1 /*
2 * Copyright (c) 2020-2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24 #include "arm_compute/runtime/CL/functions/CLQLSTMLayer.h"
25
26 #include "arm_compute/core/KernelDescriptors.h"
27 #include "arm_compute/core/QuantizationInfo.h"
28 #include "arm_compute/core/Utils.h"
29 #include "arm_compute/core/Validate.h"
30 #include "arm_compute/core/utils/misc/InfoHelpers.h"
31 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
32 #include "arm_compute/runtime/CL/CLScheduler.h"
33 #include "src/core/CL/kernels/CLFillBorderKernel.h"
34 #include "src/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h"
35 #include "src/core/helpers/WindowHelpers.h"
36 #include "src/gpu/cl/kernels/ClGemmLowpReductionKernel.h"
37
38 #include "src/common/utils/Log.h"
39
40 namespace arm_compute
41 {
42 using namespace arm_compute::utils::info_helpers;
43 using namespace arm_compute::opencl::kernels;
44 namespace
45 {
validate_mm(GEMMLowpOutputStageInfo & gemmlowp_info,const ITensorInfo * mm_input,const ITensorInfo * mm_weights,const ITensorInfo * bias,float gemmlowp_scale,const TensorInfo * mm_res_info,const TensorInfo * outstage_tensor_info)46 Status validate_mm(GEMMLowpOutputStageInfo &gemmlowp_info, const ITensorInfo *mm_input, const ITensorInfo *mm_weights, const ITensorInfo *bias,
47 float gemmlowp_scale, const TensorInfo *mm_res_info, const TensorInfo *outstage_tensor_info)
48 {
49 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(mm_input, mm_weights, nullptr, mm_res_info));
50 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(gemmlowp_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift));
51 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(mm_res_info, bias, outstage_tensor_info, gemmlowp_info));
52 return Status{};
53 }
54 } // namespace
55
validate(const ITensorInfo & src,const ITensorInfo & dst)56 Status CLQLSTMLayer::TensorCopyKernel::validate(const ITensorInfo &src, const ITensorInfo &dst)
57 {
58 ARM_COMPUTE_RETURN_ERROR_ON(src.tensor_shape().num_dimensions() > max_dimension_supported);
59 ARM_COMPUTE_RETURN_ERROR_ON(dst.tensor_shape().num_dimensions() > max_dimension_supported);
60 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(&src, &dst);
61 ARM_COMPUTE_RETURN_ERROR_ON(dst.tensor_shape().y() != src.tensor_shape().y());
62 return Status{};
63 }
64
configure(ICLTensor & src,ICLTensor & dst)65 void CLQLSTMLayer::TensorCopyKernel::configure(ICLTensor &src, ICLTensor &dst)
66 {
67 ARM_COMPUTE_ERROR_THROW_ON(CLQLSTMLayer::TensorCopyKernel::validate(*src.info(), *dst.info()));
68 _src = &src;
69 _dst = &dst;
70 _row_size = std::min(_src->info()->tensor_shape().x(), _dst->info()->tensor_shape().x());
71 _window = calculate_max_window(*_src->info(), Steps());
72 }
73
run()74 void CLQLSTMLayer::TensorCopyKernel::run()
75 {
76 auto &q = CLScheduler::get().queue();
77
78 _src->map(q, true);
79 _dst->map(q, true);
80
81 Iterator input_iter{ _src, _window };
82 Iterator output_iter{ _dst, _window };
83
84 execute_window_loop(_window, [&](const Coordinates &)
85 {
86 memcpy(output_iter.ptr(), input_iter.ptr(), _row_size);
87 },
88 input_iter, output_iter);
89
90 _src->unmap(q);
91 _dst->unmap(q);
92 }
93
CLQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)94 CLQLSTMLayer::CLQLSTMLayer(std::shared_ptr<IMemoryManager> memory_manager)
95 : _input_to_input_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
96 _recurrent_to_input_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
97 _input_to_forget_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
98 _recurrent_to_forget_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
99 _input_to_cell_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
100 _recurrent_to_cell_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
101 _input_to_output_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
102 _recurrent_to_output_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
103 _projection_reduction(std::make_unique<ClGemmLowpMatrixAReductionKernel>()),
104 _layer_norms(),
105 _copy_output()
106 {
107 for(auto &norm : _layer_norms)
108 {
109 norm = std::make_unique<CLQLSTMLayerNormalizationKernel>();
110 }
111
112 _memory_group = MemoryGroup(std::move(memory_manager));
113 }
114
115 CLQLSTMLayer::~CLQLSTMLayer() = default;
116
configure_layer_norm(LayerNormGate g,const ICLTensor * in)117 void CLQLSTMLayer::configure_layer_norm(LayerNormGate g, const ICLTensor *in)
118 {
119 ARM_COMPUTE_ERROR_ON(!_has_layer_norm);
120
121 CLTensor *out = &get_layer_norm_output(g);
122 _memory_group.manage(out);
123 out->allocator()->init(*(in->info()));
124
125 get_layer_norm(g).configure(in, out, get_layer_norm_weight(g), get_layer_norm_bias(g));
126 }
127
validate_layer_norm(const ITensorInfo & in,const ITensorInfo & weight,const ITensorInfo & bias)128 Status CLQLSTMLayer::validate_layer_norm(const ITensorInfo &in, const ITensorInfo &weight, const ITensorInfo &bias)
129 {
130 // Output quantization scale will be different, but ignored here
131 // since it will be configured at configure() stage.
132 const TensorInfo out
133 {
134 in
135 };
136 return CLQLSTMLayerNormalizationKernel::validate(&in, &out, &weight, &bias);
137 }
138
configure_mm(const CLCompileContext & compile_context,CLGEMMLowpMatrixMultiplyCore & mm,CLGEMMLowpOutputStage & outstage,GEMMLowpOutputStageInfo & gemmlowp_info,const ICLTensor * mm_input,const ICLTensor * mm_weights,const ICLTensor * bias,CLTensor * mm_res,CLTensor * outstage_res,float gemmlowp_scale,const TensorInfo & mm_res_info,const TensorInfo & outstage_tensor_info)139 void CLQLSTMLayer::configure_mm(const CLCompileContext &compile_context, CLGEMMLowpMatrixMultiplyCore &mm, CLGEMMLowpOutputStage &outstage, GEMMLowpOutputStageInfo &gemmlowp_info,
140 const ICLTensor *mm_input, const ICLTensor *mm_weights, const ICLTensor *bias,
141 CLTensor *mm_res, CLTensor *outstage_res, float gemmlowp_scale,
142 const TensorInfo &mm_res_info, const TensorInfo &outstage_tensor_info)
143 {
144 _memory_group.manage(mm_res);
145 _memory_group.manage(outstage_res);
146
147 mm_res->allocator()->init(mm_res_info);
148 outstage_res->allocator()->init(outstage_tensor_info);
149
150 // Configure matrix-multiplication
151 mm.configure(compile_context, mm_input, mm_weights, nullptr, mm_res);
152
153 // Configure output stage
154 quantization::calculate_quantized_multiplier(gemmlowp_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift);
155 outstage.configure(compile_context, mm_res, bias, outstage_res, gemmlowp_info);
156 mm_res->allocator()->allocate();
157 }
158
configure(const ICLTensor * input,const ICLTensor * input_to_forget_weights,const ICLTensor * input_to_cell_weights,const ICLTensor * input_to_output_weights,const ICLTensor * recurrent_to_forget_weights,const ICLTensor * recurrent_to_cell_weights,const ICLTensor * recurrent_to_output_weights,const ICLTensor * forget_gate_bias,const ICLTensor * cell_bias,const ICLTensor * output_gate_bias,ICLTensor * cell_state_in,ICLTensor * output_state_in,ICLTensor * cell_state_out,ICLTensor * output_state_out,ICLTensor * output,const LSTMParams<ICLTensor> & lstm_params)159 void CLQLSTMLayer::configure(const ICLTensor *input,
160 const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
161 const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
162 const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
163 ICLTensor *cell_state_in, ICLTensor *output_state_in,
164 ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output,
165 const LSTMParams<ICLTensor> &lstm_params)
166 {
167 configure(CLKernelLibrary::get().get_compile_context(), input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
168 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias,
169 cell_state_in, output_state_in, cell_state_out, output_state_out, output, lstm_params);
170 }
171
configure(const CLCompileContext & compile_context,const ICLTensor * input,const ICLTensor * input_to_forget_weights,const ICLTensor * input_to_cell_weights,const ICLTensor * input_to_output_weights,const ICLTensor * recurrent_to_forget_weights,const ICLTensor * recurrent_to_cell_weights,const ICLTensor * recurrent_to_output_weights,const ICLTensor * forget_gate_bias,const ICLTensor * cell_bias,const ICLTensor * output_gate_bias,ICLTensor * cell_state_in,ICLTensor * output_state_in,ICLTensor * cell_state_out,ICLTensor * output_state_out,ICLTensor * output,const LSTMParams<ICLTensor> & lstm_params)172 void CLQLSTMLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input,
