| 1 | /* Copyright 2020 The TensorFlow Authors. All Rights Reserved. |
|---|---|
| 2 | |
| 3 | Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | you may not use this file except in compliance with the License. |
| 5 | You may obtain a copy of the License at |
| 6 | |
| 7 | http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | |
| 9 | Unless required by applicable law or agreed to in writing, software |
| 10 | distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | See the License for the specific language governing permissions and |
| 13 | limitations under the License. |
| 14 | ==============================================================================*/ |
| 15 | #include "tensorflow/lite/kernels/lstm_eval.h" |
| 16 | |
| 17 | #include <math.h> |
| 18 | #include <string.h> |
| 19 | |
| 20 | #include <algorithm> |
| 21 | #include <cstdint> |
| 22 | #include <memory> |
| 23 | #include <vector> |
| 24 | |
| 25 | #include "ruy/matrix.h" // from @ruy |
| 26 | #include "ruy/mul_params.h" // from @ruy |
| 27 | #include "ruy/profiler/instrumentation.h" // from @ruy |
| 28 | #include "ruy/ruy.h" // from @ruy |
| 29 | #include "tensorflow/lite/c/builtin_op_data.h" |
| 30 | #include "tensorflow/lite/c/common.h" |
| 31 | #include "tensorflow/lite/kernels/cpu_backend_context.h" |
| 32 | #include "tensorflow/lite/kernels/internal/compatibility.h" |
| 33 | #include "tensorflow/lite/kernels/internal/kernel_utils.h" |
| 34 | #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h" |
| 35 | #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" |
| 36 | #include "tensorflow/lite/kernels/internal/tensor_utils.h" |
| 37 | #include "tensorflow/lite/kernels/op_macros.h" |
| 38 | |
| 39 | namespace tflite { |
| 40 | namespace ops { |
| 41 | namespace builtin { |
| 42 | namespace lstm_eval { |
| 43 | namespace { |
| 44 | |
| 45 | void MatrixBatchVectorMultiplyAccumulate( |
| 46 | const float* matrix, const float* vector, const float* result, |
| 47 | float* output, int m_rows, int m_cols, int n_batch, |
| 48 | CpuBackendContext* cpu_backend_context) { |
| 49 | tflite::FullyConnectedParams float_fc_params; |
| 50 | float_fc_params.float_activation_min = std::numeric_limits<float>::lowest(); |
| 51 | float_fc_params.float_activation_max = std::numeric_limits<float>::max(); |
| 52 | float_fc_params.lhs_cacheable = true; |
| 53 | float_fc_params.rhs_cacheable = false; |
| 54 | |
| 55 | tflite::RuntimeShape weight_shape({m_rows, m_cols}); |
| 56 | tflite::RuntimeShape input_shape({n_batch, m_cols}); |
| 57 | tflite::RuntimeShape output_shape({n_batch, m_rows}); |
| 58 | if (n_batch == 1) { |
| 59 | tflite::optimized_ops::FullyConnected( |
| 60 | float_fc_params, input_shape, vector, weight_shape, matrix, |
| 61 | output_shape, result, output_shape, output, cpu_backend_context); |
| 62 | } else { |
| 63 | tflite::optimized_ops::FullyConnected( |
| 64 | float_fc_params, input_shape, vector, weight_shape, matrix, |
| 65 | output_shape, nullptr, output_shape, output, cpu_backend_context); |
| 66 | for (int i = 0; i < m_rows * n_batch; ++i) { |
| 67 | output[i] += result[i]; |
| 68 | } |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | void ComputeRowSums( |
| 73 | int32_t* input_to_input_row_sums, int32_t* input_to_forget_row_sums, |
| 74 | int32_t* input_to_cell_row_sums, int32_t* input_to_output_row_sums, |
| 75 | int32_t* aux_input_to_input_row_sums, int32_t* aux_input_to_forget_row_sums, |
| 76 | int32_t* aux_input_to_cell_row_sums, int32_t* aux_input_to_output_row_sums, |
| 77 | int32_t* recurrent_to_input_row_sums, int32_t* recurrent_to_forget_row_sums, |
| 78 | int32_t* recurrent_to_cell_row_sums, int32_t* recurrent_to_output_row_sums, |
| 79 | int32_t* projection_weights_row_sums, int32_t* row_sums, int n_cell, |
| 80 | int n_input, int n_aux_input, int n_output, |
| 81 | const int8_t* input_to_input_weights_ptr, |
| 82 | const int8_t* input_to_forget_weights_ptr, |
| 83 | const int8_t* input_to_cell_weights_ptr, |
| 84 | const int8_t* input_to_output_weights_ptr, |
| 85 | const int8_t* aux_input_to_input_weights_ptr, |
| 86 | const int8_t* aux_input_to_forget_weights_ptr, |
| 87 | const int8_t* aux_input_to_cell_weights_ptr, |
| 88 | const int8_t* aux_input_to_output_weights_ptr, |
| 89 | const int8_t* recurrent_to_input_weights_ptr, |
| 90 | const int8_t* recurrent_to_forget_weights_ptr, |
| 91 | const int8_t* recurrent_to_cell_weights_ptr, |
| 92 | const int8_t* recurrent_to_output_weights_ptr, |
| 93 | const int8_t* projection_weights_ptr, bool use_cifg, |
| 94 | const float* aux_input_ptr) { |
| 95 | // Compute the row sums for dequantization |
| 96 | if (!use_cifg) { |
| 97 | tensor_utils::ReductionSumVector(input_to_input_weights_ptr, |
| 98 | input_to_input_row_sums, n_cell, n_input); |
| 99 | } |
| 100 | tensor_utils::ReductionSumVector(input_to_forget_weights_ptr, |
| 101 | input_to_forget_row_sums, n_cell, n_input); |
| 102 | tensor_utils::ReductionSumVector(input_to_cell_weights_ptr, |
| 103 | input_to_cell_row_sums, n_cell, n_input); |
| 104 | tensor_utils::ReductionSumVector(input_to_output_weights_ptr, |
| 105 | input_to_output_row_sums, n_cell, n_input); |
| 106 | |
| 107 | if (aux_input_ptr) { |
| 108 | if (!use_cifg) { |
| 109 | tensor_utils::ReductionSumVector(aux_input_to_input_weights_ptr, |
| 110 | aux_input_to_input_row_sums, n_cell, |
| 111 | n_aux_input); |
| 112 | } |
| 113 | tensor_utils::ReductionSumVector(aux_input_to_forget_weights_ptr, |
| 114 | aux_input_to_forget_row_sums, n_cell, |
| 115 | n_aux_input); |
| 116 | tensor_utils::ReductionSumVector(aux_input_to_cell_weights_ptr, |
| 117 | aux_input_to_cell_row_sums, n_cell, |
| 118 | n_aux_input); |
| 119 | tensor_utils::ReductionSumVector(aux_input_to_output_weights_ptr, |
| 120 | aux_input_to_output_row_sums, n_cell, |
| 121 | n_aux_input); |
| 122 | } |
| 123 | if (!use_cifg) { |
| 124 | tensor_utils::ReductionSumVector(recurrent_to_input_weights_ptr, |
| 125 | recurrent_to_input_row_sums, n_cell, |
| 126 | n_output); |
| 127 | } |
| 128 | tensor_utils::ReductionSumVector(recurrent_to_forget_weights_ptr, |
| 129 | recurrent_to_forget_row_sums, n_cell, |
| 130 | n_output); |
| 131 | tensor_utils::ReductionSumVector(recurrent_to_cell_weights_ptr, |
| 132 | recurrent_to_cell_row_sums, n_cell, |
| 133 | n_output); |
| 134 | tensor_utils::ReductionSumVector(recurrent_to_output_weights_ptr, |
| 135 | recurrent_to_output_row_sums, n_cell, |
| 136 | n_output); |
| 137 | |
| 138 | if (projection_weights_ptr != nullptr) { |
| 139 | tensor_utils::ReductionSumVector( |
| 140 | projection_weights_ptr, projection_weights_row_sums, n_output, n_cell); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | inline float GetTensorScale(const TfLiteTensor* tensor) { |
| 145 | return tensor == nullptr ? 1.0f : tensor->params.scale; |
| 146 | } |
| 147 | |
| 148 | // LINT.IfChange |
| 149 | // Calculates a single LSTM gate. |
| 150 | // |
| 151 | // Implements the following formula: (* is matrix multiply) |
| 152 | // gate = activate(W_input * input + W_aux * aux_input + |
| 153 | // W_peephole * cell + W_recurrent * prev_output + bias) |
| 154 | // with layer norm: |
| 155 | // gate = activate(W_norm * normalize(...) + bias) // not adding bias inside |
| 156 | // |
| 157 | // Activation is sigmoid except for the "cell" gate (configurable, usually tanh) |
| 158 | // |
| 159 | // Parameters: |
| 160 | // Input vectors (to LSTM): | Size: | Optional? |
| 161 | // input | n_input | |
| 162 | // aux_input | n_aux_input | y (bidir LSTM) |
| 163 | // Input vectors (persistent states): |
| 164 | // output_state | n_output | |
| 165 | // cell_state | n_cell | |
| 166 | // 'Constant' inputs: |
| 167 | // input_to_gate_weights | n_cell * n_input | |
| 168 | // aux_input_to_gate_weights | n_cell * n_aux_input | y (bidir LSTM) |
| 169 | // recurrent_to_gate_weights | n_cell * n_output | |
| 170 | // cell_to_gate_weights | n_cell | y (peephole) |
| 171 | // gate_bias | n_cell | |
| 172 | // layer_norm_coefficients | n_cell | y (layer norm) |
| 173 | // Output vector: |
| 174 | // gate | n_cell | |
| 175 | // Scalar parameters: |
| 176 | // n_batch - batch size / number of vectors |
| 177 | // n_input, n_aux_input, n_output, n_cell - size of vectors. |
| 178 | // activation - activation to use. |
| 179 | // is_input_all_zeros, is_aux_input_all_zeros - if input vectors are all zero. |
| 180 | // use_layer_norm - if doing layer norm LSTM. |
| 181 | inline void CalculateLstmGateFloat( |
| 182 | const float* input, const float* input_to_gate_weights, |
| 183 | const float* aux_input, const float* aux_input_to_gate_weights, |
| 184 | const float* output_state, const float* recurrent_to_gate_weights, |
| 185 | const float* cell_state, const float* cell_to_gate_weights, |
| 186 | const float* layer_norm_coefficients, const float* gate_bias, |
| 187 | const int n_batch, const int n_input, const int n_aux_input, |
| 188 | const int n_output, const int n_cell, |
| 189 | const TfLiteFusedActivation activation, float* gate, |
| 190 | const bool is_input_all_zeros, const bool is_aux_input_all_zeros, |
| 191 | float* output, CpuBackendContext* context) { |
| 192 | const bool use_peephole = (cell_to_gate_weights != nullptr); |
| 193 | const bool use_layer_norm = (layer_norm_coefficients != nullptr); |
| 194 | |
| 195 | // Initialize scratch buffers with bias for regular lstm or initialize with |
| 196 | // zero for layer norm lstm. |
| 197 | if (use_layer_norm) { |
| 198 | std::fill_n(gate, n_cell * n_batch, 0.0f); |
| 199 | } else { |
| 200 | tensor_utils::VectorBatchVectorAssign(gate_bias, n_cell, n_batch, gate); |
| 201 | } |
| 202 | // For each batch and cell: compute input_weight * input. |
| 203 | // Skip if input is all zeros. |
| 204 | float* accumulation_buffer = gate; |
| 205 | if (!is_input_all_zeros) { |
| 206 | MatrixBatchVectorMultiplyAccumulate(input_to_gate_weights, input, |
| 207 | accumulation_buffer, output, n_cell, |
| 208 | n_input, n_batch, context); |
| 209 | std::swap(accumulation_buffer, output); |
| 210 | } |
| 211 | // For each batch and cell: compute aux_input_weight * aux_input. |
| 212 | // Skip if auxiliary input is not available or all zeros. |
| 213 | if (!is_aux_input_all_zeros) { |
| 214 | MatrixBatchVectorMultiplyAccumulate(aux_input_to_gate_weights, aux_input, |
| 215 | accumulation_buffer, output, n_cell, |
| 216 | n_aux_input, n_batch, context); |
| 217 | std::swap(accumulation_buffer, output); |
| 218 | } |
| 219 | // For each batch and cell: compute recurrent_weight * output_state. |
| 220 | MatrixBatchVectorMultiplyAccumulate(recurrent_to_gate_weights, output_state, |
| 221 | accumulation_buffer, output, n_cell, |
| 222 | n_output, n_batch, context); |
| 223 | // For each batch and cell: compute cell_weight .* cell_state (peephole LSTM) |
| 224 | if (use_peephole) { |
| 225 | tensor_utils::VectorBatchVectorCwiseProductAccumulate( |
| 226 | cell_to_gate_weights, n_cell, cell_state, n_batch, output); |
| 227 | } |
| 228 | // Do layer normalization (if layer norm LSTM) |
| 229 | if (use_layer_norm) { |
| 230 | tensor_utils::MeanStddevNormalization(output, output, n_cell, n_batch); |
| 231 | tensor_utils::VectorBatchVectorCwiseProduct(layer_norm_coefficients, n_cell, |
| 232 | output, n_batch, output); |
| 233 | tensor_utils::VectorBatchVectorAdd(gate_bias, n_cell, n_batch, output); |
| 234 | } |
| 235 | // Apply activation |
| 236 | tensor_utils::ApplyActivationToVector(output, n_batch * n_cell, activation, |
| 237 | gate); |
| 238 | } |
| 239 | |
| 240 | // Updates the LSTM cell state, used by both float and hybrid LSTM versions. |
| 241 | // |
| 242 | // Implements the following formula: |
| 243 | // cell_state_new = clip(forget_gate * cell_state + input_gate * cell_gate) |
| 244 | // |
| 245 | // With CIFG LSTM, input gate is replaced by (1-forget_gate). |
| 246 | // |
| 247 | // Parameters: |
| 248 | // - n_batch, n_cell: sizes of vectors |
| 249 | // - cell_state: input/output vector, size n_batch*n_cell |
| 250 | // - input_gate: input vector, size n_batch*n_cell. |
| 251 | // - forget_gate: input/scratch vector, size n_batch*n_cell, modified with CIFG |
| 252 | // - cell_gate: input vector, size n_batch*n_cell. |
| 253 | // - use_cifg: use 1-forget_gate instead of input_gate. |
| 254 | // - clip: if > 0, clip the resulting cell state to [-clip, +clip]. |
| 255 | void UpdateLstmCellFloat(int n_batch, int n_cell, float* cell_state, |
| 256 | const float* input_gate, float* forget_gate, |
| 257 | const float* cell_gate, bool use_cifg, float clip) { |
| 258 | tensor_utils::VectorVectorCwiseProduct(forget_gate, cell_state, |
| 259 | n_batch * n_cell, cell_state); |
| 260 | |
| 261 | if (use_cifg) { |
| 262 | // With CIFG, input_gate = 1-forget_gate. Use the forget_gate array as |
| 263 | // scratch, as input_gate array is not allocated in this case. (Be careful |
| 264 | // not to write to the scratch before reading the forget gate data.) |
| 265 | float* scratch = forget_gate; |
| 266 | tensor_utils::Sub1Vector(forget_gate, n_batch * n_cell, scratch); |
| 267 | tensor_utils::VectorVectorCwiseProductAccumulate( |
| 268 | cell_gate, scratch, n_batch * n_cell, cell_state); |
| 269 | } else { |
| 270 | tensor_utils::VectorVectorCwiseProductAccumulate( |
| 271 | cell_gate, input_gate, n_batch * n_cell, cell_state); |
| 272 | } |
| 273 | if (clip > 0.0f) { |
| 274 | tensor_utils::CwiseClipping(cell_state, n_batch * n_cell, clip); |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | // Calculates the output state tensor of an LSTM step. |
| 279 | // |
| 280 | // Implements the following formula: |
| 281 | // output_no_projection = output_gate .* activate(cell_state) |
| 282 | // (elementwise vector product) |
| 283 | // If no projection is used: |
| 284 | // output = output_state = output_no_projection |
| 285 | // With projection: |
| 286 | // output = output_state = clip(W*output_no_projection + bias) |
| 287 | // |
| 288 | // Output might not have a different 'stride' than n_batch, so we need to copy. |
| 289 | // |
| 290 | // Parameters: |
| 291 | // - n_batch: batches: the number of distinct vectors in each array. |
| 292 | // - n_cell, n_output: sizes of vectors. |
| 293 | // - cell_state, output_gate: input vectors, size n_batch*n_cell. |
| 294 | // - projection_weights, projection_weights_scale, projection_bias: |
| 295 | // constant inputs, describing projection matrix and bias. |
| 296 | // - proj_clip: if > 0, clip the output of the projection. |
| 297 | // - output_state: output vector, size n_batch*n_output. Must be contigous. |
| 298 | // - scratch: scratch area to store output_no_projection. Size n_batch*n_cell. |
| 299 | // - projection_bias_scratch: scratch area to store projection_bias. Size |
| 300 | // n_batch*n_cell. |
| 301 | // - context: the CpuBackendContext for use with matrix multiplications. |
| 302 | void CalculateLstmOutputFloat(int n_batch, int n_cell, int n_output, |
| 303 | const float* cell_state, const float* output_gate, |
| 304 | TfLiteFusedActivation activation, |
| 305 | const float* projection_weights, |
| 306 | const float* projection_bias, |
| 307 | const float proj_clip, float* output_state, |
| 308 | float* scratch, float* projection_bias_scratch, |
| 309 | CpuBackendContext* context) { |
| 310 | tensor_utils::ApplyActivationToVector(cell_state, n_batch * n_cell, |
| 311 | activation, scratch); |
| 312 | tensor_utils::VectorVectorCwiseProduct(output_gate, scratch, n_batch * n_cell, |
| 313 | scratch); |
| 314 | |
| 315 | const bool use_projection = (projection_weights != nullptr); |
| 316 | const bool use_projection_bias = (projection_bias != nullptr); |
| 317 | |
| 318 | if (use_projection) { |
| 319 | if (use_projection_bias) { |
| 320 | tensor_utils::VectorBatchVectorAssign(projection_bias, n_output, n_batch, |
| 321 | projection_bias_scratch); |
| 322 | } else { |
| 323 | std::fill_n(projection_bias_scratch, n_batch * n_output, 0.0f); |
| 324 | } |
| 325 | MatrixBatchVectorMultiplyAccumulate(projection_weights, scratch, |
| 326 | projection_bias_scratch, output_state, |
| 327 | n_output, n_cell, n_batch, context); |
| 328 | if (proj_clip > 0.0f) { |
| 329 | tensor_utils::CwiseClipping(output_state, n_batch * n_output, proj_clip); |
| 330 | } |
| 331 | } else { |
| 332 | std::copy_n(scratch, n_batch * n_output, output_state); |
| 333 | } |
| 334 | } |
| 335 | // LINT.ThenChange(../tools/optimize/calibration/builtin_logging_ops/lstm.cc,\ |
| 336 | // ../experimental/kernels/fp16/lstm_eval.cc) |
| 337 | |
| 338 | // Calculates a single LSTM gate, hybrid version. |
| 339 | // Implements the same functionality as CalculateLstmGateFloat. |
| 340 | void CalculateLstmGateHybrid( |
| 341 | // Input and weights |
| 342 | const int8_t* input, const float* input_sf, const int32_t* input_zp, |
| 343 | const int8_t* input_to_gate_weights, |
| 344 | const uint8_t* input_to_gate_weights_ledger, |
| 345 | const float input_to_gate_weights_scale, int32_t* input_to_gate_row_sums, |
| 346 | // Aux input and weights |
| 347 | const int8_t* aux_input, const float* aux_input_sf, |
| 348 | const int32_t* aux_input_zp, const int8_t* aux_input_to_gate_weights, |
| 349 | const float aux_input_to_gate_weights_scale, |
| 350 | int32_t* aux_input_to_gate_row_sums, |
| 351 | // Output state and weights |
| 352 | const int8_t* output_state, const float* output_state_sf, |
| 353 | const int32_t* output_state_zp, const int8_t* recurrent_to_gate_weights, |
