Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [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/runtime/NEON/functions/NEWinogradLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/Utils.h" |
| 27 | #include "arm_compute/core/Validate.h" |
| 28 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 29 | #include "support/ToolchainSupport.h" |
| 30 | |
| 31 | namespace |
| 32 | { |
| 33 | inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) |
| 34 | { |
| 35 | const int in_width = input->info()->dimension(0); |
| 36 | const int in_height = input->info()->dimension(1); |
| 37 | const int in_batches = input->info()->dimension(3); |
| 38 | const int in_channels = input->info()->dimension(2); |
| 39 | return Tensor4DShape({ in_batches, in_height, in_width, in_channels }); |
| 40 | } |
| 41 | } /* namespace */ |
| 42 | |
| 43 | namespace arm_compute |
| 44 | { |
| 45 | NEWinogradLayer::NEWinogradLayer(std::shared_ptr<IMemoryManager> memory_manager) |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 46 | : _memory_group(std::move(memory_manager)), _winograd_kernel(), _permute_input(), _permute_weights(), _permute_output(), _workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), |
| 47 | _weights_hwio(), _input(), _weights(), _output(), _reshaped_kernel(false), _conv() |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 48 | { |
| 49 | } /* arm_compute */ |
| 50 | |
| 51 | void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info) |
| 52 | { |
| 53 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 54 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
| 55 | ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(1) != 3 || weights->info()->dimension(0) != 3, "Only 3x3 kernels are supported"); |
| 56 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 57 | |
| 58 | if(biases != nullptr) |
| 59 | { |
| 60 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 61 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 62 | } |
| 63 | |
| 64 | _weights = weights; |
| 65 | _input = input; |
| 66 | _output = output; |
| 67 | |
| 68 | // Get parameters from conv_info |
| 69 | unsigned int stride_x = 0; |
| 70 | unsigned int stride_y = 0; |
| 71 | std::tie(stride_x, stride_y) = conv_info.stride(); |
| 72 | ARM_COMPUTE_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); |
| 73 | |
| 74 | // Get convolved dimensions |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 75 | auto padding = PADDING_VALID; |
| 76 | const int in_channels = input->info()->dimension(2); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 77 | const int out_channels = output->info()->dimension(2); |
| 78 | const int weights_width = weights->info()->dimension(0); |
| 79 | const int weights_height = weights->info()->dimension(1); |
| 80 | |
| 81 | const KernelShape kernel_shape({ out_channels, weights_height, weights_width, in_channels }); |
| 82 | const Tensor4DShape in_shape(internal_get_input_shape(input)); |
| 83 | |
| 84 | // Get the memory required to instantiate a new Winograd operator. |
| 85 | constexpr size_t kstore_alignment = 64; |
Pablo Tello | 3d4968a | 2017-12-04 15:03:35 +0000 | [diff] [blame] | 86 | const size_t kernel_storage_per_thread = NEWinogradLayerKernel::get_kernel_storage_size(kernel_shape); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 87 | _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_per_thread + kstore_alignment - 1) }, 1, DataType::U8)); |
| 88 | _memory_group.manage(&_kernel_storage); |
| 89 | |
| 90 | // Get workbench size and allocate memory |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 91 | |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 92 | constexpr size_t wspace_alignment = 64; |
Pablo Tello | 3d4968a | 2017-12-04 15:03:35 +0000 | [diff] [blame] | 93 | const size_t ws_size = NEWinogradLayerKernel::get_working_space_size(in_shape, kernel_shape, padding); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 94 | _workspace.allocator()->init(TensorInfo(TensorShape{ (ws_size + wspace_alignment - 1) }, 1, DataType::U8)); |
| 95 | _memory_group.manage(&_workspace); |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 96 | _memory_group.manage(&_input_nhwc); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 97 | _kernel_storage.allocator()->allocate(); |
| 98 | _workspace.allocator()->allocate(); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 99 | |
| 100 | // Create Winograd operator object |
| 101 | _conv = support::cpp14::make_unique<Winograd3x3F32>(kernel_shape, in_shape, padding, _kernel_storage.buffer()); |
