• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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 #include "src/core/CL/kernels/CLIm2ColKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/CL/OpenCL.h"
30 #include "arm_compute/core/Helpers.h"
31 #include "arm_compute/core/TensorInfo.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
34 #include "src/core/AccessWindowStatic.h"
35 #include "src/core/CL/CLValidate.h"
36 #include "src/core/helpers/AutoConfiguration.h"
37 #include "src/core/helpers/WindowHelpers.h"
38 #include "support/StringSupport.h"
39 
40 #include <cmath>
41 #include <tuple>
42 #include <utility>
43 
44 namespace arm_compute
45 {
46 using namespace misc::shape_calculator;
47 
48 namespace
49 {
50 struct Im2ColConfiguration
51 {
52     std::string           kernel_name{};
53     std::set<std::string> build_options{};
54     unsigned int          num_elems_processed_per_iteration{};
55     bool                  is_padding_required_nchw{};
56 };
57 
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)58 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
59                           unsigned int num_groups)
60 {
61     const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
62 
63     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
64     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
65     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(input->data_type()) && has_bias);
66     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
67     ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
68     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
69     ARM_COMPUTE_RETURN_ERROR_ON(num_groups == 0);
70     ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::NHWC && num_groups > 1);
71     ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(channel_idx) % num_groups) != 0);
72 
73     // Since there's no implicit padding added, check the total input spatial dimensions (with conv paddings) are big enough for the kernel dimensions
74     const unsigned int width_idx    = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
75     const unsigned int height_idx   = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
76     const unsigned     total_width  = input->dimension(width_idx) + conv_info.pad_left() + conv_info.pad_right();
77     const unsigned     total_height = input->dimension(height_idx) + conv_info.pad_top() + conv_info.pad_bottom();
78     ARM_COMPUTE_RETURN_ERROR_ON((total_width < kernel_dims.width) || (total_height < kernel_dims.height));
79 
80     if(output->total_size() > 0)
81     {
82         const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups));
83         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output);
84         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
85         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
86     }
87 
88     return Status{};
89 }
90 
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_elems_processed_per_iteration,bool is_padding_required_nchw,unsigned int num_groups)91 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
92                                                         unsigned int num_elems_processed_per_iteration, bool is_padding_required_nchw, unsigned int num_groups)
93 {
94     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
95 
96     // Output tensor auto initialization if not yet initialized
97     TensorShape expected_output_shape = compute_im2col_conv_shape(input, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups);
98 
99     auto_init_if_empty(*output, input->clone()->set_tensor_shape(expected_output_shape));
100 
101     const DataLayout   data_layout  = input->data_layout();
102     const unsigned int width_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
103     const unsigned int height_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
104     const unsigned int input_width  = input->dimension(width_idx);
105     const unsigned int input_height = input->dimension(height_idx);
106 
107     // Configure the execute window based on the selected optimal OpenCL kernel
108     bool   window_changed = false;
109     Window win;
110 
111     if(data_layout == DataLayout::NHWC)
112     {
113         win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
114 
115         const int xin_start = 0;
116         const int xin_end   = input->dimension(0);
117         const int yin_start = 0;
118         const int yin_end   = input->dimension(1);
119 
120         const int xout_start = 0;
121         const int xout_end   = output->dimension(0);
122         const int yout_start = 0;
123         const int yout_end   = output->dimension(1);
124 
125         AccessWindowStatic input_access(input, xin_start, yin_start, xin_end, yin_end);
126         AccessWindowStatic output_access(output, xout_start, yout_start, xout_end, yout_end);
127         window_changed = window_changed || update_window_and_padding(win, input_access, output_access);
128     }
129     else
130     {
131         if(is_padding_required_nchw)
132         {
133             const BorderSize border(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
134             win = calculate_max_window(*input,
135                                        Steps(num_elems_processed_per_iteration * conv_info.stride().first, conv_info.stride().second));
136             AccessWindowStatic input_access(input,
137                                             -border.left,
138                                             -border.top,
