Sheri Zhang | d813bab | 2021-04-30 16:53:41 +0100 | [diff] [blame] | 1 | /// |
SiCong Li | 0a39483 | 2022-03-21 15:34:21 +0000 | [diff] [blame] | 2 | /// Copyright (c) 2021-2022 Arm Limited. |
Sheri Zhang | d813bab | 2021-04-30 16:53:41 +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 | |
| 25 | namespace 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 | |
SiCong Li | 0a39483 | 2022-03-21 15:34:21 +0000 | [diff] [blame] | 32 | With regard to convolution layers, Compute Library supports the following data layouts for input and output tensors: |
Sheri Zhang | d813bab | 2021-04-30 16:53:41 +0100 | [diff] [blame] | 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 |
Sheri Zhang | 5dda217 | 2021-10-15 19:54:17 +0100 | [diff] [blame] | 36 | - NDHWC: New data layout for supporting 3D operators |
Sheri Zhang | d813bab | 2021-04-30 16:53:41 +0100 | [diff] [blame] | 37 | |
Sheri Zhang | 5dda217 | 2021-10-15 19:54:17 +0100 | [diff] [blame] | 38 | , where N = batch, C = channel, H = height, W = width, D = depth. |
Sheri Zhang | d813bab | 2021-04-30 16:53:41 +0100 | [diff] [blame] | 39 | |
SiCong Li | 0a39483 | 2022-03-21 15:34:21 +0000 | [diff] [blame] | 40 | Note: The right-most letter represents the fastest changing dimension, which is the "lower dimension". |
| 41 | The 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 | |
| 47 | For 2d Conv, the weight / filter tensors are arranged in 4 dimensions: Height (H), Width (W), Input channel (I), Output channel (O) |
| 48 | For 3d Conv, the additional Depth dimension means exactly the same as the Depth in the input / output layout. |
| 49 | |
| 50 | The 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 | |
| 57 | Therefore, 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 | |
Sheri Zhang | d813bab | 2021-04-30 16:53:41 +0100 | [diff] [blame] | 63 | */ |
| 64 | } // namespace |