Manuel Bottini | 9c9b70b | 2019-07-01 17:35:56 +0100 | [diff] [blame] | 1 | /* |
Giorgio Arena | 1856ff7 | 2020-02-07 13:46:45 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 ARM Limited. |
Manuel Bottini | 9c9b70b | 2019-07-01 17:35:56 +0100 | [diff] [blame] | 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/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/Window.h" |
| 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 35 | |
| 36 | #include "support/ToolchainSupport.h" |
| 37 | |
| 38 | using namespace arm_compute; |
| 39 | |
| 40 | namespace arm_compute |
| 41 | { |
| 42 | namespace |
| 43 | { |
| 44 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, |
| 45 | int min, int max) |
| 46 | { |
| 47 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); |
Giorgio Arena | 1856ff7 | 2020-02-07 13:46:45 +0000 | [diff] [blame] | 48 | ARM_COMPUTE_RETURN_ERROR_ON(min > max); |
Manuel Bottini | 9c9b70b | 2019-07-01 17:35:56 +0100 | [diff] [blame] | 49 | |
| 50 | // Check biases if exist |
| 51 | if(bias != nullptr) |
| 52 | { |
| 53 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); |
| 54 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); |
| 55 | ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0)); |
| 56 | } |
| 57 | |
| 58 | if(output->total_size() != 0) |
| 59 | { |
| 60 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QSYMM16); |
| 61 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input); |
| 62 | } |
| 63 | |
| 64 | return Status{}; |
| 65 | } |
| 66 | |
| 67 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output) |
| 68 | { |
| 69 | constexpr unsigned int num_elems_processed_per_iteration = 4; |
| 70 | |
| 71 | // Output auto inizialitation if not yet initialized |
| 72 | auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QSYMM16)); |
| 73 | |
| 74 | // Configure kernel window |
| 75 | Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); |
| 76 | |
| 77 | AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); |
| 78 | |
| 79 | bool window_changed = update_window_and_padding(win, input_access); |
| 80 | |
| 81 | if(output->total_size() != 0) |
| 82 | { |
| 83 | Window win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); |
| 84 | AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration); |
| 85 | window_changed = window_changed || update_window_and_padding(win_out, output_result_access); |
| 86 | |
| 87 | output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); |
| 88 | } |
| 89 | |
| 90 | if(bias != nullptr) |
| 91 | { |
| 92 | AccessWindowStatic bias_access(bias, 0, 0, ceil_to_multiple(bias->dimension(0), num_elems_processed_per_iteration), bias->tensor_shape()[1]); |
| 93 | window_changed = window_changed || update_window_and_padding(win, bias_access); |
| 94 | } |
| 95 | |
| 96 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 97 | return std::make_pair(err, win); |
| 98 | } |
| 99 | } // namespace |
| 100 | |
| 101 | class Coordinates; |
| 102 | } // namespace arm_compute |
| 103 | |
| 104 | CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() |
| 105 | : _input(nullptr), _bias(nullptr), _output(nullptr) |
| 106 | { |
| 107 | } |
| 108 | |
| 109 | Status CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, |
| 110 | int min, int max) |
| 111 | { |
| 112 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 113 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max)); |
| 114 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), |
| 115 | (bias != nullptr) ? bias->clone().get() : nullptr, |
| 116 | output->clone().get()) |
| 117 | .first); |
| 118 | |
| 119 | return Status{}; |
| 120 | } |
| 121 | |
| 122 | void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, |
| 123 | int result_fixedpoint_multiplier, int result_shift, |
| 124 | int min, int max) |
| 125 | { |
| 126 | // Perform validate step |
| 127 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 128 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), |
| 129 | min, max)); |
| 130 | |
| 131 | _input = input; |
| 132 | _bias = bias; |
| 133 | _output = output; |
| 134 | |
| 135 | // Set the arguments to pass at compile time |
| 136 | CLBuildOptions build_opts; |
| 137 | build_opts.add_option("-DRESULT_FIXEDPOINT_MULTIPLIER=" + support::cpp11::to_string(result_fixedpoint_multiplier)); |
| 138 | build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift)); |
Giorgio Arena | 1856ff7 | 2020-02-07 13:46:45 +0000 | [diff] [blame] | 139 | build_opts.add_option_if((min > -32768), "-DMIN_BOUND=" + support::cpp11::to_string(min)); |
| 140 | build_opts.add_option_if((max < 32767), "-DMAX_BOUND=" + support::cpp11::to_string(max)); |
Manuel Bottini | 9c9b70b | 2019-07-01 17:35:56 +0100 | [diff] [blame] | 141 | build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); |
| 142 | |
| 143 | // Create kernel |
| 144 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down_fixedpoint_qsymm16", build_opts.options())); |
| 145 | |
| 146 | // Configure kernel window |
| 147 | auto win_config = validate_and_configure_window(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info()); |
| 148 | ARM_COMPUTE_ERROR_THROW_ON(win_config.first); |
| 149 | ICLKernel::configure_internal(win_config.second); |
| 150 | } |
| 151 | |
| 152 | void CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window, cl::CommandQueue &queue) |
| 153 | { |
| 154 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 155 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 156 | |
| 157 | // Create input window |
| 158 | Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); |
| 159 | Window slice = collapsed.first_slice_window_3D(); |
| 160 | |
| 161 | // Setup bias slice |
| 162 | unsigned int idx1 = num_arguments_per_3D_tensor(); |
| 163 | if(_bias != nullptr) |
| 164 | { |
| 165 | Window biases_slice(slice); |
| 166 | biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 167 | biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 168 | add_1D_tensor_argument(idx1, _bias, biases_slice); |
| 169 | } |
| 170 | |
| 171 | do |
| 172 | { |
| 173 | unsigned int idx = 0; |
| 174 | add_3D_tensor_argument(idx, _input, slice); |
| 175 | add_3D_tensor_argument(idx1, _output, slice); |
Georgios Pinitas | 275f99c | 2019-08-23 12:44:11 +0100 | [diff] [blame] | 176 | enqueue(queue, *this, slice, lws_hint()); |
Manuel Bottini | 9c9b70b | 2019-07-01 17:35:56 +0100 | [diff] [blame] | 177 | } |
| 178 | while(collapsed.slide_window_slice_3D(slice)); |
| 179 | } |