139                                             ceil_to_multiple(input_width + border.right, kernel_dims.width * num_elems_processed_per_iteration),
140                                             input_height + border.bottom);
141             window_changed = window_changed || update_window_and_padding(win, input_access);
142         }
143         else
144         {
145             // For the generic case, CLIm2ColKernel doesn't need padding (we do not read out-of-bounds elements) so
146             // update_window_and_padding() can be skipped
147             win = calculate_max_window(*input, Steps());
148         }
149     }
150     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
151     // set the Z dimension's step same size as the whole dimension so that one can't split across the Z dimension
152     win.set_dimension_step(Window::DimZ, win[Window::DimZ].end() - win[Window::DimZ].start());
153 
154     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
155     return std::make_pair(err, win);
156 }
157 
configure_opencl_kernel(const ITensorInfo * input,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)158 Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *input, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, unsigned int num_groups)
159 {
160     const DataLayout   data_layout   = input->data_layout();
161     const DataType     data_type     = input->data_type();
162     const unsigned int width_idx     = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
163     const unsigned int height_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
164     const unsigned int channel_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
165     const unsigned int input_width   = input->dimension(width_idx);
166     const unsigned int input_height  = input->dimension(height_idx);
167     const unsigned int input_channel = input->dimension(channel_idx);
168 
169     const std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
170 
171     // Im2Col configuration
172     std::string                   kernel_name = "im2col_generic_";
173     CLBuildOptions                build_opts;
174     unsigned int                  num_elems_processed_per_iteration = 1;
175     bool                          is_padding_required_nchw          = false;
176     const UniformQuantizationInfo qinfo                             = input->quantization_info().uniform();
177 
178     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
179     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->element_size()));
180     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
181     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
182     build_opts.add_option("-DCONVOLVED_WIDTH=" + support::cpp11::to_string(convolved_dims.first));
183     build_opts.add_option("-DCONVOLVED_HEIGHT=" + support::cpp11::to_string(convolved_dims.second));
184     build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
185     build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
186     build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
187     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
188     build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
189     build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
190     build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width));
191     build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input_height));
192     build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input_channel));
193     build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
194     build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
195     build_opts.add_option_if(num_groups > 1, "-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
196     build_opts.add_option_if_else(is_data_type_quantized(data_type), "-DPAD_VALUE=" + support::cpp11::to_string(qinfo.offset), "-DPAD_VALUE=0");
197     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
198 
199     if(data_layout == DataLayout::NHWC)
200     {
201         num_elems_processed_per_iteration = std::min(2U, input_channel);
202         is_padding_required_nchw          = false;
203 
204         // Only the 3x3 and 9x9 cases are optimized for NHWC
205         if(kernel_dims == Size2D(3U, 3U))
206         {
207             kernel_name = "im2col3x3_";
208         }
209         else if(kernel_dims == Size2D(9U, 9U))
210         {
211             kernel_name = "im2col9x9_";
212         }
213 
214         // Get boundary vector (the first/last vector with potentially a partial vector size) size
215         // If input_channel is a multiple of num_elems_processed_per_iteration, the boundary vec size is the (full) vector size
216         // otherwise, the boundary vec size is the (partial) remainder vector size
217         const unsigned int vec_size          = num_elems_processed_per_iteration;
218         const unsigned int partial_vec_size  = input_channel % vec_size;
219         const unsigned int boundary_vec_size = vec_size - ((vec_size - partial_vec_size) % vec_size);
220         build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vec_size));
221         build_opts.add_option("-DBOUNDARY_VECTOR_SIZE=" + support::cpp11::to_string(boundary_vec_size));
222     }
223     else
224     {
225         if(dilation == Size2D(1U, 1U))
226         {
227             const bool squared_im2col = kernel_dims.width == kernel_dims.height;
228             if(squared_im2col)
229             {
230                 // Check if we can run an optimized im2col for NCHW
231                 switch(kernel_dims.width)
