1/// 2/// Copyright (c) 2021-2022 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 25namespace arm_compute 26{ 27/** 28@page data_layout_support Data Layout Support 29 30@section data_layout_support_supported_data_layout Supported Data Layouts 31 32With regard to convolution layers, Compute Library supports the following data layouts for input and output tensors: 33 34- NHWC: The native layout of Compute Library that delivers the best performance where channels are in the fastest changing dimension 35- NCHW: Legacy layout where width is in the fastest changing dimension 36- NDHWC: New data layout for supporting 3D operators 37 38, where N = batch, C = channel, H = height, W = width, D = depth. 39 40Note: The right-most letter represents the fastest changing dimension, which is the "lower dimension". 41The corresponding @ref TensorShape for each of the data layout would be initialized as: 42 43- NHWC: TensorShape(C, W, H, N) 44- NCHW: TensorShape(W, H, C, N) 45- NDHWC: TensorShape(C, W, H, D, N) 46 47For 2d Conv, the weight / filter tensors are arranged in 4 dimensions: Height (H), Width (W), Input channel (I), Output channel (O) 48For 3d Conv, the additional Depth dimension means exactly the same as the Depth in the input / output layout. 49 50The layout of weight tensors change with that of the input / output tensors, and the dimensions can be mapped as: 51 52- Weight Height -> Height 53- Weight Width -> Width 54- Weight Input channel -> Channel 55- Weight Output channel -> Batch 56 57Therefore, the corresponding weight layouts for each input / output layout are: 58 59- (input/output tensor) NHWC: (weight tensor) OHWI 60- (input/output tensor) NCHW: (weight tensor) OIHW 61- (input/output tensor) NDHWC: (weight tensor) ODHWI 62 63*/ 64} // namespace 65