aboutsummaryrefslogtreecommitdiff
path: root/ConversionUtils_1_3.hpp
blob: 4714b84ba4586520505f2069fa1383722d56079c (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
//
// Copyright © 2020 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#pragma once

#include "ConversionUtils_1_2.hpp"

using Half = half_float::half;

namespace armnn_driver
{

using namespace armnn;
using namespace android::nn;

template<typename HalPolicy,
         typename HalOperation = typename HalPolicy::Operation,
         typename HalModel     = typename HalPolicy::Model>
bool ConvertElu(const HalOperation& operation, const HalModel& model, ConversionData& data)
{
    using HalOperandType = typename HalPolicy::OperandType;

    LayerInputHandle input0 = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data);
    if (!input0.IsValid())
    {
        return Fail("%s: Operation has invalid inputs", __func__);
    }

    // Determine data type of input tensor
    HalOperandType inputType;
    if (!GetOperandType<HalPolicy>(operation, 0, model, inputType))
    {
        return Fail("%s: Operation has invalid inputs", __func__);
    }

    ActivationDescriptor desc;
    desc.m_Function = ActivationFunction::Elu;

    // Read alpha
    if (inputType == HalOperandType::TENSOR_FLOAT16)
    {
        Half alpha;

        if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT16, alpha, model, data))
        {
            return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__);
        }

        desc.m_A = static_cast<float>(alpha);
    }
    else if (inputType == HalOperandType::TENSOR_FLOAT32)
    {
        if (!GetInputScalar<HalPolicy>(operation, 1, HalOperandType::FLOAT32, desc.m_A, model, data))
        {
            return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__);
        }
    }
    else
    {
        return Fail("%s: Unsupported input tensor type: %d", __func__, inputType);
    }

    return ::ConvertToActivation<HalPolicy>(operation, __func__, desc, model, data);
}

template<typename HalPolicy,
         typename HalOperation = typename HalPolicy::Operation,
         typename HalModel     = typename HalPolicy::Model>
bool ConvertQuantizedLstm(const HalOperation& operation, const HalModel& model, ConversionData& data)
{
    using HalOperand     = typename HalPolicy::Operand;
    using HalOperandType = typename HalPolicy::OperandType;

    ALOGV("HalPolicy::ConvertQuantizedLstm()");

    //Inputs:
    // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize]
    //    specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128.
    LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data);
    if (!input.IsValid())
    {
        return Fail("%s: Could not read input 0: input", __func__);
    }

    // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, of shape [batch_size, output_size].
    LayerInputHandle outputStatePrevTimeStep = ConvertToLayerInputHandle<HalPolicy>(operation, 18, model, data);
    if (!outputStatePrevTimeStep.IsValid())
    {
        return Fail("%s: Could not read input 18: outputStatePrevTimeStep", __func__);
    }

    // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape [batch_size, num_units].
    LayerInputHandle cellStatePrevTimeStep = ConvertToLayerInputHandle<HalPolicy>(operation, 19, model, data);
    if (!cellStatePrevTimeStep.IsValid())
    {
        return Fail("%s: Could not read input 19: cellStatePrevTimeStep", __func__);
    }

    // Get the mandatory input tensors:

    // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, input_size].
    const ConstTensorPin inputToForgetWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 2, model, data);

    // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    // [num_units, input_size].
    const ConstTensorPin inputToCellWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 3, model, data);

    // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, input_size].
    const ConstTensorPin inputToOutputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 4, model, data);

    // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, output_size].
    const ConstTensorPin recurrentToForgetWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 6, model, data);

    // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, output_size].
    const ConstTensorPin recurrentToCellWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 7, model, data);

    // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, output_size].
    const ConstTensorPin recurrentToOutputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 8, model, data);

    // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units].
    const ConstTensorPin forgetGateBiasPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 13, model, data);

    // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units].
    const ConstTensorPin cellBiasPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 14, model, data);

    // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units].
    const ConstTensorPin outputGateBiasPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation, 15, model, data);

    if (!inputToForgetWeightsPin.IsValid() ||
        !inputToCellWeightsPin.IsValid() ||
        !inputToOutputWeightsPin.IsValid() ||
        !recurrentToForgetWeightsPin.IsValid() ||
        !recurrentToCellWeightsPin.IsValid() ||
        !recurrentToOutputWeightsPin.IsValid() ||
        !forgetGateBiasPin.IsValid() ||
        !cellBiasPin.IsValid() ||
        !outputGateBiasPin.IsValid())
    {
        return Fail("%s: Operation has invalid tensor inputs", __func__);
    }

    // Get the optional input tensors:

    // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, input_size], where “num_units” corresponds to the number of cell units.
    const ConstTensorPin inputToInputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         1,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e.,
    //     “num_units”), or the second dimension of the “projection_weights”, if defined.
    const ConstTensorPin recurrentToInputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         5,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape
    // [num_units].
    const ConstTensorPin cellToInputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         9,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape
    // [num_units].
    const ConstTensorPin cellToForgetWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         10,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape
    // [num_units].
    const ConstTensorPin cellToOutputWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         11,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units].
    const ConstTensorPin inputGateBiasPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         12,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape
    //     [output_size, num_units].
    const ConstTensorPin projectionWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         16,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [output_size].
    const ConstTensorPin projectionBiasPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         17,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional())
        || (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional())
        || (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional())
        || (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional())
        || (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional())
        || (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional())
        || (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional())
        || (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional()))
    {
        return Fail("%s: Operation has invalid tensor inputs", __func__);
    }


    // Get the optional normalization tensors

    // 20: The input layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM.
    //     Used to rescale normalized inputs to activation at input gate.
    const ConstTensorPin inputLayerNormWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         20,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 21: The forget layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM
    //     Used to rescale normalized inputs to activation at forget gate.
    const ConstTensorPin forgetLayerNormWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         21,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 22: The cell layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM.
    //     Used to rescale normalized inputs to activation at cell gate.
    const ConstTensorPin cellLayerNormWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         22,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    // 23: The output layer normalization weights. A 1-D tensor of shape [num_units].
    //     Used to rescale normalized inputs to activation at output gate.
    const ConstTensorPin outputLayerNormWeightsPin =
        ConvertOperationInputToConstTensorPin<HalPolicy>(operation,
                                                         23,
                                                         model,
                                                         data,
                                                         g_DontPermute,
                                                         nullptr,
                                                         true);

    if ((!inputLayerNormWeightsPin.IsValid() && !inputLayerNormWeightsPin.IsOptional())
        || (!forgetLayerNormWeightsPin.IsValid() && !forgetLayerNormWeightsPin.IsOptional())
        || (!cellLayerNormWeightsPin.IsValid() && !cellLayerNormWeightsPin.IsOptional())
        || (!outputLayerNormWeightsPin.IsValid() && !outputLayerNormWeightsPin.IsOptional()))
    {
        return Fail("%s: Operation has invalid tensor inputs", __func__);
    }

    // Get the optional input scalars:
    // 24: The cell clip:  If provided the cell state is clipped by this value prior to the cell output activation.
    // 25: The projection clip: If provided and projection is enabled, this is used for clipping the projected values.

    // Get the mandatory input scalars:
    // 26: The scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate.
    // 27: The scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate.
    // 28: The scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate.
    // 29: The scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate.
    // 30: The zero point of the hidden state, i.e. input to projection.
    // 31: The scale of the hidden state, i.e. input to projection.
    float cellClip, projClip, matMulInputGate, matMulForgetGate, matMulCellGate, matMulOutputGate, projInputScale;
    int projInputZeroPoint;

