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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
|
/*
* Copyright (c) 2017-2021 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/Types.h"
#include "arm_compute/runtime/CL/CLTensor.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 CLCompileContext;
class CLFillBorderKernel;
class CLDepthwiseConvolutionLayerNativeKernel;
class CLDepthwiseConvolutionLayer3x3NCHWKernel;
class CLDepthwiseConvolutionLayer3x3NHWCKernel;
class ICLTensor;
/** Function to execute a depthwise convolution
*/
class CLDepthwiseConvolutionLayer : public IFunction
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
/** 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;
/** Default destructor */
~CLDepthwiseConvolutionLayer();
/** Initialize the function's source, destination, weights and convolution information.
*
* @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP16/FP32. Data layout supported: NHWC, NCHW
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
* Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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,
ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
/** Initialize the function's source, destination, weights and convolution information.
*
* @param[in] compile_context The compile context to be used.
* @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP16/FP32. Data layout supported: NHWC, NCHW
* @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
* Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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(const CLCompileContext &compile_context, 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 CLDepthwiseConvolutionLayer
*
* @param[in] input Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/FP16/FP32. Data layout supported: NHWC, NCHW
* @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
* Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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] 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(), const Size2D &dilation = Size2D(1U, 1U));
// Inherited methods overriden:
void run() override;
void prepare() override;
private:
/** Static function to choose the best depthwise convolution function for @ref CLDepthwiseConvolutionLayer
*
* @param[in] input Source tensor info. Data type supported: QASYMM8/FP16/FP32. Data layout supported: NHWC, NCHW
* @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM].
* Data type supported: Same as @p input or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*
* @return a Depthwise Convolution Function
*/
static DepthwiseConvolutionFunction get_depthwiseconvolution_function(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(), const Size2D &dilation = Size2D(1U, 1U));
/** 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 CLFillBorderKernel (if pad_x or pad_y > 0)
*
*/
class CLDepthwiseConvolutionLayerInternal3x3 : public IFunction
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayerInternal3x3(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayerInternal3x3(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
/** Default move constructor */
CLDepthwiseConvolutionLayerInternal3x3(CLDepthwiseConvolutionLayerInternal3x3 &&) = default;
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayerInternal3x3 &operator=(const CLDepthwiseConvolutionLayerInternal3x3 &) = delete;
/** Default move assignment operator */
CLDepthwiseConvolutionLayerInternal3x3 &operator=(CLDepthwiseConvolutionLayerInternal3x3 &&) = 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 or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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));
/** Initialize the function's source, destination, conv and border_size.
*
* @param[in] compile_context The compile context to be used.
* @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 or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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(const CLCompileContext &compile_context, 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 or QASYMM8/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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] 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(), const Size2D &dilation = Size2D(1U, 1U));
// Inherited methods overriden:
void run() override;
void prepare() override;
void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
{
_memory_group = MemoryGroup(std::move(memory_manager));
};
private:
MemoryGroup _memory_group;
std::unique_ptr<CLDepthwiseConvolutionLayer3x3NCHWKernel> _kernel_nchw;
std::unique_ptr<CLDepthwiseConvolutionLayer3x3NHWCKernel> _kernel_nhwc;
std::unique_ptr<CLFillBorderKernel> _border_handler;
CLPermute _permute_input_to_nchw;
CLPermute _permute_weights_to_nchw;
CLPermute _permute_output_to_nhwc;
CLTensor _permuted_input;
CLTensor _permuted_weights;
CLTensor _permuted_output;
CLTensor _output_multipliers;
CLTensor _output_shifts;
const ITensor *_original_weights;
const ITensor *_input;
const ITensor *_output;
bool _needs_permute;
bool _is_prepared;
bool _is_quantized;
bool _is_nhwc;
};
/** Basic function to execute a generic depthwise convolution. This function calls the following OpenCL kernels:
*
* -# @ref CLDepthwiseConvolutionLayerNativeKernel
* -# @ref CLPermute (x 3) if the data layout is NCHW
*
*/
class CLDepthwiseConvolutionLayerGeneric : public IFunction
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayerGeneric(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayerGeneric(const CLDepthwiseConvolutionLayerGeneric &) = delete;
/** Default move constructor */
CLDepthwiseConvolutionLayerGeneric(CLDepthwiseConvolutionLayerGeneric &&) = default;
/** Prevent instances of this class from being copied (As this class contains pointers) */
CLDepthwiseConvolutionLayerGeneric &operator=(const CLDepthwiseConvolutionLayerGeneric &) = delete;
/** Default move assignment operator */
CLDepthwiseConvolutionLayerGeneric &operator=(CLDepthwiseConvolutionLayerGeneric &&) = default;
/** Initialize the function's source, destination, weights and convolution information.
*
* @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/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 or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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));
/** Initialize the function's source, destination, weights and convolution information.
*
* @param[in] compile_context The compile context to be used.
* @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/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 or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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(const CLCompileContext &compile_context, 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 CLDepthwiseConvolutionLayerGeneric
*
* @param[in] input Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/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 or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8.
* @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/QASYMM8_SIGNED.
* @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;
void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)
{
_memory_group = MemoryGroup(std::move(memory_manager));
};
private:
MemoryGroup _memory_group;
std::unique_ptr<CLDepthwiseConvolutionLayerNativeKernel> _dwc_native_kernel;
CLPermute _permute_input_to_nhwc;
CLPermute _permute_weights_to_nhwc;
CLPermute _permute_output_to_nchw;
CLTensor _permuted_input;
CLTensor _permuted_weights;
CLTensor _permuted_output;
CLTensor _output_multipliers;
CLTensor _output_shifts;
const ITensor *_original_weights;
const ITensor *_input;
const ITensor *_output;
bool _needs_permute;
bool _is_prepared;
bool _is_quantized;
};
std::shared_ptr<IMemoryManager> _memory_manager;
DepthwiseConvolutionFunction _depth_conv_func;
CLDepthwiseConvolutionLayerInternal3x3 _func_3x3;
CLDepthwiseConvolutionLayerGeneric _func_generic;
};
} // namespace arm_compute
#endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H */
|