| 354 | const uint8_t* recurrent_to_gate_weights_ledger, |
| 355 | const float recurrent_to_gate_weights_scale, |
| 356 | int32_t* recurrent_to_gate_row_sums, |
| 357 | // Cell state and weights (peephole LSTM) |
| 358 | const float* cell_state, const int8_t* cell_to_gate_weights, |
| 359 | const float cell_to_gate_weights_scale, |
| 360 | // Layer normalization coefficients (layer norm LSTM) + gate bias |
| 361 | const float* layer_norm_coefficients, const float* gate_bias, |
| 362 | // Array sizes |
| 363 | const int n_batch, const int n_input, const int n_aux_input, |
| 364 | const int n_output, const int n_cell, |
| 365 | const TfLiteFusedActivation activation, |
| 366 | // Output |
| 367 | float* gate, |
| 368 | // Parameters for performance optimizations |
| 369 | const bool is_input_all_zeros, const bool is_aux_input_all_zeros, |
| 370 | const bool is_output_state_all_zeros, bool* compute_row_sums, |
| 371 | CpuBackendContext* context, |
| 372 | // Scratch arrays |
| 373 | float* scratch0, // size: n_batch |
| 374 | float* scratch1, // size: n_cell, only used if peephole LSTM |
| 375 | int32_t* accum_scratch // For MatrixBatchVectorMultiplyAccumulate |
| 376 | ) { |
| 377 | const bool use_peephole = (cell_to_gate_weights != nullptr); |
| 378 | const bool use_layer_norm = (layer_norm_coefficients != nullptr); |
| 379 | |
| 380 | // Initialize scratch buffers with bias for regular lstm or initialize with |
| 381 | // zero for layer norm lstm. |
| 382 | if (use_layer_norm) { |
| 383 | std::fill_n(gate, n_cell * n_batch, 0.0f); |
| 384 | } else { |
| 385 | tensor_utils::VectorBatchVectorAssign(gate_bias, n_cell, n_batch, gate); |
| 386 | } |
| 387 | // For each batch and cell: compute input_weight * input. |
| 388 | // Skip if input is all zeros. |
| 389 | if (!is_input_all_zeros) { |
| 390 | if (input_to_gate_weights_ledger != nullptr) { |
| 391 | std::vector<float> scales(n_batch); |
| 392 | for (int i = 0; i < n_batch; i++) { |
| 393 | scales[i] = input_to_gate_weights_scale * input_sf[i]; |
| 394 | } |
| 395 | tensor_utils::SparseMatrixBatchVectorMultiplyAccumulate( |
| 396 | input_to_gate_weights, input_to_gate_weights_ledger, n_cell, n_input, |
| 397 | input, scales.data(), n_batch, gate); |
| 398 | |
| 399 | } else { |
| 400 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 401 | input_to_gate_weights, n_cell, n_input, input, |
| 402 | input_to_gate_weights_scale, input_sf, n_batch, gate, |
| 403 | /*per_channel_scale=*/nullptr, input_zp, accum_scratch, |
| 404 | input_to_gate_row_sums, compute_row_sums, scratch0, context); |
| 405 | } |
| 406 | } |
| 407 | // For each batch and cell: compute aux_input_weight * aux_input. |
| 408 | // Skip if auxiliary input is not available or all zeros. |
| 409 | if (!is_aux_input_all_zeros) { |
| 410 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 411 | aux_input_to_gate_weights, n_cell, n_aux_input, aux_input, |
| 412 | aux_input_to_gate_weights_scale, aux_input_sf, n_batch, gate, |
| 413 | /*per_channel_scale=*/nullptr, aux_input_zp, accum_scratch, |
| 414 | aux_input_to_gate_row_sums, compute_row_sums, scratch0, context); |
| 415 | } |
| 416 | // For each batch and cell: compute recurrent_weight * output_state. |
| 417 | // Skip if output state is all zeros. |
| 418 | if (!is_output_state_all_zeros) { |
| 419 | if (recurrent_to_gate_weights_ledger != nullptr) { |
| 420 | std::vector<float> scales(n_batch); |
| 421 | for (int i = 0; i < n_batch; i++) { |
| 422 | scales[i] = recurrent_to_gate_weights_scale * input_sf[i]; |
| 423 | } |
| 424 | tensor_utils::SparseMatrixBatchVectorMultiplyAccumulate( |
| 425 | recurrent_to_gate_weights, recurrent_to_gate_weights_ledger, n_cell, |
| 426 | n_output, output_state, scales.data(), n_batch, gate); |
| 427 | } else { |
| 428 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 429 | recurrent_to_gate_weights, n_cell, n_output, output_state, |
| 430 | recurrent_to_gate_weights_scale, output_state_sf, n_batch, gate, |
| 431 | /*per_channel_scale=*/nullptr, output_state_zp, accum_scratch, |
| 432 | recurrent_to_gate_row_sums, compute_row_sums, scratch0, context); |
| 433 | } |
| 434 | } |
| 435 | // For each batch and cell: compute cell_weight .* cell_state (peephole LSTM) |
| 436 | if (use_peephole) { |
| 437 | float* recovered_cell_weights = scratch1; |
| 438 | tensor_utils::VectorScalarMultiply(cell_to_gate_weights, n_cell, |
| 439 | cell_to_gate_weights_scale, |
| 440 | recovered_cell_weights); |
| 441 | tensor_utils::VectorBatchVectorCwiseProductAccumulate( |
| 442 | recovered_cell_weights, n_cell, cell_state, n_batch, gate); |
| 443 | } |
| 444 | // Do layer normalization (if layer norm LSTM) |
| 445 | if (use_layer_norm) { |
| 446 | tensor_utils::MeanStddevNormalization(gate, gate, n_cell, n_batch); |
| 447 | tensor_utils::VectorBatchVectorCwiseProduct(layer_norm_coefficients, n_cell, |
| 448 | gate, n_batch, gate); |
| 449 | tensor_utils::VectorBatchVectorAdd(gate_bias, n_cell, n_batch, gate); |
| 450 | } |
| 451 | // Apply activation |
| 452 | tensor_utils::ApplyActivationToVector(gate, n_cell * n_batch, activation, |
| 453 | gate); |
| 454 | } |
| 455 | |
| 456 | // Calculates the output state tensor of an LSTM step. See Float version too. |
| 457 | // |
| 458 | // Parameters: |
| 459 | // - n_batch: batches: the number of distinct vectors in each array. |
| 460 | // - n_cell, n_output: sizes of vectors. |
| 461 | // - cell_state, output_gate: input vectors, size n_batch*n_cell. |
| 462 | // - projection_weights, projection_weights_scale, projection_bias: |
| 463 | // constant inputs, describing projection matrix and bias. |
| 464 | // - proj_clip: if > 0, clip the output of the projection. |
| 465 | // - output_state: output vector, size n_batch*n_output. Must be contigous. |
| 466 | // - asymmetric_quantize_inputs: parameter to control quantization. |
| 467 | // - projection_weights_row_sums, compute_row_sums, context: Data for optimized |
| 468 | // MatrixBatchVectorMultiplyAccumulate. |
| 469 | // - scratch0: scratch area of size n_batch*n_cell |
| 470 | // - scratch1: scratch area of size n_batch*n_cell |
| 471 | // - scratch2: scratch area of size n_batch |
| 472 | // - scratch3: scratch area of size n_batch |
| 473 | // - scratch4: scratch area used by MatrixBatchVectorMultiplyAccumulate |
| 474 | void CalculateLstmOutputHybrid( |
| 475 | int n_batch, int n_cell, int n_output, const float* cell_state, |
| 476 | const float* output_gate, TfLiteFusedActivation activation, |
| 477 | const int8_t* projection_weights, const uint8_t* projection_weights_ledger, |
| 478 | float projection_weights_scale, const float* projection_bias, |
| 479 | const float proj_clip, float* output_state, bool asymmetric_quantize_inputs, |
| 480 | int32_t* projection_weights_row_sums, bool* compute_row_sums, |
| 481 | CpuBackendContext* context, float* scratch0, int8_t* scratch1, |
| 482 | float* scratch2, int32_t* scratch3, int32_t* scratch4) { |
| 483 | tensor_utils::ApplyActivationToVector(cell_state, n_batch * n_cell, |
| 484 | activation, scratch0); |
| 485 | tensor_utils::VectorVectorCwiseProduct(output_gate, scratch0, |
| 486 | n_batch * n_cell, scratch0); |
| 487 | |
| 488 | const bool use_projection = (projection_weights != nullptr); |
| 489 | const bool use_projection_bias = (projection_bias != nullptr); |
| 490 | |
| 491 | if (use_projection) { |
| 492 | if (use_projection_bias) { |
| 493 | tensor_utils::VectorBatchVectorAssign(projection_bias, n_output, n_batch, |
| 494 | output_state); |
| 495 | } else { |
| 496 | std::fill_n(output_state, n_batch * n_output, 0.0f); |
| 497 | } |
| 498 | if (!tensor_utils::IsZeroVector(scratch0, n_batch * n_cell)) { |
| 499 | // Save quantization and matmul computation for all zero output. |
| 500 | tensor_utils::BatchQuantizeFloats(scratch0, n_batch, n_cell, scratch1, |
| 501 | scratch2, scratch3, |
| 502 | asymmetric_quantize_inputs); |
| 503 | if (projection_weights_ledger != nullptr) { |
| 504 | std::vector<float> scales(n_batch); |
| 505 | for (int i = 0; i < n_batch; i++) { |
| 506 | scales[i] = projection_weights_scale * scratch2[i]; |
| 507 | } |
| 508 | tensor_utils::SparseMatrixBatchVectorMultiplyAccumulate( |
| 509 | projection_weights, projection_weights_ledger, n_output, n_cell, |
| 510 | scratch1, scales.data(), n_batch, output_state); |
| 511 | } else { |
| 512 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 513 | projection_weights, n_output, n_cell, scratch1, |
| 514 | projection_weights_scale, scratch2, n_batch, output_state, |
| 515 | /*per_channel_scale=*/nullptr, scratch3, scratch4, |
| 516 | projection_weights_row_sums, compute_row_sums, scratch2, context); |
| 517 | } |
| 518 | } |
| 519 | if (proj_clip > 0.0f) { |
| 520 | tensor_utils::CwiseClipping(output_state, n_batch * n_output, proj_clip); |
| 521 | } |
| 522 | } else { |
| 523 | std::copy_n(scratch0, n_batch * n_output, output_state); |
| 524 | } |
| 525 | } |
| 526 | |
| 527 | // Calculates a single LSTM gate, int8x8_16 version. |
| 528 | // Implements the same functionality as CalculateLstmGateFloat. |
| 529 | void CalculateLstmGateInteger8x8_16( |
| 530 | // Input and weights |
| 531 | const int8_t* input, const int8_t* input_to_gate_weights, |
| 532 | const int32_t* input_to_gate_bias, const int32_t input_to_gate_scale_a, |
| 533 | const int32_t input_to_gate_scale_b, |
| 534 | // Output state and weights |
| 535 | const int8_t* output_state, const int8_t* recurrent_to_gate_weights, |
| 536 | const int32_t* recurrent_to_gate_bias, |
| 537 | const int32_t recurrent_to_gate_scale_a, |
| 538 | const int32_t recurrent_to_gate_scale_b, |
| 539 | // Cell state and weights |
| 540 | const int16_t* cell_state, const int16_t* cell_to_gate_weights, |
| 541 | const int32_t cell_to_gate_scale_a, const int32_t cell_to_gate_scale_b, |
| 542 | // Layer normalization parameters (layer norm LSTM) |
| 543 | const int16_t* layer_norm_coefficients, const int32_t* layer_norm_bias, |
| 544 | const int32_t layer_norm_input_scale_a, |
| 545 | const int32_t layer_norm_input_scale_b, |
| 546 | const int32_t layer_norm_variance_guard, |
| 547 | // Array sizes |
| 548 | const int n_batch, const int n_input, const int n_output, const int n_cell, |
| 549 | const TfLiteFusedActivation activation, |
| 550 | // Output |
| 551 | int16_t* gate, |
| 552 | // Parameters for performance optimizations |
| 553 | CpuBackendContext* context, |
| 554 | // Scratch arrays |
| 555 | int32_t* scratch5) { |
| 556 | const bool use_peephole = (cell_to_gate_weights != nullptr); |
| 557 | const bool use_layer_norm = (layer_norm_coefficients != nullptr); |
| 558 | |
| 559 | // Initialize scratch buffers with zeros. Note that unlike float and hybrid |
| 560 | // versions, bias is only used in layer normalization. |
| 561 | std::fill_n(gate, n_batch * n_cell, 0); |
| 562 | // For each batch and cell: compute input_weight * input. |
| 563 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 564 | input, input_to_gate_bias, input_to_gate_weights, input_to_gate_scale_a, |
| 565 | input_to_gate_scale_b, n_batch, n_input, n_cell, 0, scratch5, gate, |
| 566 | context); |
| 567 | // Note: no aux_input. |
| 568 | |
| 569 | // For each batch and cell: compute recurrent_weight * output_state. |
| 570 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 571 | output_state, recurrent_to_gate_bias, recurrent_to_gate_weights, |
| 572 | recurrent_to_gate_scale_a, recurrent_to_gate_scale_b, n_batch, n_output, |
| 573 | n_cell, 0, scratch5, gate, context); |
| 574 | // For each batch and cell: compute cell_weight * cell_state (peephole LSTM) |
| 575 | if (use_peephole) { |
| 576 | tensor_utils::VectorBatchVectorCwiseProductAccumulate( |
| 577 | cell_to_gate_weights, n_output, cell_state, n_batch, |
| 578 | cell_to_gate_scale_a, cell_to_gate_scale_b, gate); |
| 579 | } |
| 580 | // Do layer normalization (if layer norm LSTM) |
| 581 | if (use_layer_norm) { |
| 582 | tensor_utils::ApplyLayerNorm( |
| 583 | gate, layer_norm_coefficients, layer_norm_bias, |
| 584 | layer_norm_input_scale_a, layer_norm_input_scale_b, |
| 585 | layer_norm_variance_guard, n_batch, n_cell, gate); |
| 586 | } |
| 587 | // Apply activation |
| 588 | switch (activation) { |
| 589 | case kTfLiteActSigmoid: |
| 590 | tensor_utils::ApplySigmoid(gate, n_batch, n_cell, gate); |
| 591 | break; |
| 592 | case kTfLiteActTanh: |
| 593 | tensor_utils::ApplyTanh(3, gate, n_batch, n_cell, gate); |
| 594 | break; |
| 595 | default: |
| 596 | // Only Sigmoid or Tanh is used. |
| 597 | TFLITE_ASSERT_FALSE; |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | // Updates the LSTM cell state, used by both integer LSTM versions. |
| 602 | // Also see UpdateLstmCellFloat. |
| 603 | // |
| 604 | // Parameters: |
| 605 | // - n_batch, n_cell: sizes of vectors |
| 606 | // - cell_state: input/output vector, size n_batch*n_cell |
| 607 | // - cell_state_scale: scaling factor of cell state. |
| 608 | // - input_gate: input vector, size n_batch*n_cell. |
| 609 | // - forget_gate: input/scratch vector, size n_batch*n_cell, always modified. |
| 610 | // - cell_gate: input vector, size n_batch*n_cell. |
| 611 | // - use_cifg: use 1-forget_gate instead of input_gate. |
| 612 | // - clip: if > 0, clip the resulting cell state to [-clip, +clip]. |
| 613 | void UpdateLstmCellInteger(int n_batch, int n_cell, int16_t* cell_state, |
| 614 | int32_t cell_state_scale, const int16_t* input_gate, |
| 615 | int16_t* forget_gate, const int16_t* cell_gate, |
| 616 | bool use_cifg, int16_t clip) { |
| 617 | // Use the forget_gate array as scratch, as input_gate array is not allocated |
| 618 | // in CIFG case. (Be careful not to write to the scratch before reading the |
| 619 | // forget gate data.) |
| 620 | int16_t* scratch = forget_gate; |
| 621 | |
| 622 | tensor_utils::CwiseMul(forget_gate, cell_state, n_batch, n_cell, 15, |
| 623 | cell_state); |
| 624 | if (use_cifg) { |
| 625 | tensor_utils::Sub1Vector(forget_gate, n_batch * n_cell, scratch); |
| 626 | tensor_utils::CwiseMul(scratch, cell_gate, n_batch, n_cell, |
| 627 | 30 + cell_state_scale, scratch); |
| 628 | } else { |
| 629 | tensor_utils::CwiseMul(input_gate, cell_gate, n_batch, n_cell, |
| 630 | 30 + cell_state_scale, scratch); |
| 631 | } |
| 632 | tensor_utils::CwiseAdd(cell_state, scratch, n_batch, n_cell, cell_state); |
| 633 | |
| 634 | if (clip > 0) { |
| 635 | tensor_utils::CwiseClipping(cell_state, n_batch * n_cell, clip); |
| 636 | } |
| 637 | } |
| 638 | |
| 639 | // Calculates the output state tensor of an LSTM step. See Float and hybrid |
| 640 | // versions as well. |
| 641 | // |
| 642 | // Parameters: |
| 643 | // - n_batch: batches: the number of distinct vectors in each array. |
| 644 | // - n_cell, n_output: sizes of vectors. |
| 645 | // - cell_state, output_gate: input vectors, size n_batch*n_cell. |
| 646 | // - cell_state_scale: scaling of cell_state. |
| 647 | // - hidden_scale_[a|b]: effective scale of cell_state.*output_gate |
| 648 | // - hidden_zp: zero_point for cell_state.*output_gate |
| 649 | // - projection_weights, proj_scale_[a|b], projection_bias: |
| 650 | // constant inputs, describing projection matrix and bias. |
| 651 | // - output_state_zp: zero point of output_state. (Input, calibrated value.) |
| 652 | // - quantized_proj_clip: if > 0, clip the output of the projection. |
| 653 | // - output_state: output vector, size n_batch*n_output. Must be contigous. |
| 654 | // - context: data for optimized MatrixBatchVectorMultiplyAccumulate. |
| 655 | // - scratch0: scratch area of size n_batch*n_cell |
| 656 | // - scratch1: scratch area of size n_batch*n_cell |
| 657 | // - scratch2: scratch area used by MatrixBatchVectorMultiplyAccumulate |
| 658 | void CalculateLstmOutputInteger8x8_16( |
| 659 | int n_batch, int n_cell, int n_output, const int16_t* cell_state, |
| 660 | int32_t cell_state_scale, const int16_t* output_gate, |
| 661 | int32_t hidden_scale_a, int32_t hidden_scale_b, int32_t hidden_zp, |
| 662 | const int8_t* projection_weights, int32_t proj_scale_a, |
| 663 | int32_t proj_scale_b, const int32_t* projection_bias, |
| 664 | int32_t output_state_zp, int8_t quantized_proj_clip, int8_t* output_state, |
| 665 | CpuBackendContext* context, int16_t* scratch0, int8_t* scratch1, |
| 666 | int32_t* scratch2) { |
| 667 | // Note: unlike float/hybrid, the activation is always Tanh. |
| 668 | tensor_utils::ApplyTanh(15 + cell_state_scale, cell_state, n_batch, n_cell, |
| 669 | scratch0); |
| 670 | tensor_utils::CwiseMul(output_gate, scratch0, hidden_scale_a, hidden_scale_b, |
| 671 | n_batch, n_cell, hidden_zp, scratch1); |
| 672 | |
| 673 | const bool use_projection = (projection_weights != nullptr); |
| 674 | |
| 675 | if (use_projection) { |
| 676 | // Note: no bias like in float/hybrid |
| 677 | std::fill_n(output_state, n_batch * n_output, 0); |
| 678 | tensor_utils::MatrixBatchVectorMultiplyAccumulate( |
| 679 | scratch1, projection_bias, projection_weights, proj_scale_a, |
| 680 | proj_scale_b, n_batch, n_cell, n_output, output_state_zp, scratch2, |
| 681 | output_state, context); |
| 682 | if (quantized_proj_clip > 0) { |
| 683 | tensor_utils::CwiseClipping(output_state, n_batch * n_output, |
| 684 | quantized_proj_clip); |
| 685 | } |
| 686 | } else { |
| 687 | std::copy_n(scratch1, n_batch * n_output, output_state); |
| 688 | } |
| 689 | } |
| 690 | |
| 691 | // Calculates a single LSTM gate, int8x8_8 version. |
| 692 | // Implements the same functionality as CalculateLstmGateFloat. |
