,
353                               int32_t paddingTop, int32_t /*paddingBottom*/, int32_t strideWidth,
354                               int32_t strideHeight, int32_t dilationWidthFactor,
355                               int32_t dilationHeightFactor, int32_t activation, uint8_t* outputData,
356                               const Shape& outputShape) {
357     NNTRACE_TRANS("convQuant8PerChannel");
358 
359     uint32_t numBatches = getSizeOfDimension(inputShape, 0);
360     uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
361     uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
362     uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
363     uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
364     uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
365     uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
366     uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
367     uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
368     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
369 
370     int32_t inputOffset = -inputShape.offset;
371     int32_t outputOffset = outputShape.offset;
372 
373     auto realMultiplier = std::vector<double>(outputDepth, .0f);
374     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
375     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
376 
377     for (uint32_t i = 0; i < outputDepth; ++i) {
378         Shape filterChannelShape = filterShape;
379         filterChannelShape.scale = filterScales[i];
380         Shape biasChannelShape = biasShape;
381         biasChannelShape.scale = filterScales[i] * inputShape.scale;
382         NN_RET_CHECK(GetQuantizedConvolutionMultiplier(
383                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
384         int exponent;
385         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &exponent));
386         outputShift[i] = -exponent;
387     }
388 
389     int32_t output_activation_min = 0, output_activation_max = 0;
390     CalculateActivationRangeUint8(activation, outputShape, &output_activation_min,
391                                   &output_activation_max);
392     const uint8_t* inputBase = inputData;
393     uint8_t* outPtr = outputData;
394     for (uint32_t b = 0; b < numBatches; b++) {
395         for (uint32_t h = 0; h < outputHeight; h++) {
396             for (uint32_t w = 0; w < outputWidth; w++) {
397                 const int8_t* filterBase = filterData;
398 
399                 for (uint32_t d = 0; d < outputDepth; d++) {
400                     int32_t wInputOrigin = static_cast<int32_t>(w) * strideWidth - paddingLeft;
401                     int32_t hInputOrigin = static_cast<int32_t>(h) * strideHeight - paddingTop;
402                     int32_t sum = 0.0f;
403 
404                     for (uint32_t i = 0; i < filterHeight; i++) {
405                         for (uint32_t j = 0; j < filterWidth; j++) {
406                             for (uint32_t k = 0; k < filterDepth; k++) {
407                                 int32_t hInput = hInputOrigin +
408                                                  dilationHeightFactor * static_cast<int32_t>(i);
409                                 int32_t wInput = wInputOrigin +
410                                                  dilationWidthFactor * static_cast<int32_t>(j);
411                                 uint32_t dInput = k;
412                                 if (hInput >= 0 && hInput < static_cast<int32_t>(inputHeight) &&
413                                     wInput >= 0 && wInput < static_cast<int32_t>(inputWidth)) {
414                                     uint32_t filterIndex =
415                                             i * filterWidth * filterDepth + j * filterDepth + k;
416                                     uint32_t inputIndex = hInput * inputWidth * inputDepth +
417                                                           wInput * inputDepth + dInput;
418                                     sum += (static_cast<int32_t>(filterBase[filterIndex])) *
419                                            (static_cast<int32_t>(inputBase[inputIndex]) +
420                                             inputOffset);
421                                 }
422                             }
423                         }
424                     }
425                     sum += biasData[d];
426                     sum = tflite::MultiplyByQuantizedMultiplier(sum, outputMultiplier[d],
427                                                                 -outputShift[d]);
428                     sum += outputOffset;
429                     sum = std::max(std::min(sum, output_activation_max), output_activation_min);
