xref: /aosp_15_r20/external/ComputeLibrary/src/gpu/cl/kernels/ClIm2ColKernel.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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/gpu/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/Cast.h"
39 #include "support/StringSupport.h"
40 
41 #include <cmath>
42 #include <tuple>
43 #include <utility>
44 
45 namespace arm_compute
46 {
47 using namespace misc::shape_calculator;
48 namespace opencl
49 {
50 namespace kernels
51 {
52 namespace
53 {
54 struct Im2ColConfiguration
55 {
56     std::string           kernel_name{};
57     std::set<std::string> build_options{};
58     unsigned int          num_elems_processed_per_iteration{};
59     bool                  is_padding_required_nchw{};
60 };
61 
validate_arguments(const ITensorInfo * src,const ITensorInfo * dst,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)62 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
63                           unsigned int num_groups)
64 {
65     const unsigned int channel_idx = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::CHANNEL);
66 
67     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(src);
68     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
69     ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized(src->data_type()) && has_bias);
70     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(dst);
71     ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
72     ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::UNKNOWN);
73     ARM_COMPUTE_RETURN_ERROR_ON(num_groups == 0);
74     ARM_COMPUTE_RETURN_ERROR_ON(src->data_layout() == DataLayout::NHWC && num_groups > 1);
75     ARM_COMPUTE_RETURN_ERROR_ON((src->dimension(channel_idx) % num_groups) != 0);
76 
77     // Since there's no implicit padding added, check the total input spatial dimensions (with conv paddings) are big enough for the kernel dimensions
78     const unsigned int width_idx    = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH);
79     const unsigned int height_idx   = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT);
80     const unsigned     total_width  = src->dimension(width_idx) + conv_info.pad_left() + conv_info.pad_right();
81     const unsigned     total_height = src->dimension(height_idx) + conv_info.pad_top() + conv_info.pad_bottom();
82     ARM_COMPUTE_RETURN_ERROR_ON((total_width < kernel_dims.width) || (total_height < kernel_dims.height));
83 
84     if(dst->total_size() > 0)
85     {
86         const TensorInfo tensor_info_output = dst->clone()->set_tensor_shape(compute_im2col_conv_shape(src, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups));
87         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(dst, &tensor_info_output);
88         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst);
89         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(src, dst);
90     }
91 
92     return Status{};
93 }
94 
validate_and_configure_window(ITensorInfo * src,ITensorInfo * dst,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)95 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
96                                                         unsigned int num_elems_processed_per_iteration, bool is_padding_required_nchw, unsigned int num_groups)
97 {
98     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
99 
100     // Output tensor auto initialization if not yet initialized
101     TensorShape expected_output_shape = compute_im2col_conv_shape(src, kernel_dims, conv_info, has_bias, dilation, num_groups == 1, num_groups);
102 
103     auto_init_if_empty(*dst, src->clone()->set_tensor_shape(expected_output_shape));
104 
105     const DataLayout   data_layout  = src->data_layout();
106     const unsigned int width_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
107     const unsigned int height_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
108     const unsigned int input_width  = src->dimension(width_idx);
109     const unsigned int input_height = src->dimension(height_idx);
110 
111     // Configure the execute window based on the selected optimal OpenCL kernel
112     bool   window_changed = false;
113     Window win;
114 
115     if(data_layout == DataLayout::NHWC)
116     {
117         win = calculate_max_window(*src, Steps(num_elems_processed_per_iteration));
118     }
119     else
120     {
121         if(is_padding_required_nchw)
122         {
123             const BorderSize border(conv_info.pad_top(), conv_info.pad_right(), conv_info.pad_bottom(), conv_info.pad_left());
124             win = calculate_max_window(*src,
125                                        Steps(num_elems_processed_per_iteration * conv_info.stride().first, conv_info.stride().second));
126             AccessWindowStatic input_access(src,
127                                             -border.left,
128                                             -border.top,
129                                             ceil_to_multiple(input_width + border.right, kernel_dims.width * num_elems_processed_per_iteration),
130                                             input_height + border.bottom);
131             window_changed = window_changed || update_window_and_padding(win, input_access);
132         }
133         else
134         {
135             // For the generic case, CLIm2ColKernel doesn't need padding (we do not read out-of-bounds elements) so
136             // update_window_and_padding() can be skipped
137             win = calculate_max_window(*src, Steps());
138         }
139     }
140 
141     // set the Z dimension's step same size as the whole dimension so that one can't split across the Z dimension
142     win.set_dimension_step(Window::DimZ, win[Window::DimZ].end() - win[Window::DimZ].start());
143 
144     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
145     return std::make_pair(err, win);
146 }
147 
configure_opencl_kernel(const ITensorInfo * src,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)148 Im2ColConfiguration configure_opencl_kernel(const ITensorInfo *src, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation, unsigned int num_groups)
149 {
150     const DataLayout   data_layout   = src->data_layout();
151     const DataType     data_type     = src->data_type();
152     const unsigned int width_idx     = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
153     const unsigned int height_idx    = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
154     const unsigned int channel_idx   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
155     const unsigned int input_width   = src->dimension(width_idx);
156     const unsigned int input_height  = src->dimension(height_idx);