| 693 | void CalculateLstmGateInteger8x8_8( |
| 694 | // Inputs and weights |
| 695 | const int8_t* input, int32_t input_zp, const int8_t* input_to_gate_weight, |
| 696 | const int32_t input_to_gate_scale_a, const int32_t input_to_gate_scale_b, |
| 697 | const int32_t input_times_weights_scale_a, |
| 698 | const int32_t input_times_weights_scale_b, |
| 699 | const int32_t input_times_weights_zp, |
| 700 | // Output state and weights |
| 701 | const int8_t* output_state, const int32_t output_state_zp, |
| 702 | const int8_t* recurrent_to_gate_weight, |
| 703 | const int32_t recurrent_to_gate_scale_a, |
| 704 | const int32_t recurrent_to_gate_scale_b, |
| 705 | const int32_t output_state_times_weights_scale_a, |
| 706 | const int32_t output_state_times_weights_scale_b, |
| 707 | const int32_t output_state_times_weights_zp, |
| 708 | // Layer normalization parameters (layer norm LSTM) |
| 709 | const int16_t* layer_norm_gate_weight, |
| 710 | const int32_t layer_norm_gate_scale_a, |
| 711 | const int32_t layer_norm_gate_scale_b, const int32_t* gate_bias, |
| 712 | // Array sizes |
| 713 | const int n_batch, const int n_input, const int n_output, const int n_cell, |
| 714 | const TfLiteFusedActivation activation, |
| 715 | // Output |
| 716 | int16_t* gate, |
| 717 | // Scratch arrays, both sized n_batch*n_cell |
| 718 | int8_t* scratch0, int8_t* scratch1) { |
| 719 | // Multiply input * input_weights => scratch0 |
| 720 | tensor_utils::MatrixBatchVectorMultiply( |
| 721 | input, input_zp, input_to_gate_weight, input_to_gate_scale_a, |
| 722 | input_to_gate_scale_b, n_batch, n_input, n_cell, scratch0, |
| 723 | input_times_weights_zp); |
| 724 | // Multiply output_state * recurrent_weights => scratch1 |
| 725 | tensor_utils::MatrixBatchVectorMultiply( |
| 726 | output_state, output_state_zp, recurrent_to_gate_weight, |
| 727 | recurrent_to_gate_scale_a, recurrent_to_gate_scale_b, n_batch, n_output, |
| 728 | n_cell, scratch1, output_state_times_weights_zp); |
| 729 | // Add scratch0 + scratch1 => gate |
| 730 | tensor_utils::TwoGateSaturatingAdd( |
| 731 | scratch0, input_times_weights_zp, scratch1, output_state_times_weights_zp, |
| 732 | input_times_weights_scale_a, input_times_weights_scale_b, |
| 733 | output_state_times_weights_scale_a, output_state_times_weights_scale_b, |
| 734 | n_batch, n_cell, gate); |
| 735 | // Apply layer normalization. |
| 736 | tensor_utils::ApplyLayerNormFloat( |
| 737 | gate, layer_norm_gate_weight, layer_norm_gate_scale_a, |
| 738 | layer_norm_gate_scale_b, gate_bias, n_batch, n_cell, gate); |
| 739 | // Apply activation. |
| 740 | switch (activation) { |
| 741 | case kTfLiteActSigmoid: |
| 742 | tensor_utils::ApplySigmoidFloat(gate, n_batch, n_cell, gate); |
| 743 | break; |
| 744 | case kTfLiteActTanh: |
| 745 | tensor_utils::ApplyTanhFloat(gate, n_batch, n_cell, -12, gate); |
| 746 | break; |
| 747 | default: |
| 748 | // Only Sigmoid or Tanh is used. |
| 749 | TFLITE_ASSERT_FALSE; |
| 750 | } |
| 751 | } |
| 752 | |
| 753 | // Calculates the output state tensor of an LSTM step. See Float and hybrid |
| 754 | // versions as well. |
| 755 | // |
| 756 | // Parameters: |
| 757 | // - n_batch: batches: the number of distinct vectors in each array. |
| 758 | // - n_cell, n_output: sizes of vectors. |
| 759 | // - cell_state, output_gate: input vectors, size n_batch*n_cell. |
| 760 | // - projection_weights, proj_scale_[a|b], projection_bias: |
| 761 | // constant inputs, describing projection matrix and bias. |
| 762 | // - output_state_zp: zero point of the output state. |
| 763 | // - quantized_proj_clip: if > 0, clip the output of the projection. |
| 764 | // - output_state: output vector, size n_batch*n_output. Must be contigous. |
| 765 | // - scratch: scratch area of size n_batch*n_cell |
| 766 | void CalculateLstmOutputInteger8x8_8( |
| 767 | int n_batch, int n_cell, int n_output, const int16_t* cell_state, |
| 768 | const int16_t* output_gate, const int8_t* projection_weights, |
| 769 | int32_t proj_scale_a, int32_t proj_scale_b, const int32_t* projection_bias, |
| 770 | int32_t output_state_zp, int32_t quantized_proj_clip, int8_t* output_state, |
| 771 | int16_t* scratch) { |
| 772 | // Note: unlike float/hybrid, the activation is always Tanh. |
| 773 | tensor_utils::ApplyTanhFloat(cell_state, n_batch, n_cell, -15, scratch); |
| 774 | tensor_utils::CwiseMul(output_gate, scratch, n_batch, n_cell, 15 + 15 - 15, |
| 775 | scratch); |
| 776 | // Note: no bias like in float/hybrid |
| 777 | tensor_utils::MatrixBatchVectorMultiply( |
| 778 | scratch, projection_weights, proj_scale_a, proj_scale_b, projection_bias, |
| 779 | n_batch, n_cell, n_output, output_state_zp, output_state); |
| 780 | if (quantized_proj_clip > 0) { |
| 781 | tensor_utils::CwiseClipping(output_state, n_batch * n_output, |
| 782 | quantized_proj_clip); |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | // Performs an LSTM batch inference step for input specified by input_ptr. |
| 787 | // The LSTM cell is specified by the pointers to its weights (*_weights_ptr) and |
| 788 | // biases (*_bias_ptr), and buffers (*_scratch), along with additional |
| 789 | // parameters: |
| 790 | // - params: various LSTM params including activation, clipping, etc., |
| 791 | // - n_batch: size of batch, |
| 792 | // - n_cell: number of cells (or units), |
| 793 | // - n_input: the input size, |
| 794 | // - n_aux_input: the auxiliary input size. |
| 795 | // - n_output: the output size. |
| 796 | // - output_batch_leading_dim: the leading dimension of the output buffer. |
| 797 | // - context: the CpuBackendContext for use with matrix multiplications. |
| 798 | // |
| 799 | // Input of size 'n_batch * n_input': |
| 800 | // input_ptr |
| 801 | // Input of size 'n_batch * n_aux_input': |
| 802 | // aux_input_ptr - optional (can be nullptr) |
| 803 | // |
| 804 | // LSTM weights: |
| 805 | // Input weights of size 'n_cell * n_input': |
| 806 | // input_to_input_weights - optional |
| 807 | // input_to_forget_weights |
| 808 | // input_to_cell_weights |
| 809 | // input_to_output_weights |
| 810 | // Auxiliary input weights of size 'n_cell * n_aux_input': |
| 811 | // aux_input_to_input_weights - optional |
| 812 | // aux_input_to_forget_weights - optional |
| 813 | // aux_input_to_cell_weights - optional |
| 814 | // aux_input_to_output_weights - optional |
| 815 | // Recurrent weights of size 'n_cell * n_output': |
| 816 | // recurrent_to_input_weights - optional |
| 817 | // recurrent_to_forget_weights |
| 818 | // recurrent_to_cell_weights |
| 819 | // recurrent_to_input_weights |
| 820 | // Peephole weights of size 'n_cell', representing diagonal matrices. |
| 821 | // cell_to_input_weights - optional |
| 822 | // cell_to_cell_weights - optional |
| 823 | // cell_to_output_weights - optional |
| 824 | // Projection weights of size 'n_output * n_cell' |
| 825 | // projection_weights_ptr - optional |
| 826 | // Gate biases of size 'n_cell': |
| 827 | // input_gate_bias_ptr - optional |
| 828 | // forget_gate_bias_ptr |
| 829 | // cell_gate_bias_ptr |
| 830 | // output_gate_bias_ptr |
| 831 | // |
| 832 | // Layer norm coefficients of size 'n_cell', representing diagonal matrices. |
| 833 | // input_layer_norm_coefficients_ptr - optional |
| 834 | // forget_layer_norm_coefficients_ptr - optional |
| 835 | // cell_layer_norm_coefficients_ptr - optional |
| 836 | // output_layer_norm_coefficients_ptr - optional |
| 837 | // |
| 838 | // The pointers to the cell and output state and the output are updated. |
| 839 | // |
| 840 | // The pointers input_ptr, aux_input_ptr, and output_ptr point to data aligned |
| 841 | // in batch_major order, and each step processes batch_size many inputs from |
| 842 | // input_ptr, and updates batch_size many cell and output states. |
| 843 | // |
| 844 | // The output_batch_dim is output.shape[-1], i.e. the outermost dimension of the |
| 845 | // output tensor, and in most cases will be equal to n_output. It is usually not |
| 846 | // when we want to store the LSTM output into a slice of the output tensor, e.g. |
| 847 | // for bidirectional LSTMs with merge_outputs. In this case, the batched |
| 848 | // operations cannot be used since they assume that the batched outputs are |
| 849 | // contiguous, and we manually loop over the batched outputs. |
| 850 | // LINT.IfChange |
| 851 | inline void LstmStepFloat( |
| 852 | const float* input_ptr, const float* input_to_input_weights_ptr, |
| 853 | const float* input_to_forget_weights_ptr, |
| 854 | const float* input_to_cell_weights_ptr, |
| 855 | const float* input_to_output_weights_ptr, const float* aux_input_ptr, |
| 856 | const float* aux_input_to_input_weights_ptr, |
| 857 | const float* aux_input_to_forget_weights_ptr, |
| 858 | const float* aux_input_to_cell_weights_ptr, |
| 859 | const float* aux_input_to_output_weights_ptr, |
| 860 | const float* recurrent_to_input_weights_ptr, |
| 861 | const float* recurrent_to_forget_weights_ptr, |
| 862 | const float* recurrent_to_cell_weights_ptr, |
| 863 | const float* recurrent_to_output_weights_ptr, |
| 864 | const float* cell_to_input_weights_ptr, |
| 865 | const float* cell_to_forget_weights_ptr, |
| 866 | const float* cell_to_output_weights_ptr, |
| 867 | const float* input_layer_norm_coefficients_ptr, |
| 868 | const float* forget_layer_norm_coefficients_ptr, |
| 869 | const float* cell_layer_norm_coefficients_ptr, |
| 870 | const float* output_layer_norm_coefficients_ptr, |
| 871 | const float* input_gate_bias_ptr, const float* forget_gate_bias_ptr, |
| 872 | const float* cell_gate_bias_ptr, const float* output_gate_bias_ptr, |
| 873 | const float* projection_weights_ptr, const float* projection_bias_ptr, |
| 874 | const TfLiteLSTMParams* params, int n_batch, int n_cell, int n_input, |
| 875 | int n_aux_input, int n_output, int output_batch_leading_dim, |
| 876 | float* output_state_ptr, float* cell_state_ptr, float* scratch0, |
| 877 | float* scratch1, float* scratch2, float* scratch3, float* scratch4, |
| 878 | float* output_ptr, CpuBackendContext* context) { |
| 879 | ruy::profiler::ScopeLabel label("LstmStepFloat"); |
| 880 | // Since we have already checked that weights are all there or none, we can |
| 881 | // check the existence of only one to the get the condition. |
| 882 | const bool use_cifg = (input_to_input_weights_ptr == nullptr); |
| 883 | |
| 884 | // Make named scratch buffers. |
| 885 | float* input_gate_scratch = scratch0; |
| 886 | float* forget_gate_scratch = scratch1; |
| 887 | float* cell_gate_scratch = scratch2; |
| 888 | float* output_gate_scratch = scratch3; |
| 889 | float* accumulation_scratch_buffer = scratch4; |
| 890 | |
| 891 | // Check if inputs are all zeros so we can skip some computations. |
| 892 | const bool is_input_all_zeros = |
| 893 | tensor_utils::IsZeroVector(input_ptr, n_batch * n_input); |
| 894 | const bool is_aux_input_all_zeros = |
| 895 | (aux_input_ptr == nullptr || |
| 896 | tensor_utils::IsZeroVector(aux_input_ptr, n_batch * n_aux_input)); |
| 897 | |
| 898 | if (!use_cifg) { |
| 899 | // Calculate the input gate. (If not CIFG.) |
| 900 | CalculateLstmGateFloat(input_ptr, input_to_input_weights_ptr, aux_input_ptr, |
| 901 | aux_input_to_input_weights_ptr, output_state_ptr, |
| 902 | recurrent_to_input_weights_ptr, |
| 903 | |
| 904 | cell_state_ptr, cell_to_input_weights_ptr, |
| 905 | input_layer_norm_coefficients_ptr, |
| 906 | input_gate_bias_ptr, n_batch, n_input, n_aux_input, |
| 907 | n_output, n_cell, |
| 908 | /*activation=*/kTfLiteActSigmoid, input_gate_scratch, |
| 909 | is_input_all_zeros, is_aux_input_all_zeros, |
| 910 | accumulation_scratch_buffer, context); |
| 911 | } |
| 912 | // Calculate the forget gate. |
| 913 | CalculateLstmGateFloat( |
| 914 | input_ptr, input_to_forget_weights_ptr, aux_input_ptr, |
| 915 | aux_input_to_forget_weights_ptr, output_state_ptr, |
| 916 | recurrent_to_forget_weights_ptr, |
| 917 | |
| 918 | cell_state_ptr, cell_to_forget_weights_ptr, |
| 919 | forget_layer_norm_coefficients_ptr, forget_gate_bias_ptr, n_batch, |
| 920 | n_input, n_aux_input, n_output, n_cell, |
| 921 | /*activation=*/kTfLiteActSigmoid, forget_gate_scratch, is_input_all_zeros, |
| 922 | is_aux_input_all_zeros, accumulation_scratch_buffer, context); |
| 923 | // Calculate the cell update gate. |
| 924 | CalculateLstmGateFloat( |
| 925 | input_ptr, input_to_cell_weights_ptr, aux_input_ptr, |
| 926 | aux_input_to_cell_weights_ptr, output_state_ptr, |
| 927 | recurrent_to_cell_weights_ptr, |
| 928 | |
| 929 | /*cell_state=*/nullptr, |
| 930 | /*cell_to_gate_weights=*/nullptr, cell_layer_norm_coefficients_ptr, |
| 931 | cell_gate_bias_ptr, n_batch, n_input, n_aux_input, n_output, n_cell, |
| 932 | params->activation, cell_gate_scratch, is_input_all_zeros, |
| 933 | is_aux_input_all_zeros, accumulation_scratch_buffer, context); |
| 934 | // Update the cell state. |
| 935 | UpdateLstmCellFloat(n_batch, n_cell, cell_state_ptr, input_gate_scratch, |
| 936 | forget_gate_scratch, cell_gate_scratch, use_cifg, |
| 937 | params->cell_clip); |
| 938 | // Calculate output gate. |
| 939 | CalculateLstmGateFloat( |
| 940 | input_ptr, input_to_output_weights_ptr, aux_input_ptr, |
| 941 | aux_input_to_output_weights_ptr, output_state_ptr, |
| 942 | recurrent_to_output_weights_ptr, |
| 943 | |
| 944 | cell_state_ptr, cell_to_output_weights_ptr, |
| 945 | output_layer_norm_coefficients_ptr, output_gate_bias_ptr, n_batch, |
| 946 | n_input, n_aux_input, n_output, n_cell, |
| 947 | /*activation=*/kTfLiteActSigmoid, output_gate_scratch, is_input_all_zeros, |
| 948 | is_aux_input_all_zeros, accumulation_scratch_buffer, context); |
| 949 | // Update the output state. |
| 950 | CalculateLstmOutputFloat(n_batch, n_cell, n_output, cell_state_ptr, |
| 951 | output_gate_scratch, params->activation, |
| 952 | projection_weights_ptr, projection_bias_ptr, |
| 953 | params->proj_clip, output_state_ptr, scratch2, |
| 954 | accumulation_scratch_buffer, context); |
| 955 | // Copy output state to the output. Note that the output's rows may not be |
| 956 | // contiguous (output_batch_leading_dim != n_output). |
| 957 | for (int b = 0; b < n_batch; b++) { |
| 958 | std::copy_n(output_state_ptr + b * n_output, n_output, |
| 959 | output_ptr + b * output_batch_leading_dim); |
| 960 | } |
| 961 | } |
| 962 | // LINT.ThenChange(../tools/optimize/calibration/builtin_logging_ops/lstm.cc,\ |
| 963 | // ../experimental/kernels/fp16/lstm_eval.cc) |
| 964 | |
| 965 | // Same as above but with quantized weight matrices. In detail: |
| 966 | // Input of size 'n_batch * n_input': |
| 967 | // input_ptr |
| 968 | // Input of size 'n_batch * n_aux_input': |
| 969 | // aux_input_ptr - optional (can be nullptr) |
| 970 | // |
| 971 | // LSTM weights: |
| 972 | // Quantized input weights of size 'n_cell * n_input': |
| 973 | // input_to_input_weights - optional |
| 974 | // input_to_forget_weights |
| 975 | // input_to_cell_weights |
| 976 | // input_to_input_weights |
| 977 | // Quantized auxiliary input weights of size 'n_cell * n_aux_input': |
| 978 | // aux_input_to_input_weights - optional |
| 979 | // aux_input_to_forget_weights - optional |
| 980 | // aux_input_to_cell_weights - optional |
| 981 | // aux_input_to_output_weights - optional |
| 982 | // Quantized recurrent weights of size 'n_cell * n_output': |
| 983 | // recurrent_to_input_weights - optional |
| 984 | // recurrent_to_forget_weights |
| 985 | // recurrent_to_cell_weights |
| 986 | // recurrent_to_input_weights |
| 987 | // Quantized peephole weights of size 'n_cell', representing diagonal matrices. |
| 988 | // cell_to_input_weights - optional |
| 989 | // cell_to_cell_weights - optional |
| 990 | // cell_to_output_weights - optional |
| 991 | // Quantized projection weights of size 'n_output * n_cell' |
| 992 | // projection_weights_ptr - optional |
| 993 | // Weight scales (scalars) for each of the weights above. |
| 994 | // input_to_input_weights_scale - optional |
| 995 | // input_to_forget_weights_scale |
| 996 | // input_to_cell_weights_scale |
| 997 | // input_to_output_weights_scale |
| 998 | // aux_input_to_input_weights_scale - optional |
| 999 | // aux_input_to_forget_weights_scale - optional |
| 1000 | // aux_input_to_cell_weights_scale - optional |
| 1001 | // aux_input_to_output_weights_scale - optional |
| 1002 | // recurrent_to_input_weights_scale - optional |
| 1003 | // recurrent_to_forget_weights_scale |
| 1004 | // recurrent_to_cell_weights_scale |
| 1005 | // recurrent_to_output_weights_scale |
| 1006 | // cell_to_input_weights_scale, |
| 1007 | // cell_to_forget_weights_scale, |
| 1008 | // cell_to_output_weights_scale, |
| 1009 | // projection_weights_scale - optional |
| 1010 | // Gate biases of size 'n_cell': |
| 1011 | // input_gate_bias_ptr - optional |
| 1012 | // forget_gate_bias_ptr |
| 1013 | // cell_gate_bias_ptr |
| 1014 | // output_gate_bias_ptr |
| 1015 | // |
| 1016 | // Layer norm coefficients of size 'n_cell', representing diagonal matrices. |
| 1017 | // input_layer_norm_coefficients_ptr - optional |
| 1018 | // forget_layer_norm_coefficients_ptr - optional |
| 1019 | // cell_layer_norm_coefficients_ptr - optional |
| 1020 | // output_layer_norm_coefficients_ptr - optional |
| 1021 | // |
| 1022 | // Temporary pre-allocated storage for quantized values: |
| 1023 | // quantized_input_ptr (same size as input_ptr) |
| 1024 | // quantized_output_state_ptr (same size as output_state_ptr) |
