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 "QuantizedLSTM.h"
20 
21 #include <public/gemmlowp.h>
22 #include <tensorflow/lite/kernels/internal/reference/legacy_reference_ops.h>
23 
24 #include <algorithm>
25 #include <vector>
26 
27 #include "CpuExecutor.h"
28 #include "CpuOperationUtils.h"
29 #include "Tracing.h"
30 
31 namespace android {
32 namespace nn {
33 
34 namespace {
35 
36 template <typename T>
GetBuffer(RunTimeOperandInfo * operand)37 inline T* GetBuffer(RunTimeOperandInfo* operand) {
38     return reinterpret_cast<T*>(operand->buffer);
39 }
40 
41 template <typename T>
GetBuffer(const RunTimeOperandInfo * operand)42 inline const T* GetBuffer(const RunTimeOperandInfo* operand) {
43     return reinterpret_cast<const T*>(operand->buffer);
44 }
45 
46 using tflite::Dims;
47 
48 // The function below is taken from TF Lite implementation in order to decouple
49 // NN API from TF Lite dependency. Original function, with a description of its
50 // parameters and types can be found by this link:
51 // https://github.com/tensorflow/tensorflow/blob/0d697e5fc4c05c699eea0764364104ea500ccc68/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h#L1926
52 //
53 // clang-format off
54 template <int StateIntegerBits>
quantizedLstmStep(const uint8_t * input_data_uint8,const Dims<4> & input_dims,const uint8_t * prev_activ_data_uint8,const Dims<4> & prev_activ_dims,const uint8_t * weights_data_uint8,const Dims<4> & weights_dims,const int32_t * bias_data_int32,const Dims<4> & bias_dims,const int16_t * prevCellState_data_int16,const Dims<4> & prevCellState_dims,int16_t * output_state_data_int16,const Dims<4> & output_state_dims,uint8_t * output_activ_data_uint8,const Dims<4> & output_activ_dims,uint8_t * concat_temp_data_uint8,const Dims<4> & concat_temp_dims,int16_t * activ_temp_data_int16,const Dims<4> & activ_temp_dims,int32_t weights_zero_point,int32_t accum_multiplier,int accum_shift)55 void quantizedLstmStep(const uint8_t* input_data_uint8, const Dims<4>& input_dims,
56                        const uint8_t* prev_activ_data_uint8,
57                        const Dims<4>& prev_activ_dims, const uint8_t* weights_data_uint8,
58                        const Dims<4>& weights_dims, const int32_t* bias_data_int32,
59                        const Dims<4>& bias_dims, const int16_t* prevCellState_data_int16,
60                        const Dims<4>& prevCellState_dims, int16_t* output_state_data_int16,
61                        const Dims<4>& output_state_dims, uint8_t* output_activ_data_uint8,
62                        const Dims<4>& output_activ_dims, uint8_t* concat_temp_data_uint8,
63                        const Dims<4>& concat_temp_dims, int16_t* activ_temp_data_int16,
64                        const Dims<4>& activ_temp_dims, int32_t weights_zero_point,
65                        int32_t accum_multiplier, int accum_shift) {
66   // Gather dimensions information, and perform consistency checks.
67   const int outer_size =
68       MatchingFlatSizeSkipDim(input_dims, 0, prev_activ_dims, prevCellState_dims,
69                               output_state_dims, output_activ_dims);
70   TFLITE_CHECK_EQ(ArraySize(weights_dims, 2), 1);
71   TFLITE_CHECK_EQ(ArraySize(weights_dims, 3), 1);
72   const int input_depth = ArraySize(input_dims, 0);
73   const int prev_activ_depth = ArraySize(prev_activ_dims, 0);
74   const int total_input_depth = prev_activ_depth + input_depth;
75   TFLITE_CHECK_EQ(ArraySize(weights_dims, 0), total_input_depth);
76   TFLITE_CHECK_EQ(MatchingArraySize(bias_dims, 1, bias_dims, 2, bias_dims, 3),
77                   1);
78   const int intern_activ_depth =
79       MatchingArraySize(weights_dims, 1, bias_dims, 0);
80   TFLITE_CHECK_EQ(intern_activ_depth % 4, 0);
81   const int output_depth =
82       MatchingArraySize(prevCellState_dims, 0, prev_activ_dims, 0,
83                         output_state_dims, 0, output_activ_dims, 0);
84   TFLITE_CHECK_EQ(output_depth, intern_activ_depth / 4);
85   const int fc_batches = FlatSizeSkipDim(activ_temp_dims, 0);
86   const int fc_output_depth =
87       MatchingArraySize(weights_dims, 1, activ_temp_dims, 0);
88   const int fc_accum_depth = ArraySize(weights_dims, 0);
89   TFLITE_CHECK_EQ(fc_output_depth, 4 * output_depth);
90 
91   // Depth-concatenate prev_activ and input data together.