173 const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights,
174 const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights,
175 const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias,
176 ICLTensor *cell_state_in, ICLTensor *output_state_in,
177 ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output,
178 const LSTMParams<ICLTensor> &lstm_params)
179 {
180 ARM_COMPUTE_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
181 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
182 forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in,
183 cell_state_out, output_state_out, output);
184
185 ARM_COMPUTE_LOG_PARAMS(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
186 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights,
187 forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in,
188 cell_state_out, output_state_out, output, lstm_params);
189 // Set lstm parameters
190 LSTMParams<ITensorInfo> lstm_params_info{};
191 build_lstm_params_tensor_info(lstm_params, &lstm_params_info);
192
193 // Validate
194 ARM_COMPUTE_ERROR_THROW_ON(CLQLSTMLayer::validate(input->info(), input_to_forget_weights->info(), input_to_cell_weights->info(), input_to_output_weights->info(),
195 recurrent_to_forget_weights->info(), recurrent_to_cell_weights->info(), recurrent_to_output_weights->info(),
196 forget_gate_bias->info(), cell_bias->info(), output_gate_bias->info(),
197 cell_state_in->info(), output_state_in->info(), cell_state_out->info(), output_state_out->info(), output->info(),
198 lstm_params_info));
199
200 const int batch_size = input->info()->dimension(1);
201 const int num_units = input_to_output_weights->info()->dimension(1);
202 const int output_size = output_state_out->info()->dimension(_out_state_output_size_dimension_idx);
203
204 const UniformQuantizationInfo qinput = input->info()->quantization_info().uniform();
205 const UniformQuantizationInfo qcell_state_in = cell_state_in->info()->quantization_info().uniform();
206 const UniformQuantizationInfo qoutput_state_in = output_state_in->info()->quantization_info().uniform();
207
208 _projection_bias = lstm_params.projection_bias();
209 _input_to_forget_weights = input_to_forget_weights;
210 _input_to_cell_weights = input_to_cell_weights;
211 _input_to_output_weights = input_to_output_weights;
212 _recurrent_to_forget_weights = recurrent_to_forget_weights;
213 _recurrent_to_cell_weights = recurrent_to_cell_weights;
214 _recurrent_to_output_weights = recurrent_to_output_weights;
215 _projection_weights = lstm_params.projection_weights();
216
217 // Layer normalization
218 _has_layer_norm = lstm_params.use_layer_norm();
219 if(_has_layer_norm)
220 {
221 set_layer_norm_weight(lstm_params.forget_layer_norm_weights(), LayerNormGate::Forget);
222 set_layer_norm_weight(lstm_params.cell_layer_norm_weights(), LayerNormGate::Cell);
223 set_layer_norm_weight(lstm_params.input_layer_norm_weights(), LayerNormGate::Input);
224 set_layer_norm_weight(lstm_params.output_layer_norm_weights(), LayerNormGate::Output);
225
226 set_layer_norm_bias(forget_gate_bias, LayerNormGate::Forget);
227 set_layer_norm_bias(cell_bias, LayerNormGate::Cell);
228 set_layer_norm_bias(lstm_params.input_gate_bias(), LayerNormGate::Input);
229 set_layer_norm_bias(output_gate_bias, LayerNormGate::Output);
230 }
231
232 _has_cifg = lstm_params.has_cifg_opt();
233 _has_projection = lstm_params.has_projection();
234 _has_peephole = lstm_params.has_peephole_opt();
235
236 // Calculate and decompose effective scales for optimizing matmul calculation
237 const int32_t cell_shift = log2(qcell_state_in.scale);
238
239 // Calculate quantized parameters for clipping.
240 int16_t quantized_cell_clip = 0;
241 if(lstm_params.cell_clip() > 0.0f)
242 {
243 quantized_cell_clip = quantize_qsymm16(lstm_params.cell_clip(), qcell_state_in);
244 }
245 _has_cell_clipping = quantized_cell_clip > 0;
246
247 // Precompute effective bias for optimizing the matmul computations.
248 if(!_has_cifg)
249 {
250 _input_to_input_weights = lstm_params.input_to_input_weights();
251 _recurrent_to_input_weights = lstm_params.recurrent_to_input_weights();
252
253 _input_to_input_reduction->configure(compile_context, _input_to_input_weights->info(), _input_to_input_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true));
254 _recurrent_to_input_reduction->configure(compile_context, _recurrent_to_input_weights->info(), _recurrent_to_input_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false,
255 -qoutput_state_in.offset, true));
256 }
257 _input_to_forget_reduction->configure(compile_context, input_to_forget_weights->info(), _input_to_forget_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true));
258 _recurrent_to_forget_reduction->configure(compile_context, recurrent_to_forget_weights->info(), _recurrent_to_forget_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false,
259 -qoutput_state_in.offset, true));
260 _input_to_cell_reduction->configure(compile_context, input_to_cell_weights->info(), _input_to_cell_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true));
261 _recurrent_to_cell_reduction->configure(compile_context, recurrent_to_cell_weights->info(), _recurrent_to_cell_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset,
262 true));
263 _input_to_output_reduction->configure(compile_context, input_to_output_weights->info(), _input_to_output_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true));
264 _recurrent_to_output_reduction->configure(compile_context, recurrent_to_output_weights->info(), _recurrent_to_output_eff_bias.info(), GEMMLowpReductionKernelInfo(num_units, false,
265 -qoutput_state_in.offset, true));
266 if(_has_projection)
267 {
268 _projection_reduction->configure(compile_context, _projection_weights->info(), _projection_eff_bias.info(), GEMMLowpReductionKernelInfo(output_size, false, lstm_params.hidden_state_zero(), true));
269 if(_projection_bias != nullptr)
270 {
271 _projection_bias_add.configure(compile_context, _projection_bias, &_projection_eff_bias, &_projection_eff_bias, ConvertPolicy::SATURATE);
272 }
273 }
274
275 // Pre-transpose weights to be used in GEMM.
276 _transpose_input_to_forget_weights.configure(compile_context, input_to_forget_weights, &_input_to_forget_weights_transposed);
277 _transpose_input_to_cell_weights.configure(compile_context, input_to_cell_weights, &_input_to_cell_weights_transposed);
278 _transpose_input_to_output_weights.configure(compile_context, input_to_output_weights, &_input_to_output_weights_transposed);
279 _transpose_recurrent_to_forget_weights.configure(compile_context, recurrent_to_forget_weights, &_recurrent_to_forget_weights_transposed);
280 _transpose_recurrent_to_cell_weights.configure(compile_context, recurrent_to_cell_weights, &_recurrent_to_cell_weights_transposed);
281 _transpose_recurrent_to_output_weights.configure(compile_context, recurrent_to_output_weights, &_recurrent_to_output_weights_transposed);
282 if(!_has_cifg)
283 {
284 _transpose_input_to_input_weights.configure(compile_context, lstm_params.input_to_input_weights(), &_input_to_input_weights_transposed);
285 _transpose_recurrent_to_input_weights.configure(compile_context, lstm_params.recurrent_to_input_weights(), &_recurrent_to_input_weights_transposed);
286 }
287 if(_has_projection)
288 {
289 _transpose_projection_weights.configure(compile_context, _projection_weights, &_projection_weights_transposed);
290 }
291
292 GEMMLowpOutputStageInfo gemmlowp_info;
293 gemmlowp_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
294 gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int16_t>::lowest();
295 gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int16_t>::max();
296 gemmlowp_info.output_data_type = DataType::QSYMM16;
297
298 const TensorInfo mm_out_info(TensorShape(num_units, batch_size), 1, DataType::S32);
299 // Forget gate.
