Sang-Hoon Park | 1a531fa | 2020-03-26 13:57:57 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2020 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 "src/core/utils/quantization/AsymmHelpers.cpp" |
| 30 | |
Georgios Pinitas | afc630f | 2020-03-30 14:09:27 +0100 | [diff] [blame] | 31 | #include "support/ToolchainSupport.h" |
| 32 | |
Sang-Hoon Park | 1a531fa | 2020-03-26 13:57:57 +0000 | [diff] [blame] | 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace test |
| 36 | { |
| 37 | namespace validation |
| 38 | { |
| 39 | namespace reference |
| 40 | { |
| 41 | SimpleTensor<float> qlstm_layer_normalization_float_compute(SimpleTensor<float> src, SimpleTensor<float> weight, SimpleTensor<float> bias) |
| 42 | { |
| 43 | SimpleTensor<float> output = mean_std_normalization_layer(src); |
Michele Di Giorgio | 9428a18 | 2020-03-30 14:10:20 +0100 | [diff] [blame] | 44 | output = pixel_wise_multiplication<float, float, float>(output, weight, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO, DataType::F32); |
Sang-Hoon Park | 1a531fa | 2020-03-26 13:57:57 +0000 | [diff] [blame] | 45 | return arithmetic_operation(ArithmeticOperation::ADD, output, bias, DataType::F32, ConvertPolicy::SATURATE); |
| 46 | } |
| 47 | |
| 48 | SimpleTensor<int16_t> qlstm_layer_normalization(const SimpleTensor<int16_t> &src, const SimpleTensor<int16_t> &weight, const SimpleTensor<int32_t> &bias) |
| 49 | { |
| 50 | ARM_COMPUTE_ERROR_ON(src.shape().num_dimensions() > 2); |
| 51 | |
| 52 | SimpleTensor<float> converted_src{ src.shape(), DataType::F32 }; |
| 53 | SimpleTensor<float> converted_weight{ weight.shape(), DataType::F32 }; |
| 54 | SimpleTensor<float> converted_bias{ bias.shape(), DataType::F32 }; |
| 55 | |
| 56 | const auto iq_info = src.quantization_info().uniform(); |
| 57 | int output_multiplier{}; |
| 58 | int output_shift{}; |
| 59 | quantization::calculate_quantized_multiplier(iq_info.scale, &output_multiplier, &output_shift); |
| 60 | |
| 61 | const float layer_norm_scale = output_multiplier * std::pow(2, static_cast<double>(output_shift - 31)); |
| 62 | const float bias_scale = std::pow(2., -10) * layer_norm_scale; |
| 63 | |
| 64 | for(int i = 0; i < src.num_elements(); i++) |
| 65 | { |
| 66 | converted_src[i] = static_cast<float>(src[i]); |
| 67 | } |
| 68 | |
| 69 | for(int i = 0; i < bias.num_elements(); i++) |
| 70 | { |
| 71 | converted_bias[i] = static_cast<float>(bias[i]) * bias_scale; |
| 72 | } |
| 73 | |
| 74 | for(int i = 0; i < weight.num_elements(); i++) |
| 75 | { |
| 76 | converted_weight[i] = weight[i] * layer_norm_scale; |
| 77 | } |
| 78 | |
| 79 | SimpleTensor<float> output_float = qlstm_layer_normalization_float_compute(converted_src, converted_weight, converted_bias); |
| 80 | SimpleTensor<int16_t> output{ output_float.shape(), DataType::QSYMM16 }; |
| 81 | |
| 82 | for(int i = 0; i < output.num_elements(); i++) |
| 83 | { |
Georgios Pinitas | afc630f | 2020-03-30 14:09:27 +0100 | [diff] [blame] | 84 | const auto output_val_s32 = static_cast<int32_t>(support::cpp11::round(output_float[i] * std::pow(2, 12))); |
Sang-Hoon Park | 1a531fa | 2020-03-26 13:57:57 +0000 | [diff] [blame] | 85 | output[i] = utility::clamp<int32_t, int16_t>(output_val_s32, std::numeric_limits<int16_t>::min()); |
| 86 | } |
| 87 | |
| 88 | return output; |
| 89 | } |
| 90 | } // namespace reference |
| 91 | } // namespace validation |
| 92 | } // namespace test |
| 93 | } // namespace arm_compute |