157     const unsigned int input_channel = src->dimension(channel_idx);
158 
159     const std::pair<unsigned int, unsigned int> convolved_dims = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
160 
161     // Im2Col configuration
162     std::string                   kernel_name = "im2col_generic_";
163     CLBuildOptions                build_opts;
164     unsigned int                  num_elems_processed_per_iteration = 1;
165     bool                          is_padding_required_nchw          = false;
166     const UniformQuantizationInfo qinfo                             = src->quantization_info().uniform();
167 
168     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
169     build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(src->element_size()));
170     build_opts.add_option("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
171     build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
172     build_opts.add_option("-DCONVOLVED_WIDTH=" + support::cpp11::to_string(convolved_dims.first));
173     build_opts.add_option("-DCONVOLVED_HEIGHT=" + support::cpp11::to_string(convolved_dims.second));
174     build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
175     build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
176     build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
177     build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
178     build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
179     build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
180     build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width));
181     build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input_height));
182     build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input_channel));
183     build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
184     build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
185     build_opts.add_option_if(num_groups > 1, "-DNUM_GROUPS=" + support::cpp11::to_string(num_groups));
186     build_opts.add_option_if_else(is_data_type_quantized(data_type), "-DPAD_VALUE=" + support::cpp11::to_string(qinfo.offset), "-DPAD_VALUE=0");
187     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
188 
189     if(data_layout == DataLayout::NHWC)
190     {
191         num_elems_processed_per_iteration = std::min(2U, input_channel);
192         is_padding_required_nchw          = false;
193 
194         // Only the 3x3 and 9x9 cases are optimized for NHWC
195         if(kernel_dims == Size2D(3U, 3U))
196         {
197             kernel_name = "im2col3x3_";
198             build_opts.add_option("-DIM2COL_3X3");
199         }
200         else if(kernel_dims == Size2D(9U, 9U))
201         {
202             kernel_name = "im2col9x9_";
203             build_opts.add_option("-DIM2COL_9X9");
204         }
205         else
206         {
207             build_opts.add_option("-DIM2COL_GENERIC");
208         }
209 
210         // Get boundary vector (the first/last vector with potentially a partial vector size) size
211         // If input_channel is a multiple of num_elems_processed_per_iteration, the boundary vec size is the (full) vector size
212         // otherwise, the boundary vec size is the (partial) remainder vector size
213         const unsigned int vec_size          = num_elems_processed_per_iteration;
214         const unsigned int partial_vec_size  = input_channel % vec_size;
215         const unsigned int boundary_vec_size = vec_size - ((vec_size - partial_vec_size) % vec_size);
216         build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vec_size));
217         build_opts.add_option("-DBOUNDARY_VECTOR_SIZE=" + support::cpp11::to_string(boundary_vec_size));
218     }
219     else
220     {
221         if(dilation == Size2D(1U, 1U))
222         {
223             const bool squared_im2col = kernel_dims.width == kernel_dims.height;
224             if(squared_im2col)
225             {
226                 // Check if we can run an optimized im2col for NCHW
227                 switch(kernel_dims.width)
228                 {
229                     case 1:
230                         // Optimized im2col1x1 if stride_x = 1 and conv_info.has_padding() = false
231                         if(conv_info.stride().first == 1 && !conv_info.has_padding())
232                         {
233                             kernel_name                       = "im2col1x1_stridex1_";
234                             num_elems_processed_per_iteration = 4;
235                             is_padding_required_nchw          = true;
236                         }
237                         break;
238                     case 3:
239                         kernel_name                       = "im2col3x3_";
240                         num_elems_processed_per_iteration = 1;
241                         is_padding_required_nchw          = true;
242                         break;
243                     case 5:
244                         kernel_name                       = "im2col5x5_";
245                         num_elems_processed_per_iteration = 1;
246                         is_padding_required_nchw          = true;
247                         break;
248                     case 11:
249                         // Optimized im2col11x11 if pad_x = pad_y = 0
250                         if(!conv_info.has_padding())
251                         {
252                             kernel_name                       = "im2col11x11_padx0_pady0_";
253                             num_elems_processed_per_iteration = 1;
254                             is_padding_required_nchw          = true;
255                         }
256                         break;
257                     default:
258                         kernel_name                       = "im2col_generic_";
259                         num_elems_processed_per_iteration = 1;
260                         is_padding_required_nchw          = false;
261                         break;
262                 }
263             }
264             else if(kernel_dims.width > 1 && !conv_info.has_padding())
265             {
266                 kernel_name                       = "im2col_generic_padx0_pady0_";
267                 num_elems_processed_per_iteration = 1;
268                 is_padding_required_nchw          = false;
269 
270                 // Optimized im2col is performed using one or more vector operations with the specified vector size
271                 // and a remainder. For example, for 5x5 convolutions, im2col is performed using vectors of size 4
272                 // and scalars; for 7x7 convolutions, using vectors of size 4 and vectors of size 3.