300 const TensorInfo forget_gate_outstage_info(mm_out_info.tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0));
301 const float input_to_forget_scale = input_to_forget_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.forget_intermediate_scale();
302 configure_mm(compile_context, _mm_input_to_forget, _input_to_forget_outstage, gemmlowp_info,
303 input, &_input_to_forget_weights_transposed, &_input_to_forget_eff_bias,
304 &_mm_input_to_forget_res, &_input_to_forget_outstage_res, input_to_forget_scale,
305 mm_out_info, forget_gate_outstage_info);
306
307 const float recurrent_to_forget_scale = recurrent_to_forget_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.forget_intermediate_scale();
308 configure_mm(compile_context, _mm_recurrent_to_forget, _recurrent_to_forget_outstage, gemmlowp_info,
309 output_state_in, &_recurrent_to_forget_weights_transposed, &_recurrent_to_forget_eff_bias,
310 &_mm_recurrent_to_forget_res, &_recurrent_to_forget_outstage_res, recurrent_to_forget_scale,
311 mm_out_info, forget_gate_outstage_info);
312
313 _accumulate_input_recurrent_forget.configure(compile_context, &_input_to_forget_outstage_res, &_recurrent_to_forget_outstage_res, &_recurrent_to_forget_outstage_res,
314 ConvertPolicy::SATURATE);
315 _input_to_forget_outstage_res.allocator()->allocate();
316
317 if(_has_peephole)
318 {
319 _mul_cell_to_forget_res.allocator()->init(TensorInfo(cell_state_in->info()->tensor_shape(), 1, DataType::S32));
320 _memory_group.manage(&_mul_cell_to_forget_res);
321 _pixelwise_mul_cell_to_forget.configure(compile_context, cell_state_in, lstm_params.cell_to_forget_weights(), &_mul_cell_to_forget_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
322 _cell_to_forget_outstage_res.allocator()->init(TensorInfo(_mul_cell_to_forget_res.info()->tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0)));
323 _memory_group.manage(&_cell_to_forget_outstage_res);
324 const float cell_to_forget_scale = std::pow(2, cell_shift) * lstm_params.cell_to_forget_weights()->info()->quantization_info().uniform().scale / lstm_params.forget_intermediate_scale();
325 quantization::calculate_quantized_multiplier(cell_to_forget_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift);
326 _cell_to_forget_outstage.configure(compile_context, &_mul_cell_to_forget_res, nullptr, &_cell_to_forget_outstage_res, gemmlowp_info);
327 _mul_cell_to_forget_res.allocator()->allocate();
328 _accumulate_cell_forget.configure(compile_context, &_recurrent_to_forget_outstage_res, &_cell_to_forget_outstage_res, &_recurrent_to_forget_outstage_res,
329 ConvertPolicy::SATURATE);
330 _cell_to_forget_outstage_res.allocator()->allocate();
331 }
332
333 CLTensor *forget_activation_input = &_recurrent_to_forget_outstage_res;
334
335 if(_has_layer_norm)
336 {
337 configure_layer_norm(LayerNormGate::Forget, &_recurrent_to_forget_outstage_res);
338 _recurrent_to_forget_outstage_res.allocator()->allocate();
339 forget_activation_input = &get_layer_norm_output(LayerNormGate::Forget);
340 }
341
342 // Output quantization info of Sigmoid and Tanh activations
343 const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0);
344
345 const TensorInfo forget_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
346 _memory_group.manage(&_forget_gate);
347 _forget_gate.allocator()->init(forget_gate_info);
348 _forget_gate_sigmoid.configure(compile_context, forget_activation_input, &_forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
349 forget_activation_input->allocator()->allocate();
350
351 // Modulation gate.
352 const TensorInfo cell_outstage_info(mm_out_info.tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0));
353 const float input_to_cell_scale = input_to_cell_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.cell_intermediate_scale();
354 configure_mm(compile_context, _mm_input_to_cell, _input_to_cell_outstage, gemmlowp_info,
355 input, &_input_to_cell_weights_transposed, &_input_to_cell_eff_bias,
356 &_mm_input_to_cell_res, &_input_to_cell_outstage_res, input_to_cell_scale,
357 mm_out_info, cell_outstage_info);
358
359 const float recurrent_to_cell_scale = recurrent_to_cell_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.cell_intermediate_scale();
360 configure_mm(compile_context, _mm_recurrent_to_cell, _recurrent_to_cell_outstage, gemmlowp_info,
361 output_state_in, &_recurrent_to_cell_weights_transposed, &_recurrent_to_cell_eff_bias,
362 &_mm_recurrent_to_cell_res, &_recurrent_to_cell_outstage_res, recurrent_to_cell_scale,
363 mm_out_info, cell_outstage_info);
364
365 _accumulate_input_recurrent_modulation.configure(compile_context, &_input_to_cell_outstage_res, &_recurrent_to_cell_outstage_res, &_recurrent_to_cell_outstage_res,
366 ConvertPolicy::SATURATE);
367 _input_to_cell_outstage_res.allocator()->allocate();
368
369 CLTensor *cell_activation_input = &_recurrent_to_cell_outstage_res;
370
371 if(_has_layer_norm)
372 {
373 configure_layer_norm(LayerNormGate::Cell, &_recurrent_to_cell_outstage_res);
374 _recurrent_to_cell_outstage_res.allocator()->allocate();
375 cell_activation_input = &get_layer_norm_output(LayerNormGate::Cell);
376 }
377
378 const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
379 _memory_group.manage(&_cell_gate);
380 _cell_gate.allocator()->init(cell_gate_info);
381 _cell_gate_tanh.configure(compile_context, cell_activation_input, &_cell_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
382 cell_activation_input->allocator()->allocate();
383
384 // Input gate.
385 const TensorInfo input_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
386 _input_gate.allocator()->init(input_gate_info);
387 _memory_group.manage(&_input_gate);
388 if(_has_cifg)
389 {
390 _ones.allocator()->init(*_forget_gate.info());
391 _input_gate_sub.configure(compile_context, &_ones, &_forget_gate, &_input_gate, ConvertPolicy::SATURATE);
392 _ones.allocator()->allocate();
393 }
394 else
395 {
396 const TensorInfo input_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0));
397 const float input_to_input_scale = _input_to_input_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.input_intermediate_scale();
398 configure_mm(compile_context, _mm_input_to_input, _input_to_input_outstage, gemmlowp_info,
399 input, &_input_to_input_weights_transposed, &_input_to_input_eff_bias,
400 &_mm_input_to_input_res, &_input_to_input_outstage_res, input_to_input_scale,
401 mm_out_info, input_outstage_info);
402
403 const float recurrent_to_input_scale = _recurrent_to_input_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.input_intermediate_scale();
404 configure_mm(compile_context, _mm_recurrent_to_input, _recurrent_to_input_outstage, gemmlowp_info,
405 output_state_in, &_recurrent_to_input_weights_transposed, &_recurrent_to_input_eff_bias,
406 &_mm_recurrent_to_input_res, &_recurrent_to_input_outstage_res, recurrent_to_input_scale,
407 mm_out_info, input_outstage_info);
408 _accumulate_input_recurrent_input.configure(compile_context, &_input_to_input_outstage_res, &_recurrent_to_input_outstage_res, &_recurrent_to_input_outstage_res,
409 ConvertPolicy::SATURATE);
410 _input_to_input_outstage_res.allocator()->allocate();
411
412 if(_has_peephole)
413 {
414 _mul_cell_to_input_res.allocator()->init(TensorInfo(cell_state_in->info()->tensor_shape(), 1, DataType::S32));
415 _memory_group.manage(&_mul_cell_to_input_res);
416 _pixelwise_mul_cell_to_input.configure(compile_context, cell_state_in, lstm_params.cell_to_input_weights(), &_mul_cell_to_input_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
417 const float cell_to_input_scale = std::pow(2, cell_shift) * lstm_params.cell_to_input_weights()->info()->quantization_info().uniform().scale / lstm_params.input_intermediate_scale();
418 quantization::calculate_quantized_multiplier(cell_to_input_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift);
419 _cell_to_input_outstage_res.allocator()->init(TensorInfo(_mul_cell_to_input_res.info()->tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0)));
420 _memory_group.manage(&_cell_to_input_outstage_res);
421 _cell_to_input_outstage.configure(compile_context, &_mul_cell_to_input_res, nullptr, &_cell_to_input_outstage_res, gemmlowp_info);
422 _mul_cell_to_input_res.allocator()->allocate();
423 _accumulate_cell_input.configure(&_recurrent_to_input_outstage_res, &_cell_to_input_outstage_res, &_recurrent_to_input_outstage_res, ConvertPolicy::SATURATE);
424 _cell_to_input_outstage_res.allocator()->allocate();
425 }
426
427 CLTensor *input_activation_input = &_recurrent_to_input_outstage_res;
428
429 if(_has_layer_norm)
430 {
431 configure_layer_norm(LayerNormGate::Input, &_recurrent_to_input_outstage_res);
432 _recurrent_to_input_outstage_res.allocator()->allocate();
433 input_activation_input = &get_layer_norm_output(LayerNormGate::Input);
434 }
435
436 _input_gate_sigmoid.configure(compile_context, input_activation_input, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
437 input_activation_input->allocator()->allocate();
438 }
439 // Cell.