| 1025 | // quantized_output_scratch (same size as cell_state_ptr) |
| 1026 | // Temporary pre-allocated storage for recovered values: |
| 1027 | // recovered_cell_weights (same size as cell_to_*_weights) |
| 1028 | // |
| 1029 | // Outputs: |
| 1030 | // output_state_ptr - size 'n_batch * n_output' |
| 1031 | // cell_state_ptr - size 'n_batch * n_cell' |
| 1032 | // output_ptr - size 'n_batch * output_batch_leading_dim' |
| 1033 | inline void LstmStepHybrid( |
| 1034 | const float* input_ptr, const int8_t* input_to_input_weights_ptr, |
| 1035 | const uint8_t* input_to_input_weights_ledger_ptr, |
| 1036 | float input_to_input_weights_scale, |
| 1037 | const int8_t* input_to_forget_weights_ptr, |
| 1038 | const uint8_t* input_to_forget_weights_ledger_ptr, |
| 1039 | float input_to_forget_weights_scale, |
| 1040 | const int8_t* input_to_cell_weights_ptr, |
| 1041 | const uint8_t* input_to_cell_weights_ledger_ptr, |
| 1042 | float input_to_cell_weights_scale, |
| 1043 | const int8_t* input_to_output_weights_ptr, |
| 1044 | const uint8_t* input_to_output_weights_ledger_ptr, |
| 1045 | float input_to_output_weights_scale, const float* aux_input_ptr, |
| 1046 | const int8_t* aux_input_to_input_weights_ptr, |
| 1047 | float aux_input_to_input_weights_scale, |
| 1048 | const int8_t* aux_input_to_forget_weights_ptr, |
| 1049 | float aux_input_to_forget_weights_scale, |
| 1050 | const int8_t* aux_input_to_cell_weights_ptr, |
| 1051 | float aux_input_to_cell_weights_scale, |
| 1052 | const int8_t* aux_input_to_output_weights_ptr, |
| 1053 | float aux_input_to_output_weights_scale, |
| 1054 | const int8_t* recurrent_to_input_weights_ptr, |
| 1055 | const uint8_t* recurrent_to_input_weights_ledger_ptr, |
| 1056 | float recurrent_to_input_weights_scale, |
| 1057 | const int8_t* recurrent_to_forget_weights_ptr, |
| 1058 | const uint8_t* recurrent_to_forget_weights_ledger_ptr, |
| 1059 | float recurrent_to_forget_weights_scale, |
| 1060 | const int8_t* recurrent_to_cell_weights_ptr, |
| 1061 | const uint8_t* recurrent_to_cell_weights_ledger_ptr, |
| 1062 | float recurrent_to_cell_weights_scale, |
| 1063 | const int8_t* recurrent_to_output_weights_ptr, |
| 1064 | const uint8_t* recurrent_to_output_weights_ledger_ptr, |
| 1065 | float recurrent_to_output_weights_scale, |
| 1066 | const int8_t* cell_to_input_weights_ptr, float cell_to_input_weights_scale, |
| 1067 | const int8_t* cell_to_forget_weights_ptr, |
| 1068 | float cell_to_forget_weights_scale, |
| 1069 | const int8_t* cell_to_output_weights_ptr, |
| 1070 | float cell_to_output_weights_scale, |
| 1071 | const float* input_layer_norm_coefficients_ptr, |
| 1072 | const float* forget_layer_norm_coefficients_ptr, |
| 1073 | const float* cell_layer_norm_coefficients_ptr, |
| 1074 | const float* output_layer_norm_coefficients_ptr, |
| 1075 | const float* input_gate_bias_ptr, const float* forget_gate_bias_ptr, |
| 1076 | const float* cell_gate_bias_ptr, const float* output_gate_bias_ptr, |
| 1077 | const int8_t* projection_weights_ptr, |
| 1078 | const uint8_t* projection_weights_ledger_ptr, |
| 1079 | float projection_weights_scale, const float* projection_bias_ptr, |
| 1080 | const TfLiteLSTMParams* params, int n_batch, int n_cell, int n_input, |
| 1081 | int n_aux_input, int n_output, int output_batch_leading_dim, |
| 1082 | float* scratch0, float* scratch1, float* scratch2, float* scratch3, |
| 1083 | float* input_sf, float* aux_input_sf, float* output_state_sf, |
| 1084 | float* scaling_factors_scratch, float* recovered_cell_weights, |
| 1085 | int8_t* quantized_input_ptr, int8_t* quantized_aux_input_ptr, |
| 1086 | int8_t* quantized_output_state_ptr, int8_t* quantized_output_scratch, |
| 1087 | float* output_state_ptr, float* cell_state_ptr, int32_t* accum_scratch_ptr, |
| 1088 | float* output_ptr, int32_t* input_zp, int32_t* aux_input_zp, |
| 1089 | int32_t* output_state_zp, int32_t* row_sums, int row_sums_size, |
| 1090 | bool* compute_row_sums, bool asymmetric_quantize_inputs, |
| 1091 | CpuBackendContext* context) { |
| 1092 | ruy::profiler::ScopeLabel label("LstmStepHybrid"); |
| 1093 | // Since we have already checked that weights are all there or none, we |
| 1094 | // can check the existence of only one to the get the condition. |
| 1095 | const bool use_cifg = (input_to_input_weights_ptr == nullptr); |
| 1096 | // Make named scratch buffers for the different gates. |
| 1097 | float* input_gate_scratch = scratch0; |
| 1098 | float* forget_gate_scratch = scratch1; |
| 1099 | float* cell_gate_scratch = scratch2; |
| 1100 | float* output_gate_scratch = scratch3; |
| 1101 | |
| 1102 | int32_t* input_to_input_row_sums = nullptr; |
| 1103 | int32_t* input_to_forget_row_sums = nullptr; |
| 1104 | int32_t* input_to_cell_row_sums = nullptr; |
| 1105 | int32_t* input_to_output_row_sums = nullptr; |
| 1106 | int32_t* aux_input_to_input_row_sums = nullptr; |
| 1107 | int32_t* aux_input_to_forget_row_sums = nullptr; |
| 1108 | int32_t* aux_input_to_cell_row_sums = nullptr; |
| 1109 | int32_t* aux_input_to_output_row_sums = nullptr; |
| 1110 | int32_t* recurrent_to_input_row_sums = nullptr; |
| 1111 | int32_t* recurrent_to_forget_row_sums = nullptr; |
| 1112 | int32_t* recurrent_to_cell_row_sums = nullptr; |
| 1113 | int32_t* recurrent_to_output_row_sums = nullptr; |
| 1114 | int32_t* projection_weights_row_sums = nullptr; |
| 1115 | |
| 1116 | if (asymmetric_quantize_inputs) { |
| 1117 | int num_row_sums = use_cifg ? 6 : 8; |
| 1118 | if (aux_input_ptr != nullptr) { |
| 1119 | num_row_sums += use_cifg ? 3 : 4; |
| 1120 | } |
| 1121 | if (projection_weights_ptr != nullptr) { |
| 1122 | num_row_sums += ceil(static_cast<float>(n_output) / n_cell); |
| 1123 | } |
| 1124 | TF_LITE_ASSERT(row_sums_size == num_row_sums); |
| 1125 | input_to_input_row_sums = row_sums; |
| 1126 | input_to_forget_row_sums = |
| 1127 | use_cifg ? input_to_input_row_sums : input_to_input_row_sums + n_cell; |
| 1128 | input_to_cell_row_sums = input_to_forget_row_sums + n_cell; |
| 1129 | input_to_output_row_sums = input_to_cell_row_sums + n_cell; |
| 1130 | if (aux_input_ptr != nullptr) { |
| 1131 | aux_input_to_input_row_sums = input_to_output_row_sums + n_cell; |
| 1132 | aux_input_to_forget_row_sums = use_cifg |
| 1133 | ? aux_input_to_input_row_sums |
| 1134 | : aux_input_to_input_row_sums + n_cell; |
| 1135 | aux_input_to_cell_row_sums = aux_input_to_forget_row_sums + n_cell; |
| 1136 | aux_input_to_output_row_sums = aux_input_to_cell_row_sums + n_cell; |
| 1137 | } |
| 1138 | recurrent_to_input_row_sums = aux_input_ptr |
| 1139 | ? aux_input_to_output_row_sums + n_cell |
| 1140 | : input_to_output_row_sums + n_cell; |
| 1141 | recurrent_to_forget_row_sums = use_cifg |
| 1142 | ? recurrent_to_input_row_sums |
| 1143 | : recurrent_to_input_row_sums + n_cell; |
| 1144 | recurrent_to_cell_row_sums = recurrent_to_forget_row_sums + n_cell; |
| 1145 | recurrent_to_output_row_sums = recurrent_to_cell_row_sums + n_cell; |
| 1146 | if (projection_weights_ptr != nullptr) { |
| 1147 | projection_weights_row_sums = recurrent_to_output_row_sums + n_cell; |
| 1148 | } |
| 1149 | if (*compute_row_sums) { |
| 1150 | ComputeRowSums( |
| 1151 | input_to_input_row_sums, input_to_forget_row_sums, |
| 1152 | input_to_cell_row_sums, input_to_output_row_sums, |
| 1153 | aux_input_to_input_row_sums, aux_input_to_forget_row_sums, |
| 1154 | aux_input_to_cell_row_sums, aux_input_to_output_row_sums, |
| 1155 | recurrent_to_input_row_sums, recurrent_to_forget_row_sums, |
| 1156 | recurrent_to_cell_row_sums, recurrent_to_output_row_sums, |
| 1157 | projection_weights_row_sums, row_sums, n_cell, n_input, n_aux_input, |
| 1158 | n_output, input_to_input_weights_ptr, input_to_forget_weights_ptr, |
| 1159 | input_to_cell_weights_ptr, input_to_output_weights_ptr, |
| 1160 | aux_input_to_input_weights_ptr, aux_input_to_forget_weights_ptr, |
| 1161 | aux_input_to_cell_weights_ptr, aux_input_to_output_weights_ptr, |
| 1162 | recurrent_to_input_weights_ptr, recurrent_to_forget_weights_ptr, |
| 1163 | recurrent_to_cell_weights_ptr, recurrent_to_output_weights_ptr, |
| 1164 | projection_weights_ptr, use_cifg, aux_input_ptr); |
| 1165 | *compute_row_sums = false; |
| 1166 | } |
| 1167 | } |
| 1168 | |
| 1169 | // Check if inputs are all zeros so we can skip some computations. |
| 1170 | const bool is_input_all_zeros = |
| 1171 | tensor_utils::IsZeroVector(input_ptr, n_batch * n_input); |
| 1172 | const bool is_aux_input_all_zeros = |
| 1173 | (aux_input_ptr == nullptr || |
| 1174 | tensor_utils::IsZeroVector(aux_input_ptr, n_batch * n_aux_input)); |
| 1175 | const bool is_output_state_all_zeros = |
| 1176 | tensor_utils::IsZeroVector(output_state_ptr, n_batch * n_output); |
| 1177 | // Quantize inputs. |
| 1178 | if (!is_input_all_zeros) { |
| 1179 | tensor_utils::BatchQuantizeFloats(input_ptr, n_batch, n_input, |
| 1180 | quantized_input_ptr, input_sf, input_zp, |
| 1181 | asymmetric_quantize_inputs); |
| 1182 | } |
| 1183 | if (!is_aux_input_all_zeros) { |
| 1184 | tensor_utils::BatchQuantizeFloats(aux_input_ptr, n_batch, n_aux_input, |
| 1185 | quantized_aux_input_ptr, aux_input_sf, |
| 1186 | aux_input_zp, asymmetric_quantize_inputs); |
| 1187 | } |
| 1188 | if (!is_output_state_all_zeros) { |
| 1189 | tensor_utils::BatchQuantizeFloats( |
| 1190 | output_state_ptr, n_batch, n_output, quantized_output_state_ptr, |
| 1191 | output_state_sf, output_state_zp, asymmetric_quantize_inputs); |
| 1192 | } |
| 1193 | if (!use_cifg) { |
| 1194 | // Calculate the input gate. (If not CIFG.) |
| 1195 | CalculateLstmGateHybrid( |
| 1196 | quantized_input_ptr, input_sf, input_zp, input_to_input_weights_ptr, |
| 1197 | input_to_input_weights_ledger_ptr, input_to_input_weights_scale, |
| 1198 | input_to_input_row_sums, quantized_aux_input_ptr, aux_input_sf, |
| 1199 | aux_input_zp, aux_input_to_input_weights_ptr, |
| 1200 | aux_input_to_input_weights_scale, aux_input_to_input_row_sums, |
| 1201 | quantized_output_state_ptr, output_state_sf, output_state_zp, |
| 1202 | recurrent_to_input_weights_ptr, recurrent_to_input_weights_ledger_ptr, |
| 1203 | recurrent_to_input_weights_scale, recurrent_to_input_row_sums, |
| 1204 | cell_state_ptr, cell_to_input_weights_ptr, cell_to_input_weights_scale, |
| 1205 | input_layer_norm_coefficients_ptr, input_gate_bias_ptr, n_batch, |
| 1206 | n_input, n_aux_input, n_output, n_cell, kTfLiteActSigmoid, |
| 1207 | input_gate_scratch, is_input_all_zeros, is_aux_input_all_zeros, |
| 1208 | is_output_state_all_zeros, compute_row_sums, context, |
| 1209 | scaling_factors_scratch, recovered_cell_weights, accum_scratch_ptr); |
| 1210 | } |
| 1211 | // Calculate the forget gate. |
| 1212 | CalculateLstmGateHybrid( |
| 1213 | quantized_input_ptr, input_sf, input_zp, input_to_forget_weights_ptr, |
| 1214 | input_to_forget_weights_ledger_ptr, input_to_forget_weights_scale, |
| 1215 | input_to_forget_row_sums, quantized_aux_input_ptr, aux_input_sf, |
| 1216 | aux_input_zp, aux_input_to_forget_weights_ptr, |
| 1217 | aux_input_to_forget_weights_scale, aux_input_to_forget_row_sums, |
| 1218 | quantized_output_state_ptr, output_state_sf, output_state_zp, |
| 1219 | recurrent_to_forget_weights_ptr, recurrent_to_forget_weights_ledger_ptr, |
| 1220 | recurrent_to_forget_weights_scale, recurrent_to_forget_row_sums, |
| 1221 | cell_state_ptr, cell_to_forget_weights_ptr, cell_to_forget_weights_scale, |
| 1222 | forget_layer_norm_coefficients_ptr, forget_gate_bias_ptr, n_batch, |
| 1223 | n_input, n_aux_input, n_output, n_cell, kTfLiteActSigmoid, |
| 1224 | forget_gate_scratch, is_input_all_zeros, is_aux_input_all_zeros, |
| 1225 | is_output_state_all_zeros, compute_row_sums, context, |
| 1226 | scaling_factors_scratch, recovered_cell_weights, accum_scratch_ptr); |
| 1227 | // Calculate the cell update gate. |
| 1228 | CalculateLstmGateHybrid( |
| 1229 | quantized_input_ptr, input_sf, input_zp, input_to_cell_weights_ptr, |
| 1230 | input_to_cell_weights_ledger_ptr, input_to_cell_weights_scale, |
| 1231 | input_to_cell_row_sums, quantized_aux_input_ptr, aux_input_sf, |
| 1232 | aux_input_zp, aux_input_to_cell_weights_ptr, |
| 1233 | aux_input_to_cell_weights_scale, aux_input_to_cell_row_sums, |
| 1234 | quantized_output_state_ptr, output_state_sf, output_state_zp, |
| 1235 | recurrent_to_cell_weights_ptr, recurrent_to_cell_weights_ledger_ptr, |
| 1236 | recurrent_to_cell_weights_scale, recurrent_to_cell_row_sums, |
| 1237 | /*cell_state=*/nullptr, /*cell_to_gate_weights=*/nullptr, |
| 1238 | /*cell_to_gate_weights_scale=*/0.0f, cell_layer_norm_coefficients_ptr, |
| 1239 | cell_gate_bias_ptr, n_batch, n_input, n_aux_input, n_output, n_cell, |
| 1240 | params->activation, cell_gate_scratch, is_input_all_zeros, |
| 1241 | is_aux_input_all_zeros, is_output_state_all_zeros, compute_row_sums, |
| 1242 | context, scaling_factors_scratch, recovered_cell_weights, |
| 1243 | accum_scratch_ptr); |
| 1244 | // Update the cell state. |
| 1245 | UpdateLstmCellFloat(n_batch, n_cell, cell_state_ptr, input_gate_scratch, |
| 1246 | forget_gate_scratch, cell_gate_scratch, use_cifg, |
| 1247 | params->cell_clip); |
| 1248 | // Calculate the output gate. |
| 1249 | CalculateLstmGateHybrid( |
| 1250 | quantized_input_ptr, input_sf, input_zp, input_to_output_weights_ptr, |
| 1251 | input_to_output_weights_ledger_ptr, input_to_output_weights_scale, |
| 1252 | input_to_output_row_sums, quantized_aux_input_ptr, aux_input_sf, |
| 1253 | aux_input_zp, aux_input_to_output_weights_ptr, |
| 1254 | aux_input_to_output_weights_scale, aux_input_to_output_row_sums, |
| 1255 | quantized_output_state_ptr, output_state_sf, output_state_zp, |
| 1256 | recurrent_to_output_weights_ptr, recurrent_to_output_weights_ledger_ptr, |
| 1257 | recurrent_to_output_weights_scale, recurrent_to_output_row_sums, |
| 1258 | cell_state_ptr, cell_to_output_weights_ptr, cell_to_output_weights_scale, |
| 1259 | output_layer_norm_coefficients_ptr, output_gate_bias_ptr, n_batch, |
| 1260 | n_input, n_aux_input, n_output, n_cell, kTfLiteActSigmoid, |
| 1261 | output_gate_scratch, is_input_all_zeros, is_aux_input_all_zeros, |
| 1262 | is_output_state_all_zeros, compute_row_sums, context, |
| 1263 | scaling_factors_scratch, recovered_cell_weights, accum_scratch_ptr); |
| 1264 | // Update the output state. |
| 1265 | CalculateLstmOutputHybrid( |
| 1266 | n_batch, n_cell, n_output, cell_state_ptr, output_gate_scratch, |
| 1267 | params->activation, projection_weights_ptr, projection_weights_ledger_ptr, |
| 1268 | projection_weights_scale, projection_bias_ptr, params->proj_clip, |
| 1269 | output_state_ptr, asymmetric_quantize_inputs, projection_weights_row_sums, |
| 1270 | compute_row_sums, context, scratch2, quantized_output_scratch, input_sf, |
| 1271 | input_zp, accum_scratch_ptr); |
| 1272 | // Copy output state to the output. Note that the output's rows may not be |
| 1273 | // contiguous (output_batch_leading_dim != n_output). |
| 1274 | for (int b = 0; b < n_batch; b++) { |
| 1275 | std::copy_n(output_state_ptr + b * n_output, n_output, |
| 1276 | output_ptr + b * output_batch_leading_dim); |
| 1277 | } |
| 1278 | } |
| 1279 | |
| 1280 | // Fully quantized lstm kernel for 16 bit gate matmul output. |
| 1281 | // |
| 1282 | // Input tensor of size n_batch * n_input: |
| 1283 | // input_ptr |
| 1284 | // |
| 1285 | // LSTM weights: |
| 1286 | // Quantized input weights of size 'n_cell * n_input': |
| 1287 | // input_to_input_weight_ptr - optional |
| 1288 | // input_to_forget_weight_ptr - optional |
| 1289 | // input_to_cell_weight_ptr - optional |
| 1290 | // input_to_output_weight_ptr - optional |
| 1291 | // |
| 1292 | // Quantized recurrent weights of size 'n_cell * n_output': |
| 1293 | // recurrent_to_input_weight_ptr - optional |
| 1294 | // recurrent_to_forget_weights_ptr |
| 1295 | // recurrent_to_cell_weights_ptr |
| 1296 | // recurrent_to_input_weights_ptr |
| 1297 | // |
| 1298 | // Quantized peephole weights of size 'n_cell', representing diagonal matrices. |
| 1299 | // cell_to_input_weights - optional |
| 1300 | // cell_to_cell_weights - optional |
| 1301 | // cell_to_output_weights - optional |
| 1302 | // |
| 1303 | // Quantized projection weights of size 'n_output * n_cell' |
| 1304 | // projection_weight_ptr - optional |
| 1305 | // |
| 1306 | // Weight scales (scalars) for each of the weights above. |
| 1307 | // effective_input_to_input_scale_a - optional |
| 1308 | // effective_input_to_input_scale_b - optional |
| 1309 | // effective_input_to_forget_scale_a |
| 1310 | // effective_input_to_forget_scale_b |
| 1311 | // effective_input_to_cell_scale_a |
| 1312 | // effective_input_to_cell_scale_b |
| 1313 | // effective_input_to_output_scale_a |
| 1314 | // effective_input_to_output_scale_b |
| 1315 | // effective_recurrent_to_input_scale_a - optional |
| 1316 | // effective_recurrent_to_input_scale_b - optional |
| 1317 | // effective_recurrent_to_forget_scale_a |
| 1318 | // effective_recurrent_to_forget_scale_b |
| 1319 | // effective_recurrent_to_cell_scale_a |
| 1320 | // effective_recurrent_to_cell_scale_b |
| 1321 | // effective_recurrent_to_output_scale_a |
| 1322 | // effective_recurrent_to_output_scale_b |
| 1323 | // effective_proj_scale_a - optional |
| 1324 | // effective_proj_scale_b - optional |
| 1325 | // |
| 1326 | // Gate biases of size 'n_cell': |
| 1327 | // input_gate_bias_ptr - optional |
| 1328 | // forget_gate_bias_ptr |
| 1329 | // cell_gate_bias_ptr |
| 1330 | // output_gate_bias_ptr |
| 1331 | // |
