Michele Di Giorgio | 32982d8 | 2017-07-07 14:44:43 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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/core/NEON/kernels/NEDequantizationLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/TensorInfo.h" |
| 29 | #include "arm_compute/core/Utils.h" |
| 30 | #include "arm_compute/core/Validate.h" |
| 31 | #include "arm_compute/core/Window.h" |
| 32 | |
| 33 | #include <arm_neon.h> |
| 34 | |
| 35 | using namespace arm_compute; |
| 36 | |
| 37 | NEDequantizationLayerKernel::NEDequantizationLayerKernel() |
| 38 | : _input(nullptr), _output(nullptr), _min(nullptr), _max(nullptr) |
| 39 | { |
| 40 | } |
| 41 | |
| 42 | void NEDequantizationLayerKernel::configure(const ITensor *input, ITensor *output, const float *min, const float *max) |
| 43 | { |
| 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 45 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
| 46 | ARM_COMPUTE_ERROR_ON_NULLPTR(min); |
| 47 | ARM_COMPUTE_ERROR_ON_NULLPTR(max); |
| 48 | |
| 49 | // Output tensor auto initialization if not yet initialized |
| 50 | auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, DataType::F32, 0); |
| 51 | |
| 52 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); |
| 53 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 54 | |
| 55 | _input = input; |
| 56 | _output = output; |
| 57 | _min = min; |
| 58 | _max = max; |
| 59 | |
| 60 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 61 | |
| 62 | // Configure window |
| 63 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 64 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 65 | update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, num_elems_processed_per_iteration), output_access); |
| 66 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 67 | |
| 68 | INEKernel::configure(win); |
| 69 | } |
| 70 | |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 71 | void NEDequantizationLayerKernel::run(const Window &window, const ThreadInfo &info) |
Michele Di Giorgio | 32982d8 | 2017-07-07 14:44:43 +0100 | [diff] [blame] | 72 | { |
Moritz Pflanzer | c186b57 | 2017-09-07 09:48:04 +0100 | [diff] [blame] | 73 | ARM_COMPUTE_UNUSED(info); |
Michele Di Giorgio | 32982d8 | 2017-07-07 14:44:43 +0100 | [diff] [blame] | 74 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 75 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 76 | |
| 77 | Iterator input(_input, window); |
| 78 | Iterator output(_output, window); |
| 79 | |
| 80 | const float32x4_t vmin = vdupq_n_f32(*_min); |
| 81 | const float range = *_max - *_min; |
| 82 | const float32x4_t scaling = vdupq_n_f32(range / 255.0f); |
| 83 | |
| 84 | // Uniformly map values to range 8bit integers, i.e. [min, max] -> [0, 255] |
| 85 | execute_window_loop(window, [&](const Coordinates & id) |
| 86 | { |
| 87 | const uint8x8_t val_u8 = vld1_u8(reinterpret_cast<uint8_t *>(input.ptr())); |
| 88 | const uint16x8_t val_u16 = vmovl_u8(val_u8); |
| 89 | const uint32x4_t val_u32_low = vmovl_u16(vget_low_u16(val_u16)); |
| 90 | const uint32x4_t val_u32_high = vmovl_u16(vget_high_u16(val_u16)); |
| 91 | float32x4_t val_low = vcvtq_f32_u32(val_u32_low); |
| 92 | float32x4_t val_high = vcvtq_f32_u32(val_u32_high); |
| 93 | |
| 94 | // Dequantize -> (q / 255.0 * range) + min |
| 95 | val_low = vmulq_f32(val_low, scaling); |
| 96 | val_high = vmulq_f32(val_high, scaling); |
| 97 | val_low = vaddq_f32(val_low, vmin); |
| 98 | val_high = vaddq_f32(val_high, vmin); |
| 99 | |
| 100 | const float32x4x2_t dequantized = vuzpq_f32(val_low, val_high); |
| 101 | vst2q_f32(reinterpret_cast<float *>(output.ptr()), dequantized); |
| 102 | }, |
| 103 | input, output); |
| 104 | } |