440 // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication
441 _pixelwise_mul_forget_cell.configure(compile_context, &_forget_gate, cell_state_in, &_forget_gate, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
442 const float cell_gate_scale = _cell_gate.info()->quantization_info().uniform().scale;
443 const float mul_input_cell_scale = cell_gate_scale * std::pow(2, 15 + cell_shift);
444 const TensorInfo mul_input_cell_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(mul_input_cell_scale, 0));
445 _memory_group.manage(&_mul_input_cell_res);
446 _mul_input_cell_res.allocator()->init(mul_input_cell_info);
447 _pixelwise_mul_input_cell.configure(compile_context, &_input_gate, &_cell_gate, &_mul_input_cell_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
448 _cell_gate.allocator()->allocate();
449 _add_forget_cell.configure(compile_context, &_forget_gate, &_mul_input_cell_res, cell_state_out, ConvertPolicy::SATURATE);
450 _mul_input_cell_res.allocator()->allocate();
451 _forget_gate.allocator()->allocate();
452 if(_has_cell_clipping)
453 {
454 _cell_clip.configure(compile_context, cell_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_cell_clip, quantized_cell_clip));
455 }
456 // Output gate.
457 const TensorInfo output_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0));
458 const float input_to_output_scale = input_to_output_weights->info()->quantization_info().uniform().scale * qinput.scale / lstm_params.output_intermediate_scale();
459 configure_mm(compile_context, _mm_input_to_output, _input_to_output_outstage, gemmlowp_info,
460 input, &_input_to_output_weights_transposed, &_input_to_output_eff_bias,
461 &_mm_input_to_output_res, &_input_to_output_outstage_res, input_to_output_scale,
462 mm_out_info, output_outstage_info);
463
464 const float recurrent_to_output_scale = recurrent_to_output_weights->info()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.output_intermediate_scale();
465 configure_mm(compile_context, _mm_recurrent_to_output, _recurrent_to_output_outstage, gemmlowp_info,
466 output_state_in, &_recurrent_to_output_weights_transposed, &_recurrent_to_output_eff_bias,
467 &_mm_recurrent_to_output_res, &_recurrent_to_output_outstage_res, recurrent_to_output_scale,
468 mm_out_info, output_outstage_info);
469
470 _accumulate_input_recurrent_output.configure(compile_context, &_recurrent_to_output_outstage_res, &_input_to_output_outstage_res, &_recurrent_to_output_outstage_res,
471 ConvertPolicy::SATURATE);
472 _input_to_output_outstage_res.allocator()->allocate();
473
474 if(_has_peephole)
475 {
476 // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication
477 // Here we are not using the output stage because all operations are done in float
478 _mul_cell_to_output_res.allocator()->init(TensorInfo(cell_state_out->info()->tensor_shape(), 1, DataType::S32));
479 _memory_group.manage(&_mul_cell_to_output_res);
480 _pixelwise_mul_cell_to_output.configure(compile_context, cell_state_out, lstm_params.cell_to_output_weights(), &_mul_cell_to_output_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
481
482 const float cell_to_output_scale = std::pow(2, cell_shift) * lstm_params.cell_to_output_weights()->info()->quantization_info().uniform().scale / lstm_params.output_intermediate_scale();
483 quantization::calculate_quantized_multiplier(cell_to_output_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift);
484 _cell_to_output_outstage_res.allocator()->init(TensorInfo(_mul_cell_to_output_res.info()->tensor_shape(), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0)));
485 _memory_group.manage(&_cell_to_output_outstage_res);
486 _cell_to_output_outstage.configure(compile_context, &_mul_cell_to_output_res, nullptr, &_cell_to_output_outstage_res, gemmlowp_info);
487 _mul_cell_to_output_res.allocator()->allocate();
488
489 _accumulate_cell_to_output.configure(compile_context, &_recurrent_to_output_outstage_res, &_cell_to_output_outstage_res, &_recurrent_to_output_outstage_res,
490 ConvertPolicy::SATURATE);
491 _cell_to_output_outstage_res.allocator()->allocate();
492 }
493
494 CLTensor *output_activation_input = &_recurrent_to_output_outstage_res;
495
496 if(_has_layer_norm)
497 {
498 configure_layer_norm(LayerNormGate::Output, &_recurrent_to_output_outstage_res);
499 _recurrent_to_output_outstage_res.allocator()->allocate();
500 output_activation_input = &get_layer_norm_output(LayerNormGate::Output);
501 }
502
503 const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
504 _memory_group.manage(&_output_gate);
505 _output_gate.allocator()->init(output_gate_info);
506 _output_gate_sigmoid.configure(compile_context, output_activation_input, &_output_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC));
507 output_activation_input->allocator()->allocate();
508
509 // Hidden.
510 _hidden_tanh.configure(compile_context, cell_state_out, &_input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f));
511 // TODO(COMPMID-3396): Perform multiplication in the quantized domain in CLPixelWiseMultiplication
512 _memory_group.manage(&_hidden_mul_res);
513 const TensorInfo hidden_mul_res(_input_gate.info()->tensor_shape(), 1, DataType::S32);
514 _hidden_mul_res.allocator()->init(hidden_mul_res);
515 _pixelwise_mul_hidden.configure(compile_context, &_output_gate, &_input_gate, &_hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO);
516 _output_gate.allocator()->allocate();
517 _input_gate.allocator()->allocate();
518 const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15);
519 quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift, /* ignore_epsilon */ true);
520 gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero();
521 gemmlowp_info.output_data_type = output_state_in->info()->data_type();
522
523 _projection_tensor_copy_required = (num_units != output_size);
524 ICLTensor *hidden_gate_result = output_state_out;
525
526 _memory_group.manage(&_hidden_gate);
527
528 if(_projection_tensor_copy_required)
529 {
530 _hidden_gate.allocator()->init(*output_state_out->info());
531 _hidden_gate.info()->set_tensor_shape(_hidden_mul_res.info()->tensor_shape());
532 hidden_gate_result = &_hidden_gate;
533 }
534
535 _hidden_outstage.configure(compile_context, &_hidden_mul_res, nullptr, hidden_gate_result, gemmlowp_info);
536 _hidden_mul_res.allocator()->allocate();
537
538 // Projection.