92   uint8_t const* concat_input_arrays_data[2] = {input_data_uint8,
93                                                 prev_activ_data_uint8};
94   Dims<4> const* concat_input_arrays_dims[2] = {&input_dims, &prev_activ_dims};
95   tflite::reference_ops::Concatenation<tflite::FusedActivationFunctionType::kNone, uint8_t>(
96       0, concat_input_arrays_data, concat_input_arrays_dims, 2,
97       concat_temp_data_uint8, concat_temp_dims);
98 
99   // Implementation of the fully connected node inside the LSTM cell.
100   // The operands are 8-bit integers, the accumulators are internally 32bit
101   // integers, and the output is 16-bit fixed-point with 3 integer bits so
102   // the output range is [-2^3, 2^3] == [-8, 8]. The rationale for that
103   // is explained in the function comment above.
104   for (int b = 0; b < fc_batches; ++b) {
105     for (int out_c = 0; out_c < fc_output_depth; ++out_c) {
106       // Internal accumulation.
107       // Initialize accumulator with the bias-value.
108       int32_t accum = bias_data_int32[out_c];
109       // Accumulation loop.
110       for (int d = 0; d < fc_accum_depth; ++d) {
111         int16_t input_val = concat_temp_data_uint8[b * fc_accum_depth + d] - 128;
112         int16_t weights_val =
113             weights_data_uint8[out_c * fc_accum_depth + d] - weights_zero_point;
114         accum += input_val * weights_val;
115       }
116       // Down-scale the final int32 accumulator to the scale used by our
117       // (16-bit, using 3 integer bits) fixed-point format. The quantized
118       // multiplier and shift here have been pre-computed offline
119       // (e.g. by toco).
120       accum =
121           tflite::MultiplyByQuantizedMultiplier(accum, accum_multiplier, accum_shift);
122       // Saturate, cast to int16, and store to the temporary activations array.
123       accum = std::max(-32768, std::min(32767, accum));
124       activ_temp_data_int16[out_c + fc_output_depth * b] = accum;
125     }
126   }
127 
128   // Rest of the LSTM cell: tanh and logistic math functions, and some adds
129   // and muls, all done in 16-bit fixed-point.
130   for (int b = 0; b < outer_size; ++b) {
131     for (int c = 0; c < output_depth; ++c) {
132       // Define the fixed-point data types that we will use here. All use
133       // int16 as the underlying integer type i.e. all are 16-bit fixed-point.
134       // They only differ by the number of integral vs. fractional bits,
135       // determining the range of values that they can represent.
136       //
137       // F0 uses 0 integer bits, range [-1, 1].
138       // This is the return type of math functions such as tanh, logistic,
139       // whose range is in [-1, 1].
140       using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
141       // F3 uses 3 integer bits, range [-8, 8].
142       // This is the range of the previous fully-connected node's output,
143       // which is our input here.
144       using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
145       // FS uses StateIntegerBits integer bits, range [-2^StateIntegerBits,
146       // 2^StateIntegerBits]. It's used to represent the internal state, whose
147       // number of integer bits is currently dictated by the model. See comment
148       // on the StateIntegerBits template parameter above.
149       using FS = gemmlowp::FixedPoint<std::int16_t, StateIntegerBits>;
150       // Implementation of input gate, using fixed-point logistic function.
151       F3 input_gate_input = F3::FromRaw(
152           activ_temp_data_int16[b * fc_output_depth + 0 * output_depth + c]);
153       F0 input_gate_output = gemmlowp::logistic(input_gate_input);
154       // Implementation of input modulation gate, using fixed-point tanh
155       // function.
156       F3 input_modulation_gate_input = F3::FromRaw(
157           activ_temp_data_int16[b * fc_output_depth + 1 * output_depth + c]);
158       F0 input_modulation_gate_output =
159           gemmlowp::tanh(input_modulation_gate_input);
160       // Implementation of forget gate, using fixed-point logistic function.
