Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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/CL/kernels/CLPixelWiseMultiplicationKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 28 | #include "arm_compute/core/CL/ICLTensor.h" |
| 29 | #include "arm_compute/core/CL/OpenCL.h" |
| 30 | #include "arm_compute/core/Error.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Validate.h" |
| 34 | #include "arm_compute/core/Window.h" |
| 35 | |
| 36 | #include <cmath> |
| 37 | #include <cstdlib> |
| 38 | #include <set> |
| 39 | #include <string> |
| 40 | |
| 41 | using namespace arm_compute; |
| 42 | |
| 43 | CLPixelWiseMultiplicationKernel::CLPixelWiseMultiplicationKernel() |
| 44 | : _input1(nullptr), _input2(nullptr), _output(nullptr) |
| 45 | { |
| 46 | } |
| 47 | |
| 48 | void CLPixelWiseMultiplicationKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float scale, |
| 49 | ConvertPolicy overflow_policy, RoundingPolicy rounding_policy) |
| 50 | { |
Georgios Pinitas | f0dea70 | 2017-07-03 18:17:28 +0100 | [diff] [blame^] | 51 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); |
| 52 | |
| 53 | // Auto initialize output if not initialized |
| 54 | { |
| 55 | set_shape_if_empty(*output->info(), input1->info()->tensor_shape()); |
| 56 | |
| 57 | if(input1->info()->data_type() == DataType::S16 || input2->info()->data_type() == DataType::S16) |
| 58 | { |
| 59 | set_format_if_unknown(*output->info(), Format::S16); |
| 60 | } |
| 61 | else if(input1->info()->data_type() == DataType::F32 || input2->info()->data_type() == DataType::F32) |
| 62 | { |
| 63 | set_format_if_unknown(*output->info(), Format::F32); |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input1, input2, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 68 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); |
| 69 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input2, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); |
| 70 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16, DataType::F16, DataType::F32); |
| 71 | ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_type() == DataType::U8 && (input1->info()->data_type() != DataType::U8 || input2->info()->data_type() != DataType::U8), |
| 72 | "Output can only be U8 if both inputs are U8"); |
| 73 | ARM_COMPUTE_ERROR_ON_MSG(scale < 0, "Scale cannot be negative. "); |
| 74 | |
| 75 | _input1 = input1; |
| 76 | _input2 = input2; |
| 77 | _output = output; |
| 78 | |
| 79 | int scale_int = -1; |
| 80 | // Extract sign, exponent and mantissa |
| 81 | int exponent = 0; |
| 82 | float normalized_mantissa = std::frexp(scale, &exponent); |
| 83 | // Use int scaling if factor is equal to 1/2^n for 0 <= n <= 15 |
| 84 | // frexp returns 0.5 as mantissa which means that the exponent will be in the range of -1 <= e <= 14 |
| 85 | // Moreover, it will be negative as we deal with 1/2^n |
| 86 | if((normalized_mantissa == 0.5f) && (-14 <= exponent) && (exponent <= 1)) |
| 87 | { |
| 88 | // Store the positive exponent. We know that we compute 1/2^n |
| 89 | // Additionally we need to subtract 1 to compensate that frexp used a mantissa of 0.5 |
| 90 | scale_int = std::abs(exponent - 1); |
| 91 | } |
| 92 | |
| 93 | std::string data_type; |
| 94 | std::string compute_type; |
| 95 | // Check if it has float inputs and output |
| 96 | if(is_data_type_float(input1->info()->data_type()) || is_data_type_float(input2->info()->data_type())) |
| 97 | { |
| 98 | scale_int = -1; |
| 99 | compute_type = (DataType::F32 == input1->info()->data_type() || DataType::F32 == input2->info()->data_type()) ? "float" : "half"; |
| 100 | data_type = "DATA_TYPE_FLOAT"; |
| 101 | } |
| 102 | else |
| 103 | { |
| 104 | compute_type = (DataType::S16 == input1->info()->data_type() || DataType::S16 == input2->info()->data_type()) ? "int" : "ushort"; |
| 105 | data_type = "DATA_TYPE_INT"; |
| 106 | } |
| 107 | |
| 108 | // Construct kernel name |
| 109 | std::string kernel_name = "pixelwise_mul"; |
| 110 | kernel_name += (scale_int >= 0) ? "_int" : "_float"; |
| 111 | |
| 112 | // Set kernel build options |
| 113 | std::set<std::string> build_opts; |
| 114 | build_opts.emplace((overflow_policy == ConvertPolicy::WRAP || is_data_type_float(output->info()->data_type())) ? "-DWRAP" : "-DSATURATE"); |
| 115 | build_opts.emplace((rounding_policy == RoundingPolicy::TO_ZERO) ? "-DROUND=_rtz" : "-DROUND=_rte"); |
| 116 | build_opts.emplace("-DDATA_TYPE_IN1=" + get_cl_type_from_data_type(input1->info()->data_type())); |
| 117 | build_opts.emplace("-DDATA_TYPE_IN2=" + get_cl_type_from_data_type(input2->info()->data_type())); |
| 118 | build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type())); |
| 119 | build_opts.emplace("-DDATA_TYPE_RES=" + compute_type); |
| 120 | build_opts.emplace("-D" + data_type); |
| 121 | |
| 122 | // Create kernel |
| 123 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts)); |
| 124 | |
| 125 | // Set scale argument |
| 126 | unsigned int idx = 3 * num_arguments_per_2D_tensor(); //Skip the inputs and output parameters |
| 127 | |
| 128 | if(scale_int >= 0) |
| 129 | { |
| 130 | _kernel.setArg(idx++, scale_int); |
| 131 | } |
| 132 | else |
| 133 | { |
| 134 | _kernel.setArg(idx++, scale); |
| 135 | } |
| 136 | |
| 137 | // Configure kernel window |
| 138 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 139 | |
| 140 | Window win = calculate_max_window(*input1->info(), Steps(num_elems_processed_per_iteration)); |
| 141 | |
| 142 | AccessWindowHorizontal input1_access(input1->info(), 0, num_elems_processed_per_iteration); |
| 143 | AccessWindowHorizontal input2_access(input2->info(), 0, num_elems_processed_per_iteration); |
| 144 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 145 | |
| 146 | update_window_and_padding(win, input1_access, input2_access, output_access); |
| 147 | |
| 148 | ValidRegion valid_region = intersect_valid_regions(input1->info()->valid_region(), |
| 149 | input2->info()->valid_region()); |
| 150 | output_access.set_valid_region(win, valid_region); |
| 151 | |
| 152 | ICLKernel::configure(win); |
| 153 | } |
| 154 | |
| 155 | void CLPixelWiseMultiplicationKernel::run(const Window &window, cl::CommandQueue &queue) |
| 156 | { |
| 157 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 158 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 159 | |
| 160 | Window slice = window.first_slice_window_2D(); |
| 161 | |
| 162 | do |
| 163 | { |
| 164 | unsigned int idx = 0; |
| 165 | add_2D_tensor_argument(idx, _input1, slice); |
| 166 | add_2D_tensor_argument(idx, _input2, slice); |
| 167 | add_2D_tensor_argument(idx, _output, slice); |
| 168 | enqueue(queue, *this, slice); |
| 169 | } |
| 170 | while(window.slide_window_slice_2D(slice)); |
| 171 | } |