539 if(_has_projection)
540 {
541 const TensorInfo projection_outstage_info(*output_state_out->info());
542 const UniformQuantizationInfo qprojection = _projection_weights->info()->quantization_info().uniform();
543 const float projection_scale = qprojection.scale * lstm_params.hidden_state_scale() / qoutput_state_in.scale;
544 gemmlowp_info.gemmlowp_offset = qoutput_state_in.offset;
545 gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int8_t>::lowest();
546 gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int8_t>::max();
547 gemmlowp_info.output_data_type = DataType::QASYMM8_SIGNED;
548
549 TensorInfo projection_mm_out_info{ mm_out_info };
550 projection_mm_out_info.set_tensor_shape(TensorShape(output_size, batch_size));
551
552 configure_mm(compile_context, _mm_projection, _projection_outstage, gemmlowp_info,
553 hidden_gate_result, &_projection_weights_transposed, &_projection_eff_bias,
554 &_mm_projection_res, &_projection_outstage_res, projection_scale,
555 projection_mm_out_info, projection_outstage_info);
556
557 ICLTensor *accumulate_destination = output_state_out;
558
559 if(_projection_tensor_copy_required)
560 {
561 _hidden_gate.allocator()->allocate();
562 _projection_accumulate_res.allocator()->init(*output_state_in->info());
563 _projection_accumulate_res.info()->set_tensor_shape(_projection_outstage_res.info()->tensor_shape());
564 _projection_output_to_accumulate_copy.configure(*output_state_in, _projection_accumulate_res);
565 accumulate_destination = &_projection_accumulate_res;
566 }
567
568 _accumulate_projection.configure(compile_context, &_projection_outstage_res, accumulate_destination, accumulate_destination, ConvertPolicy::SATURATE);
569 _projection_outstage_res.allocator()->allocate();
570
571 if(_projection_tensor_copy_required)
572 {
573 _projection_accumulate_to_output_copy.configure(_projection_accumulate_res, *output_state_out);
574 _projection_accumulate_res.allocator()->allocate();
575 }
576
577 int8_t quantized_projection_clip{ 0 };
578 if(lstm_params.projection_clip() > 0.0f)
579 {
580 quantized_projection_clip = utility::clamp<int8_t>(lstm_params.projection_clip() / qprojection.scale, -128, 127);
581 }
582
583 if(quantized_projection_clip > 0)
584 {
585 _projection_clip.configure(compile_context, output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_projection_clip,
586 quantized_projection_clip));
587 _has_projection_clipping = true;
588 }
589 }
590 else
591 {
592 if(_projection_tensor_copy_required)
593 {
594 _hidden_to_output_copy.configure(_hidden_gate, *output_state_out);
595 _hidden_gate.allocator()->allocate();
596 }
597 }
598
599 // Copy output_state_out to output
600 _copy_output.configure(compile_context, output_state_out, output);
601 }
602
validate(const ITensorInfo * input,const ITensorInfo * input_to_forget_weights,const ITensorInfo * input_to_cell_weights,const ITensorInfo * input_to_output_weights,const ITensorInfo * recurrent_to_forget_weights,const ITensorInfo * recurrent_to_cell_weights,const ITensorInfo * recurrent_to_output_weights,const ITensorInfo * forget_gate_bias,const ITensorInfo * cell_bias,const ITensorInfo * output_gate_bias,const ITensorInfo * cell_state_in,const ITensorInfo * output_state_in,const ITensorInfo * cell_state_out,const ITensorInfo * output_state_out,const ITensorInfo * output,const LSTMParams<ITensorInfo> & lstm_params)603 Status CLQLSTMLayer::validate(const ITensorInfo *input,
604 const ITensorInfo *input_to_forget_weights, const ITensorInfo *input_to_cell_weights, const ITensorInfo *input_to_output_weights,
605 const ITensorInfo *recurrent_to_forget_weights, const ITensorInfo *recurrent_to_cell_weights, const ITensorInfo *recurrent_to_output_weights,
606 const ITensorInfo *forget_gate_bias, const ITensorInfo *cell_bias, const ITensorInfo *output_gate_bias,
607 const ITensorInfo *cell_state_in, const ITensorInfo *output_state_in,
608 const ITensorInfo *cell_state_out, const ITensorInfo *output_state_out, const ITensorInfo *output,
609 const LSTMParams<ITensorInfo> &lstm_params)
610 {
611 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, input_to_forget_weights, input_to_cell_weights, input_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights,
612 recurrent_to_output_weights, forget_gate_bias, cell_bias, output_gate_bias, cell_state_in, output_state_in,
613 cell_state_out, output_state_out, output);
614
615 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8_SIGNED);
616 ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->num_dimensions() != 2, "Input must have exactly 2 dimensions");
617
618 const unsigned int input_size = input->dimension(0);
619 const unsigned int batch_size = input->dimension(1);
620 const unsigned int num_units = input_to_output_weights->dimension(1);
621 const unsigned int output_size = output_state_out->dimension(_out_state_output_size_dimension_idx);
622
623 ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->num_dimensions() != 2);
624 ARM_COMPUTE_RETURN_ERROR_ON(input_to_output_weights->dimension(0) != input_size);
625 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_to_output_weights, input_to_forget_weights, input_to_cell_weights);
626 ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->num_dimensions() != 2);
627 ARM_COMPUTE_RETURN_ERROR_ON(recurrent_to_output_weights->dimension(1) != num_units);
628 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(recurrent_to_output_weights, recurrent_to_forget_weights, recurrent_to_cell_weights);
629 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_to_forget_weights, 1, DataType::QSYMM8);
630 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_to_forget_weights, input_to_cell_weights, input_to_output_weights,
631 recurrent_to_forget_weights, recurrent_to_cell_weights, recurrent_to_output_weights);
632
633 ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->num_dimensions() != 1);
634 ARM_COMPUTE_RETURN_ERROR_ON(forget_gate_bias->dimension(0) != num_units);
635 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(forget_gate_bias, cell_bias, output_gate_bias);
636 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(forget_gate_bias, 1, DataType::S32);
637 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, cell_bias, output_gate_bias);
638
639 ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->num_dimensions() != 2);
640 ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->dimension(0) != num_units);
641 ARM_COMPUTE_RETURN_ERROR_ON(cell_state_in->dimension(1) != batch_size);
642 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(cell_state_in, 1, DataType::QSYMM16);
643
644 ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->num_dimensions() != 2);
645 ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->dimension(0) != output_size);
646 ARM_COMPUTE_RETURN_ERROR_ON(output_state_in->dimension(1) != batch_size);
647 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output_state_in);
648
649 // Check whether peephole weights are all there or none
650 if(lstm_params.has_peephole_opt())
651 {
652 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights());
653 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_forget_weights(), 1, DataType::QSYMM16);
654 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->num_dimensions() != 1);
655 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_to_forget_weights()->dimension(0) != num_units);
656 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights());
657 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_output_weights());
658
659 if(!lstm_params.has_cifg_opt())
660 {
661 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.cell_to_input_weights());
662 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_input_weights());
663 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(lstm_params.cell_to_forget_weights(), lstm_params.cell_to_input_weights());
664 }
665 }
666
667 const UniformQuantizationInfo qinput = input->quantization_info().uniform();
668 const UniformQuantizationInfo qcell_state_in = cell_state_in->quantization_info().uniform();
669 const UniformQuantizationInfo qoutput_state_in = output_state_in->quantization_info().uniform();
670
671 // Calculate and decompose effective scales for optimizing matmul calculation
672 const int32_t cell_shift = log2(qcell_state_in.scale);
673 ARM_COMPUTE_RETURN_ERROR_ON(cell_shift > -9);
674
675 // Calculate quantized parameters for clipping.
676 int16_t quantized_cell_clip = 0;
677 if(lstm_params.cell_clip() > 0.0f)
678 {
679 quantized_cell_clip = quantize_qsymm16(lstm_params.cell_clip(), qcell_state_in);
680 }
681
682 // Precompute effective bias for optimizing the matmul computations.