161       F3 forget_gate_input = F3::FromRaw(
162           activ_temp_data_int16[b * fc_output_depth + 2 * output_depth + c]);
163       F0 forget_gate_output = gemmlowp::logistic(forget_gate_input);
164       // Implementation of output gate, using fixed-point logistic function.
165       F3 output_gate_input = F3::FromRaw(
166           activ_temp_data_int16[b * fc_output_depth + 3 * output_depth + c]);
167       F0 output_gate_output = gemmlowp::logistic(output_gate_input);
168       // Implementation of internal multiplication nodes, still in fixed-point.
169       F0 input_times_input_modulation =
170           input_gate_output * input_modulation_gate_output;
171       FS prevCellState = FS::FromRaw(prevCellState_data_int16[b * output_depth + c]);
172       FS prevCellState_times_forget_state = forget_gate_output * prevCellState;
173       // Implementation of internal addition node, saturating.
174       FS new_state = gemmlowp::SaturatingAdd(
175           gemmlowp::Rescale<StateIntegerBits>(input_times_input_modulation),
176           prevCellState_times_forget_state);
177       // Implementation of last internal Tanh node, still in fixed-point.
178       // Since a Tanh fixed-point implementation is specialized for a given
179       // number or integer bits, and each specialization can have a substantial
180       // code size, and we already used above a Tanh on an input with 3 integer
181       // bits, and per the table in the above function comment there is no
182       // significant accuracy to be lost by clamping to [-8, +8] for a
183       // 3-integer-bits representation, let us just do that. This helps people
184       // porting this to targets where code footprint must be minimized.
185       F3 new_state_f3 = gemmlowp::Rescale<3>(new_state);
186       F0 output_activ_int16 = output_gate_output * gemmlowp::tanh(new_state_f3);
187       // Store the new internal state back to memory, as 16-bit integers.
188       // Note: here we store the original value with StateIntegerBits, not
189       // the rescaled 3-integer-bits value fed to tanh.
190       output_state_data_int16[b * output_depth + c] = new_state.raw();
191       // Down-scale the output activations to 8-bit integers, saturating,
192       // and store back to memory.
193       int16_t rescaled_output_activ =
194           gemmlowp::RoundingDivideByPOT(output_activ_int16.raw(), 8);
195       int16_t clamped_output_activ =
196           std::max<int16_t>(-128, std::min<int16_t>(127, rescaled_output_activ));
197       output_activ_data_uint8[b * output_depth + c] =
198           128 + clamped_output_activ;
199     }
200   }
201 }
202 // clang-format on
203 
204 // The function assigns a 2D matrix to a submatrix of the weights at a given row
205 // and column offsets.
assignWeightsSubmatrix(const RunTimeOperandInfo * submatrix,const int32_t offset_row,const int32_t offset_column,const std::vector<uint32_t> & weightsDims,uint8_t * weights)206 void assignWeightsSubmatrix(const RunTimeOperandInfo* submatrix, const int32_t offset_row,
207                             const int32_t offset_column, const std::vector<uint32_t>& weightsDims,
208                             uint8_t* weights) {
209     const uint8_t* submatrixValues = GetBuffer<uint8_t>(submatrix);
210     const std::vector<uint32_t> submatrixDims = submatrix->shape().dimensions;
211     for (uint32_t i = 0; i < submatrixDims[0] * submatrixDims[1]; ++i) {
212         const uint32_t row = i / submatrixDims[1];
213         const uint32_t column = i % submatrixDims[1];
214         weights[(row + offset_row) * weightsDims[1] + column + offset_column] = submatrixValues[i];
215     }
216 }
217 
218 }  // namespace
219 
QuantizedLSTMCell(const Operation & operation,RunTimeOperandInfo * operands)220 QuantizedLSTMCell::QuantizedLSTMCell(const Operation& operation, RunTimeOperandInfo* operands) {
221     input_ = GetInput(operation, operands, kInputTensor);
222 
223     inputToInputWeights_ = GetInput(operation, operands, kInputToInputWeightsTensor);
224     inputToForgetWeights_ = GetInput(operation, operands, kInputToForgetWeightsTensor);
