diff options
| author | SiCong Li <sicong.li@arm.com> | 2022-03-21 15:34:21 +0000 |
|---|---|---|
| committer | SiCong Li <sicong.li@arm.com> | 2022-04-25 10:50:35 +0000 |
| commit | 0a3948394e7e77344201b8732e9c20fcb5fa9a38 (patch) | |
| tree | d80a559557f8ed6d573d9156bf1517d9ea85c3d0 | |
| parent | f55cca5f17da72004108f92047d3177b7cdb1a76 (diff) | |
| download | ComputeLibrary-0a3948394e7e77344201b8732e9c20fcb5fa9a38.tar.gz | |
Document data layout of weight tensors in convolution layers
Resolves COMPMID-5187
Signed-off-by: SiCong Li <sicong.li@arm.com>
Change-Id: I4fddd1f1e7134896a40f62553d705fa5e411e00b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7405
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
| -rw-r--r-- | docs/user_guide/data_layout.dox | 28 |
1 files changed, 25 insertions, 3 deletions
diff --git a/docs/user_guide/data_layout.dox b/docs/user_guide/data_layout.dox index ae69bbf457..711b85f08c 100644 --- a/docs/user_guide/data_layout.dox +++ b/docs/user_guide/data_layout.dox @@ -1,5 +1,5 @@ /// -/// Copyright (c) 2021 Arm Limited. +/// Copyright (c) 2021-2022 Arm Limited. /// /// SPDX-License-Identifier: MIT /// @@ -29,8 +29,7 @@ namespace arm_compute @section data_layout_support_supported_data_layout Supported Data Layouts -Compute Library supports the following data layouts and -the right-most letter represents the fastest changing dimension: +With regard to convolution layers, Compute Library supports the following data layouts for input and output tensors: - NHWC: The native layout of Compute Library that delivers the best performance where channels are in the fastest changing dimension - NCHW: Legacy layout where width is in the fastest changing dimension @@ -38,5 +37,28 @@ the right-most letter represents the fastest changing dimension: , where N = batch, C = channel, H = height, W = width, D = depth. +Note: The right-most letter represents the fastest changing dimension, which is the "lower dimension". +The corresponding @ref TensorShape for each of the data layout would be initialized as: + +- NHWC: TensorShape(C, W, H, N) +- NCHW: TensorShape(W, H, C, N) +- NDHWC: TensorShape(C, W, H, D, N) + +For 2d Conv, the weight / filter tensors are arranged in 4 dimensions: Height (H), Width (W), Input channel (I), Output channel (O) +For 3d Conv, the additional Depth dimension means exactly the same as the Depth in the input / output layout. + +The layout of weight tensors change with that of the input / output tensors, and the dimensions can be mapped as: + +- Weight Height -> Height +- Weight Width -> Width +- Weight Input channel -> Channel +- Weight Output channel -> Batch + +Therefore, the corresponding weight layouts for each input / output layout are: + +- (input/output tensor) NHWC: (weight tensor) OHWI +- (input/output tensor) NCHW: (weight tensor) OIHW +- (input/output tensor) NDHWC: (weight tensor) ODHWI + */ } // namespace |