683 const TensorInfo eff_bias_info(TensorShape(num_units), 1, DataType::S32);
684 const TensorInfo projection_eff_bias_info(TensorShape(output_size), 1, DataType::S32);
685 if(!lstm_params.has_cifg_opt())
686 {
687 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(lstm_params.input_to_input_weights(), &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)));
688 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(lstm_params.recurrent_to_input_weights(), &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset,
689 true)));
690 }
691 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(input_to_forget_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)));
692 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(recurrent_to_forget_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)));
693 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(input_to_cell_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)));
694 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(recurrent_to_cell_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)));
695 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(input_to_output_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qinput.offset, true)));
696 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(recurrent_to_output_weights, &eff_bias_info, GEMMLowpReductionKernelInfo(num_units, false, -qoutput_state_in.offset, true)));
697 if(lstm_params.has_projection())
698 {
699 ARM_COMPUTE_RETURN_ON_ERROR(ClGemmLowpMatrixAReductionKernel::validate(lstm_params.projection_weights(), &projection_eff_bias_info, GEMMLowpReductionKernelInfo(output_size, false,
700 lstm_params.hidden_state_zero(),
701 true)));
702 if(lstm_params.projection_bias() != nullptr)
703 {
704 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.projection_bias(), 1, DataType::S32);
705 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(lstm_params.projection_bias(), &projection_eff_bias_info,
706 &projection_eff_bias_info, ConvertPolicy::SATURATE));
707 }
708 }
709
710 const TensorInfo input_weights_transposed(TensorShape(num_units, input_size), 1, input_to_forget_weights->data_type(), input_to_forget_weights->quantization_info());
711 const TensorInfo recurrent_weights_transposed(TensorShape(num_units, output_size), 1, recurrent_to_forget_weights->data_type(), recurrent_to_forget_weights->quantization_info());
712
713 // Validate weights transpose
714 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(input_to_forget_weights, &input_weights_transposed));
715 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(input_to_cell_weights, &input_weights_transposed));
716 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(input_to_output_weights, &input_weights_transposed));
717 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(recurrent_to_forget_weights, &recurrent_weights_transposed));
718 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(recurrent_to_cell_weights, &recurrent_weights_transposed));
719 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(recurrent_to_output_weights, &recurrent_weights_transposed));
720 if(!lstm_params.has_cifg_opt())
721 {
722 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(lstm_params.input_to_input_weights(), &input_weights_transposed));
723 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(lstm_params.recurrent_to_input_weights(), &recurrent_weights_transposed));
724 }
725 if(lstm_params.has_projection())
726 {
727 const TensorInfo projection_weights_transposed(TensorShape(output_size, num_units), 1, lstm_params.projection_weights()->data_type(), lstm_params.projection_weights()->quantization_info());
728 ARM_COMPUTE_RETURN_ON_ERROR(CLTranspose::validate(lstm_params.projection_weights(), &projection_weights_transposed));
729 }
730
731 GEMMLowpOutputStageInfo gemmlowp_info;
732 gemmlowp_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
733 gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int16_t>::lowest();
734 gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int16_t>::max();
735 gemmlowp_info.output_data_type = DataType::QSYMM16;
736
737 const bool has_layer_norm = lstm_params.use_layer_norm();
738
739 // Forget gate.
740 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.forget_intermediate_scale() == 0);
741 const TensorInfo forget_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.forget_intermediate_scale(), 0));
742 const TensorInfo mm_out_info(TensorShape(num_units, batch_size), 1, DataType::S32);
743 const float input_to_forget_scale = input_to_forget_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.forget_intermediate_scale();
744 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_forget_scale, &mm_out_info, &forget_outstage_info));
745
746 const float recurrent_to_forget_scale = recurrent_to_forget_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.forget_intermediate_scale();
747 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, output_state_in, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_forget_scale, &mm_out_info, &forget_outstage_info));
748
749 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_outstage_info, &forget_outstage_info, &forget_outstage_info, ConvertPolicy::SATURATE));
750
751 if(lstm_params.has_peephole_opt())
752 {
753 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_forget_weights(), 1, DataType::QSYMM16);
754 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_forget_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE,
755 RoundingPolicy::TO_ZERO));
756 const float cell_to_forget_scale = std::pow(2, cell_shift) * lstm_params.cell_to_forget_weights()->quantization_info().uniform().scale / lstm_params.forget_intermediate_scale();
757 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_forget_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift));
758 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&mm_out_info, nullptr, &forget_outstage_info, gemmlowp_info));
759 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_outstage_info, &forget_outstage_info, &forget_outstage_info, ConvertPolicy::SATURATE));
760 }
761
762 if(has_layer_norm)
763 {
764 const ITensorInfo *w_info = lstm_params.forget_layer_norm_weights();
765 const ITensorInfo *b_info = forget_gate_bias;
766 ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(forget_outstage_info, *w_info, *b_info));
767 }
768
769 // Output quantization info of Sigmoid and Tanh activations
770 const QuantizationInfo sigmoid_tanh_outqinfo(1.f / 32768.f, 0);
771
772 const TensorInfo forget_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
773 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_outstage_info, &forget_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
774
775 // Modulation gate.
776 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.cell_intermediate_scale() == 0);
777 const TensorInfo cell_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.cell_intermediate_scale(), 0));
778 const float input_to_cell_scale = input_to_cell_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.cell_intermediate_scale();
779 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_cell_scale, &mm_out_info, &cell_outstage_info));
780
781 const float recurrent_to_cell_scale = recurrent_to_cell_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.cell_intermediate_scale();
782 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, output_state_in, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_cell_scale, &mm_out_info, &cell_outstage_info));
783
784 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_outstage_info, &cell_outstage_info, &cell_outstage_info, ConvertPolicy::SATURATE));
785
786 if(has_layer_norm)
787 {
788 const ITensorInfo *w_info = lstm_params.cell_layer_norm_weights();
789 const ITensorInfo *b_info = cell_bias;
790 ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(cell_outstage_info, *w_info, *b_info));
791 }
792
793 const TensorInfo cell_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
794 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_outstage_info, &cell_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)));
795
796 // Input gate.
797 const TensorInfo input_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
798 if(lstm_params.has_cifg_opt())
799 {
800 ARM_COMPUTE_RETURN_ERROR_ON_MSG(lstm_params.input_gate_bias() != nullptr, "Input gate bias must not be present when CIFG is used");
801 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticSubtraction::validate(&input_gate_info, &forget_gate_info, &forget_gate_info, ConvertPolicy::SATURATE));
802 }
803 else
804 {
805 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.input_gate_bias());
806 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_to_forget_weights, lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights());
807 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_to_forget_weights, lstm_params.input_to_input_weights());
808 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(recurrent_to_forget_weights, lstm_params.recurrent_to_input_weights());
809 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(forget_gate_bias, lstm_params.input_gate_bias());
810 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(forget_gate_bias, lstm_params.input_gate_bias());
811
812 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.input_intermediate_scale() == 0);
813 const TensorInfo input_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.input_intermediate_scale(), 0));
814 const float input_to_input_scale = lstm_params.input_to_input_weights()->quantization_info().uniform().scale * qinput.scale / lstm_params.input_intermediate_scale();
815 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_input_scale, &mm_out_info, &input_outstage_info));
816
817 const float recurrent_to_input_scale = lstm_params.recurrent_to_input_weights()->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.input_intermediate_scale();
818 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, output_state_in, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_input_scale, &mm_out_info, &input_outstage_info));
819
820 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_outstage_info, &input_outstage_info, &input_outstage_info, ConvertPolicy::SATURATE));
821
822 if(lstm_params.has_peephole_opt())
823 {
824 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_in, lstm_params.cell_to_input_weights(), &mm_out_info, 1.f, ConvertPolicy::SATURATE,
825 RoundingPolicy::TO_ZERO));
826 const float cell_to_input_scale = std::pow(2, cell_shift) * lstm_params.cell_to_input_weights()->quantization_info().uniform().scale / lstm_params.input_intermediate_scale();
827 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_input_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift));
828 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&mm_out_info, &eff_bias_info, &input_outstage_info, gemmlowp_info));
829 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_outstage_info, &input_outstage_info, &input_outstage_info, ConvertPolicy::SATURATE));
830 }
831
832 if(has_layer_norm)
833 {
834 const ITensorInfo *w_info = lstm_params.input_layer_norm_weights();
835 const ITensorInfo *b_info = lstm_params.input_gate_bias();
836 ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(cell_outstage_info, *w_info, *b_info));
837 }
838
839 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_outstage_info, &input_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 1.f, 1.f)));
840 }
841 // Cell.
842 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&forget_gate_info, cell_state_in, &forget_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
843 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&input_gate_info, cell_state_in, &cell_gate_info, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
844 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate_info, &cell_gate_info, cell_state_out, ConvertPolicy::SATURATE));
845 if(quantized_cell_clip > 0)
846 {
847 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(cell_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_cell_clip,
848 quantized_cell_clip)));
849 }
850 // Output gate.