    if (!GetInputScalar<HalPolicy>(operation, 24, HalOperandType::FLOAT32, cellClip, model, data, true) ||
        !GetInputScalar<HalPolicy>(operation, 25, HalOperandType::FLOAT32, projClip, model, data, true) ||
        !GetInputScalar<HalPolicy>(operation, 26, HalOperandType::FLOAT32, matMulInputGate, model, data) ||
        !GetInputScalar<HalPolicy>(operation, 27, HalOperandType::FLOAT32, matMulForgetGate, model, data) ||
        !GetInputScalar<HalPolicy>(operation, 28, HalOperandType::FLOAT32, matMulCellGate, model, data) ||
        !GetInputScalar<HalPolicy>(operation, 29, HalOperandType::FLOAT32, matMulOutputGate, model, data) ||
        !GetInputScalar<HalPolicy>(operation, 30, HalOperandType::FLOAT32, projInputScale, model, data) ||
        !GetInputScalar<HalPolicy>(operation, 31, HalOperandType::FLOAT32, projInputZeroPoint, model, data))
    {
        return Fail("%s: Operation has invalid scalar inputs", __func__);
    }

    // Outputs:
    // 0: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, of shape [batch_size,
    // output_size].
    const HalOperand* outputStateOut = GetOutputOperand<HalPolicy>(operation, 0, model);
    if (!outputStateOut)
    {
        return Fail("%s: Could not read output 0: outputStateOut", __func__);
    }

    // 1: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape [batch_size, num_units].
    const HalOperand* cellStateOut = GetOutputOperand<HalPolicy>(operation, 1, model);
    if (!cellStateOut)
    {
        return Fail("%s: Could not read output 1: cellStateOut", __func__);
    }

    // 2: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, of shape [batch_size, output_size].
    // This is effectively the same as the current “output state (out)” value.
    const HalOperand* output = GetOutputOperand<HalPolicy>(operation, 2, model);
    if (!output)
    {
        return Fail("%s: Could not read output 2: output", __func__);
    }

    // set the params structure for the AddLstmLayer call
    LstmInputParams params;
    params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr();
    params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr();
    params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr();
    params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr();
    params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr();
    params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr();
    params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr();
    params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr();
    params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr();
    params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr();
    params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr();
    params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr();
    params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr();
    params.m_CellBias = cellBiasPin.GetConstTensorPtr();
    params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr();
    params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr();
    params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr();
    params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr();
    params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr();
    params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr();
    params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr();

    // set the layer descriptor
    QLstmDescriptor desc;
    desc.m_CellClip = cellClip;
    desc.m_ProjectionClip = projClip;
    desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr ||
                          params.m_RecurrentToInputWeights == nullptr ||
                          params.m_InputGateBias == nullptr);
    desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr ||
                              params.m_CellToOutputWeights != nullptr);
    desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr);
    desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr ||
                               params.m_ForgetLayerNormWeights != nullptr ||
                               params.m_CellLayerNormWeights != nullptr ||
                               params.m_OutputLayerNormWeights != nullptr);
    desc.m_InputIntermediateScale = matMulInputGate;
    desc.m_ForgetIntermediateScale = matMulForgetGate;
    desc.m_CellIntermediateScale = matMulCellGate;
    desc.m_OutputIntermediateScale = matMulOutputGate;
    desc.m_HiddenStateScale = projInputScale;
    desc.m_HiddenStateZeroPoint = projInputZeroPoint;

    // validate the optional input groups
    if (desc.m_CifgEnabled &&
        (params.m_InputToInputWeights != nullptr ||
         params.m_RecurrentToInputWeights != nullptr ||
         params.m_InputGateBias != nullptr))
    {
        return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights,"
                    " and input gate bias must be provided", __func__);
    }

    if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr)
    {
        return Fail("%s: projection bias should not be provided without projection weights", __func__);
    }

    if (desc.m_PeepholeEnabled &&
        (params.m_CellToForgetWeights == nullptr ||
         params.m_CellToOutputWeights == nullptr ||
         (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr)))
    {
        return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided"
                    " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__);
    }

    if (desc.m_LayerNormEnabled &&
        (params.m_ForgetLayerNormWeights == nullptr ||
         params.m_CellLayerNormWeights == nullptr ||
         params.m_OutputLayerNormWeights == nullptr ||
         (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr)))
    {
        return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be"
                    " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__);
    }