| 1332 | // Layer norm coefficients of size 'n_cell', representing diagonal matrices. |
| 1333 | // layer_norm_input_weight_ptr - optional |
| 1334 | // layer_norm_forget_weight_ptr - optional |
| 1335 | // layer_norm_cell_weight_ptr - optional |
| 1336 | // layer_norm_output_weight_ptr - optional |
| 1337 | // |
| 1338 | // Layer norm scales of size 'n_cell'. |
| 1339 | // layer_norm_input_scale_a - optional |
| 1340 | // layer_norm_input_scale_b - optional |
| 1341 | // layer_norm_forget_scale_a - optional |
| 1342 | // layer_norm_forget_scale_b - optional |
| 1343 | // layer_norm_cell_scale_a - optional |
| 1344 | // layer_norm_cell_scale_b - optional |
| 1345 | // layer_norm_output_scale_a - optional |
| 1346 | // layer_norm_output_scale_b - optional |
| 1347 | // |
| 1348 | // Scalar values: |
| 1349 | // quantized_cell_clip: quantized clip value for cell. |
| 1350 | // quantized_proj_clip: quantized clip value for projection. |
| 1351 | // cell_state_scale: the power of two scale for cell state. |
| 1352 | // |
| 1353 | // Zero points: |
| 1354 | // output_state_zp: zero point of output state |
| 1355 | // hidden_zp: zero point for hidden state. |
| 1356 | // |
| 1357 | // Temporary pre-allocated storage for the calculation. Each is of size n_cell * |
| 1358 | // n_batch. |
| 1359 | // scratch0 |
| 1360 | // scratch1 |
| 1361 | // scratch2 |
| 1362 | // scratch3 |
| 1363 | // scratch4 |
| 1364 | // scratch5: this scratch buffer is created purely for optimizing the |
| 1365 | // MatrixBatchVectorMultiplyAccumulate. |
| 1366 | // |
| 1367 | // Outputs: |
| 1368 | // output_state_ptr - size 'n_batch * n_output' |
| 1369 | // cell_state_ptr - size 'n_batch * n_cell' |
| 1370 | // output_ptr - size 'n_batch * n_output' |
| 1371 | // TODO(b/159947023): scratch0 is not used if (!cifg). Don't allocate then. |
| 1372 | inline void LstmStepInteger8x8_16( |
| 1373 | const int8_t* input_ptr, const int8_t* input_to_input_weight_ptr, |
| 1374 | int32_t effective_input_to_input_scale_a, |
| 1375 | int32_t effective_input_to_input_scale_b, |
| 1376 | const int8_t* input_to_forget_weight_ptr, |
| 1377 | int32_t effective_input_to_forget_scale_a, |
| 1378 | int32_t effective_input_to_forget_scale_b, |
| 1379 | const int8_t* input_to_cell_weight_ptr, |
| 1380 | int32_t effective_input_to_cell_scale_a, |
| 1381 | int32_t effective_input_to_cell_scale_b, |
| 1382 | const int8_t* input_to_output_weight_ptr, |
| 1383 | int32_t effective_input_to_output_scale_a, |
| 1384 | int32_t effective_input_to_output_scale_b, |
| 1385 | const int8_t* recurrent_to_input_weight_ptr, |
| 1386 | int32_t effective_recurrent_to_input_scale_a, |
| 1387 | int32_t effective_recurrent_to_input_scale_b, |
| 1388 | const int8_t* recurrent_to_forget_weight_ptr, |
| 1389 | int32_t effective_recurrent_to_forget_scale_a, |
| 1390 | int32_t effective_recurrent_to_forget_scale_b, |
| 1391 | const int8_t* recurrent_to_cell_weight_ptr, |
| 1392 | int32_t effective_recurrent_to_cell_scale_a, |
| 1393 | int32_t effective_recurrent_to_cell_scale_b, |
| 1394 | const int8_t* recurrent_to_output_weight_ptr, |
| 1395 | int32_t effective_recurrent_to_output_scale_a, |
| 1396 | int32_t effective_recurrent_to_output_scale_b, |
| 1397 | const int16_t* cell_to_input_weight_ptr, |
| 1398 | int32_t effective_cell_to_input_scale_a, |
| 1399 | int32_t effective_cell_to_input_scale_b, |
| 1400 | const int16_t* cell_to_forget_weight_ptr, |
| 1401 | int32_t effective_cell_to_forget_scale_a, |
| 1402 | int32_t effective_cell_to_forget_scale_b, |
| 1403 | const int16_t* cell_to_output_weight_ptr, |
| 1404 | int32_t effective_cell_to_output_scale_a, |
| 1405 | int32_t effective_cell_to_output_scale_b, |
| 1406 | const int8_t* projection_weight_ptr, int32_t effective_proj_scale_a, |
| 1407 | int32_t effective_proj_scale_b, int32_t hidden_zp, |
| 1408 | int32_t effective_hidden_scale_a, int32_t effective_hidden_scale_b, |
| 1409 | const int16_t* layer_norm_input_weight_ptr, |
| 1410 | int32_t layer_norm_input_scale_a, int32_t layer_norm_input_scale_b, |
| 1411 | const int16_t* layer_norm_forget_weight_ptr, |
| 1412 | int32_t layer_norm_forget_scale_a, int32_t layer_norm_forget_scale_b, |
| 1413 | const int16_t* layer_norm_cell_weight_ptr, int32_t layer_norm_cell_scale_a, |
| 1414 | int32_t layer_norm_cell_scale_b, |
| 1415 | const int16_t* layer_norm_output_weight_ptr, |
| 1416 | int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b, |
| 1417 | const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr, |
| 1418 | const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr, |
| 1419 | int16_t quantized_cell_clip, int8_t quantized_proj_clip, |
| 1420 | int32_t cell_state_scale, int32_t input_variance_guard, |
| 1421 | int32_t forget_variance_guard, int32_t cell_variance_guard, |
| 1422 | int32_t output_variance_guard, |
| 1423 | const int32_t* input_to_forget_effective_bias, |
| 1424 | const int32_t* recurrent_to_forget_effective_bias, |
| 1425 | const int32_t* input_to_cell_effective_bias, |
| 1426 | const int32_t* recurrent_to_cell_effective_bias, |
| 1427 | const int32_t* input_to_output_effective_bias, |
| 1428 | const int32_t* recurrent_to_output_effective_bias, |
| 1429 | const int32_t* input_to_input_effective_bias, |
| 1430 | const int32_t* recurrent_to_input_effective_bias, |
| 1431 | const int32_t* projection_effective_bias, int n_batch, int n_cell, |
| 1432 | int n_input, int n_output, int8_t* output_state_ptr, |
| 1433 | int32_t output_state_zp, int16_t* cell_state_ptr, int8_t* output_ptr, |
| 1434 | int16_t* scratch0, int16_t* scratch1, int16_t* scratch2, int16_t* scratch3, |
| 1435 | int8_t* scratch4, int32_t* scratch5, CpuBackendContext* context) { |
| 1436 | ruy::profiler::ScopeLabel label("LstmStepInteger8x8_16"); |
| 1437 | // Make named scratch buffers for the different gates. |
| 1438 | int16_t* input_gate_scratch = scratch0; |
| 1439 | int16_t* forget_gate_scratch = scratch1; |
| 1440 | int16_t* cell_gate_scratch = scratch2; |
| 1441 | int16_t* output_gate_scratch = scratch3; |
| 1442 | |
| 1443 | // Since we have already checked that weights are all there or none, we |
| 1444 | // can check the existence of only one to the get the condition. |
| 1445 | const bool use_cifg = (input_to_input_weight_ptr == nullptr); |
| 1446 | |
| 1447 | // Check for nullptrs. |
| 1448 | TFLITE_DCHECK(input_to_forget_effective_bias); |
| 1449 | TFLITE_DCHECK(recurrent_to_forget_effective_bias); |
| 1450 | TFLITE_DCHECK(input_to_cell_effective_bias); |
| 1451 | TFLITE_DCHECK(recurrent_to_cell_effective_bias); |
| 1452 | TFLITE_DCHECK(input_to_output_effective_bias); |
| 1453 | TFLITE_DCHECK(recurrent_to_output_effective_bias); |
| 1454 | if (!use_cifg) { |
| 1455 | TFLITE_DCHECK(input_to_input_effective_bias); |
| 1456 | TFLITE_DCHECK(recurrent_to_input_effective_bias); |
| 1457 | } |
| 1458 | const bool use_projection = (projection_weight_ptr != nullptr); |
| 1459 | if (use_projection) { |
| 1460 | TFLITE_DCHECK(projection_effective_bias); |
| 1461 | } |
| 1462 | if (!use_cifg) { |
| 1463 | // Calculate the input gate. (If not CIFG.) |
| 1464 | CalculateLstmGateInteger8x8_16( |
| 1465 | input_ptr, input_to_input_weight_ptr, input_to_input_effective_bias, |
| 1466 | effective_input_to_input_scale_a, effective_input_to_input_scale_b, |
| 1467 | output_state_ptr, recurrent_to_input_weight_ptr, |
| 1468 | recurrent_to_input_effective_bias, effective_recurrent_to_input_scale_a, |
| 1469 | effective_recurrent_to_input_scale_b, cell_state_ptr, |
| 1470 | cell_to_input_weight_ptr, effective_cell_to_input_scale_a, |
| 1471 | effective_cell_to_input_scale_b, layer_norm_input_weight_ptr, |
| 1472 | input_gate_bias_ptr, layer_norm_input_scale_a, layer_norm_input_scale_b, |
| 1473 | input_variance_guard, n_batch, n_input, n_output, n_cell, |
| 1474 | kTfLiteActSigmoid, input_gate_scratch, context, scratch5); |
| 1475 | } |
| 1476 | // Calculate the forget gate. |
| 1477 | CalculateLstmGateInteger8x8_16( |
| 1478 | input_ptr, input_to_forget_weight_ptr, input_to_forget_effective_bias, |
| 1479 | effective_input_to_forget_scale_a, effective_input_to_forget_scale_b, |
| 1480 | output_state_ptr, recurrent_to_forget_weight_ptr, |
| 1481 | recurrent_to_forget_effective_bias, effective_recurrent_to_forget_scale_a, |
| 1482 | effective_recurrent_to_forget_scale_b, cell_state_ptr, |
| 1483 | cell_to_forget_weight_ptr, effective_cell_to_forget_scale_a, |
| 1484 | effective_cell_to_forget_scale_b, layer_norm_forget_weight_ptr, |
| 1485 | forget_gate_bias_ptr, layer_norm_forget_scale_a, |
| 1486 | layer_norm_forget_scale_b, forget_variance_guard, n_batch, n_input, |
| 1487 | n_output, n_cell, kTfLiteActSigmoid, forget_gate_scratch, context, |
| 1488 | scratch5); |
| 1489 | // Calculate the cell update gate. |
| 1490 | CalculateLstmGateInteger8x8_16( |
| 1491 | input_ptr, input_to_cell_weight_ptr, input_to_cell_effective_bias, |
| 1492 | effective_input_to_cell_scale_a, effective_input_to_cell_scale_b, |
| 1493 | output_state_ptr, recurrent_to_cell_weight_ptr, |
| 1494 | recurrent_to_cell_effective_bias, effective_recurrent_to_cell_scale_a, |
| 1495 | effective_recurrent_to_cell_scale_b, cell_state_ptr, |
| 1496 | /*cell_to_gate_weights=*/nullptr, /*cell_to_gate_scale_a=*/0, |
| 1497 | /*cell_to_gate_scale_b=*/0, layer_norm_cell_weight_ptr, |
| 1498 | cell_gate_bias_ptr, layer_norm_cell_scale_a, layer_norm_cell_scale_b, |
| 1499 | cell_variance_guard, n_batch, n_input, n_output, n_cell, kTfLiteActTanh, |
| 1500 | cell_gate_scratch, context, scratch5); |
| 1501 | // Update the cell state. |
| 1502 | UpdateLstmCellInteger(n_batch, n_cell, cell_state_ptr, cell_state_scale, |
| 1503 | input_gate_scratch, forget_gate_scratch, |
| 1504 | cell_gate_scratch, use_cifg, quantized_cell_clip); |
| 1505 | // Calculate the output gate. |
| 1506 | CalculateLstmGateInteger8x8_16( |
| 1507 | input_ptr, input_to_output_weight_ptr, input_to_output_effective_bias, |
| 1508 | effective_input_to_output_scale_a, effective_input_to_output_scale_b, |
| 1509 | output_state_ptr, recurrent_to_output_weight_ptr, |
| 1510 | recurrent_to_output_effective_bias, effective_recurrent_to_output_scale_a, |
| 1511 | effective_recurrent_to_output_scale_b, cell_state_ptr, |
| 1512 | cell_to_output_weight_ptr, effective_cell_to_output_scale_a, |
| 1513 | effective_cell_to_output_scale_b, layer_norm_output_weight_ptr, |
| 1514 | output_gate_bias_ptr, layer_norm_output_scale_a, |
| 1515 | layer_norm_output_scale_b, output_variance_guard, n_batch, n_input, |
| 1516 | n_output, n_cell, kTfLiteActSigmoid, output_gate_scratch, context, |
| 1517 | scratch5); |
| 1518 | // Update the output state. |
| 1519 | CalculateLstmOutputInteger8x8_16( |
| 1520 | n_batch, n_cell, n_output, cell_state_ptr, cell_state_scale, |
| 1521 | output_gate_scratch, effective_hidden_scale_a, effective_hidden_scale_b, |
| 1522 | hidden_zp, projection_weight_ptr, effective_proj_scale_a, |
| 1523 | effective_proj_scale_b, projection_effective_bias, output_state_zp, |
| 1524 | quantized_proj_clip, output_state_ptr, context, scratch0, scratch4, |
| 1525 | scratch5); |
| 1526 | // Copy output state to the output. Note that unlike float or hybrid, output |
| 1527 | // is always contiguous. |
| 1528 | std::copy_n(output_state_ptr, n_batch * n_output, output_ptr); |
| 1529 | } |
| 1530 | |
| 1531 | // Fully quantized lstm kernel for 8 bit gate matmul output. |
| 1532 | // |
| 1533 | // Input tensor of size n_batch * n_input: |
| 1534 | // input_ptr |
| 1535 | // |
| 1536 | // LSTM weights: |
| 1537 | // Quantized input weights of size 'n_cell * n_input': |
| 1538 | // input_to_input_weight_ptr - optional |
| 1539 | // input_to_forget_weight_ptr - optional |
| 1540 | // input_to_cell_weight_ptr - optional |
| 1541 | // input_to_output_weight_ptr - optional |
| 1542 | // |
| 1543 | // Quantized recurrent weights of size 'n_cell * n_output': |
| 1544 | // recurrent_to_input_weight_ptr - optional |
| 1545 | // recurrent_to_forget_weights_ptr |
| 1546 | // recurrent_to_cell_weights_ptr |
| 1547 | // recurrent_to_input_weights_ptr |
| 1548 | // |
| 1549 | // Quantized peephole weights of size 'n_cell', representing diagonal matrices. |
| 1550 | // cell_to_input_weights - optional |
| 1551 | // cell_to_cell_weights - optional |
| 1552 | // cell_to_output_weights - optional |
| 1553 | // |
| 1554 | // Quantized projection weights of size 'n_output * n_cell' |
| 1555 | // projection_weight_ptr - optional |
| 1556 | // |
| 1557 | // Weight scales (scalars) for each of the weights above. |
| 1558 | // effective_input_to_input_scale_a - optional |
| 1559 | // effective_input_to_input_scale_b - optional |
| 1560 | // effective_input_to_forget_scale_a |
| 1561 | // effective_input_to_forget_scale_b |
| 1562 | // effective_input_to_cell_scale_a |
| 1563 | // effective_input_to_cell_scale_b |
| 1564 | // effective_input_to_output_scale_a |
| 1565 | // effective_input_to_output_scale_b |
| 1566 | // effective_recurrent_to_input_scale_a - optional |
| 1567 | // effective_recurrent_to_input_scale_b - optional |
| 1568 | // effective_recurrent_to_forget_scale_a |
| 1569 | // effective_recurrent_to_forget_scale_b |
| 1570 | // effective_recurrent_to_cell_scale_a |
| 1571 | // effective_recurrent_to_cell_scale_b |
| 1572 | // effective_recurrent_to_output_scale_a |
| 1573 | // effective_recurrent_to_output_scale_b |
| 1574 | // effective_proj_scale_a - optional |
| 1575 | // effective_proj_scale_b - optional |
| 1576 | // |
| 1577 | // Gate biases of size 'n_cell': |
| 1578 | // input_gate_bias_ptr - optional |
| 1579 | // forget_gate_bias_ptr |
| 1580 | // cell_gate_bias_ptr |
| 1581 | // output_gate_bias_ptr |
| 1582 | // |
| 1583 | // Layer norm coefficients of size 'n_cell', representing diagonal matrices. |
| 1584 | // layer_norm_input_weight_ptr - optional |
| 1585 | // layer_norm_forget_weight_ptr - optional |
| 1586 | // layer_norm_cell_weight_ptr - optional |
| 1587 | // layer_norm_output_weight_ptr - optional |
| 1588 | // |
| 1589 | // Layer norm scales of size 'n_cell'. |
| 1590 | // layer_norm_input_scale_a - optional |
| 1591 | // layer_norm_input_scale_b - optional |
| 1592 | // layer_norm_forget_scale_a - optional |
| 1593 | // layer_norm_forget_scale_b - optional |
| 1594 | // layer_norm_cell_scale_a - optional |
| 1595 | // layer_norm_cell_scale_b - optional |
| 1596 | // layer_norm_output_scale_a - optional |
| 1597 | // layer_norm_output_scale_b - optional |
| 1598 | // |
| 1599 | // Scalar values: |
| 1600 | // quantized_cell_clip: quantized clip value for cell. |
| 1601 | // quantized_proj_clip: quantized clip value for projection. |
| 1602 | // cell_state_scale: the power of two scale for cell state. |
| 1603 | // |
| 1604 | // Zero points: |
| 1605 | // input_zp: zero point for input tensor. |
| 1606 | // output_state_zp: zero point of output state. |
| 1607 | // hidden_zp: zero point for hidden state. |
| 1608 | // |
| 1609 | // Temporary pre-allocated storage for the calculation. Each is of size n_cell * |
| 1610 | // n_batch. |
| 1611 | // scratch0 |
| 1612 | // scratch1 |
| 1613 | // scratch2 |
| 1614 | // scratch3 |
| 1615 | // scratch4 |
| 1616 | // scratch5 |
| 1617 | // scratch6 |
| 1618 | // scratch7 |
| 1619 | // |
| 1620 | // Outputs: |
| 1621 | // output_state_ptr - size 'n_batch * n_output' |
| 1622 | // cell_state_ptr - size 'n_batch * n_cell' |
| 1623 | // output_ptr - size 'n_batch * n_output' |
| 1624 | // |
| 1625 | // Can move zero point calculation into Prepare() for better perfomance. |
| 1626 | // TODO(b/159947023): scratch5 is unused, remove. |
| 1627 | inline void LstmStepInteger8x8_8( |
| 1628 | const int8_t* input_ptr, int32_t input_zp, |
| 1629 | const int8_t* input_to_input_weight_ptr, |
| 1630 | int32_t effective_input_to_input_scale_a, |
| 1631 | int32_t effective_input_to_input_scale_b, |
| 1632 | const int8_t* input_to_forget_weight_ptr, |
| 1633 | int32_t effective_input_to_forget_scale_a, |
| 1634 | int32_t effective_input_to_forget_scale_b, |
| 1635 | const int8_t* input_to_cell_weight_ptr, |
| 1636 | int32_t effective_input_to_cell_scale_a, |
| 1637 | int32_t effective_input_to_cell_scale_b, |
| 1638 | const int8_t* input_to_output_weight_ptr, |
| 1639 | int32_t effective_input_to_output_scale_a, |
| 1640 | int32_t effective_input_to_output_scale_b, |
| 1641 | const int8_t* recurrent_to_input_weight_ptr, |
| 1642 | int32_t effective_recurrent_to_input_scale_a, |
| 1643 | int32_t effective_recurrent_to_input_scale_b, |
| 1644 | const int8_t* recurrent_to_forget_weight_ptr, |
| 1645 | int32_t effective_recurrent_to_forget_scale_a, |
| 1646 | int32_t effective_recurrent_to_forget_scale_b, |
| 1647 | const int8_t* recurrent_to_cell_weight_ptr, |
| 1648 | int32_t effective_recurrent_to_cell_scale_a, |
| 1649 | int32_t effective_recurrent_to_cell_scale_b, |
| 1650 | const int8_t* recurrent_to_output_weight_ptr, |
| 1651 | int32_t effective_recurrent_to_output_scale_a, |
| 1652 | int32_t effective_recurrent_to_output_scale_b, |
| 1653 | const int8_t* cell_to_input_weight_ptr, |
| 1654 | int32_t effective_cell_to_input_scale_a, |
| 1655 | int32_t effective_cell_to_input_scale_b, |
| 1656 | const int8_t* cell_to_forget_weight_ptr, |
| 1657 | int32_t effective_cell_to_forget_scale_a, |
| 1658 | int32_t effective_cell_to_forget_scale_b, |
| 1659 | const int8_t* cell_to_output_weight_ptr, |
| 1660 | int32_t effective_cell_to_output_scale_a, |
| 1661 | int32_t effective_cell_to_output_scale_b, |
| 1662 | const int8_t* projection_weight_ptr, int32_t effective_proj_scale_a, |