273                 // Using the vector size of 4 is always safe since OpenCL supports vectors of size 2 and 3.
274                 // Using the vector size of 8, however, may be faster.
275                 // For 2x2 convolutions, use vectors of size 2. (For 3x3 convolutions, im2col_kernel3x3_padx0_pady0
276                 // is used instead.)
277                 const size_t vector_size           = std::min(static_cast<size_t>(4), kernel_dims.width);
278                 const size_t width_mod_vector_size = kernel_dims.width % vector_size;
279                 build_opts.add_option("-DVECTOR_SIZE=" + support::cpp11::to_string(vector_size));
280                 build_opts.add_option("-DWIDTH_MOD_VECTOR_SIZE=" + support::cpp11::to_string(width_mod_vector_size));
281             }
282         }
283     }
284 
285     // Append the data layout to the kernel_name
286     kernel_name += lower_string(string_from_data_layout(data_layout));
287 
288     Im2ColConfiguration im2col_config;
289     im2col_config.kernel_name                       = kernel_name;
290     im2col_config.build_options                     = build_opts.options();
291     im2col_config.num_elems_processed_per_iteration = num_elems_processed_per_iteration;
292     im2col_config.is_padding_required_nchw          = is_padding_required_nchw;
293 
294     return im2col_config;
295 }
296 } // namespace
297 
ClIm2ColKernel()298 ClIm2ColKernel::ClIm2ColKernel()
299     : _data_layout(DataLayout::UNKNOWN), _convolved_dims(), _num_elems_processed_per_iteration(1), _kernel_dims(), _conv_info(), _num_groups()
300 {
301     _type = CLKernelType::ELEMENTWISE;
302 }
303 
configure(const ClCompileContext & compile_context,ITensorInfo * src,ITensorInfo * dst,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)304 void ClIm2ColKernel::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias,
305                                const Size2D &dilation,
306                                unsigned int  num_groups)
307 {
308     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
309     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups));
310 
311     auto padding_info = get_padding_info({ src, dst });
312     _data_layout      = src->data_layout();
313 
314     const unsigned int width_idx    = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
315     const unsigned int height_idx   = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
316     const unsigned int input_width  = src->dimension(width_idx);
317     const unsigned int input_height = src->dimension(height_idx);
318 
319     // Select and configure the optimal OpenCL kernel to run.