430                     outPtr[d] = static_cast<uint8_t>(sum);
431                     filterBase += filterHeight * filterWidth * filterDepth;
432                 }
433                 outPtr += outputDepth;
434             }
435         }
436         inputBase += inputHeight * inputWidth * inputDepth;
437     }
438 
439     return true;
440 }
441 
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,int32_t paddingTop,int32_t,int32_t strideWidth,int32_t strideHeight,int32_t dilationWidthFactor,int32_t dilationHeightFactor,int32_t activation,int8_t * outputData,const Shape & outputShape)442 bool convQuant8PerChannelNhwc(const int8_t* inputData, const Shape& inputShape,
443                               const int8_t* filterData, const Shape& filterShape,
444                               const float* filterScales, const int32_t* biasData,
445                               const Shape& biasShape, int32_t paddingLeft, int32_t /*paddingRight*/,
446                               int32_t paddingTop, int32_t /*paddingBottom*/, int32_t strideWidth,
447                               int32_t strideHeight, int32_t dilationWidthFactor,
448                               int32_t dilationHeightFactor, int32_t activation, int8_t* outputData,
449                               const Shape& outputShape) {
450     NNTRACE_TRANS("convQuant8SignedPerChannel");
451 
452     [[maybe_unused]] uint32_t numBatches = getSizeOfDimension(inputShape, 0);
453     [[maybe_unused]] uint32_t inputHeight = getSizeOfDimension(inputShape, 1);
454     [[maybe_unused]] uint32_t inputWidth = getSizeOfDimension(inputShape, 2);
455     [[maybe_unused]] uint32_t inputDepth = getSizeOfDimension(inputShape, 3);
456     [[maybe_unused]] uint32_t filterHeight = getSizeOfDimension(filterShape, 1);
457     [[maybe_unused]] uint32_t filterWidth = getSizeOfDimension(filterShape, 2);
458     [[maybe_unused]] uint32_t filterDepth = getSizeOfDimension(filterShape, 3);
459     [[maybe_unused]] uint32_t outputHeight = getSizeOfDimension(outputShape, 1);
460     [[maybe_unused]] uint32_t outputWidth = getSizeOfDimension(outputShape, 2);
461     uint32_t outputDepth = getSizeOfDimension(outputShape, 3);
462 
463     [[maybe_unused]] int32_t inputOffset = -inputShape.offset;
464     [[maybe_unused]] int32_t outputOffset = outputShape.offset;
465 
466     auto realMultiplier = std::vector<double>(outputDepth, .0f);
467     auto outputMultiplier = std::vector<int32_t>(outputDepth, 0);
468     auto outputShift = std::vector<int32_t>(outputDepth, .0f);
469 
470     for (uint32_t i = 0; i < outputDepth; ++i) {
471         Shape filterChannelShape = filterShape;
472         filterChannelShape.scale = filterScales[i];
473         Shape biasChannelShape = biasShape;
474         biasChannelShape.scale = filterScales[i] * inputShape.scale;
475         NN_RET_CHECK(GetQuantizedConvolutionMultiplier(
476                 inputShape, filterChannelShape, biasChannelShape, outputShape, &realMultiplier[i]));
477         NN_RET_CHECK(QuantizeMultiplier(realMultiplier[i], &outputMultiplier[i], &outputShift[i]));
478     }
479 
480     int32_t output_activation_min = 0, output_activation_max = 0;
481     CalculateActivationRangeInt8(activation, outputShape, &output_activation_min,
482                                  &output_activation_max);
483 
484     tflite::ConvParams convParams;
485     convParams.input_offset = -inputShape.offset;
486     convParams.output_offset = outputShape.offset;
487     convParams.stride_height = strideHeight;
488     convParams.stride_width = strideWidth;
489     convParams.dilation_height_factor = dilationHeightFactor;
490     convParams.dilation_width_factor = dilationWidthFactor;
491     convParams.padding_values.height = paddingTop;
492     convParams.padding_values.width = paddingLeft;
493     convParams.quantized_activation_min = output_activation_min;
494     convParams.quantized_activation_max = output_activation_max;
495 
496     NNTRACE_COMP_SWITCH("reference_integer_ops::ConvPerChannel");
497     tflite::reference_integer_ops::ConvPerChannel(
498             convParams, outputMultiplier.data(), outputShift.data(),
499             convertShapeToTflshape(inputShape), inputData, convertShapeToTflshape(filterShape),
500             filterData, convertShapeToTflshape(biasShape), biasData,
501             convertShapeToTflshape(outputShape), outputData);
502     return true;
503 }
504 
505 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)506 bool convQuant8PerChannel(const T* inputData, const Shape& inputShape, const int8_t* filterData,
