1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
/*
* Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
#define __ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__
#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h"
#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h"
#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
#include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
#include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
#include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
#include "arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/runtime/CL/CLTensor.h"
#include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
#include "arm_compute/runtime/CL/functions/CLPermute.h"
#include "arm_compute/runtime/IFunction.h"
#include "arm_compute/runtime/MemoryGroup.h"
namespace arm_compute
{
class ICLTensor;
/** Basic function to execute a depthwise convolution for kernel size 3x3xC (when data layout NCHW) or Cx3x3 (when data layout NHWC). This function calls the following OpenCL kernels:
*
* -# @ref CLDepthwiseConvolutionLayer3x3NCHWKernel (if data_layout == NCHW)
* -# @ref CLDepthwiseConvolutionLayer3x3NHWCKernel (if data_layout == NHWC)
* -# @ref CLDepthwiseConvolutionLayerReshapeWeightsKernel (if data_layout == NHWC)
* -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
*
*/
class CLDepthwiseConvolutionLayer3x3 : public IFunction
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayer3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayer3x3(const CLDepthwiseConvolutionLayer3x3 &) = delete;
/** Default move constructor */
CLDepthwiseConvolutionLayer3x3(CLDepthwiseConvolutionLayer3x3 &&) = default;
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayer3x3 &operator=(const CLDepthwiseConvolutionLayer3x3 &) = delete;
/** Default move assignment operator */
CLDepthwiseConvolutionLayer3x3 &operator=(CLDepthwiseConvolutionLayer3x3 &&) = default;
/** Initialize the function's source, destination, conv and border_size.
*
* @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input.
* @param[out] output Destination tensor. Data type supported: same as @p input.
* @param[in] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
* @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*/
void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
/** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3
*
* @param[in] input Source tensor info. Data type supported: QASYMM8 for all layouts, F16/F32 for NCHW.
* @param[in] weights Weights tensor info. A 3D tensor with shape [3, 3, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input, S32 when input is QASYMM8.
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @param[in] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
* @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 QASYMM8 supported.
* @param[in] gpu_target (Optional) GPU target to validate the kernel for. Defaults to midgard.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*
* @return a status
*/
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1,
ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U));
// Inherited methods overriden:
void run() override;
void prepare() override;
private:
MemoryGroup _memory_group;
std::unique_ptr<ICLDepthwiseConvolutionLayer3x3Kernel> _kernel;
CLFillBorderKernel _border_handler;
CLPermute _permute_input_to_nchw;
CLPermute _permute_weights_to_nchw;
CLPermute _permute_output_to_nhwc;
CLDepthwiseConvolutionLayerReshapeWeightsKernel _reshape_weights;
CLTensor _permuted_input;
CLTensor _permuted_weights;
CLTensor _permuted_output;
const ITensor *_original_weights;
bool _needs_permute;
bool _needs_weights_reshape;
bool _is_prepared;
};
/** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
*
* -# @ref CLDepthwiseIm2ColKernel
* -# @ref CLGEMMMatrixVectorMultiplyKernel
* -# @ref CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel
* -# @ref CLFillBorderKernel (if pad_x or pad_y > 0)
*
*/
class CLDepthwiseConvolutionLayer : public IFunction
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayer();
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete;
/** Default move constructor */
CLDepthwiseConvolutionLayer(CLDepthwiseConvolutionLayer &&) = default;
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayer &operator=(const CLDepthwiseConvolutionLayer &) = delete;
/** Default move assignment operator */
CLDepthwiseConvolutionLayer &operator=(CLDepthwiseConvolutionLayer &&) = default;
/** Initialize the function's source, destination, weights and convolution information.
*
* @param[in, out] input Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input, S32 when input is QASYMM8.
* @param[out] output Destination tensor. Data type supported: same as @p input.
* @param[in] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
* @param[in] act_info (Optional) Activation layer information in case of a fused activation.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*/
void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
/** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer
*
* @param[in] input Source tensor info. Data type supported: QASYMM8/F32.
* @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
* @param[in] biases Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
* Data type supported: Same as @p input, S32 when input is QASYMM8.
* @param[in] output Destination tensor. Data type supported: same as @p input.
* @param[in] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
* @param[in] act_info (Optional) Activation layer information in case of a fused activation.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*
* @return a status
*/
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
// Inherited methods overriden:
void run() override;
void prepare() override;
private:
CLDepthwiseIm2ColKernel _im2col_kernel;
CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel _weights_reshape_kernel;
CLGEMMMatrixVectorMultiplyKernel _v2mm_kernel;
CLDepthwiseVectorToTensorKernel _vector_to_tensor_kernel;
CLDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
CLActivationLayer _activationlayer_function;
CLFillBorderKernel _v2mm_input_fill_border;
CLFillBorderKernel _v2mm_weights_fill_border;
CLTensor _input_reshaped;
CLTensor _weights_reshaped;
CLTensor _v2mm_output;
CLTensor _output_reshaped;
bool _is_prepared;
bool _is_quantized;
bool _is_activationlayer_enabled;
const ICLTensor *_original_weights;
std::unique_ptr<IFunction> _optimised_function;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__ */
|
