Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 1 | /* |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 2 | * Copyright (c) 2019-2020 Arm Limited. |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +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 | */ |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 24 | #include "arm_compute/core/CL/OpenCL.h" |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 25 | #include "arm_compute/core/Types.h" |
| 26 | #include "arm_compute/runtime/CL/CLHelpers.h" |
| 27 | #include "arm_compute/runtime/CL/CLScheduler.h" |
Inki Dae | ea2ce17 | 2020-04-09 10:01:44 +0900 | [diff] [blame] | 28 | #include "arm_compute/runtime/CL/Utils.h" |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 29 | #include "arm_compute/runtime/CL/functions/CLPermute.h" |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 30 | #include "utils/Utils.h" |
| 31 | |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 32 | using namespace arm_compute; |
| 33 | using namespace utils; |
| 34 | |
| 35 | namespace |
| 36 | { |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 37 | } // namespace |
| 38 | |
| 39 | class CLCacheExample : public Example |
| 40 | { |
| 41 | public: |
| 42 | CLCacheExample() = default; |
| 43 | |
| 44 | bool do_setup(int argc, char **argv) override |
| 45 | { |
| 46 | std::cout << "Once the program has run and created the file cache.bin, rerun with --restore_cache." << std::endl; |
| 47 | CLScheduler::get().default_init(); |
Gian Marco Iodice | f3622be | 2019-07-29 14:27:16 +0100 | [diff] [blame] | 48 | |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 49 | if(argc > 1) |
| 50 | { |
| 51 | std::string argv1 = argv[1]; |
| 52 | std::transform(argv1.begin(), argv1.end(), argv1.begin(), ::tolower); |
| 53 | if(argv1 == "--restore_cache") |
| 54 | { |
| 55 | // Load the precompiled kernels from a file into the kernel library, in this way the next time they are needed |
| 56 | // compilation won't be required. |
| 57 | restore_program_cache_from_file(); |
| 58 | } |
| 59 | else |
| 60 | { |
| 61 | std::cout << "Unkown option " << argv1 << std::endl; |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | // Initialise shapes |
| 66 | init_tensor(TensorShape(8U, 4U, 2U), tensor_nchw, DataType::U8, DataLayout::NCHW); |
| 67 | init_tensor(TensorShape(2U, 8U, 4U), tensor_nhwc, DataType::U8, DataLayout::NHWC); |
| 68 | init_tensor(TensorShape(8U, 4U, 2U), tensor_nchw_result, DataType::U8, DataLayout::NCHW); |
| 69 | |
| 70 | // Create the permutation vector to turn a NCHW tensor to NHWC. |
| 71 | // The input tensor is NCHW, which means that the fastest changing coordinate is W=8U. |
| 72 | // For permutation vectors the fastest changing coordinate is the one on the left too. |
| 73 | // Each element in the permutation vector specifies a mapping from the source tensor to the destination one, thus if we |
| 74 | // use 2U in the permutation vector's first element we are telling the function to move the channels to the fastest |
| 75 | // changing coordinate in the destination tensor. |
| 76 | |
| 77 | const PermutationVector vector_nchw_to_nhwc(2U, 0U, 1U); |
| 78 | permute_nhwc.configure(&tensor_nchw, &tensor_nhwc, vector_nchw_to_nhwc); |
| 79 | |
| 80 | // Allocate and fill tensors |
| 81 | tensor_nhwc.allocator()->allocate(); |
| 82 | tensor_nchw.allocator()->allocate(); |
| 83 | fill_tensor(tensor_nchw); |
| 84 | |
| 85 | // Demostrate autoconfigure for the output tensor |
| 86 | const PermutationVector vector_nhwc_to_nchw(1U, 2U, 0U); |
| 87 | permute_nchw.configure(&tensor_nhwc, &tensor_nchw_result, vector_nhwc_to_nchw); |
| 88 | tensor_nchw_result.allocator()->allocate(); |
| 89 | |
Pablo Tello | f6f23ea | 2019-07-05 14:00:30 +0100 | [diff] [blame] | 90 | // Save the opencl kernels to a file |
| 91 | save_program_cache_to_file(); |
| 92 | |
| 93 | return true; |
| 94 | } |
| 95 | void do_run() override |
| 96 | { |
| 97 | permute_nhwc.run(); |
| 98 | permute_nchw.run(); |
| 99 | } |
| 100 | void do_teardown() override |
| 101 | { |
| 102 | } |
| 103 | |
| 104 | private: |
| 105 | void validate_result(CLTensor &reference, CLTensor &result) |
| 106 | { |
| 107 | reference.map(); |
| 108 | result.map(); |
| 109 | Window window; |
| 110 | window.use_tensor_dimensions(reference.info()->tensor_shape()); |
| 111 | Iterator it_ref(&reference, window); |
| 112 | Iterator it_res(&result, window); |
| 113 | execute_window_loop(window, [&](const Coordinates &) |
| 114 | { |
| 115 | assert(*reinterpret_cast<unsigned char *>(it_ref.ptr()) == *reinterpret_cast<unsigned char *>(it_res.ptr())); |
| 116 | }, |
| 117 | it_ref, it_res); |
| 118 | reference.unmap(); |
| 119 | result.unmap(); |
| 120 | } |
| 121 | |
| 122 | void fill_tensor(CLTensor &tensor) |
| 123 | { |
| 124 | tensor.map(); |
| 125 | Window window; |
| 126 | window.use_tensor_dimensions(tensor.info()->tensor_shape()); |
| 127 | Iterator it_tensor(&tensor, window); |
| 128 | unsigned char val(0); |
| 129 | execute_window_loop(window, [&](const Coordinates &) |
| 130 | { |
| 131 | *reinterpret_cast<unsigned char *>(it_tensor.ptr()) = val++; |
| 132 | }, |
| 133 | it_tensor); |
| 134 | tensor.unmap(); |
| 135 | } |
| 136 | void init_tensor(const TensorShape shape, CLTensor &tensor, DataType type, DataLayout layout) |
| 137 | { |
| 138 | tensor.allocator()->init(TensorInfo(shape, 1, type).set_data_layout(layout)); |
| 139 | } |
| 140 | |
| 141 | CLTensor tensor_nchw{}; |
| 142 | CLTensor tensor_nhwc{}; |
| 143 | CLTensor tensor_nchw_result{}; |
| 144 | CLPermute permute_nhwc{}; |
| 145 | CLPermute permute_nchw{}; |
| 146 | }; |
| 147 | |
| 148 | /** Main program creating an example that demostrates how to load precompiled kernels from a file. |
| 149 | * |
| 150 | * @param[in] argc Number of arguments |
| 151 | * @param[in] argv Arguments |
| 152 | */ |
| 153 | int main(int argc, char **argv) |
| 154 | { |
| 155 | return utils::run_example<CLCacheExample>(argc, argv); |
| 156 | } |