| 1663 | int32_t effective_proj_scale_b, const int16_t* layer_norm_input_weight_ptr, |
| 1664 | int32_t layer_norm_input_scale_a, int32_t layer_norm_input_scale_b, |
| 1665 | const int16_t* layer_norm_forget_weight_ptr, |
| 1666 | int32_t layer_norm_forget_scale_a, int32_t layer_norm_forget_scale_b, |
| 1667 | const int16_t* layer_norm_cell_weight_ptr, int32_t layer_norm_cell_scale_a, |
| 1668 | int32_t layer_norm_cell_scale_b, |
| 1669 | const int16_t* layer_norm_output_weight_ptr, |
| 1670 | int32_t layer_norm_output_scale_a, int32_t layer_norm_output_scale_b, |
| 1671 | const int32_t* input_gate_bias_ptr, const int32_t* forget_gate_bias_ptr, |
| 1672 | const int32_t* cell_gate_bias_ptr, const int32_t* output_gate_bias_ptr, |
| 1673 | const int32_t* projection_bias_ptr, const TfLiteLSTMParams* params, |
| 1674 | const int32_t* intermediate_scale_a, const int32_t* intermediate_scale_b, |
| 1675 | const int32_t* intermediate_zp, int16_t quantized_cell_clip, |
| 1676 | int8_t quantized_proj_clip, int n_batch, int n_cell, int n_input, |
| 1677 | int n_output, int output_batch_leading_dim, int8_t* output_state_ptr, |
| 1678 | int32_t output_state_zp, int16_t* cell_state_ptr, int8_t* output_ptr, |
| 1679 | int8_t* scratch0, int8_t* scratch1, int16_t* scratch2, int16_t* scratch3, |
| 1680 | int16_t* scratch4, int16_t* scratch5, int16_t* scratch6, |
| 1681 | int16_t* scratch7) { |
| 1682 | // TODO(b/159066113): scratch5 is unused, remove. |
| 1683 | |
| 1684 | ruy::profiler::ScopeLabel label("LstmStepInteger8x8_8"); |
| 1685 | // Make named scratch buffers for the different gates. |
| 1686 | int16_t* forget_gate_scratch = scratch2; |
| 1687 | int16_t* cell_gate_scratch = scratch3; |
| 1688 | int16_t* output_gate_scratch = scratch4; |
| 1689 | // no-CIFG is not supported here |
| 1690 | |
| 1691 | // Calculate the forget gate. |
| 1692 | CalculateLstmGateInteger8x8_8( |
| 1693 | input_ptr, input_zp, input_to_forget_weight_ptr, |
| 1694 | effective_input_to_forget_scale_a, effective_input_to_forget_scale_b, |
| 1695 | intermediate_scale_a[2], intermediate_scale_b[2], intermediate_zp[4], |
| 1696 | output_state_ptr, output_state_zp, recurrent_to_forget_weight_ptr, |
| 1697 | effective_recurrent_to_forget_scale_a, |
| 1698 | effective_recurrent_to_forget_scale_b, intermediate_scale_a[3], |
| 1699 | intermediate_scale_b[3], intermediate_zp[5], layer_norm_forget_weight_ptr, |
| 1700 | layer_norm_forget_scale_a, layer_norm_forget_scale_b, |
| 1701 | forget_gate_bias_ptr, n_batch, n_input, n_output, n_cell, |
| 1702 | kTfLiteActSigmoid, forget_gate_scratch, scratch0, scratch1); |
| 1703 | // Calculate the cell update gate. |
| 1704 | CalculateLstmGateInteger8x8_8( |
| 1705 | input_ptr, input_zp, input_to_cell_weight_ptr, |
| 1706 | effective_input_to_cell_scale_a, effective_input_to_cell_scale_b, |
| 1707 | intermediate_scale_a[4], intermediate_scale_b[4], intermediate_zp[7], |
| 1708 | output_state_ptr, output_state_zp, recurrent_to_cell_weight_ptr, |
| 1709 | effective_recurrent_to_cell_scale_a, effective_recurrent_to_cell_scale_b, |
| 1710 | intermediate_scale_a[5], intermediate_scale_b[5], intermediate_zp[8], |
| 1711 | layer_norm_cell_weight_ptr, layer_norm_cell_scale_a, |
| 1712 | layer_norm_cell_scale_b, cell_gate_bias_ptr, n_batch, n_input, n_output, |
| 1713 | n_cell, kTfLiteActTanh, cell_gate_scratch, scratch0, scratch1); |
| 1714 | // Update the cell state. |
| 1715 | UpdateLstmCellInteger(n_batch, n_cell, cell_state_ptr, |
| 1716 | /*cell_state_scale=*/-15, /*input_gate=*/nullptr, |
| 1717 | forget_gate_scratch, cell_gate_scratch, |
| 1718 | /*use_cifg=*/true, quantized_cell_clip); |
| 1719 | // Calculate the output gate. |
| 1720 | CalculateLstmGateInteger8x8_8( |
| 1721 | input_ptr, input_zp, input_to_output_weight_ptr, |
| 1722 | effective_input_to_output_scale_a, effective_input_to_output_scale_b, |
| 1723 | intermediate_scale_a[6], intermediate_scale_b[6], intermediate_zp[10], |
| 1724 | output_state_ptr, output_state_zp, recurrent_to_output_weight_ptr, |
| 1725 | effective_recurrent_to_output_scale_a, |
| 1726 | effective_recurrent_to_output_scale_b, intermediate_scale_a[11], |
| 1727 | intermediate_scale_b[7], intermediate_zp[7], layer_norm_output_weight_ptr, |
| 1728 | layer_norm_output_scale_a, layer_norm_output_scale_b, |
| 1729 | output_gate_bias_ptr, n_batch, n_input, n_output, n_cell, |
| 1730 | kTfLiteActSigmoid, output_gate_scratch, scratch0, scratch1); |
| 1731 | // Update the output state. |
| 1732 | CalculateLstmOutputInteger8x8_8( |
| 1733 | n_batch, n_cell, n_output, cell_state_ptr, output_gate_scratch, |
| 1734 | projection_weight_ptr, effective_proj_scale_a, effective_proj_scale_b, |
| 1735 | projection_bias_ptr, output_state_zp, quantized_proj_clip, |
| 1736 | output_state_ptr, scratch2); |
| 1737 | // Copy output state to the output. Note that unlike float or hybrid, output |
| 1738 | // is always contigous. |
| 1739 | std::copy_n(output_state_ptr, n_batch * n_output, output_ptr); |
| 1740 | } |
| 1741 | |
| 1742 | } // namespace |
| 1743 | |
| 1744 | // LINT.IfChange |
| 1745 | TfLiteStatus EvalFloat( |
| 1746 | const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, |
| 1747 | const TfLiteTensor* input_to_forget_weights, |
| 1748 | const TfLiteTensor* input_to_cell_weights, |
| 1749 | const TfLiteTensor* input_to_output_weights, |
| 1750 | const TfLiteTensor* recurrent_to_input_weights, |
| 1751 | const TfLiteTensor* recurrent_to_forget_weights, |
| 1752 | const TfLiteTensor* recurrent_to_cell_weights, |
| 1753 | const TfLiteTensor* recurrent_to_output_weights, |
| 1754 | const TfLiteTensor* cell_to_input_weights, |
| 1755 | const TfLiteTensor* cell_to_forget_weights, |
| 1756 | const TfLiteTensor* cell_to_output_weights, |
| 1757 | const TfLiteTensor* input_layer_norm_coefficients, |
| 1758 | const TfLiteTensor* forget_layer_norm_coefficients, |
| 1759 | const TfLiteTensor* cell_layer_norm_coefficients, |
| 1760 | const TfLiteTensor* output_layer_norm_coefficients, |
| 1761 | const TfLiteTensor* aux_input, |
| 1762 | const TfLiteTensor* aux_input_to_input_weights, |
| 1763 | const TfLiteTensor* aux_input_to_forget_weights, |
| 1764 | const TfLiteTensor* aux_input_to_cell_weights, |
| 1765 | const TfLiteTensor* aux_input_to_output_weights, |
| 1766 | const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, |
| 1767 | const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias, |
| 1768 | const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias, |
| 1769 | const TfLiteLSTMParams* params, bool forward_sequence, bool time_major, |
| 1770 | int output_offset, TfLiteTensor* scratch_buffer, TfLiteTensor* output_state, |
| 1771 | TfLiteTensor* cell_state, TfLiteTensor* output, |
| 1772 | CpuBackendContext* context) { |
| 1773 | TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3); |
| 1774 | int max_time, n_batch; |
| 1775 | if (input->dims->size == 3) { |
| 1776 | max_time = (time_major) ? input->dims->data[0] : input->dims->data[1]; |
| 1777 | n_batch = (time_major) ? input->dims->data[1] : input->dims->data[0]; |
| 1778 | } else { |
| 1779 | max_time = 1; |
| 1780 | n_batch = input->dims->data[0]; |
| 1781 | } |
| 1782 | const int n_input = input->dims->data[input->dims->size - 1]; |
| 1783 | const int aux_input_size = |
| 1784 | (aux_input) ? aux_input->dims->data[aux_input->dims->size - 1] : 0; |
| 1785 | |
| 1786 | // n_cell and n_output will be the same size when there is no projection. |
| 1787 | const int n_cell = input_to_output_weights->dims->data[0]; |
| 1788 | const int n_output = recurrent_to_output_weights->dims->data[1]; |
| 1789 | |
| 1790 | // Since we have already checked that weights are all there or none, we can |
| 1791 | // check the existence of only one to the get the condition. |
| 1792 | const bool use_cifg = (input_to_input_weights == nullptr); |
| 1793 | |
| 1794 | // Index the scratch buffers pointers to the global scratch buffer. |
| 1795 | float* scratch_buffer_ptr = GetTensorData<float>(scratch_buffer); |
| 1796 | float* input_gate_scratch = nullptr; |
| 1797 | float* cell_gate_scratch = nullptr; |
| 1798 | float* forget_gate_scratch = nullptr; |
| 1799 | float* output_gate_scratch = nullptr; |
| 1800 | float* accumulation_scratch_buffer = nullptr; |
| 1801 | if (use_cifg) { |
| 1802 | cell_gate_scratch = scratch_buffer_ptr; |
| 1803 | forget_gate_scratch = scratch_buffer_ptr + n_cell * n_batch; |
| 1804 | output_gate_scratch = scratch_buffer_ptr + 2 * n_cell * n_batch; |
| 1805 | accumulation_scratch_buffer = scratch_buffer_ptr + 3 * n_cell * n_batch; |
| 1806 | } else { |
| 1807 | input_gate_scratch = scratch_buffer_ptr; |
| 1808 | cell_gate_scratch = scratch_buffer_ptr + n_cell * n_batch; |
| 1809 | forget_gate_scratch = scratch_buffer_ptr + 2 * n_cell * n_batch; |
| 1810 | output_gate_scratch = scratch_buffer_ptr + 3 * n_cell * n_batch; |
| 1811 | accumulation_scratch_buffer = scratch_buffer_ptr + 4 * n_cell * n_batch; |
| 1812 | } |
| 1813 | |
| 1814 | const int output_batch_leading_dim = |
| 1815 | output->dims->data[output->dims->size - 1]; |
| 1816 | if (time_major) { |
| 1817 | // Loop through the sequence. |
| 1818 | const int input_step = n_batch * n_input; |
| 1819 | const int output_step = n_batch * output_batch_leading_dim; |
| 1820 | for (int t = 0; t < max_time; t++) { |
| 1821 | // If this is the forward_sequence, step forward, otherwise step |
| 1822 | // backwards. |
| 1823 | const int t_rel = forward_sequence ? t : max_time - t - 1; |
| 1824 | const float* input_ptr = GetTensorData<float>(input) + t_rel * input_step; |
| 1825 | const float* aux_input_ptr = nullptr; |
| 1826 | if (aux_input) { |
| 1827 | aux_input_ptr = GetTensorData<float>(aux_input) + t_rel * input_step; |
| 1828 | } |
| 1829 | float* output_ptr = |
| 1830 | GetTensorData<float>(output) + t_rel * output_step + output_offset; |
| 1831 | |
| 1832 | LstmStepFloat( |
| 1833 | input_ptr, GetTensorData<float>(input_to_input_weights), |
| 1834 | GetTensorData<float>(input_to_forget_weights), |
| 1835 | GetTensorData<float>(input_to_cell_weights), |
| 1836 | GetTensorData<float>(input_to_output_weights), aux_input_ptr, |
| 1837 | GetTensorData<float>(aux_input_to_input_weights), |
| 1838 | GetTensorData<float>(aux_input_to_forget_weights), |
| 1839 | GetTensorData<float>(aux_input_to_cell_weights), |
| 1840 | GetTensorData<float>(aux_input_to_output_weights), |
| 1841 | GetTensorData<float>(recurrent_to_input_weights), |
| 1842 | GetTensorData<float>(recurrent_to_forget_weights), |
| 1843 | GetTensorData<float>(recurrent_to_cell_weights), |
| 1844 | GetTensorData<float>(recurrent_to_output_weights), |
| 1845 | GetTensorData<float>(cell_to_input_weights), |
| 1846 | GetTensorData<float>(cell_to_forget_weights), |
| 1847 | GetTensorData<float>(cell_to_output_weights), |
| 1848 | GetTensorData<float>(input_layer_norm_coefficients), |
| 1849 | GetTensorData<float>(forget_layer_norm_coefficients), |
| 1850 | GetTensorData<float>(cell_layer_norm_coefficients), |
| 1851 | GetTensorData<float>(output_layer_norm_coefficients), |
| 1852 | GetTensorData<float>(input_gate_bias), |
| 1853 | GetTensorData<float>(forget_gate_bias), |
| 1854 | GetTensorData<float>(cell_gate_bias), |
| 1855 | GetTensorData<float>(output_gate_bias), |
| 1856 | GetTensorData<float>(projection_weights), |
| 1857 | GetTensorData<float>(projection_bias), params, n_batch, n_cell, |
| 1858 | n_input, aux_input_size, n_output, output_batch_leading_dim, |
| 1859 | GetTensorData<float>(output_state), GetTensorData<float>(cell_state), |
| 1860 | input_gate_scratch, forget_gate_scratch, cell_gate_scratch, |
| 1861 | output_gate_scratch, accumulation_scratch_buffer, output_ptr, |
| 1862 | context); |
| 1863 | } |
| 1864 | } else { |
| 1865 | for (int b = 0; b < n_batch; b++) { |
| 1866 | const int input_step = n_input; |
| 1867 | const int output_step = output_batch_leading_dim; |
| 1868 | for (int t = 0; t < max_time; t++) { |
| 1869 | // If this is the forward_sequence, step forward, otherwise step |
| 1870 | // backwards. |
| 1871 | const int t_rel = forward_sequence ? t : max_time - t - 1; |
| 1872 | const int time_offset = b * max_time + t_rel; |
| 1873 | const float* input_ptr = |
| 1874 | GetTensorData<float>(input) + time_offset * input_step; |
| 1875 | const float* aux_input_ptr = nullptr; |
| 1876 | if (aux_input) { |
| 1877 | aux_input_ptr = |
| 1878 | GetTensorData<float>(aux_input) + time_offset * input_step; |
| 1879 | } |
| 1880 | float* output_ptr = GetTensorData<float>(output) + |
| 1881 | time_offset * output_step + output_offset; |
| 1882 | |
| 1883 | // Offset the {output,cell}_state pointers to the right batch. |
| 1884 | float* output_state_ptr = |
| 1885 | GetTensorData<float>(output_state) + b * output_batch_leading_dim; |
| 1886 | float* cell_state_ptr = GetTensorData<float>(cell_state) + b * n_cell; |
| 1887 | // Offset the scratch pointers to the right batch. |
| 1888 | float* input_gate_scratch_ptr = |
| 1889 | input_gate_scratch ? input_gate_scratch + b * n_cell : nullptr; |
| 1890 | float* forget_gate_scratch_ptr = forget_gate_scratch + b * n_cell; |
| 1891 | float* cell_gate_scratch_ptr = cell_gate_scratch + b * n_cell; |
| 1892 | float* output_gate_scratch_ptr = output_gate_scratch + b * n_cell; |
| 1893 | |
| 1894 | LstmStepFloat( |
| 1895 | input_ptr, GetTensorData<float>(input_to_input_weights), |
| 1896 | GetTensorData<float>(input_to_forget_weights), |
| 1897 | GetTensorData<float>(input_to_cell_weights), |
| 1898 | GetTensorData<float>(input_to_output_weights), aux_input_ptr, |
| 1899 | GetTensorData<float>(aux_input_to_input_weights), |
| 1900 | GetTensorData<float>(aux_input_to_forget_weights), |
| 1901 | GetTensorData<float>(aux_input_to_cell_weights), |
| 1902 | GetTensorData<float>(aux_input_to_output_weights), |
| 1903 | GetTensorData<float>(recurrent_to_input_weights), |
| 1904 | GetTensorData<float>(recurrent_to_forget_weights), |
| 1905 | GetTensorData<float>(recurrent_to_cell_weights), |
| 1906 | GetTensorData<float>(recurrent_to_output_weights), |
| 1907 | GetTensorData<float>(cell_to_input_weights), |
| 1908 | GetTensorData<float>(cell_to_forget_weights), |
| 1909 | GetTensorData<float>(cell_to_output_weights), |
| 1910 | GetTensorData<float>(input_layer_norm_coefficients), |
| 1911 | GetTensorData<float>(forget_layer_norm_coefficients), |
| 1912 | GetTensorData<float>(cell_layer_norm_coefficients), |
| 1913 | GetTensorData<float>(output_layer_norm_coefficients), |
| 1914 | GetTensorData<float>(input_gate_bias), |
| 1915 | GetTensorData<float>(forget_gate_bias), |
| 1916 | GetTensorData<float>(cell_gate_bias), |
| 1917 | GetTensorData<float>(output_gate_bias), |
| 1918 | GetTensorData<float>(projection_weights), |
| 1919 | GetTensorData<float>(projection_bias), params, /*n_batch=*/1, |
| 1920 | n_cell, n_input, aux_input_size, n_output, output_batch_leading_dim, |
| 1921 | output_state_ptr, cell_state_ptr, input_gate_scratch_ptr, |
| 1922 | forget_gate_scratch_ptr, cell_gate_scratch_ptr, |
| 1923 | output_gate_scratch_ptr, accumulation_scratch_buffer, output_ptr, |
| 1924 | context); |
| 1925 | } |
| 1926 | } |
| 1927 | } |
| 1928 | return kTfLiteOk; |
| 1929 | } |
| 1930 | // LINT.ThenChange(//tensorflow/lite/tools/optimize/calibration/builtin_logging_ops/lstm.cc) |
| 1931 | |
| 1932 | TfLiteStatus EvalHybrid( |
| 1933 | const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, |
| 1934 | const TfLiteTensor* input_to_input_weights_ledger, |
| 1935 | const TfLiteTensor* input_to_forget_weights, |
| 1936 | const TfLiteTensor* input_to_forget_weights_ledger, |
| 1937 | const TfLiteTensor* input_to_cell_weights, |
| 1938 | const TfLiteTensor* input_to_cell_weights_ledger, |
| 1939 | const TfLiteTensor* input_to_output_weights, |
| 1940 | const TfLiteTensor* input_to_output_weights_ledger, |
| 1941 | const TfLiteTensor* recurrent_to_input_weights, |
| 1942 | const TfLiteTensor* recurrent_to_input_weights_ledger, |
| 1943 | const TfLiteTensor* recurrent_to_forget_weights, |
| 1944 | const TfLiteTensor* recurrent_to_forget_weights_ledger, |
| 1945 | const TfLiteTensor* recurrent_to_cell_weights, |
| 1946 | const TfLiteTensor* recurrent_to_cell_weights_ledger, |
| 1947 | const TfLiteTensor* recurrent_to_output_weights, |
| 1948 | const TfLiteTensor* recurrent_to_output_weights_ledger, |
| 1949 | const TfLiteTensor* cell_to_input_weights, |
| 1950 | const TfLiteTensor* cell_to_forget_weights, |
| 1951 | const TfLiteTensor* cell_to_output_weights, |
| 1952 | const TfLiteTensor* input_layer_norm_coefficients, |
| 1953 | const TfLiteTensor* forget_layer_norm_coefficients, |
| 1954 | const TfLiteTensor* cell_layer_norm_coefficients, |
| 1955 | const TfLiteTensor* output_layer_norm_coefficients, |
| 1956 | const TfLiteTensor* aux_input, |
| 1957 | const TfLiteTensor* aux_input_to_input_weights, |
| 1958 | const TfLiteTensor* aux_input_to_forget_weights, |
| 1959 | const TfLiteTensor* aux_input_to_cell_weights, |
| 1960 | const TfLiteTensor* aux_input_to_output_weights, |
| 1961 | const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, |
| 1962 | const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias, |
| 1963 | const TfLiteTensor* projection_weights, |
| 1964 | const TfLiteTensor* projection_weights_ledger, |
| 1965 | const TfLiteTensor* projection_bias, const TfLiteLSTMParams* params, |
| 1966 | bool forward_sequence, bool time_major, int output_offset, |