225     inputToCellWeights_ = GetInput(operation, operands, kInputToCellWeightsTensor);
226     inputToOutputWeights_ = GetInput(operation, operands, kInputToOutputWeightsTensor);
227 
228     recurrentToInputWeights_ = GetInput(operation, operands, kRecurrentToInputWeightsTensor);
229     recurrentToForgetWeights_ = GetInput(operation, operands, kRecurrentToForgetWeightsTensor);
230     recurrentToCellWeights_ = GetInput(operation, operands, kRecurrentToCellWeightsTensor);
231     recurrentToOutputWeights_ = GetInput(operation, operands, kRecurrentToOutputWeightsTensor);
232 
233     inputGateBias_ = GetInput(operation, operands, kInputGateBiasTensor);
234     forgetGateBias_ = GetInput(operation, operands, kForgetGateBiasTensor);
235     cellGateBias_ = GetInput(operation, operands, kCellGateBiasTensor);
236     outputGateBias_ = GetInput(operation, operands, kOutputGateBiasTensor);
237 
238     prevCellState_ = GetInput(operation, operands, kPrevCellStateTensor);
239     prevOutput_ = GetInput(operation, operands, kPrevOutputTensor);
240 
241     cellStateOut_ = GetOutput(operation, operands, kCellStateOutTensor);
242     output_ = GetOutput(operation, operands, kOutputTensor);
243 }
244 
prepare(const Operation & operation,RunTimeOperandInfo * operands,Shape * cellStateOutShape,Shape * outputShape)245 bool QuantizedLSTMCell::prepare(const Operation& operation, RunTimeOperandInfo* operands,
246                                 Shape* cellStateOutShape, Shape* outputShape) {
247     auto input = GetInput(operation, operands, kInputTensor);
248     NN_RET_CHECK_EQ(NumDimensions(input), 2);
249     NN_RET_CHECK_EQ(input->scale, 1. / 128.0);
250     NN_RET_CHECK_EQ(input->zeroPoint, 128);
251     const uint32_t numBatches = SizeOfDimension(input, 0);
252     const uint32_t inputSize = SizeOfDimension(input, 1);
253 
254     auto prevOutput = GetInput(operation, operands, kPrevOutputTensor);
255     NN_RET_CHECK_EQ(NumDimensions(prevOutput), 2);
256     NN_RET_CHECK_EQ(SizeOfDimension(prevOutput, 0), numBatches);
257     NN_RET_CHECK_EQ(prevOutput->scale, 1. / 128.0);
258     NN_RET_CHECK_EQ(prevOutput->zeroPoint, 128);
259     const uint32_t outputSize = SizeOfDimension(prevOutput, 1);
260 
261     auto inputToInputWeights = GetInput(operation, operands, kInputToInputWeightsTensor);
262     const float weightsScale = inputToInputWeights->scale;
263     NN_RET_CHECK(weightsScale != 0);
264     const float weightsZeroPoint = inputToInputWeights->zeroPoint;
265 
266     auto checkWeightsShape = [&](const RunTimeOperandInfo* weights, uint32_t columns) -> bool {
267         NN_RET_CHECK_EQ(NumDimensions(weights), 2);
268         NN_RET_CHECK_EQ(SizeOfDimension(weights, 0), outputSize);
269         NN_RET_CHECK_EQ(SizeOfDimension(weights, 1), columns);
270         NN_RET_CHECK_EQ(weights->scale, weightsScale);
271         NN_RET_CHECK_EQ(weights->zeroPoint, weightsZeroPoint);
272         return true;
273     };
274 
275     auto inputToForgetWeights = GetInput(operation, operands, kInputToForgetWeightsTensor);
276     auto inputToCellWeights = GetInput(operation, operands, kInputToCellWeightsTensor);
277     auto inputToOutputWeights = GetInput(operation, operands, kInputToOutputWeightsTensor);
278     NN_RET_CHECK(checkWeightsShape(inputToInputWeights, inputSize));
279     NN_RET_CHECK(checkWeightsShape(inputToForgetWeights, inputSize));
280     NN_RET_CHECK(checkWeightsShape(inputToCellWeights, inputSize));
281     NN_RET_CHECK(checkWeightsShape(inputToOutputWeights, inputSize));
282 
283     auto recurrentToInputWeights = GetInput(operation, operands, kRecurrentToInputWeightsTensor);
284     auto recurrentToForgetWeights = GetInput(operation, operands, kRecurrentToForgetWeightsTensor);
285     auto recurrentToCellWeights = GetInput(operation, operands, kRecurrentToCellWeightsTensor);
286     auto recurrentToOutputWeights = GetInput(operation, operands, kRecurrentToOutputWeightsTensor);
287     NN_RET_CHECK(checkWeightsShape(recurrentToInputWeights, outputSize));
288     NN_RET_CHECK(checkWeightsShape(recurrentToForgetWeights, outputSize));
289     NN_RET_CHECK(checkWeightsShape(recurrentToCellWeights, outputSize));