851 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.output_intermediate_scale() == 0);
852 const TensorInfo output_outstage_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, QuantizationInfo(lstm_params.output_intermediate_scale(), 0));
853 const float input_to_output_scale = input_to_output_weights->quantization_info().uniform().scale * qinput.scale / lstm_params.output_intermediate_scale();
854 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, input, &input_weights_transposed, &eff_bias_info, input_to_output_scale, &mm_out_info, &output_outstage_info));
855
856 const float recurrent_to_output_scale = recurrent_to_output_weights->quantization_info().uniform().scale * qoutput_state_in.scale / lstm_params.output_intermediate_scale();
857 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, output_state_in, &recurrent_weights_transposed, &eff_bias_info, recurrent_to_output_scale, &mm_out_info, &output_outstage_info));
858
859 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE));
860 if(lstm_params.has_peephole_opt())
861 {
862 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lstm_params.cell_to_output_weights(), 1, DataType::QSYMM16);
863 // TODO(COMPMID-3395): Perform multiplication in the quantized domain in NEPixelWiseMultiplicationKernel
864 // Here we are not using the output stage because all operations are done in float
865 // const float cell_to_output_scale = std::pow(2, cell_shift) * lstm_params.cell_to_output_weights()->quantization_info().uniform().scale / lstm_params.output_intermediate_scale();
866 // ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(cell_to_output_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift));
867 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(cell_state_out, lstm_params.cell_to_output_weights(), &output_outstage_info, 1.f, ConvertPolicy::SATURATE,
868 RoundingPolicy::TO_ZERO));
869 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_outstage_info, &output_outstage_info, &output_outstage_info, ConvertPolicy::SATURATE));
870 }
871
872 if(has_layer_norm)
873 {
874 const ITensorInfo *w_info = lstm_params.output_layer_norm_weights();
875 const ITensorInfo *b_info = output_gate_bias;
876 ARM_COMPUTE_RETURN_ON_ERROR(validate_layer_norm(output_outstage_info, *w_info, *b_info));
877 }
878
879 const TensorInfo output_gate_info(TensorShape(num_units, batch_size), 1, DataType::QSYMM16, sigmoid_tanh_outqinfo);
880 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_outstage_info, &output_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
881
882 // Hidden.
883 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(cell_state_out, &input_gate_info, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH, 1.f, 1.f)));
884 const TensorInfo hidden_mul_res(TensorShape(num_units, batch_size), 1, DataType::S32);
885 const TensorInfo hidden_out_info(TensorShape(num_units, batch_size), 1, DataType::QASYMM8_SIGNED);
886
887 ARM_COMPUTE_RETURN_ERROR_ON(lstm_params.hidden_state_scale() == 0);
888 ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplication::validate(&output_gate_info, &input_gate_info, &hidden_mul_res, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO));
889 const float hidden_state_scale = std::pow(2, -15) / lstm_params.hidden_state_scale() * std::pow(2, -15);
890 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(hidden_state_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift, /* ignore_epsilon */ true));
891 gemmlowp_info.gemmlowp_offset = lstm_params.hidden_state_zero();
892 gemmlowp_info.output_data_type = hidden_out_info.data_type();
893 ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOutputStage::validate(&hidden_mul_res, nullptr, &hidden_out_info, gemmlowp_info));
894
895 const bool projection_tensor_copy_required = num_units != output_size;
896
897 // Projection.
898 if(lstm_params.has_projection())
899 {
900 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(recurrent_to_forget_weights, lstm_params.projection_weights());
901 ARM_COMPUTE_RETURN_ERROR_ON(qoutput_state_in.scale == 0);
902
903 const UniformQuantizationInfo qprojection = lstm_params.projection_weights()->quantization_info().uniform();
904 const float projection_scale = qprojection.scale * lstm_params.hidden_state_scale() / qoutput_state_in.scale;
905 ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(projection_scale, &gemmlowp_info.gemmlowp_multiplier, &gemmlowp_info.gemmlowp_shift));
906 gemmlowp_info.gemmlowp_offset = qoutput_state_in.offset;
907 gemmlowp_info.gemmlowp_min_bound = std::numeric_limits<int8_t>::lowest();
908 gemmlowp_info.gemmlowp_max_bound = std::numeric_limits<int8_t>::max();
909 gemmlowp_info.output_data_type = DataType::QASYMM8_SIGNED;
910
911 const TensorInfo projection_outstage_info(*output_state_out);
912 const TensorInfo projection_weights_transposed(TensorShape(output_size, num_units), 1, lstm_params.projection_weights()->data_type(), lstm_params.projection_weights()->quantization_info());
913
914 TensorInfo projection_mm_out_info{ mm_out_info };
915 projection_mm_out_info.set_tensor_shape(TensorShape(output_size, batch_size));
916
917 ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemmlowp_info, &hidden_out_info, &projection_weights_transposed, &projection_eff_bias_info, projection_scale, &projection_mm_out_info,
918 &projection_outstage_info));
919
920 if(projection_tensor_copy_required)
921 {
922 ARM_COMPUTE_RETURN_ON_ERROR(CLQLSTMLayer::TensorCopyKernel::validate(*output_state_in, projection_outstage_info));
923 }
924
925 ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(output_state_out, output_state_out, output_state_out, ConvertPolicy::SATURATE));
926
927 if(projection_tensor_copy_required)
928 {
929 ARM_COMPUTE_RETURN_ON_ERROR(CLQLSTMLayer::TensorCopyKernel::validate(projection_outstage_info, *output_state_out));
930 }
931
932 int8_t quantized_projection_clip{ 0 };
933 if(lstm_params.projection_clip() > 0.0f)
934 {
935 quantized_projection_clip = quantize_qasymm8_signed(lstm_params.projection_clip(), qprojection);
936 }
937
938 if(quantized_projection_clip > 0)
939 {
940 ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -quantized_projection_clip,
941 quantized_projection_clip)));
942 }
943 }
944 else
945 {
946 if(projection_tensor_copy_required)
947 {
948 ARM_COMPUTE_RETURN_ON_ERROR(CLQLSTMLayer::TensorCopyKernel::validate(hidden_out_info, *output_state_out));
949 }
950 }
951
952 if(cell_state_out->total_size() > 0)
953 {
954 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(cell_state_in, cell_state_out);
955 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(cell_state_in, cell_state_out);
956 }
957
958 if(output_state_out->total_size() > 0)
959 {
960 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output_state_out);
961 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output_state_in, output_state_out);
962 }
963
964 ARM_COMPUTE_RETURN_ON_ERROR(CLCopy::validate(output_state_out, output));
965 return Status{};
966 }
967
run()968 void CLQLSTMLayer::run()
969 {
970 prepare();
971
972 // Acquire all the temporaries
973 MemoryGroupResourceScope scope_mg(_memory_group);
974
975 // Forget gate.
976 _mm_input_to_forget.run();
977 _input_to_forget_outstage.run();
978
979 _mm_recurrent_to_forget.run();
980 _recurrent_to_forget_outstage.run();
981 _accumulate_input_recurrent_forget.run();
982
983 if(_has_peephole)
984 {
985 _pixelwise_mul_cell_to_forget.run();
986 _cell_to_forget_outstage.run();
987 _accumulate_cell_forget.run();
988 }
989
990 if(_has_layer_norm)
991 {
992 CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Forget));
993 }
994
995 _forget_gate_sigmoid.run();
996
997 // Modulation gate.
998 _mm_input_to_cell.run();
999 _input_to_cell_outstage.run();
1000
1001 _mm_recurrent_to_cell.run();
1002 _recurrent_to_cell_outstage.run();
1003 _accumulate_input_recurrent_modulation.run();
1004
1005 if(_has_layer_norm)
1006 {
1007 CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Cell));
1008 }
1009
1010 _cell_gate_tanh.run();
1011
1012 // Input gate
1013 if(_has_cifg)
1014 {
1015 _input_gate_sub.run();
1016 }
1017 else
1018 {
1019 _mm_input_to_input.run();
1020 _input_to_input_outstage.run();
1021 _mm_recurrent_to_input.run();
1022 _recurrent_to_input_outstage.run();
1023 _accumulate_input_recurrent_input.run();
1024
1025 if(_has_peephole)
1026 {
1027 _pixelwise_mul_cell_to_input.run();
1028 _cell_to_input_outstage.run();
1029 _accumulate_cell_input.run();
1030 }
1031
1032 if(_has_layer_norm)
1033 {
1034 CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Input));
1035 }
1036
1037 _input_gate_sigmoid.run();
1038 }
1039
1040 // Cell.
