• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2020 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 #ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
25 #define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
26 
27 #include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
28 
29 namespace arm_compute
30 {
31 class ICLTensor;
32 
33 /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
34  */
35 class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
36 {
37 public:
38     /** Default constructor */
39     CLDepthwiseConvolutionLayer3x3NCHWKernel();
40     /** Initialize the function's source, destination, conv and border_size.
41      *
42      * @param[in]  input              Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
43      * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
44      *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
45      * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
46      *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
47      * @param[out] output             Destination tensor. Data type supported: Same as @p input.
48      * @param[in]  conv_info          Padding and stride information to use for the convolution.
49      * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
50      * @param[in]  act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
51      * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
52      * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
53      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
54      * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
55      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
56      */
57     void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
58                    unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
59                    const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
60     /** Initialize the function's source, destination, conv and border_size.
61      *
62      * @param[in]  compile_context    The compile context to be used.
63      * @param[in]  input              Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
64      * @param[in]  weights            Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
65      *                                Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
66      * @param[in]  biases             Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
67      *                                Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
68      * @param[out] output             Destination tensor. Data type supported: Same as @p input.
69      * @param[in]  conv_info          Padding and stride information to use for the convolution.
70      * @param[in]  depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
71      * @param[in]  act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported.
72      * @param[in]  dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
73      * @param[in]  output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
74      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
75      * @param[in]  output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
76      *                                the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
77      */
78     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
79                    unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
80                    const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
81     /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel
82      *
83      * @param[in] input              Source tensor info. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
84      * @param[in] weights            Weights tensor info. A 3D tensor with dimensions [3, 3, IFM].
85      *                               Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
86      * @param[in] biases             Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
87      *                               Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
88      * @param[in] output             Destination tensor. Data type supported: Same as @p input.
89      * @param[in] conv_info          Padding and stride information to use for the convolution.
90      * @param[in] depth_multiplier   (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
91      * @param[in] act_info           (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
92      * @param[in] gpu_target         (Optional) GPU target to validate the kernel for. Defaults to midgard.
93      * @param[in] dilation           (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
94      * @param[in] output_multipliers (Optional) Output multipliers tensor info for quantized computations. In case of per-channel quantization,
95      *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
96      * @param[in] output_shifts      (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
97      *                               the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
98      *
99      * @return a status
100      */
101     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
102                            unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD,
103                            const Size2D &dilation = Size2D(1U, 1U), const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
104 
105     void run(const Window &window, cl::CommandQueue &queue) override;
106     BorderSize border_size() const override;
107 
108 private:
109     unsigned int _conv_stride_x;
110     unsigned int _conv_pad_top;
111     unsigned int _conv_pad_left;
112 };
113 } // namespace arm_compute
114 #endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H */
115