290     NN_RET_CHECK(checkWeightsShape(recurrentToOutputWeights, outputSize));
291 
292     auto inputGateBias = GetInput(operation, operands, kInputGateBiasTensor);
293     const float biasScale = inputGateBias->scale;
294     NN_RET_CHECK_EQ(biasScale, weightsScale / 128.0);
295     const float biasZeroPoint = inputGateBias->zeroPoint;
296     NN_RET_CHECK_EQ(biasZeroPoint, 0);
297 
298     auto checkBiasShape = [&](const RunTimeOperandInfo* bias) -> bool {
299         NN_RET_CHECK_EQ(NumDimensions(bias), 1);
300         NN_RET_CHECK_EQ(SizeOfDimension(bias, 0), outputSize);
301         NN_RET_CHECK_EQ(bias->scale, biasScale);
302         NN_RET_CHECK_EQ(bias->zeroPoint, biasZeroPoint);
303         return true;
304     };
305 
306     auto forgetGateBias = GetInput(operation, operands, kForgetGateBiasTensor);
307     auto cellGateBias = GetInput(operation, operands, kCellGateBiasTensor);
308     auto outputGateBias = GetInput(operation, operands, kOutputGateBiasTensor);
309     NN_RET_CHECK(checkBiasShape(inputGateBias));
310     NN_RET_CHECK(checkBiasShape(forgetGateBias));
311     NN_RET_CHECK(checkBiasShape(cellGateBias));
312     NN_RET_CHECK(checkBiasShape(outputGateBias));
313 
314     auto prevCellState = GetInput(operation, operands, kPrevCellStateTensor);
315     NN_CHECK_EQ(NumDimensions(prevCellState), 2);
316     NN_CHECK_EQ(SizeOfDimension(prevCellState, 0), numBatches);
317     NN_CHECK_EQ(SizeOfDimension(prevCellState, 1), outputSize);
318     NN_CHECK_EQ(prevCellState->zeroPoint, 0);
319     // Cell state range for quantized LSTM is a function of StateIntegerBits and
320     // can be calculated as:
321     // [-2^StateIntegerBits, 2^StateIntegerBits * 32767/32768].
322     // Therefore, for a fixed StateIntegerBits parameter, cell state scale is
323     // equal to 2^StateIntegerBits * 2^(-15) = 2^(StateIntegerBits - 15) and
324     // therefore:
325     // StateIntegerBits = log2(cell state scale) + 15
326     int stateScaleLog2Rounded;
327     NN_CHECK(tflite::CheckedLog2(prevCellState->scale, &stateScaleLog2Rounded));
328     const int stateIntegerBits = 15 + stateScaleLog2Rounded;
329     // We only support StateIntegerBits == 4
330     NN_CHECK(stateIntegerBits == 4);
331 
332     *cellStateOutShape = prevCellState->shape();
333     *outputShape = prevOutput->shape();
334     return true;
335 }
336 
337 // The function contatenates 8 input weight matrices into one. Resulting matrix
338 // has a shape [4 * outputSize, outputSize + inputSize]. The matrix is
339 // constructed as follows:
340 // +-----------------------------------+
341 // | recurrentToInput  | inputToInput  |
342 // |-------------------+---------------|
343 // | recurrentToCell   | inputToCell   |
344 // |-------------------+---------------|
345 // | recurrentToForget | inputToForget |
346 // |-------------------+---------------|
347 // | recurrentToOutput | inputToOutput |
348 // +-----------------------------------+
concatenateWeights(const std::vector<uint32_t> & weightsDims,uint8_t * weights)349 void QuantizedLSTMCell::concatenateWeights(const std::vector<uint32_t>& weightsDims,
350                                            uint8_t* weights) {
351     const int outputSize = SizeOfDimension(inputToInputWeights_, 0);
352 
353     assignWeightsSubmatrix(inputToInputWeights_, 0 * outputSize, outputSize, weightsDims, weights);
354     assignWeightsSubmatrix(inputToCellWeights_, 1 * outputSize, outputSize, weightsDims, weights);
355     assignWeightsSubmatrix(inputToForgetWeights_, 2 * outputSize, outputSize, weightsDims, weights);
356     assignWeightsSubmatrix(inputToOutputWeights_, 3 * outputSize, outputSize, weightsDims, weights);
357     assignWeightsSubmatrix(recurrentToInputWeights_, 0 * outputSize, 0, weightsDims, weights);
358     assignWeightsSubmatrix(recurrentToCellWeights_, 1 * outputSize, 0, weightsDims, weights);
359     assignWeightsSubmatrix(recurrentToForgetWeights_, 2 * outputSize, 0, weightsDims, weights);
360     assignWeightsSubmatrix(recurrentToOutputWeights_, 3 * outputSize, 0, weightsDims, weights);
361 }
362 
363 // The function concatenate four bias vectors of shape [outputSize] into one
364 // vector of shape [4 * outputSize].