    // Basic parameters
    LstmInputParamsInfo paramsInfo;
    paramsInfo.m_InputToForgetWeights     = &(params.m_InputToForgetWeights->GetInfo());
    paramsInfo.m_InputToCellWeights       = &(params.m_InputToCellWeights->GetInfo());
    paramsInfo.m_InputToOutputWeights     = &(params.m_InputToOutputWeights->GetInfo());
    paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo());
    paramsInfo.m_RecurrentToCellWeights   = &(params.m_RecurrentToCellWeights->GetInfo());
    paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo());
    paramsInfo.m_ForgetGateBias           = &(params.m_ForgetGateBias->GetInfo());
    paramsInfo.m_CellBias                 = &(params.m_CellBias->GetInfo());
    paramsInfo.m_OutputGateBias           = &(params.m_OutputGateBias->GetInfo());

    // Inputs
    const TensorInfo& inputInfo = input.GetTensorInfo();
    const TensorInfo& outputStatePrevTimeStepInfo = outputStatePrevTimeStep.GetTensorInfo();
    const TensorInfo& cellStatePrevTimeStepInfo = cellStatePrevTimeStep.GetTensorInfo();

    // Outputs
    TensorInfo outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut);
    TensorInfo outputInfo = GetTensorInfoForOperand(*output);
    const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut);

    // Optional parameters
    if (!desc.m_CifgEnabled)
    {
        paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo());
        paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo());
        if (desc.m_PeepholeEnabled)
        {
            paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo());
        }
        paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo());
    }


    if (desc.m_ProjectionEnabled)
    {
        paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo());
        if (params.m_ProjectionBias != nullptr)
        {
            paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo());
        }
    }
    else
    {
        // If Projection is disabled, override non-const outputs to change the quant info with hidden params, then
        // create a new const TensorInfo based on this
        outputStateOutInfo.SetQuantizationScale(projInputScale);
        outputStateOutInfo.SetQuantizationOffset(projInputZeroPoint);
        outputInfo.SetQuantizationScale(projInputScale);
        outputInfo.SetQuantizationOffset(projInputZeroPoint);
    }

    const TensorInfo constOutputStateOutInfo(outputStateOutInfo);
    const TensorInfo constOutputInfo(outputInfo);

    if (desc.m_PeepholeEnabled)
    {
        paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo());
        paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo());
    }

    if (desc.m_LayerNormEnabled)
    {
        if(!desc.m_CifgEnabled)
        {
            paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo());
        }
        paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo());
        paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo());
        paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo());
    }

    // Check if the layer is supported

    if (IsDynamicTensor(constOutputStateOutInfo) ||
        IsDynamicTensor(cellStateOutInfo)   ||
        IsDynamicTensor(constOutputInfo))
    {
        return Fail("%s: Dynamic output tensors are not supported %d %d %d %d", __func__,
                    IsDynamicTensor(constOutputStateOutInfo), IsDynamicTensor(cellStateOutInfo),
                    IsDynamicTensor(constOutputInfo));
    }

    bool isSupported = false;
    FORWARD_LAYER_SUPPORT_FUNC(__func__,
                               IsQLstmSupported,
                               data.m_Backends,
                               isSupported,
                               inputInfo,
                               outputStatePrevTimeStepInfo,
                               cellStatePrevTimeStepInfo,
                               constOutputStateOutInfo,
                               cellStateOutInfo,
                               constOutputInfo,
                               desc,
                               paramsInfo);
    if (!isSupported)
    {
        return false;
    }

    // Add the layer
    IConnectableLayer* layer = data.m_Network->AddQLstmLayer(desc, params, "QLstm");

    input.Connect(layer->GetInputSlot(0));
    outputStatePrevTimeStep.Connect(layer->GetInputSlot(1));
    cellStatePrevTimeStep.Connect(layer->GetInputSlot(2));

    return ( SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, 0, model, data,
                                                     &constOutputStateOutInfo) &&
             SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 1, *layer, 1, model, data) &&
             SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 2, *layer, 2, model, data, &constOutputInfo));
}

} // armnn_driver namespace