232                 {
233                     case 1:
234                         // Optimized im2col1x1 if stride_x = 1 and conv_info.has_padding() = false
235                         if(conv_info.stride().first == 1 && !conv_info.has_padding())
236                         {
237                             kernel_name                       = "im2col1x1_stridex1_";
238                             num_elems_processed_per_iteration = 4;
239                             is_padding_required_nchw          = true;
240                         }
241                         break;
242                     case 3:
243                         kernel_name                       = "im2col3x3_";
244                         num_elems_processed_per_iteration = 1;
245                         is_padding_required_nchw          = true;
246                         break;
247                     case 5:
248                         kernel_name                       = "im2col5x5_";
249                         num_elems_processed_per_iteration = 1;
250                         is_padding_required_nchw          = true;
251                         break;
252                     case 11:
253                         // Optimized im2col11x11 if pad_x = pad_y = 0
254                         if(!conv_info.has_padding())
255                         {
256                             kernel_name                       = "im2col11x11_padx0_pady0_";
257                             num_elems_processed_per_iteration = 1;
258                             is_padding_required_nchw          = true;
259                         }
260                         break;
261                     default:
262                         kernel_name                       = "im2col_generic_";
263                         num_elems_processed_per_iteration = 1;
264                         is_padding_required_nchw          = false;
265                         break;
266                 }
267             }
268             else if(kernel_dims.width > 1 && !conv_info.has_padding())
269             {
270                 kernel_name                       = "im2col_generic_padx0_pady0_";
271                 num_elems_processed_per_iteration = 1;
272                 is_padding_required_nchw          = false;
273 
274                 // Optimized im2col is performed using one or more vector operations with the specified vector size
275                 // and a remainder. For example, for 5x5 convolutions, im2col is performed using vectors of size 4
276                 // and scalars; for 7x7 convolutions, using vectors of size 4 and vectors of size 3.
277                 // Using the vector size of 4 is always safe since OpenCL supports vectors of size 2 and 3.
278                 // Using the vector size of 8, however, may be faster.
279                 // For 2x2 convolutions, use vectors of size 2. (For 3x3 convolutions, im2col_kernel3x3_padx0_pady0
280                 // is used instead.)
281                 const size_t vector_size           = std::min(static_cast<size_t>(4), kernel_dims.width);
282                 const size_t width_mod_vector_size = kernel_dims.width % vector_size;
283                 build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vector_size));
284                 build_opts.add_option("-DWIDTH_MOD_VECTOR_SIZE=" + support::cpp11::to_string(width_mod_vector_size));
285             }
286         }
287     }
288 
289     // Append the data layout to the kernel_name
290     kernel_name += lower_string(string_from_data_layout(data_layout));
291 
292     Im2ColConfiguration im2col_config;
293     im2col_config.kernel_name                       = kernel_name;
294     im2col_config.build_options                     = build_opts.options();
295     im2col_config.num_elems_processed_per_iteration = num_elems_processed_per_iteration;
296     im2col_config.is_padding_required_nchw          = is_padding_required_nchw;
297 
298     return im2col_config;
299 }
300 } // namespace
301 
CLIm2ColKernel()302 CLIm2ColKernel::CLIm2ColKernel()
303     : _input(nullptr), _output(nullptr), _data_layout(DataLayout::UNKNOWN), _convolved_dims(), _num_elems_processed_per_iteration(1), _kernel_dims(), _conv_info(), _num_groups()
304 {
305 }
306 
configure(const ICLTensor * input,ICLTensor * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)307 void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
308                                unsigned int num_groups)
309 {
310     configure(CLKernelLibrary::get().get_compile_context(), input, output, kernel_dims, conv_info, has_bias, dilation, num_groups);
311 }
312 
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)313 void CLIm2ColKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
314                                const Size2D &dilation,
315                                unsigned int  num_groups)
316 {
317     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
318     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation, num_groups));
319 
320     auto padding_info = get_padding_info({ input, output });
321     _data_layout      = input->info()->data_layout();
322 
323     const unsigned int width_idx    = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
324     const unsigned int height_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
325     const unsigned int input_width  = input->info()->dimension(width_idx);
326     const unsigned int input_height = input->info()->dimension(height_idx);
327 
328     // Select and configure the optimal OpenCL kernel to run.