507                           const Shape& filterShape, const float* filterScales,
508                           const int32_t* biasData, const Shape& biasShape, int32_t paddingLeft,
509                           int32_t paddingRight, int32_t paddingTop, int32_t paddingBottom,
510                           int32_t strideWidth, int32_t strideHeight, int32_t dilationWidthFactor,
511                           int32_t dilationHeightFactor, int32_t activation, bool useNchw,
512                           T* outputData, const Shape& outputShape) {
513     InputWithLayout<T> input(useNchw);
514     OutputWithLayout<T> output(useNchw);
515     NN_RET_CHECK(input.initialize(inputData, inputShape));
516     NN_RET_CHECK(output.initialize(outputData, outputShape));
517     NN_RET_CHECK(convQuant8PerChannelNhwc(
518             input.getNhwcBuffer(), input.getNhwcShape(), filterData, filterShape, filterScales,
519             biasData, biasShape, paddingLeft, paddingRight, paddingTop, paddingBottom, strideWidth,
520             strideHeight, dilationWidthFactor, dilationHeightFactor, activation,
521             output.getNhwcBuffer(), output.getNhwcShape()));
522     NN_RET_CHECK(output.commit());
523     return true;
524 }
525 
526 #undef ANDROID_NN_CONV_PARAMETERS
527 
528 }  // namespace
529 
prepare(IOperationExecutionContext * context)530 bool prepare(IOperationExecutionContext* context) {
531     Shape input = context->getInputShape(kInputTensor);
532     Shape filter = context->getInputShape(kFilterTensor);
533     Shape bias = context->getInputShape(kBiasTensor);
534 
535     if (filter.type == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
536         NN_RET_CHECK(input.type == OperandType::TENSOR_QUANT8_ASYMM ||
537                      input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED);
538     } else {
539         NN_RET_CHECK(input.type == filter.type);
540     }
541     if (input.type == OperandType::TENSOR_QUANT8_ASYMM ||
542         input.type == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
543         NN_RET_CHECK(bias.type == OperandType::TENSOR_INT32);
544     } else {
545         NN_RET_CHECK(input.type == bias.type);
546     }
547     NN_RET_CHECK_EQ(getNumberOfDimensions(input), 4u);
548     NN_RET_CHECK_EQ(getNumberOfDimensions(filter), 4u);
549     NN_RET_CHECK_EQ(getNumberOfDimensions(bias), 1u);
550 
551     Conv2dParam param;
552     NN_RET_CHECK(param.initialize(context));
553 
554     uint32_t batches = getSizeOfDimension(input, 0);
555     uint32_t height = getSizeOfDimension(input, param.useNchw ? 2 : 1);
556     uint32_t width = getSizeOfDimension(input, param.useNchw ? 3 : 2);
557     uint32_t channels_in = getSizeOfDimension(input, param.useNchw ? 1 : 3);
558     uint32_t channels_out = getSizeOfDimension(filter, 0);
559     uint32_t filterHeight = getSizeOfDimension(filter, 1);
560     uint32_t filterWidth = getSizeOfDimension(filter, 2);
561     // Only batches can be zero.
562     NN_RET_CHECK_EQ(channels_in, getSizeOfDimension(filter, 3));
563     NN_RET_CHECK_EQ(channels_out, getSizeOfDimension(bias, 0));
564     NN_RET_CHECK_GT(height, 0u);
565     NN_RET_CHECK_GT(width, 0u);
566     NN_RET_CHECK_GT(channels_in, 0u);
567     NN_RET_CHECK_GT(channels_out, 0u);
568 
569     int32_t effectiveFilterWidth = (filterWidth - 1) * param.dilation_width_factor + 1;
570     int32_t effectiveFilterHeight = (filterHeight - 1) * param.dilation_height_factor + 1;
571     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_left);
572     NN_RET_CHECK_GT(effectiveFilterWidth, param.padding_right);
573     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_top);
574     NN_RET_CHECK_GT(effectiveFilterHeight, param.padding_bottom);
575 
576     uint32_t outWidth =
577             computeOutSize(width, filterWidth, param.stride_width, param.dilation_width_factor,
578                            param.padding_left, param.padding_right);
579     uint32_t outHeight =
580             computeOutSize(height, filterHeight, param.stride_height, param.dilation_height_factor,
581                            param.padding_top, param.padding_bottom);
582 
583     Shape output = context->getOutputShape(kOutputTensor);
584     output.type = input.type;
585     if (param.useNchw) {
586         output.dimensions = {batches, channels_out, outHeight, outWidth};
587     } else {
588         output.dimensions = {batches, outHeight, outWidth, channels_out};
589     }
590     return context->setOutputShape(kOutputTensor, output);
591 }
592 
execute(IOperationExecutionContext * context)593 bool execute(IOperationExecutionContext* context) {
594     // Bypass execution in the case of zero-sized input.