| 1967 | TfLiteTensor* scratch_buffer, TfLiteTensor* input_sf, |
| 1968 | TfLiteTensor* aux_input_sf, TfLiteTensor* output_state_sf, |
| 1969 | TfLiteTensor* prod_scaling_factors, TfLiteTensor* recovered_cell_weights, |
| 1970 | TfLiteTensor* input_quantized, TfLiteTensor* aux_input_quantized, |
| 1971 | TfLiteTensor* output_state_quantized, TfLiteTensor* cell_state_quantized, |
| 1972 | TfLiteTensor* output_state, TfLiteTensor* cell_state, |
| 1973 | TfLiteTensor* output_scratch_buffer, TfLiteTensor* output, |
| 1974 | TfLiteTensor* input_zp, TfLiteTensor* aux_input_zp, |
| 1975 | TfLiteTensor* output_state_zp, TfLiteTensor* row_sums, int row_sums_size, |
| 1976 | bool* compute_row_sums, CpuBackendContext* context) { |
| 1977 | TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3); |
| 1978 | const int n_input = input->dims->data[input->dims->size - 1]; |
| 1979 | int max_time, n_batch; |
| 1980 | if (input->dims->size == 2) { |
| 1981 | max_time = 1; |
| 1982 | n_batch = input->dims->data[0]; |
| 1983 | } else { |
| 1984 | max_time = (time_major) ? input->dims->data[0] : input->dims->data[1]; |
| 1985 | n_batch = (time_major) ? input->dims->data[1] : input->dims->data[0]; |
| 1986 | } |
| 1987 | const int aux_input_size = |
| 1988 | (aux_input) ? aux_input->dims->data[aux_input->dims->size - 1] : 0; |
| 1989 | // n_cell and n_output will be the same size when there is no projection. |
| 1990 | const int n_cell = input_to_output_weights->dims->data[0]; |
| 1991 | const int n_output = recurrent_to_output_weights->dims->data[1]; |
| 1992 | |
| 1993 | // Since we have already checked that weights are all there or none, we can |
| 1994 | // check the existence of only one to get the condition. |
| 1995 | const bool use_cifg = (input_to_input_weights == nullptr); |
| 1996 | |
| 1997 | float* scratch_buffer_ptr = GetTensorData<float>(scratch_buffer); |
| 1998 | float* input_gate_scratch = nullptr; |
| 1999 | float* cell_gate_scratch = nullptr; |
| 2000 | float* forget_gate_scratch = nullptr; |
| 2001 | float* output_gate_scratch = nullptr; |
| 2002 | if (use_cifg) { |
| 2003 | cell_gate_scratch = scratch_buffer_ptr; |
| 2004 | forget_gate_scratch = scratch_buffer_ptr + n_cell * n_batch; |
| 2005 | output_gate_scratch = scratch_buffer_ptr + 2 * n_cell * n_batch; |
| 2006 | } else { |
| 2007 | input_gate_scratch = scratch_buffer_ptr; |
| 2008 | cell_gate_scratch = scratch_buffer_ptr + n_cell * n_batch; |
| 2009 | forget_gate_scratch = scratch_buffer_ptr + 2 * n_cell * n_batch; |
| 2010 | output_gate_scratch = scratch_buffer_ptr + 3 * n_cell * n_batch; |
| 2011 | } |
| 2012 | |
| 2013 | const int output_batch_leading_dim = |
| 2014 | output->dims->data[output->dims->size - 1]; |
| 2015 | |
| 2016 | int32_t* input_zp_ptr = nullptr; |
| 2017 | int32_t* aux_input_zp_ptr = nullptr; |
| 2018 | int32_t* output_state_zp_ptr = nullptr; |
| 2019 | int32_t* row_sums_ptr = nullptr; |
| 2020 | if (params->asymmetric_quantize_inputs) { |
| 2021 | input_zp_ptr = GetTensorData<int32_t>(input_zp); |
| 2022 | aux_input_zp_ptr = GetTensorData<int32_t>(aux_input_zp); |
| 2023 | output_state_zp_ptr = GetTensorData<int32_t>(output_state_zp); |
| 2024 | row_sums_ptr = GetTensorData<int32_t>(row_sums); |
| 2025 | } |
| 2026 | |
| 2027 | if (time_major) { |
| 2028 | // Feed the sequence into the LSTM step-by-step. |
| 2029 | const int input_step = n_batch * n_input; |
| 2030 | const int output_step = n_batch * output_batch_leading_dim; |
| 2031 | for (int t = 0; t < max_time; t++) { |
| 2032 | // If this is the forward_sequence, step forward, otherwise step |
| 2033 | // backwards. |
| 2034 | const int t_rel = forward_sequence ? t : max_time - t - 1; |
| 2035 | const float* input_ptr = GetTensorData<float>(input) + t_rel * input_step; |
| 2036 | const float* aux_input_ptr = nullptr; |
| 2037 | if (aux_input) { |
| 2038 | aux_input_ptr = GetTensorData<float>(aux_input) + t_rel * input_step; |
| 2039 | } |
| 2040 | float* output_ptr = |
| 2041 | GetTensorData<float>(output) + t_rel * output_step + output_offset; |
| 2042 | LstmStepHybrid( |
| 2043 | input_ptr, GetTensorData<int8_t>(input_to_input_weights), |
| 2044 | GetTensorData<uint8_t>(input_to_input_weights_ledger), |
| 2045 | GetTensorScale(input_to_input_weights), |
| 2046 | GetTensorData<int8_t>(input_to_forget_weights), |
| 2047 | GetTensorData<uint8_t>(input_to_forget_weights_ledger), |
| 2048 | GetTensorScale(input_to_forget_weights), |
| 2049 | GetTensorData<int8_t>(input_to_cell_weights), |
| 2050 | GetTensorData<uint8_t>(input_to_cell_weights_ledger), |
| 2051 | GetTensorScale(input_to_cell_weights), |
| 2052 | GetTensorData<int8_t>(input_to_output_weights), |
| 2053 | GetTensorData<uint8_t>(input_to_output_weights_ledger), |
| 2054 | GetTensorScale(input_to_output_weights), aux_input_ptr, |
| 2055 | GetTensorData<int8_t>(aux_input_to_input_weights), |
| 2056 | GetTensorScale(aux_input_to_input_weights), |
| 2057 | GetTensorData<int8_t>(aux_input_to_forget_weights), |
| 2058 | GetTensorScale(aux_input_to_forget_weights), |
| 2059 | GetTensorData<int8_t>(aux_input_to_cell_weights), |
| 2060 | GetTensorScale(aux_input_to_cell_weights), |
| 2061 | GetTensorData<int8_t>(aux_input_to_output_weights), |
| 2062 | GetTensorScale(aux_input_to_output_weights), |
| 2063 | GetTensorData<int8_t>(recurrent_to_input_weights), |
| 2064 | GetTensorData<uint8_t>(recurrent_to_input_weights_ledger), |
| 2065 | GetTensorScale(recurrent_to_input_weights), |
| 2066 | GetTensorData<int8_t>(recurrent_to_forget_weights), |
| 2067 | GetTensorData<uint8_t>(recurrent_to_forget_weights_ledger), |
| 2068 | GetTensorScale(recurrent_to_forget_weights), |
| 2069 | GetTensorData<int8_t>(recurrent_to_cell_weights), |
| 2070 | GetTensorData<uint8_t>(recurrent_to_cell_weights_ledger), |
| 2071 | GetTensorScale(recurrent_to_cell_weights), |
| 2072 | GetTensorData<int8_t>(recurrent_to_output_weights), |
| 2073 | GetTensorData<uint8_t>(recurrent_to_output_weights_ledger), |
| 2074 | GetTensorScale(recurrent_to_output_weights), |
| 2075 | GetTensorData<int8_t>(cell_to_input_weights), |
| 2076 | GetTensorScale(cell_to_input_weights), |
| 2077 | GetTensorData<int8_t>(cell_to_forget_weights), |
| 2078 | GetTensorScale(cell_to_forget_weights), |
| 2079 | GetTensorData<int8_t>(cell_to_output_weights), |
| 2080 | GetTensorScale(cell_to_output_weights), |
| 2081 | GetTensorData<float>(input_layer_norm_coefficients), |
| 2082 | GetTensorData<float>(forget_layer_norm_coefficients), |
| 2083 | GetTensorData<float>(cell_layer_norm_coefficients), |
| 2084 | GetTensorData<float>(output_layer_norm_coefficients), |
| 2085 | GetTensorData<float>(input_gate_bias), |
| 2086 | GetTensorData<float>(forget_gate_bias), |
| 2087 | GetTensorData<float>(cell_gate_bias), |
| 2088 | GetTensorData<float>(output_gate_bias), |
| 2089 | GetTensorData<int8_t>(projection_weights), |
| 2090 | GetTensorData<uint8_t>(projection_weights_ledger), |
| 2091 | GetTensorScale(projection_weights), |
| 2092 | GetTensorData<float>(projection_bias), params, n_batch, n_cell, |
| 2093 | n_input, aux_input_size, n_output, output_batch_leading_dim, |
| 2094 | input_gate_scratch, forget_gate_scratch, cell_gate_scratch, |
| 2095 | output_gate_scratch, GetTensorData<float>(input_sf), |
| 2096 | GetTensorData<float>(aux_input_sf), |
| 2097 | GetTensorData<float>(output_state_sf), |
| 2098 | GetTensorData<float>(prod_scaling_factors), |
| 2099 | GetTensorData<float>(recovered_cell_weights), |
| 2100 | GetTensorData<int8_t>(input_quantized), |
| 2101 | GetTensorData<int8_t>(aux_input_quantized), |
| 2102 | GetTensorData<int8_t>(output_state_quantized), |
| 2103 | GetTensorData<int8_t>(cell_state_quantized), |
| 2104 | GetTensorData<float>(output_state), GetTensorData<float>(cell_state), |
| 2105 | GetTensorData<int32_t>(output_scratch_buffer), output_ptr, |
| 2106 | input_zp_ptr, aux_input_zp_ptr, output_state_zp_ptr, row_sums_ptr, |
| 2107 | row_sums_size, compute_row_sums, params->asymmetric_quantize_inputs, |
| 2108 | context); |
| 2109 | } |
| 2110 | } else { |
| 2111 | for (int b = 0; b < n_batch; b++) { |
| 2112 | const int input_step = n_input; |
| 2113 | const int output_step = output_batch_leading_dim; |
| 2114 | for (int t = 0; t < max_time; t++) { |
| 2115 | // If this is the forward_sequence, step forward, otherwise step |
| 2116 | // backwards. |
| 2117 | const int t_rel = forward_sequence ? t : max_time - t - 1; |
| 2118 | const int time_offset = b * max_time + t_rel; |
| 2119 | const float* input_ptr = |
| 2120 | GetTensorData<float>(input) + time_offset * input_step; |
| 2121 | const float* aux_input_ptr = nullptr; |
| 2122 | if (aux_input) { |
| 2123 | aux_input_ptr = |
| 2124 | GetTensorData<float>(aux_input) + time_offset * input_step; |
| 2125 | } |
| 2126 | float* output_ptr = GetTensorData<float>(output) + |
| 2127 | time_offset * output_step + output_offset; |
| 2128 | |
| 2129 | // Offset the {output,cell}_state pointers to the right batch. |
| 2130 | float* output_state_ptr = |
| 2131 | GetTensorData<float>(output_state) + b * output_batch_leading_dim; |
| 2132 | float* cell_state_ptr = GetTensorData<float>(cell_state) + b * n_cell; |
| 2133 | // Offset the scratch pointers to the right batch. |
| 2134 | float* input_gate_scratch_ptr = |
| 2135 | input_gate_scratch ? input_gate_scratch + b * n_cell : nullptr; |
| 2136 | float* forget_gate_scratch_ptr = forget_gate_scratch + b * n_cell; |
| 2137 | float* cell_gate_scratch_ptr = cell_gate_scratch + b * n_cell; |
| 2138 | float* output_gate_scratch_ptr = output_gate_scratch + b * n_cell; |
| 2139 | |
| 2140 | LstmStepHybrid( |
| 2141 | input_ptr, GetTensorData<int8_t>(input_to_input_weights), |
| 2142 | GetTensorData<uint8_t>(input_to_input_weights_ledger), |
| 2143 | GetTensorScale(input_to_input_weights), |
| 2144 | GetTensorData<int8_t>(input_to_forget_weights), |
| 2145 | GetTensorData<uint8_t>(input_to_forget_weights_ledger), |
| 2146 | GetTensorScale(input_to_forget_weights), |
| 2147 | GetTensorData<int8_t>(input_to_cell_weights), |
| 2148 | GetTensorData<uint8_t>(input_to_cell_weights_ledger), |
| 2149 | GetTensorScale(input_to_cell_weights), |
| 2150 | GetTensorData<int8_t>(input_to_output_weights), |
| 2151 | GetTensorData<uint8_t>(input_to_output_weights_ledger), |
| 2152 | GetTensorScale(input_to_output_weights), aux_input_ptr, |
| 2153 | GetTensorData<int8_t>(aux_input_to_input_weights), |
| 2154 | GetTensorScale(aux_input_to_input_weights), |
| 2155 | GetTensorData<int8_t>(aux_input_to_forget_weights), |
| 2156 | GetTensorScale(aux_input_to_forget_weights), |
| 2157 | GetTensorData<int8_t>(aux_input_to_cell_weights), |
| 2158 | GetTensorScale(aux_input_to_cell_weights), |
| 2159 | GetTensorData<int8_t>(aux_input_to_output_weights), |
| 2160 | GetTensorScale(aux_input_to_output_weights), |
| 2161 | GetTensorData<int8_t>(recurrent_to_input_weights), |
| 2162 | GetTensorData<uint8_t>(recurrent_to_input_weights_ledger), |
| 2163 | GetTensorScale(recurrent_to_input_weights), |
| 2164 | GetTensorData<int8_t>(recurrent_to_forget_weights), |
| 2165 | GetTensorData<uint8_t>(recurrent_to_forget_weights_ledger), |
| 2166 | GetTensorScale(recurrent_to_forget_weights), |
| 2167 | GetTensorData<int8_t>(recurrent_to_cell_weights), |
| 2168 | GetTensorData<uint8_t>(recurrent_to_cell_weights_ledger), |
| 2169 | GetTensorScale(recurrent_to_cell_weights), |
| 2170 | GetTensorData<int8_t>(recurrent_to_output_weights), |
| 2171 | GetTensorData<uint8_t>(recurrent_to_output_weights_ledger), |
| 2172 | GetTensorScale(recurrent_to_output_weights), |
| 2173 | GetTensorData<int8_t>(cell_to_input_weights), |
| 2174 | GetTensorScale(cell_to_input_weights), |
| 2175 | GetTensorData<int8_t>(cell_to_forget_weights), |
| 2176 | GetTensorScale(cell_to_forget_weights), |
| 2177 | GetTensorData<int8_t>(cell_to_output_weights), |
| 2178 | GetTensorScale(cell_to_output_weights), |
| 2179 | GetTensorData<float>(input_layer_norm_coefficients), |
| 2180 | GetTensorData<float>(forget_layer_norm_coefficients), |
| 2181 | GetTensorData<float>(cell_layer_norm_coefficients), |
| 2182 | GetTensorData<float>(output_layer_norm_coefficients), |
| 2183 | GetTensorData<float>(input_gate_bias), |
| 2184 | GetTensorData<float>(forget_gate_bias), |
| 2185 | GetTensorData<float>(cell_gate_bias), |
| 2186 | GetTensorData<float>(output_gate_bias), |
| 2187 | GetTensorData<int8_t>(projection_weights), |
| 2188 | GetTensorData<uint8_t>(projection_weights_ledger), |
| 2189 | GetTensorScale(projection_weights), |
| 2190 | GetTensorData<float>(projection_bias), params, |
| 2191 | /*n_batch=*/1, n_cell, n_input, aux_input_size, n_output, |
| 2192 | output_batch_leading_dim, input_gate_scratch_ptr, |
| 2193 | forget_gate_scratch_ptr, cell_gate_scratch_ptr, |
| 2194 | output_gate_scratch_ptr, GetTensorData<float>(input_sf), |
| 2195 | GetTensorData<float>(aux_input_sf), |
| 2196 | GetTensorData<float>(output_state_sf), |
| 2197 | GetTensorData<float>(prod_scaling_factors), |
| 2198 | GetTensorData<float>(recovered_cell_weights), |
| 2199 | GetTensorData<int8_t>(input_quantized), |
| 2200 | GetTensorData<int8_t>(aux_input_quantized), |
| 2201 | GetTensorData<int8_t>(output_state_quantized), |
| 2202 | GetTensorData<int8_t>(cell_state_quantized), output_state_ptr, |
| 2203 | cell_state_ptr, GetTensorData<int32_t>(output_scratch_buffer), |
| 2204 | output_ptr, input_zp_ptr, aux_input_zp_ptr, output_state_zp_ptr, |
| 2205 | row_sums_ptr, row_sums_size, compute_row_sums, |
| 2206 | params->asymmetric_quantize_inputs, context); |
| 2207 | } |
| 2208 | } |
| 2209 | } |
| 2210 | |
| 2211 | return kTfLiteOk; |
| 2212 | } |
| 2213 | |
| 2214 | TfLiteStatus EvalInteger8x8_16( |
| 2215 | const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, |
| 2216 | const TfLiteTensor* input_to_forget_weights, |
| 2217 | const TfLiteTensor* input_to_cell_weights, |
| 2218 | const TfLiteTensor* input_to_output_weights, |
| 2219 | const TfLiteTensor* recurrent_to_input_weights, |
| 2220 | const TfLiteTensor* recurrent_to_forget_weights, |
| 2221 | const TfLiteTensor* recurrent_to_cell_weights, |
| 2222 | const TfLiteTensor* recurrent_to_output_weights, |
| 2223 | const TfLiteTensor* cell_to_input_weights, |
| 2224 | const TfLiteTensor* cell_to_forget_weights, |
| 2225 | const TfLiteTensor* cell_to_output_weights, |
| 2226 | const TfLiteTensor* input_layer_norm_coefficients, |
| 2227 | const TfLiteTensor* forget_layer_norm_coefficients, |
| 2228 | const TfLiteTensor* cell_layer_norm_coefficients, |
| 2229 | const TfLiteTensor* output_layer_norm_coefficients, |
| 2230 | const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, |
| 2231 | const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias, |
| 2232 | const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias, |
| 2233 | const TfLiteLSTMParams* params, bool forward_sequence, bool time_major, |
| 2234 | const lstm_eval::IntegerLstmParameter* integer_lstm_param, |
| 2235 | TfLiteTensor* output_state, TfLiteTensor* cell_state, TfLiteTensor* output, |
| 2236 | TfLiteTensor* scratch0, TfLiteTensor* scratch1, TfLiteTensor* scratch2, |
| 2237 | TfLiteTensor* scratch3, TfLiteTensor* scratch4, TfLiteTensor* scratch5, |
| 2238 | CpuBackendContext* context) { |
| 2239 | TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3); |
| 2240 | const int n_input = input->dims->data[input->dims->size - 1]; |
| 2241 | int max_time, n_batch; |
| 2242 | if (input->dims->size == 2) { |
| 2243 | max_time = 1; |
| 2244 | n_batch = input->dims->data[0]; |
| 2245 | } else { |
| 2246 | max_time = (time_major) ? input->dims->data[0] : input->dims->data[1]; |
| 2247 | n_batch = (time_major) ? input->dims->data[1] : input->dims->data[0]; |
| 2248 | } |
| 2249 | |
| 2250 | // n_cell and n_output will be the same size when there is no projection. |
| 2251 | const int n_cell = input_to_output_weights->dims->data[0]; |
| 2252 | const int n_output = recurrent_to_output_weights->dims->data[1]; |
| 2253 | |
| 2254 | // Activation zero point |
| 2255 | int output_state_zp = output_state->params.zero_point; |
| 2256 | |
| 2257 | // Get params for time/batch/sequence. |
| 2258 | const int output_batch_leading_dim = |
| 2259 | output->dims->data[output->dims->size - 1]; |
| 2260 | |
| 2261 | if (time_major) { |
| 2262 | const int input_step = n_batch * n_input; |
| 2263 | const int output_step = n_batch * output_batch_leading_dim; |
| 2264 | for (int t = 0; t < max_time; t++) { |
| 2265 | const int t_rel = t; |
| 2266 | int8_t* output_ptr = GetTensorData<int8_t>(output) + t_rel * output_step; |
| 2267 | const int8_t* input_ptr = |
| 2268 | GetTensorData<int8_t>(input) + t_rel * input_step; |
| 2269 | LstmStepInteger8x8_16( |
| 2270 | input_ptr, GetTensorData<int8_t>(input_to_input_weights), |
| 2271 | integer_lstm_param->effective_input_to_input_scale_a, |
| 2272 | integer_lstm_param->effective_input_to_input_scale_b, |
| 2273 | GetTensorData<int8_t>(input_to_forget_weights), |
| 2274 | integer_lstm_param->effective_input_to_forget_scale_a, |
| 2275 | integer_lstm_param->effective_input_to_forget_scale_b, |
| 2276 | GetTensorData<int8_t>(input_to_cell_weights), |
| 2277 | integer_lstm_param->effective_input_to_cell_scale_a, |
| 2278 | integer_lstm_param->effective_input_to_cell_scale_b, |
| 2279 | GetTensorData<int8_t>(input_to_output_weights), |
| 2280 | integer_lstm_param->effective_input_to_output_scale_a, |
| 2281 | integer_lstm_param->effective_input_to_output_scale_b, |
| 2282 | GetTensorData<int8_t>(recurrent_to_input_weights), |
