1 /*
2 * Copyright (c) 2020-2021 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
25 #include "QLSTMLayerNormalization.h"
26 #include "ArithmeticOperations.h"
27 #include "MeanStdDevNormalizationLayer.h"
28 #include "PixelWiseMultiplication.h"
29 #include "arm_compute/core/utils/misc/Utility.h"
30 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
31
32 namespace arm_compute
33 {
34 namespace test
35 {
36 namespace validation
37 {
38 namespace reference
39 {
qlstm_layer_normalization(const SimpleTensor<int16_t> & src,const SimpleTensor<int16_t> & weight,const SimpleTensor<int32_t> & bias)40 SimpleTensor<int16_t> qlstm_layer_normalization(const SimpleTensor<int16_t> &src, const SimpleTensor<int16_t> &weight, const SimpleTensor<int32_t> &bias)
41 {
42 ARM_COMPUTE_ERROR_ON(src.shape().num_dimensions() > 2);
43 SimpleTensor<int16_t> output{ src.shape(), DataType::QSYMM16 };
44
45 const auto wq_info = weight.quantization_info().uniform();
46 int output_multiplier{};
47 int output_shift{};
48 const auto s = quantization::calculate_quantized_multiplier(wq_info.scale, &output_multiplier, &output_shift);
49 output_shift *= -1;
50
51 if(!bool(s))
52 {
53 output_multiplier = 0;
54 output_shift = 0;
55 }
56
57 const uint32_t num_batch = src.shape()[1];
58 const uint32_t num_input = src.shape()[0];
59
60 for(uint32_t batch_idx = 0; batch_idx < num_batch; ++batch_idx)
61 {
62 int64_t sum{};
63 int64_t sum_sq{};
64
65 for(uint32_t input_idx = 0; input_idx < num_input; ++input_idx)
66 {
67 const auto index = batch_idx * num_input + input_idx;
68 const auto val = static_cast<int32_t>(src[index]);
69 sum += val;
70 sum_sq += val * val;
71 }
72
73 const auto temp = static_cast<int64_t>(0x100000) / num_input;
74 const auto mean = sum * 1024 / static_cast<int64_t>(num_input);
75 const auto variance = ((sum_sq * temp) - (mean * mean)) / 0x100000;
76
77 int32_t stddev_invsqrt_mul{};
78 int32_t stddev_invsqrt_shift{};
79 quantization::get_invsqrt_quantized_multiplier_exp(variance, -1, stddev_invsqrt_mul, stddev_invsqrt_shift);
80
81 for(uint32_t input_idx = 0; input_idx < num_input; ++input_idx)
82 {
83 const auto index = batch_idx * num_input + input_idx;
84 const auto val = static_cast<int32_t>(src[index]);
85 const auto shifted = (val << 10) - mean;
86 const auto rescaled = quantization::multiply_by_quantized_multiplier(shifted, stddev_invsqrt_mul, stddev_invsqrt_shift);
87 const int64_t weighted = rescaled * weight[input_idx] + bias[input_idx];
88 const auto reverse_shifted = static_cast<int32_t>((weighted + 512) >> 10);
89 auto out_val = quantization::multiply_by_quantized_multiplier(reverse_shifted, output_multiplier, output_shift + 12);
90 out_val = arm_compute::utility::clamp<decltype(out_val), int16_t>(out_val, std::numeric_limits<int16_t>::min());
91 output[index] = static_cast<int16_t>(out_val);
92 }
93 }
94 return output;
95 }
96 } // namespace reference
97 } // namespace validation
98 } // namespace test
99 } // namespace arm_compute
100