320     // This function returns the OpenCL kernel's name, the arguments to pass at compile time, the number of elements processed per iteration
321     // and the padding requirement flag
322     Im2ColConfiguration im2col_config = configure_opencl_kernel(src, kernel_dims, conv_info, has_bias, dilation, num_groups);
323 
324     // Create kernel
325     _kernel = create_kernel(compile_context, im2col_config.kernel_name, im2col_config.build_options);
326 
327     _convolved_dims                    = scaled_dimensions(input_width, input_height, kernel_dims.width, kernel_dims.height, conv_info, dilation);
328     _num_elems_processed_per_iteration = im2col_config.num_elems_processed_per_iteration;
329     _kernel_dims                       = kernel_dims; // Only needed by the Tuner
330     _conv_info                         = conv_info;   // Only needed by the Tuner
331     _num_groups                        = num_groups;
332 
333     // Configure kernel window
334     auto win_config = validate_and_configure_window(src, dst, kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
335                                                     im2col_config.is_padding_required_nchw, num_groups);
336     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
337     IClKernel::configure_internal(win_config.second);
338 
339     // Set config_id for enabling LWS tuning
340     _config_id = im2col_config.kernel_name;
341     _config_id += "_";
342     _config_id += lower_string(string_from_data_type(src->data_type()));
343     _config_id += "_";
344     _config_id += support::cpp11::to_string(num_groups);
345     _config_id += "_";
346     _config_id += support::cpp11::to_string(dst->dimension(0));
347     _config_id += "_";
348     _config_id += support::cpp11::to_string(dst->dimension(1));
349     _config_id += "_";
350     _config_id += lower_string(string_from_data_layout(_data_layout));
351 
352     ARM_COMPUTE_ERROR_ON(src->data_layout() == DataLayout::NHWC && has_padding_changed(padding_info));
353 }
354 
validate(const ITensorInfo * src,const ITensorInfo * dst,const Size2D & kernel_dims,const PadStrideInfo & conv_info,bool has_bias,const Size2D & dilation,unsigned int num_groups)355 Status ClIm2ColKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation,
356                                 unsigned int num_groups)
357 {
358     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, dst, kernel_dims, conv_info, has_bias, dilation, num_groups));
359     Im2ColConfiguration im2col_config = configure_opencl_kernel(src, kernel_dims, conv_info, has_bias, dilation, num_groups);
360     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(src->clone().get(), dst->clone().get(), kernel_dims, conv_info, has_bias, dilation, im2col_config.num_elems_processed_per_iteration,
361                                                               im2col_config.is_padding_required_nchw, num_groups)
362                                 .first);
363     return Status{};
364 }
365 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)366 void ClIm2ColKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
367 {
368     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
369     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IClKernel::window(), window);
370     ARM_COMPUTE_ERROR_ON(tensors.empty());
371 
372     // Get initial windows
373     // Collapse in order to have (SRC_DEPTH * BATCH_SIZE) on the 3rd dimension
374     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
375     window_collapsed.set_dimension_step(Window::DimZ, 1);
376 
377     auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
378     auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
379     ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
380 
381     Window window_output;
382     window_output.use_tensor_dimensions(dst->info()->tensor_shape());
383 
384     const Window first_slice_3d = window_collapsed.first_slice_window_3D();
385 
386     Window slice     = first_slice_3d;
387     Window slice_in  = first_slice_3d;
388     Window slice_out = window_output.first_slice_window_2D();
389 
390     if(_data_layout == DataLayout::NHWC)
391     {
392         const Window tmp_win     = window.collapse_if_possible(ICLKernel::window(), 3);
393         const int    num_batches = tmp_win[3].end();
394 
395         slice.set(1, Window::Dimension(0, static_cast<int>(dst->info()->tensor_shape()[1]), 1));
396         slice.set(2, Window::Dimension(0, static_cast<int>(num_batches), 1));
397     }
398     else
399     {
400         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));
401         slice.set(1, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
402         // Note: In case of NCHW the 3rd dimension is already set collapsing the input window
403     }
404 
405     // Setup input slice
406     // The dimensions of the input are increased within the OpenCL kernel
407     slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
408     slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
409     slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
410 
411     // Setup output slice
412     // The dimensions of the output are increased within the OpenCL kernel
413     slice_out.set(Window::DimX, Window::Dimension(0, 0, 0));
414     slice_out.set(Window::DimY, Window::Dimension(0, 0, 0));
415 
416     unsigned int idx = num_arguments_per_3D_tensor() + (_num_groups == 1 ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
417     _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(src->info()->strides_in_bytes()[3]));
418     _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(dst->info()->strides_in_bytes()[((_num_groups == 1) ? 2 : 3)]));
419     do
420     {
421         unsigned int idx = 0;
422         add_3D_tensor_argument(idx, src, slice_in);
423         if(_num_groups == 1)
424         {
425             add_2D_tensor_argument(idx, dst, slice_out);
426         }
427         else
428         {
429             add_3D_tensor_argument(idx, dst, slice_out);
430         }
431         enqueue(queue, *this, slice, lws_hint());
432     }
433     while(window_collapsed.slide_window_slice_3D(slice) && window_output.slide_window_slice_2D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
434 }
435 } // namespace kernels
436 } // namespace opencl
437 } // namespace arm_compute
438