595     if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
596     Conv2dParam param;
597     NN_RET_CHECK(param.initialize(context));
598     switch (context->getInputType(kInputTensor)) {
599         case OperandType::TENSOR_FLOAT32:
600             return conv(context->getInputBuffer<float>(kInputTensor),
601                         context->getInputShape(kInputTensor),
602                         context->getInputBuffer<float>(kFilterTensor),
603                         context->getInputShape(kFilterTensor),
604                         context->getInputBuffer<float>(kBiasTensor),
605                         context->getInputShape(kBiasTensor), param.padding_left,
606                         param.padding_right, param.padding_top, param.padding_bottom,
607                         param.stride_width, param.stride_height, param.dilation_width_factor,
608                         param.dilation_height_factor, param.activation, param.useNchw,
609                         context->getOutputBuffer<float>(kOutputTensor),
610                         context->getOutputShape(kOutputTensor));
611         case OperandType::TENSOR_FLOAT16:
612             return conv(context->getInputBuffer<_Float16>(kInputTensor),
613                         context->getInputShape(kInputTensor),
614                         context->getInputBuffer<_Float16>(kFilterTensor),
615                         context->getInputShape(kFilterTensor),
616                         context->getInputBuffer<_Float16>(kBiasTensor),
617                         context->getInputShape(kBiasTensor), param.padding_left,
618                         param.padding_right, param.padding_top, param.padding_bottom,
619                         param.stride_width, param.stride_height, param.dilation_width_factor,
620                         param.dilation_height_factor, param.activation, param.useNchw,
621                         context->getOutputBuffer<_Float16>(kOutputTensor),
622                         context->getOutputShape(kOutputTensor));
623         case OperandType::TENSOR_QUANT8_ASYMM:
624             if (context->getInputType(kFilterTensor) ==
625                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
626                 return convQuant8PerChannel(
627                         context->getInputBuffer<uint8_t>(kInputTensor),
628                         context->getInputShape(kInputTensor),
629                         context->getInputBuffer<int8_t>(kFilterTensor),
630                         context->getInputShape(kFilterTensor),
631                         std::get<Operand::SymmPerChannelQuantParams>(
632                                 context->getInputExtraParams(kFilterTensor))
633                                 .scales.data(),
634                         context->getInputBuffer<int32_t>(kBiasTensor),
635                         context->getInputShape(kBiasTensor), param.padding_left,
636                         param.padding_right, param.padding_top, param.padding_bottom,
637                         param.stride_width, param.stride_height, param.dilation_width_factor,
638                         param.dilation_height_factor, param.activation, param.useNchw,
639                         context->getOutputBuffer<uint8_t>(kOutputTensor),
640                         context->getOutputShape(kOutputTensor));
641             } else if (context->getInputType(kFilterTensor) == OperandType::TENSOR_QUANT8_ASYMM) {
642                 return conv(context->getInputBuffer<uint8_t>(kInputTensor),
643                             context->getInputShape(kInputTensor),
644                             context->getInputBuffer<uint8_t>(kFilterTensor),
645                             context->getInputShape(kFilterTensor),
646                             context->getInputBuffer<int32_t>(kBiasTensor),
647                             context->getInputShape(kBiasTensor), param.padding_left,
648                             param.padding_right, param.padding_top, param.padding_bottom,
649                             param.stride_width, param.stride_height, param.dilation_width_factor,
650                             param.dilation_height_factor, param.activation, param.useNchw,
651                             context->getOutputBuffer<uint8_t>(kOutputTensor),
652                             context->getOutputShape(kOutputTensor));
653             } else {
654                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
655             }
656         case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
657             if (context->getInputType(kFilterTensor) ==
658                 OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
659                 return convQuant8PerChannel(
660                         context->getInputBuffer<int8_t>(kInputTensor),
661                         context->getInputShape(kInputTensor),
662                         context->getInputBuffer<int8_t>(kFilterTensor),
663                         context->getInputShape(kFilterTensor),
664                         std::get<Operand::SymmPerChannelQuantParams>(
665                                 context->getInputExtraParams(kFilterTensor))
666                                 .scales.data(),
667                         context->getInputBuffer<int32_t>(kBiasTensor),
668                         context->getInputShape(kBiasTensor), param.padding_left,
669                         param.padding_right, param.padding_top, param.padding_bottom,
670                         param.stride_width, param.stride_height, param.dilation_width_factor,
671                         param.dilation_height_factor, param.activation, param.useNchw,
672                         context->getOutputBuffer<int8_t>(kOutputTensor),
673                         context->getOutputShape(kOutputTensor));
674             } else if (context->getInputType(kFilterTensor) ==
675                        OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
676                 return conv(context->getInputBuffer<int8_t>(kInputTensor),
677                             context->getInputShape(kInputTensor),
678                             context->getInputBuffer<int8_t>(kFilterTensor),
679                             context->getInputShape(kFilterTensor),
680                             context->getInputBuffer<int32_t>(kBiasTensor),
681                             context->getInputShape(kBiasTensor), param.padding_left,
682                             param.padding_right, param.padding_top, param.padding_bottom,
683                             param.stride_width, param.stride_height, param.dilation_width_factor,
684                             param.dilation_height_factor, param.activation, param.useNchw,
685                             context->getOutputBuffer<int8_t>(kOutputTensor),
686                             context->getOutputShape(kOutputTensor));
687             } else {
688                 NN_RET_CHECK_FAIL() << "Unsupported filter type for operation " << kOperationName;
689             }
690         default:
691             NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
692     }
693 }
694 #endif  // NN_INCLUDE_CPU_IMPLEMENTATION
695 
696 }  // namespace conv_2d
697 
698 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(CONV_2D, conv_2d::prepare, conv_2d::execute,
699                                          .allowZeroSizedInput = true);
700 
701 }  // namespace nn
702 }  // namespace android
703