| 102 | |
| 103 | // Configure the kernel, padding not needed so it's safe to call configure after allocare |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 104 | _winograd_kernel.configure(_conv.get()); |
| 105 | |
| 106 | // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map] |
| 107 | switch(weights->info()->num_dimensions()) |
| 108 | { |
| 109 | case 3: |
| 110 | { |
| 111 | _permute_weights.configure(weights, &_weights_hwio, PermutationVector(2U, 0U, 1U)); |
| 112 | break; |
| 113 | } |
| 114 | case 4: |
| 115 | { |
| 116 | _permute_weights.configure(weights, &_weights_hwio, PermutationVector(3U, 2U, 0U, 1U)); |
| 117 | break; |
| 118 | } |
| 119 | default: |
| 120 | { |
| 121 | ARM_COMPUTE_ERROR("Not supported."); |
| 122 | break; |
| 123 | } |
| 124 | } |
| 125 | // configure the kernel to transform the input tensor from NCHW -> NHWC |
| 126 | _permute_input.configure(input, &_input_nhwc, PermutationVector(2U, 0U, 1U)); |
| 127 | |
| 128 | _weights_hwio.allocator()->allocate(); |
| 129 | _input_nhwc.allocator()->allocate(); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 130 | } |
| 131 | |
| 132 | void NEWinogradLayer::run() |
| 133 | { |
| 134 | #if defined(__aarch64__) |
| 135 | _memory_group.acquire(); |
| 136 | if(!_reshaped_kernel) |
| 137 | { |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 138 | _reshaped_kernel = true; |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 139 | _permute_weights.run(); |
| 140 | _conv->transform_weights(reinterpret_cast<const float *>(_weights_hwio.buffer()), nullptr); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 141 | } |
| 142 | const Tensor4DShape in_shape(internal_get_input_shape(_input)); |
| 143 | auto padding = PADDING_VALID; |
| 144 | |
| 145 | //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 146 | _permute_input.run(); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 147 | |
| 148 | //Get ptrs into the workspace |
Pablo Tello | 3d4968a | 2017-12-04 15:03:35 +0000 | [diff] [blame] | 149 | std::pair<void *, void *> nhwc_ptrs = _conv->get_nhwc_ptrs(in_shape, padding, _workspace.buffer()); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 150 | |
| 151 | //Setup matrices ptrs and transfor the input tensor to the appropriate form before running GEMM. |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 152 | _conv->reshape_input(in_shape, padding, reinterpret_cast<float *>(_input_nhwc.buffer()), _workspace.buffer()); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 153 | |
| 154 | //Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 155 | NEScheduler::get().schedule(&_winograd_kernel, Window::DimX); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 156 | |
| 157 | //Transform the output to the appropriate form |
| 158 | _conv->reshape_output(in_shape, padding, nhwc_ptrs.first); |
| 159 | |
Pablo Tello | 02541fb | 2017-12-15 09:48:59 +0000 | [diff] [blame^] | 160 | const unsigned int out_width = _output->info()->dimension(0); |
| 161 | const unsigned int out_height = _output->info()->dimension(1); |
| 162 | const unsigned int out_channels = _output->info()->dimension(2); |
| 163 | const unsigned int out_batches = _output->info()->dimension(3); |
| 164 | |
| 165 | // We create a temporary tensor with the results in the workspace so that the we can run a function to reorder from NHWC -> NCHW |
| 166 | Tensor output_nhwc; |
| 167 | TensorInfo info(TensorShape(out_channels, out_width, out_height, out_batches), 1, _output->info()->data_type()); |
| 168 | output_nhwc.allocator()->init(info); |
| 169 | output_nhwc.allocator()->import_memory(Memory(static_cast<uint8_t *>(nhwc_ptrs.first))); |
| 170 | |
| 171 | // Reorder the convoluted output to ACL's ordering NCHW |
| 172 | _permute_output.configure(&output_nhwc, _output, PermutationVector(1U, 2U, 0U)); |
| 173 | _permute_output.run(); |
Pablo Tello | 8951933 | 2017-11-17 11:52:36 +0000 | [diff] [blame] | 174 | |
| 175 | _memory_group.release(); |
| 176 | #else /* __aarch64__ */ |
| 177 | ARM_COMPUTE_UNUSED(_winograd_kernel); |
| 178 | ARM_COMPUTE_UNUSED(_workspace); |
| 179 | ARM_COMPUTE_UNUSED(_kernel_storage); |
| 180 | ARM_COMPUTE_UNUSED(_input); |
| 181 | ARM_COMPUTE_UNUSED(_weights); |
| 182 | ARM_COMPUTE_UNUSED(_output); |
| 183 | ARM_COMPUTE_UNUSED(_reshaped_kernel); |
| 184 | ARM_COMPUTE_UNUSED(_conv); |
| 185 | ARM_COMPUTE_ERROR("Winograd only supported for aarch64, recompile with arch=arm64-v8a."); |
| 186 | #endif /* __aarch64__ */ |
| 187 | } |
| 188 | } // namespace arm_compute |