concatenateBiases(uint32_t outputSize,int32_t * bias)365 void QuantizedLSTMCell::concatenateBiases(uint32_t outputSize, int32_t* bias) {
366     memcpy(bias + 0 * outputSize, GetBuffer<int32_t>(inputGateBias_), sizeof(int32_t) * outputSize);
367     memcpy(bias + 1 * outputSize, GetBuffer<int32_t>(cellGateBias_), sizeof(int32_t) * outputSize);
368     memcpy(bias + 2 * outputSize, GetBuffer<int32_t>(forgetGateBias_),
369            sizeof(int32_t) * outputSize);
370     memcpy(bias + 3 * outputSize, GetBuffer<int32_t>(outputGateBias_),
371            sizeof(int32_t) * outputSize);
372 }
373 
eval()374 bool QuantizedLSTMCell::eval() {
375     NNTRACE_COMP("QuantizedLSTM::eval");
376 
377     Shape weightsShape;
378     weightsShape.dimensions = {4 * SizeOfDimension(prevOutput_, 1),
379                                SizeOfDimension(input_, 1) + SizeOfDimension(prevOutput_, 1)};
380     std::vector<uint8_t> weights(getNumberOfElements(weightsShape));
381     concatenateWeights(weightsShape.dimensions, weights.data());
382 
383     Shape biasShape;
384     biasShape.dimensions = {getSizeOfDimension(weightsShape, 0)};
385     std::vector<int32_t> bias(getNumberOfElements(biasShape));
386     concatenateBiases(SizeOfDimension(prevOutput_, 1), bias.data());
387 
388     Shape concatTempShape;
389     concatTempShape.dimensions = {SizeOfDimension(input_, 0), getSizeOfDimension(weightsShape, 1)};
390 
391     Shape activationTempShape;
392     activationTempShape.dimensions = {SizeOfDimension(input_, 0),
393                                       getSizeOfDimension(weightsShape, 0)};
394 
395     std::vector<uint8_t> concatTemp(getNumberOfElements(concatTempShape));
396     std::vector<int16_t> activationTemp(getNumberOfElements(activationTempShape));
397 
398     // From https://arxiv.org/pdf/1712.05877, for a fully-connected layer,
399     // accumulator multiplier is equal to:
400     // (input scale) * (weights scale) / (fully-connected output scale)
401     // In our case fully-connected output scale is fixed and equal to
402     // 2^(-12) (See LSTMCell definition in TF Lite for more details on that).
403     // But bias scale is set to (input scale) * (weights scale) (also from the
404     // paper), so we can multiply it to an inverse of the fc-output scale to get
405     // the multiplier value:
406     double realAccumMultiplier = 4096 * inputGateBias_->scale;
407     int32_t accumMultiplier;
408     int accumShift;
409     tflite::QuantizeMultiplier(realAccumMultiplier, &accumMultiplier, &accumShift);
410     quantizedLstmStep<4>(
411             // Inputs.
412             GetBuffer<const uint8_t>(input_), convertShapeToDims(input_->shape()),
413             GetBuffer<const uint8_t>(prevOutput_), convertShapeToDims(prevOutput_->shape()),
414             weights.data(), convertShapeToDims(weightsShape), bias.data(),
415             convertShapeToDims(biasShape), GetBuffer<const int16_t>(prevCellState_),
416             convertShapeToDims(prevCellState_->shape()),
417             // Outputs.
418             GetBuffer<int16_t>(cellStateOut_), convertShapeToDims(cellStateOut_->shape()),
419             GetBuffer<uint8_t>(output_), convertShapeToDims(output_->shape()), concatTemp.data(),
420             convertShapeToDims(concatTempShape), activationTemp.data(),
421             convertShapeToDims(activationTempShape), inputToInputWeights_->zeroPoint,
422             accumMultiplier, accumShift);
423     return true;
424 }
425 
426 }  // namespace nn
427 }  // namespace android
428