329     // This function returns the OpenCL kernel's name, the arguments to pass at compile time, the number of elements processed per iteration
330     // and the padding requirement flag
331     Im2ColConfiguration im2col_config = configure_opencl_kernel(input->info(), kernel_dims, conv_info, has_bias, dilation, num_groups);
332 
333     // Create kernel
334     _kernel = create_kernel(compile_context, im2col_config.kernel_name, im2col_config.build_options);
335 
336     _input                             = input;
337     _output                            = output;
338     _convolved_dims                    = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
339     _num_elems_processed_per_iteration = im2col_config.num_elems_processed_per_iteration;
340     _kernel_dims                       = kernel_dims; // Only needed by the Tuner
341     _conv_info                         = conv_info;   // Only needed by the Tuner
342     _num_groups                        = num_groups;
343 
344     // Configure kernel window
345     auto win_config = validate_and_configure_window(input->info(), output->info(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
346                                                     im2col_config.is_padding_required_nchw, num_groups);
347     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
348     ICLKernel::configure_internal(win_config.second);
349 
350     // Set config_id for enabling LWS tuning
351     _config_id = im2col_config.kernel_name;
352     _config_id += "_";
353     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
354     _config_id += "_";
355     _config_id += support::cpp11::to_string(num_groups);
356     _config_id += "_";
357     _config_id += support::cpp11::to_string(output->info()->dimension(0));
358     _config_id += "_";
359     _config_id += support::cpp11::to_string(output->info()->dimension(1));
360     _config_id += "_";
361     _config_id += lower_string(string_from_data_layout(_data_layout));
362 
363     ARM_COMPUTE_ERROR_ON(input->info()->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info));
364 }
365 
validate(const ITensorInfo * input,const ITensorInfo * output,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)366 Status CLIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
367                                 unsigned int num_groups)
368 {
369     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, kernel_dims, conv_info, has_bias, dilation, num_groups));
370     Im2ColConfiguration im2col_config = configure_opencl_kernel(input, kernel_dims, conv_info, has_bias, dilation, num_groups);
371     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
372                                                               im2col_config.is_padding_required_nchw, num_groups)
373                                 .first);
374     return Status{};
375 }
376 
run(const Window & window,cl::CommandQueue & queue)377 void CLIm2ColKernel::run(const Window &window, cl::CommandQueue &queue)
378 {
379     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
380     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
381 
382     // Get initial windows
383     // Collapse in order to have (SRC_DEPTH * BATCH_SIZE) on the 3rd dimension
384     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
385     window_collapsed.set_dimension_step(Window::DimZ, 1);
386 
387     Window window_output;
388     window_output.use_tensor_dimensions(_output->info()->tensor_shape());
389 
390     const Window first_slice_3d = window_collapsed.first_slice_window_3D();
391 
392     Window slice     = first_slice_3d;
393     Window slice_in  = first_slice_3d;
394     Window slice_out = window_output.first_slice_window_2D();
395 
396     if(_data_layout == DataLayout::NHWC)
397     {
398         const Window tmp_win     = window.collapse_if_possible(ICLKernel::window(), 3);
399         const int    num_batches = tmp_win[3].end();
400 
401         slice.set(1, Window::Dimension(0, static_cast<int>(_output->info()->tensor_shape()[1]), 1));
402         slice.set(2, Window::Dimension(0, static_cast<int>(num_batches), 1));
403     }
404     else
405     {
406         slice.set(0, Window::Dimension(0, static_cast<int>(ceil_to_multiple(_convolved_dims.first, _num_elems_processed_per_iteration)), _num_elems_processed_per_iteration));
407         slice.set(1, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
408         // Note: In case of NCHW the 3rd dimension is already set collapsing the input window
409     }
410 
411     // Setup input slice
412     // The dimensions of the input are increased within the OpenCL kernel
413     slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
414     slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
415     slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
416 
417     // Setup output slice
418     // The dimensions of the output are increased within the OpenCL kernel
419     slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
420     slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
421 
422     unsigned int idx = num_arguments_per_3D_tensor() + (_num_groups == 1 ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
423     _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
424     _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[((_num_groups == 1) ? 2 : 3)]));
425     do
426     {
427         unsigned int idx = 0;
428         add_3D_tensor_argument(idx, _input, slice_in);
429         if(_num_groups == 1)
430         {
431             add_2D_tensor_argument(idx, _output, slice_out);
432         }
433         else
434         {
435             add_3D_tensor_argument(idx, _output, slice_out);
436         }
437         enqueue(queue, *this, slice, lws_hint());
438     }
439     while(window_collapsed.slide_window_slice_3D(slice) && window_output.slide_window_slice_2D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
440 }
441 } // namespace arm_compute
442