1041 _pixelwise_mul_forget_cell.run();
1042 _pixelwise_mul_input_cell.run();
1043 _add_forget_cell.run();
1044 if(_has_cell_clipping)
1045 {
1046 _cell_clip.run();
1047 }
1048
1049 // Output gate.
1050 _mm_input_to_output.run();
1051 _input_to_output_outstage.run();
1052 _mm_recurrent_to_output.run();
1053 _recurrent_to_output_outstage.run();
1054 _accumulate_input_recurrent_output.run();
1055 if(_has_peephole)
1056 {
1057 _pixelwise_mul_cell_to_output.run();
1058 _cell_to_output_outstage.run();
1059 _accumulate_cell_to_output.run();
1060 }
1061
1062 if(_has_layer_norm)
1063 {
1064 CLScheduler::get().enqueue(get_layer_norm(LayerNormGate::Output));
1065 }
1066
1067 _output_gate_sigmoid.run();
1068
1069 // Hidden.
1070 _hidden_tanh.run();
1071 _pixelwise_mul_hidden.run();
1072 _hidden_outstage.run();
1073
1074 // Projection.
1075 if(_has_projection)
1076 {
1077 _mm_projection.run();
1078 _projection_outstage.run();
1079
1080 if(_projection_tensor_copy_required)
1081 {
1082 _projection_output_to_accumulate_copy.run();
1083 }
1084
1085 _accumulate_projection.run();
1086
1087 if(_projection_tensor_copy_required)
1088 {
1089 _projection_accumulate_to_output_copy.run();
1090 }
1091
1092 if(_has_projection_clipping)
1093 {
1094 _projection_clip.run();
1095 }
1096 }
1097 else
1098 {
1099 if(_projection_tensor_copy_required)
1100 {
1101 _hidden_to_output_copy.run();
1102 }
1103 }
1104
1105 // Copy output_state_out to output
1106 _copy_output.run();
1107 }
1108
prepare()1109 void CLQLSTMLayer::prepare()
1110 {
1111 if(!_is_prepared)
1112 {
1113 // Pre-transpose weights to be used in GEMM.
1114 _input_to_forget_weights_transposed.allocator()->allocate();
1115 _input_to_cell_weights_transposed.allocator()->allocate();
1116 _input_to_output_weights_transposed.allocator()->allocate();
1117 _recurrent_to_forget_weights_transposed.allocator()->allocate();
1118 _recurrent_to_cell_weights_transposed.allocator()->allocate();
1119 _recurrent_to_output_weights_transposed.allocator()->allocate();
1120 _transpose_input_to_forget_weights.run();
1121 _transpose_input_to_cell_weights.run();
1122 _transpose_input_to_output_weights.run();
1123 _transpose_recurrent_to_forget_weights.run();
1124 _transpose_recurrent_to_cell_weights.run();
1125 _transpose_recurrent_to_output_weights.run();
1126
1127 // Precompute effective biases
1128 if(_has_cifg)
1129 {
1130 _ones.map(true);
1131 std::fill_n(reinterpret_cast<int16_t *>(_ones.buffer()), _ones.info()->total_size() / _ones.info()->element_size(), 32767);
1132 _ones.unmap();
1133 }
1134 else
1135 {
1136 _input_to_input_eff_bias.allocator()->allocate();
1137 _recurrent_to_input_eff_bias.allocator()->allocate();
1138
1139 ITensorPack input_to_input_red_pack = { { ACL_SRC, _input_to_input_weights }, { ACL_DST, &_input_to_input_eff_bias } };
1140 CLScheduler::get().enqueue_op(*_input_to_input_reduction, input_to_input_red_pack, false);
1141
1142 ITensorPack rec_to_input_red_pack = { { ACL_SRC, _recurrent_to_input_weights }, { ACL_DST, &_recurrent_to_input_eff_bias } };
1143 CLScheduler::get().enqueue_op(*_recurrent_to_input_reduction, rec_to_input_red_pack, false);
1144
1145 _input_to_input_weights_transposed.allocator()->allocate();
1146 _recurrent_to_input_weights_transposed.allocator()->allocate();
1147 _transpose_input_to_input_weights.run();
1148 _transpose_recurrent_to_input_weights.run();
1149 _input_to_input_weights->mark_as_unused();
1150 _recurrent_to_input_weights->mark_as_unused();
1151 }
1152 _input_to_forget_eff_bias.allocator()->allocate();
1153 _recurrent_to_forget_eff_bias.allocator()->allocate();
1154 _input_to_cell_eff_bias.allocator()->allocate();
1155 _recurrent_to_cell_eff_bias.allocator()->allocate();
1156 _input_to_output_eff_bias.allocator()->allocate();
1157 _recurrent_to_output_eff_bias.allocator()->allocate();
1158
1159 ITensorPack input_to_forget_red_pack = { { ACL_SRC, _input_to_forget_weights }, { ACL_DST, &_input_to_forget_eff_bias } };
1160 CLScheduler::get().enqueue_op(*_input_to_forget_reduction, input_to_forget_red_pack, false);
1161
1162 ITensorPack rec_to_forget_red_pack = { { ACL_SRC, _recurrent_to_forget_weights }, { ACL_DST, &_recurrent_to_forget_eff_bias } };
1163 CLScheduler::get().enqueue_op(*_recurrent_to_forget_reduction, rec_to_forget_red_pack, false);
1164
1165 ITensorPack input_to_cell_red_pack = { { ACL_SRC, _input_to_cell_weights }, { ACL_DST, &_input_to_cell_eff_bias } };
1166 CLScheduler::get().enqueue_op(*_input_to_cell_reduction, input_to_cell_red_pack, false);
1167
1168 ITensorPack rec_to_cell_red_pack = { { ACL_SRC, _recurrent_to_cell_weights }, { ACL_DST, &_recurrent_to_cell_eff_bias } };
1169 CLScheduler::get().enqueue_op(*_recurrent_to_cell_reduction, rec_to_cell_red_pack, false);
1170
1171 ITensorPack input_to_output_red_pack = { { ACL_SRC, _input_to_output_weights }, { ACL_DST, &_input_to_output_eff_bias } };
1172 CLScheduler::get().enqueue_op(*_input_to_output_reduction, input_to_output_red_pack, false);
1173
1174 ITensorPack rec_to_output_red_pack = { { ACL_SRC, _recurrent_to_output_weights }, { ACL_DST, &_recurrent_to_output_eff_bias } };
1175 CLScheduler::get().enqueue_op(*_recurrent_to_output_reduction, rec_to_output_red_pack, false);
1176
1177 if(_has_projection)
1178 {
1179 _projection_eff_bias.allocator()->allocate();
1180 ITensorPack proj_red_pack{ { ACL_SRC, _projection_weights }, { ACL_DST, &_projection_eff_bias } };
1181 CLScheduler::get().enqueue_op(*_projection_reduction, proj_red_pack, false);
1182 if(_projection_bias != nullptr)
1183 {
1184 _projection_bias_add.run();
1185 _projection_bias->mark_as_unused();
1186 }
1187
1188 _projection_weights_transposed.allocator()->allocate();
1189 _transpose_projection_weights.run();
1190 _projection_weights->mark_as_unused();
1191
1192 if(!_projection_tensor_copy_required)
1193 {
1194 _hidden_gate.mark_as_unused();
1195 _projection_accumulate_res.mark_as_unused();
1196 }
1197 }
1198
1199 // Mark weights as unused
1200 _input_to_forget_weights->mark_as_unused();
1201 _input_to_cell_weights->mark_as_unused();
1202 _input_to_output_weights->mark_as_unused();
1203 _recurrent_to_forget_weights->mark_as_unused();
1204 _recurrent_to_cell_weights->mark_as_unused();
1205 _recurrent_to_output_weights->mark_as_unused();
1206
1207 CLScheduler::get().queue().finish();
1208 _is_prepared = true;
1209 }
1210 }
1211
1212 } // namespace arm_compute
1213