| 2283 | integer_lstm_param->effective_recurrent_to_input_scale_a, |
| 2284 | integer_lstm_param->effective_recurrent_to_input_scale_b, |
| 2285 | GetTensorData<int8_t>(recurrent_to_forget_weights), |
| 2286 | integer_lstm_param->effective_recurrent_to_forget_scale_a, |
| 2287 | integer_lstm_param->effective_recurrent_to_forget_scale_b, |
| 2288 | GetTensorData<int8_t>(recurrent_to_cell_weights), |
| 2289 | integer_lstm_param->effective_recurrent_to_cell_scale_a, |
| 2290 | integer_lstm_param->effective_recurrent_to_cell_scale_b, |
| 2291 | GetTensorData<int8_t>(recurrent_to_output_weights), |
| 2292 | integer_lstm_param->effective_recurrent_to_output_scale_a, |
| 2293 | integer_lstm_param->effective_recurrent_to_output_scale_b, |
| 2294 | GetTensorData<int16_t>(cell_to_input_weights), |
| 2295 | integer_lstm_param->effective_cell_to_input_scale_a, |
| 2296 | integer_lstm_param->effective_cell_to_input_scale_b, |
| 2297 | GetTensorData<int16_t>(cell_to_forget_weights), |
| 2298 | integer_lstm_param->effective_cell_to_forget_scale_a, |
| 2299 | integer_lstm_param->effective_cell_to_forget_scale_b, |
| 2300 | GetTensorData<int16_t>(cell_to_output_weights), |
| 2301 | integer_lstm_param->effective_cell_to_output_scale_a, |
| 2302 | integer_lstm_param->effective_cell_to_output_scale_b, |
| 2303 | GetTensorData<int8_t>(projection_weights), |
| 2304 | integer_lstm_param->effective_proj_scale_a, |
| 2305 | integer_lstm_param->effective_proj_scale_b, |
| 2306 | integer_lstm_param->hidden_zp, |
| 2307 | integer_lstm_param->effective_hidden_scale_a, |
| 2308 | integer_lstm_param->effective_hidden_scale_b, |
| 2309 | GetTensorData<int16_t>(input_layer_norm_coefficients), |
| 2310 | integer_lstm_param->layer_norm_input_scale_a, |
| 2311 | integer_lstm_param->layer_norm_input_scale_b, |
| 2312 | GetTensorData<int16_t>(forget_layer_norm_coefficients), |
| 2313 | integer_lstm_param->layer_norm_forget_scale_a, |
| 2314 | integer_lstm_param->layer_norm_forget_scale_b, |
| 2315 | GetTensorData<int16_t>(cell_layer_norm_coefficients), |
| 2316 | integer_lstm_param->layer_norm_cell_scale_a, |
| 2317 | integer_lstm_param->layer_norm_cell_scale_b, |
| 2318 | GetTensorData<int16_t>(output_layer_norm_coefficients), |
| 2319 | integer_lstm_param->layer_norm_output_scale_a, |
| 2320 | integer_lstm_param->layer_norm_output_scale_b, |
| 2321 | GetTensorData<int32_t>(input_gate_bias), |
| 2322 | GetTensorData<int32_t>(forget_gate_bias), |
| 2323 | GetTensorData<int32_t>(cell_gate_bias), |
| 2324 | GetTensorData<int32_t>(output_gate_bias), |
| 2325 | integer_lstm_param->quantized_cell_clip, |
| 2326 | integer_lstm_param->quantized_proj_clip, |
| 2327 | integer_lstm_param->cell_scale, |
| 2328 | integer_lstm_param->input_variance_guard, |
| 2329 | integer_lstm_param->forget_variance_guard, |
| 2330 | integer_lstm_param->cell_variance_guard, |
| 2331 | integer_lstm_param->output_variance_guard, |
| 2332 | integer_lstm_param->input_to_forget_effective_bias.get(), |
| 2333 | integer_lstm_param->recurrent_to_forget_effective_bias.get(), |
| 2334 | integer_lstm_param->input_to_cell_effective_bias.get(), |
| 2335 | integer_lstm_param->recurrent_to_cell_effective_bias.get(), |
| 2336 | integer_lstm_param->input_to_output_effective_bias.get(), |
| 2337 | integer_lstm_param->recurrent_to_output_effective_bias.get(), |
| 2338 | integer_lstm_param->input_to_input_effective_bias.get(), |
| 2339 | integer_lstm_param->recurrent_to_input_effective_bias.get(), |
| 2340 | integer_lstm_param->projection_effective_bias.get(), n_batch, n_cell, |
| 2341 | n_input, n_output, GetTensorData<int8_t>(output_state), |
| 2342 | output_state_zp, GetTensorData<int16_t>(cell_state), output_ptr, |
| 2343 | GetTensorData<int16_t>(scratch0), GetTensorData<int16_t>(scratch1), |
| 2344 | GetTensorData<int16_t>(scratch2), GetTensorData<int16_t>(scratch3), |
| 2345 | GetTensorData<int8_t>(scratch4), GetTensorData<int32_t>(scratch5), |
| 2346 | context); |
| 2347 | } |
| 2348 | } else { |
| 2349 | for (int b = 0; b < n_batch; b++) { |
| 2350 | const int input_step = n_input; |
| 2351 | const int output_step = output_batch_leading_dim; |
| 2352 | for (int t = 0; t < max_time; t++) { |
| 2353 | // If this is the forward_sequence, step forward, otherwise step |
| 2354 | // backwards. |
| 2355 | const int t_rel = forward_sequence ? t : max_time - t - 1; |
| 2356 | const int time_offset = b * max_time + t_rel; |
| 2357 | const int8_t* input_ptr = |
| 2358 | GetTensorData<int8_t>(input) + time_offset * input_step; |
| 2359 | int8_t* output_ptr = |
| 2360 | GetTensorData<int8_t>(output) + time_offset * output_step; |
| 2361 | |
| 2362 | // Offset the {output,cell}_state pointers to the right batch. |
| 2363 | int8_t* output_state_ptr = |
| 2364 | GetTensorData<int8_t>(output_state) + b * output_batch_leading_dim; |
| 2365 | int16_t* cell_state_ptr = |
| 2366 | GetTensorData<int16_t>(cell_state) + b * n_cell; |
| 2367 | |
| 2368 | LstmStepInteger8x8_16( |
| 2369 | input_ptr, GetTensorData<int8_t>(input_to_input_weights), |
| 2370 | integer_lstm_param->effective_input_to_input_scale_a, |
| 2371 | integer_lstm_param->effective_input_to_input_scale_b, |
| 2372 | GetTensorData<int8_t>(input_to_forget_weights), |
| 2373 | integer_lstm_param->effective_input_to_forget_scale_a, |
| 2374 | integer_lstm_param->effective_input_to_forget_scale_b, |
| 2375 | GetTensorData<int8_t>(input_to_cell_weights), |
| 2376 | integer_lstm_param->effective_input_to_cell_scale_a, |
| 2377 | integer_lstm_param->effective_input_to_cell_scale_b, |
| 2378 | GetTensorData<int8_t>(input_to_output_weights), |
| 2379 | integer_lstm_param->effective_input_to_output_scale_a, |
| 2380 | integer_lstm_param->effective_input_to_output_scale_b, |
| 2381 | GetTensorData<int8_t>(recurrent_to_input_weights), |
| 2382 | integer_lstm_param->effective_recurrent_to_input_scale_a, |
| 2383 | integer_lstm_param->effective_recurrent_to_input_scale_b, |
| 2384 | GetTensorData<int8_t>(recurrent_to_forget_weights), |
| 2385 | integer_lstm_param->effective_recurrent_to_forget_scale_a, |
| 2386 | integer_lstm_param->effective_recurrent_to_forget_scale_b, |
| 2387 | GetTensorData<int8_t>(recurrent_to_cell_weights), |
| 2388 | integer_lstm_param->effective_recurrent_to_cell_scale_a, |
| 2389 | integer_lstm_param->effective_recurrent_to_cell_scale_b, |
| 2390 | GetTensorData<int8_t>(recurrent_to_output_weights), |
| 2391 | integer_lstm_param->effective_recurrent_to_output_scale_a, |
| 2392 | integer_lstm_param->effective_recurrent_to_output_scale_b, |
| 2393 | GetTensorData<int16_t>(cell_to_input_weights), |
| 2394 | integer_lstm_param->effective_cell_to_input_scale_a, |
| 2395 | integer_lstm_param->effective_cell_to_input_scale_b, |
| 2396 | GetTensorData<int16_t>(cell_to_forget_weights), |
| 2397 | integer_lstm_param->effective_cell_to_forget_scale_a, |
| 2398 | integer_lstm_param->effective_cell_to_forget_scale_b, |
| 2399 | GetTensorData<int16_t>(cell_to_output_weights), |
| 2400 | integer_lstm_param->effective_cell_to_output_scale_a, |
| 2401 | integer_lstm_param->effective_cell_to_output_scale_b, |
| 2402 | GetTensorData<int8_t>(projection_weights), |
| 2403 | integer_lstm_param->effective_proj_scale_a, |
| 2404 | integer_lstm_param->effective_proj_scale_b, |
| 2405 | integer_lstm_param->hidden_zp, |
| 2406 | integer_lstm_param->effective_hidden_scale_a, |
| 2407 | integer_lstm_param->effective_hidden_scale_b, |
| 2408 | GetTensorData<int16_t>(input_layer_norm_coefficients), |
| 2409 | integer_lstm_param->layer_norm_input_scale_a, |
| 2410 | integer_lstm_param->layer_norm_input_scale_b, |
| 2411 | GetTensorData<int16_t>(forget_layer_norm_coefficients), |
| 2412 | integer_lstm_param->layer_norm_forget_scale_a, |
| 2413 | integer_lstm_param->layer_norm_forget_scale_b, |
| 2414 | GetTensorData<int16_t>(cell_layer_norm_coefficients), |
| 2415 | integer_lstm_param->layer_norm_cell_scale_a, |
| 2416 | integer_lstm_param->layer_norm_cell_scale_b, |
| 2417 | GetTensorData<int16_t>(output_layer_norm_coefficients), |
| 2418 | integer_lstm_param->layer_norm_output_scale_a, |
| 2419 | integer_lstm_param->layer_norm_output_scale_b, |
| 2420 | GetTensorData<int32_t>(input_gate_bias), |
| 2421 | GetTensorData<int32_t>(forget_gate_bias), |
| 2422 | GetTensorData<int32_t>(cell_gate_bias), |
| 2423 | GetTensorData<int32_t>(output_gate_bias), |
| 2424 | integer_lstm_param->quantized_cell_clip, |
| 2425 | integer_lstm_param->quantized_proj_clip, |
| 2426 | integer_lstm_param->cell_scale, |
| 2427 | integer_lstm_param->input_variance_guard, |
| 2428 | integer_lstm_param->forget_variance_guard, |
| 2429 | integer_lstm_param->cell_variance_guard, |
| 2430 | integer_lstm_param->output_variance_guard, |
| 2431 | integer_lstm_param->input_to_forget_effective_bias.get(), |
| 2432 | integer_lstm_param->recurrent_to_forget_effective_bias.get(), |
| 2433 | integer_lstm_param->input_to_cell_effective_bias.get(), |
| 2434 | integer_lstm_param->recurrent_to_cell_effective_bias.get(), |
| 2435 | integer_lstm_param->input_to_output_effective_bias.get(), |
| 2436 | integer_lstm_param->recurrent_to_output_effective_bias.get(), |
| 2437 | integer_lstm_param->input_to_input_effective_bias.get(), |
| 2438 | integer_lstm_param->recurrent_to_input_effective_bias.get(), |
| 2439 | integer_lstm_param->projection_effective_bias.get(), /*n_batch=*/1, |
| 2440 | n_cell, n_input, n_output, output_state_ptr, output_state_zp, |
| 2441 | cell_state_ptr, output_ptr, GetTensorData<int16_t>(scratch0), |
| 2442 | GetTensorData<int16_t>(scratch1), GetTensorData<int16_t>(scratch2), |
| 2443 | GetTensorData<int16_t>(scratch3), GetTensorData<int8_t>(scratch4), |
| 2444 | GetTensorData<int32_t>(scratch5), context); |
| 2445 | } |
| 2446 | } |
| 2447 | } |
| 2448 | |
| 2449 | return kTfLiteOk; |
| 2450 | } |
| 2451 | |
| 2452 | TfLiteStatus EvalInteger8x8_8( |
| 2453 | const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, |
| 2454 | const TfLiteTensor* input_to_forget_weights, |
| 2455 | const TfLiteTensor* input_to_cell_weights, |
| 2456 | const TfLiteTensor* input_to_output_weights, |
| 2457 | const TfLiteTensor* recurrent_to_input_weights, |
| 2458 | const TfLiteTensor* recurrent_to_forget_weights, |
| 2459 | const TfLiteTensor* recurrent_to_cell_weights, |
| 2460 | const TfLiteTensor* recurrent_to_output_weights, |
| 2461 | const TfLiteTensor* cell_to_input_weights, |
| 2462 | const TfLiteTensor* cell_to_forget_weights, |
| 2463 | const TfLiteTensor* cell_to_output_weights, |
| 2464 | const TfLiteTensor* input_layer_norm_coefficients, |
| 2465 | const TfLiteTensor* forget_layer_norm_coefficients, |
| 2466 | const TfLiteTensor* cell_layer_norm_coefficients, |
| 2467 | const TfLiteTensor* output_layer_norm_coefficients, |
| 2468 | const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, |
| 2469 | const TfLiteTensor* cell_gate_bias, const TfLiteTensor* output_gate_bias, |
| 2470 | const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias, |
| 2471 | const TfLiteLSTMParams* params, TfLiteTensor* output_state, |
| 2472 | TfLiteTensor* cell_state, TfLiteTensor* output, |
| 2473 | const lstm_eval::IntegerLstmParameter* integer_lstm_param, |
| 2474 | TfLiteTensor* scratch0, TfLiteTensor* scratch1, TfLiteTensor* scratch2, |
| 2475 | TfLiteTensor* scratch3, TfLiteTensor* scratch4, TfLiteTensor* scratch5, |
| 2476 | TfLiteTensor* scratch6, TfLiteTensor* scratch7) { |
| 2477 | TF_LITE_ASSERT(input->dims->size >= 2 && input->dims->size <= 3); |
| 2478 | const int n_input = input->dims->data[input->dims->size - 1]; |
| 2479 | int max_time, n_batch; |
| 2480 | if (input->dims->size == 2) { |
| 2481 | max_time = 1; |
| 2482 | n_batch = input->dims->data[0]; |
| 2483 | } else { |
| 2484 | max_time = input->dims->data[0]; |
| 2485 | n_batch = input->dims->data[1]; |
| 2486 | } |
| 2487 | |
| 2488 | // n_cell and n_output will be the same size when there is no projection. |
| 2489 | const int n_cell = input_to_output_weights->dims->data[0]; |
| 2490 | const int n_output = recurrent_to_output_weights->dims->data[1]; |
| 2491 | |
| 2492 | const int32_t input_zp = input->params.zero_point; |
| 2493 | const int32_t output_state_zp = output_state->params.zero_point; |
| 2494 | |
| 2495 | // Get params for time/batch/sequence. |
| 2496 | const int output_batch_leading_dim = |
| 2497 | output->dims->data[output->dims->size - 1]; |
| 2498 | const int input_step = n_batch * n_input; |
| 2499 | const int output_step = n_batch * output_batch_leading_dim; |
| 2500 | |
| 2501 | for (int t = 0; t < max_time; t++) { |
| 2502 | const int t_rel = t; |
| 2503 | int8_t* output_ptr = GetTensorData<int8_t>(output) + t_rel * output_step; |
| 2504 | // Input can be int8 asymmetric or int16 symmetric. |
| 2505 | const int8_t* input_ptr = GetTensorData<int8_t>(input) + t_rel * input_step; |
| 2506 | lstm_eval::LstmStepInteger8x8_8( |
| 2507 | input_ptr, input_zp, |
| 2508 | |
| 2509 | GetTensorData<int8_t>(input_to_input_weights), |
| 2510 | integer_lstm_param->effective_input_to_input_scale_a, |
| 2511 | integer_lstm_param->effective_input_to_input_scale_b, |
| 2512 | |
| 2513 | GetTensorData<int8_t>(input_to_forget_weights), |
| 2514 | integer_lstm_param->effective_input_to_forget_scale_a, |
| 2515 | integer_lstm_param->effective_input_to_forget_scale_b, |
| 2516 | |
| 2517 | GetTensorData<int8_t>(input_to_cell_weights), |
| 2518 | integer_lstm_param->effective_input_to_cell_scale_a, |
| 2519 | integer_lstm_param->effective_input_to_cell_scale_b, |
| 2520 | |
| 2521 | GetTensorData<int8_t>(input_to_output_weights), |
| 2522 | integer_lstm_param->effective_input_to_output_scale_a, |
| 2523 | integer_lstm_param->effective_input_to_output_scale_b, |
| 2524 | |
| 2525 | GetTensorData<int8_t>(recurrent_to_input_weights), |
| 2526 | integer_lstm_param->effective_recurrent_to_input_scale_a, |
| 2527 | integer_lstm_param->effective_recurrent_to_input_scale_b, |
| 2528 | |
| 2529 | GetTensorData<int8_t>(recurrent_to_forget_weights), |
| 2530 | integer_lstm_param->effective_recurrent_to_forget_scale_a, |
| 2531 | integer_lstm_param->effective_recurrent_to_forget_scale_b, |
| 2532 | |
| 2533 | GetTensorData<int8_t>(recurrent_to_cell_weights), |
| 2534 | integer_lstm_param->effective_recurrent_to_cell_scale_a, |
| 2535 | integer_lstm_param->effective_recurrent_to_cell_scale_b, |
| 2536 | |
| 2537 | GetTensorData<int8_t>(recurrent_to_output_weights), |
| 2538 | integer_lstm_param->effective_recurrent_to_output_scale_a, |
| 2539 | integer_lstm_param->effective_recurrent_to_output_scale_b, |
| 2540 | |
| 2541 | GetTensorData<int8_t>(cell_to_input_weights), |
| 2542 | integer_lstm_param->effective_cell_to_input_scale_a, |
| 2543 | integer_lstm_param->effective_cell_to_input_scale_b, |
| 2544 | |
| 2545 | GetTensorData<int8_t>(cell_to_forget_weights), |
| 2546 | integer_lstm_param->effective_cell_to_forget_scale_a, |
| 2547 | integer_lstm_param->effective_cell_to_forget_scale_b, |
| 2548 | |
| 2549 | GetTensorData<int8_t>(cell_to_output_weights), |
| 2550 | integer_lstm_param->effective_cell_to_output_scale_a, |
| 2551 | integer_lstm_param->effective_cell_to_output_scale_b, |
| 2552 | |
| 2553 | GetTensorData<int8_t>(projection_weights), |
| 2554 | integer_lstm_param->effective_proj_scale_a, |
| 2555 | integer_lstm_param->effective_proj_scale_b, |
| 2556 | |
| 2557 | GetTensorData<int16_t>(input_layer_norm_coefficients), |
| 2558 | integer_lstm_param->layer_norm_input_scale_a, |
| 2559 | integer_lstm_param->layer_norm_input_scale_b, |
| 2560 | |
| 2561 | GetTensorData<int16_t>(forget_layer_norm_coefficients), |
| 2562 | integer_lstm_param->layer_norm_forget_scale_a, |
| 2563 | integer_lstm_param->layer_norm_forget_scale_b, |
| 2564 | |
| 2565 | GetTensorData<int16_t>(cell_layer_norm_coefficients), |
| 2566 | integer_lstm_param->layer_norm_cell_scale_a, |
| 2567 | integer_lstm_param->layer_norm_cell_scale_b, |
| 2568 | |
| 2569 | GetTensorData<int16_t>(output_layer_norm_coefficients), |
| 2570 | integer_lstm_param->layer_norm_output_scale_a, |
| 2571 | integer_lstm_param->layer_norm_output_scale_b, |
| 2572 | |
| 2573 | GetTensorData<int32_t>(input_gate_bias), |
| 2574 | GetTensorData<int32_t>(forget_gate_bias), |
| 2575 | GetTensorData<int32_t>(cell_gate_bias), |
| 2576 | GetTensorData<int32_t>(output_gate_bias), |
| 2577 | GetTensorData<int32_t>(projection_bias), |
| 2578 | |
| 2579 | params, integer_lstm_param->intermediate_scale_a, |
| 2580 | integer_lstm_param->intermediate_scale_b, |
| 2581 | integer_lstm_param->intermediate_zp, |
| 2582 | integer_lstm_param->quantized_cell_clip, |
| 2583 | integer_lstm_param->quantized_proj_clip, n_batch, n_cell, n_input, |
| 2584 | n_output, output_batch_leading_dim, GetTensorData<int8_t>(output_state), |
| 2585 | output_state_zp, GetTensorData<int16_t>(cell_state), output_ptr, |
| 2586 | GetTensorData<int8_t>(scratch0), GetTensorData<int8_t>(scratch1), |
| 2587 | GetTensorData<int16_t>(scratch2), GetTensorData<int16_t>(scratch3), |
| 2588 | GetTensorData<int16_t>(scratch4), GetTensorData<int16_t>(scratch5), |
| 2589 | GetTensorData<int16_t>(scratch6), GetTensorData<int16_t>(scratch7)); |
| 2590 | } |
| 2591 | |
| 2592 | return kTfLiteOk; |
| 2593 | } |
| 2594 | |
| 2595 | } // namespace lstm_eval |
| 2596 | } // namespace builtin |
| 2597 | } // namespace ops |
| 2598 | } // namespace tflite |
| 2599 |
Generated on 2022-Sep-26 from project tensorflow revision f2e850b6cce
Powered by 
Code Browser 2.1
Generator usage only permitted with license.